Datasets:

babytreecc

/

ICLR2025review

like 7

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "How can the sample complexity bounds be further improved to match the practical performance observed in experiments, and is there potential for achieving instance-optimal sample complexity in the PAC best-item setting?\n\nCould the proposed DE and DEBC algorithms be adapted or extended to work with partial or full rankings instead of just identifying the best item, and what challenges might arise in such extensions?\n\nHow would the algorithms perform in settings where item correlations are dynamic or evolve over time, and what adjustments to DE/DEBC might be necessary to handle such changes effectively?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. DE and its extension, DEBC, significantly improve sample complexity for identifying the best item compared to existing algorithms, reducing the number of subset plays needed.\n\n2. The incorporation of inferred updates through item correlation in DEBC provides a robust mechanism to handle correlated item structures, leading to superior performance in certain datasets.\n\n3. The paper extensively evaluates DE and DEBC across various synthetic and real-world datasets, demonstrating their practical effectiveness and robustness across different settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work addresses the problem of identifying the best item from a set of items based on relative feedback, specifically using a method called Dynamic Elimination (DE). DE efficiently prunes sub-optimal items as it progresses, improving sample complexity compared to existing algorithms. The authors also propose an extension, Dynamic Elimination by Correlation (DEBC), which incorporates inferred updates based on item correlations. DEBC significantly outperforms DE in settings where item correlation is strong, reducing sample complexity further. Extensive experiments demonstrate that both DE and DEBC outperform existing state-of-the-art (SOTA) methods in terms of sample complexity across multiple datasets and settings. Additionally, the paper explores future directions for improving sample complexity bounds and extending the methods to partial/full rankings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While DE and DEBC perform well in practice, the theoretical sample complexity bounds provided in the paper are not as tight as their practical performance would suggest, leaving room for further theoretical refinement.\n\n2. DEBC’s performance heavily relies on the strength of item correlations, this raise potential limiting its applicability in scenarios with weak correlations. \n\n3. The paper primarily focuses on cosine similarity for item embeddings and correlations. Extending this to other similarity measures or more general settings is only briefly mentioned and not fully explored.\n\nSome cosmetics: for example on row 330: `we can can combine`." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Q1: It would be appreciated if the authors could include additional experiments conducted in real-world scenarios. These experiments should aim to demonstrate the effectiveness of the proposed methods in best item identification. Additionally, it is recommended that results be presented using widely accepted metrics such as accuracy, AUC, F1, precision, and recall. \nQ2: The authors might consider providing a more precise and explicit formulation of the item selection problem in Section 3. Furthermore, it would be valuable to discuss in greater detail how DE and DEBC can be applied in practical, real-world contexts." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1: The proposed methods enhance existing approaches by dynamically eliminating suboptimal items, significantly reducing the algorithm's complexity.\n\nS2: By incorporating correlations between items, the proposed methods extend their applicability to items initially absent from the played set.\n\nS3: The authors offer theoretical assurances regarding the sample complexity of DE and DEBC. They also demonstrate that the sample mean of an inferred update sequence serves as an unbiased estimator." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors have introduced a dynamic elimination method for item selection based on relative feedback. They propose two distinct algorithms: one that implements the fundamental dynamic elimination approach and another that incorporates item correlations into the elimination process. The paper provides theoretical analysis on both the sample complexities and the correctness of the proposed methods. Furthermore, the authors demonstrate the proposed methods' ability in terms of reducing sample complexities when compared to several baseline approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1: The DE and DEBC algorithms are designed for the task of item selection, yet the performance of best item identification is neglected. The authors dedicate substantial space to discussing the efficiency of the proposed methods in reducing sample complexity. However, the mathematical formulation lacks clarity, as the best item identification problem and the relative feedback are not thoroughly formulated.\nW2: The proposed DEBC algorithm presumes that item correlations are known to the user, raising concerns about the validity of this assumption. The authors do not address the implications of this assumption in real-world applications or indicate whether it is a common assumption in existing literature.\nW3: The current work lacks demonstration in real-world scenarios. Although the authors mention that learning to rank is crucial in fields like sociology, information retrieval, and search engine optimization, they do not provide examples of its application in these areas. Consequently, the practical applicability of this work remains uncertain." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weakness part." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper studies an important problem. \n\nThe proposed algorithm is complemented with theoretical analysis as well as extensive experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of best-item identification from relative feedback in the setting where a learner adaptively plays subsets of items and receives stochastic feedback in the form of the best item in the set. An algorithm named Dynamic Elimination (DE) is proposed, which dynamically prunes sub-optimal items from contention to efficiently identify the best item. Then the model is extended to capture the generalized linear correlation of items. An algorithm named DEBC, an extension of DE is proposed to handle this extension. The core idea is leveraging the generalized linear correlation to obtain estimates on item win rates without directly playing them by leveraging item correlation information. Extensive experiments are conducted to validate the empirical performance of the proposed algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The algorithmic contribution of this looks narrow. The core idea of DE is flexibly eliminating items once they are deemed suboptimal. This idea is old in the bandit literature. The authors can refer to Chapter 6 of [3] for some reference. Also, a simple google search would gives you a number of work on elimination algorithms. The core idea of DEBC is exploiting the generalized linear structure on the correlation among arms. This idea is also not new since linear structure has been extensively studied in linear bandits, reinforcement learning. Please refer to Part V of [3] from some details. \n\nThe proof techniques of this paper are not new, most of them are drawn from literature. Thus, this paper does not contribute to new proof techniques. To me more specific, compared to [1,2], I do not see enough new ideas in the proof. For example, the analysis of concentration, probability of event, etc., looks very normal. Could the author elaborate on the novelty of the proof? \n\nThe theoretical improvement over SOTA techniques is not clear. The improvement on the sample complexity compared with SOTA works is not stated. How does it improve the sample complexity upper bound? \n\nThe second paragraph of the related work overstated the limitations of previous works without any supporting evidence. Previous algorithms may require up to millions of samples to rank only a few items, but this possibility depends on the setting of the problem. It should not be stated as a general claim. Furthermore, this paragraph is not precise. What do you mean by often? Could you quantify it? \n\nLemma 1 is confusing. It is highlighted as a lower bound, but the sample complexity is stated using the big O notation. \n\n[1] Yisong Yue, et al. The K-armed Dueling Bandits Problem, Journal of Computer and System Sciences, 78(5): 1538–1556.\n\n[2] Björn Haddenhorst. Identification of the Generalized Condorcet Winner in Multi-dueling Bandits, NeurIPS, 2021\n\n[3] Tor Lattimore, et al. Bandit Algorithms. Cambridge press." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "It will help if the authors can have more evidence to demonstrate this paper's results' significance. \nOr if I perceive this paper wrongly, please let me know." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper proposes new algorithms for best item identification and winning rate estimation. These algorithms do have some novelty and are inspiring." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the best item selection problem from noisy multi-wise comparisons. There are totally n items, and at each round the agent can select n_s items to compare, and the comparison will return one item as the winner according to the PL model. The better item will have a higher chance to win the comparison. The problem is to find the best item with 1-\\delta confidence with the least amount of comparisons. \n\nThe authors propose a new algorithm for best item selection from multi-wise comparisons and give the worst-case, best-case, and expected sample complexity. The authors further propose a method to estimate the winning probabilities between two items without directly comparing these two items." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The significance of this paper's results is questionable. For the best item identification, the sample complexity (worst-case) is O(n*\\epsilon^{-2} *\\log(n*n_s^{-1}*\\delta^{-1})). However, in a previous paper [1], when n_s = 2 (i.e., pairwise comparisons), the sample complexity (expected) for best item identification is O(n*\\epsilon^{-2} *\\log\\delta^{-1})) (Theorem 5 of [1]), which is log(n) better than the proposed algorithm. When n_s is large enough like \\Omega(n), the sample complexity of the proposed in this paper will become the same as that in [1]. Hence, the proposed algorithm does not show superiority compared to existing algorithms, or at least be as good as existing ones. Although its sample complexity is worst case instead of in expectation, but this difference is not large enough to support the significance of the new results. \n\nBesides, the winning chance estimation and the sample complexity of best item identification seem not to be correlated enough to be put in the same paper. Putting them in one paper makes the paper's scope to be ambiguous. If the winning chance estimate is significant enough, it is better to be placed in another paper focusing on a more related topic.\n\n[1] Ren, W., Liu, J., & Shroff, N. (2020, November). The Sample Complexity of Best-$ k $ Items Selection from Pairwise Comparisons. In International Conference on Machine Learning (pp. 8051-8072). PMLR." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Algorithms for PAC optimal item selection from subset wise relative feedback based on suboptimal item dynamic elimination that outperforms SOTA benchmarks; additionally introduce the notion of inferred updates to utilize item similarity information" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024dynamic,\ntitle={Dynamic Elimination For {PAC} Optimal Item Selection From Relative Feedback},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zCJqgXnV7f},\nnote={under review}\n}" }, "abstract": { "value": "We study the problem of best-item identification from relative feedback where a learner adaptively plays subsets of items and receives stochastic feedback in the form of the best item in the set. We propose an algorithm - Dynamic Elimination (DE) - that dynamically prunes sub-optimal items from contention to efficiently identify the best item and show a strong sample complexity upper bound for it. We further formalize the notion of inferred updates to obtain estimates on item win rates without directly playing them by leveraging item correlation information. We propose the Dynamic Elimination by Correlation (DEBC) algorithm as an extension to DE with inferred updates. We show through extensive experiments that DE and DEBC significantly outperform all existing baselines across multiple datasets in various settings." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "probably approximately correct", "optimal item selection", "relative feedback", "multi armed bandits", "Plackett Luce Model", "Condorcet winner", "Bayesian updates", "active learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a2d85a2fd9a81e27e2c69845e52423c5bd04e7af.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/9f5f2b950c66f16ca32ace3f35a9c8c1df8594ae.zip" }, "title": { "value": "Dynamic Elimination For PAC Optimal Item Selection From Relative Feedback" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Have the authors tested out the proposed algorithms (or their adaptive gradient variants) on problems that appropriately reflect problems/setups that are of realistic practical interest? For instance, training transformer based architectures on standard pre-training tasks?\n2. Can the authors clarify how this algorithm's guarantees look like with (a) strong convexity, (b) non-convex but say with a PL style condition?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This is a well written paper that presents two interesting algorithms with strong theoretical guarantees. The algorithms definitely appear novel to my knowledge but I do not know of the latest developments in this area." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents two methods for stochastic optimization that achieve similar guarantees as standard gradient descent and accelerated gradient descent while being robust to the choice of learning rate involved." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The empirical section seems to be fairly underbaked in my opinion." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I am curious to know about your choice of single sample per iterate setting. I understand that this choice might be more convenient to work in theory, but could you still carry the same theory using batch size b?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well-organized, with clear explanations of each momentum component and its integration into $\\mu^2$. A convergence guarantee for stochastic optimization with a fixed step-size addresses a significant challenge in optimization.\n\n- The theoretical results are rigorously supported. Extending Theorem 4.2 to the accelerated version in Theorem 5.1 provides a pleasing completeness to the theory. In the former, the learning rate has a dependency on $T$, which prevents it from achieving the accelerated rate. This dependency is eliminated by introduction of the $\\mu^2$-ExtraSGD.\n\n- The theoretical results carry over nicely in the convex experiment, logistic regression on MNIST. $\\mu^2$SGD consistently achieves the best performance across a very wide range of learning rates ($10^{-3}$ to $10^3$). Unfortunately, I am unable to tell where each method lies in the last plot in Figure 1 ($10^{-4}$)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the $\\mu^2$ (Momentum$^2$) gradient estimator, designed to manage gradient noise in stochastic optimization, particularly in convex and smooth loss scenarios using single-sample batch sizes. This estimator integrates two momentum techniques:\n\n1. **Anytime Averaging**, which averages the query points for the gradient oracle.\n2. **STORM**, a corrected momentum method that averages gradient estimates with bias correction.\n\nBy combining these techniques, $\\mu^2$ achieves a progressively shrinking square error of the gradient estimates ($\\|\\epsilon_t\\|^2 \\propto 1/t$), contrasting with the fixed gradient error in standard stochastic optimization ($\\|\\epsilon_t\\|^2 = O(1)$). This property enables the use of a fixed step-size for convergence, removing the necessity of step-size decay in stochastic settings. Additionally, $\\mu^2$ allows the norm of the gradient estimate to serve as a stopping criterion.\n\nThe paper implements $\\mu^2$ in two algorithms:\n\n1. **$\\mu^2$-SGD**: This combines $\\mu^2$ with SGD. Although $\\mu^2$ could work with other first-order methods, the authors focus on SGD to derive theoretical guarantees. Theorem 4.2 shows that $\\mu^2$-SGD achieves optimal convergence rates for noiseless ($O(L/T)$) and noisy ($O(L/T + \\tilde{\\sigma}/\\sqrt{T})$) settings using a fixed learning rate $\\eta_{\\text{Offline}} = 1/8LT$. Remarkably, this rate does not need to change between noiseless and noisy conditions.\n\n2. **$\\mu^2$-ExtraSGD**: This accelerated variant of $\\mu^2$-SGD uses the ExtraGradient framework to achieve optimal convergence rates of $O(L/T^2)$ (noiseless) and $O(L/T^2 + \\tilde{\\sigma}/\\sqrt{T})$ (noisy) with a fixed learning rate $\\eta_{\\text{Offline}} = 1/2L$.\n\nExperiments on MNIST and CIFAR-10 datasets, using convex (logistic regression) and non-convex (deep learning) models, demonstrate $\\mu^2$-SGD's stability and performance over a range of learning rates compared to various baseline optimizers, including SGD, Polyak momentum, and individual applications of Anytime Averaging and STORM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- My main critique of this paper goes back to its main promise: with $\\mu^2$SGD, convergence using a fixed step-size is achievable in a stochastic setting. However, although the step-size is fixed, varying momentum parameters, $\\alpha$ and $\\beta$, are required to prove convergence in Theorem 4.2. The decreasing step-size nature necessary for convergence in a stochastic setting seems to be delegated to the momentum parameter here. That being said, I appreciate the theoretical results and the simplicity of the schedule for $\\alpha$. This same schedule was also used in the experiments, and it seems to work well.\n- There is no experiment on $\\mu^2$-ExtraSGD. Observing the accelerated rate in experiments could have potentially highlighted the strength of this method. I would also be curious to see how this acceleration performs compared to the Nesterov accelerated method. \n- Although the paper shows $\\mu^2$’s stability across learning rates, there is little analysis on how sensitive the algorithm might be to momentum parameter choices, especially in non-convex settings where fixed parameters are used." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. By combining two existing methods--Anytime-SGD (Cutkosky, 2019) and STORM (STochastic Recursive Momentum) (Cutkosky & Orabona, 2019), the authors propose $\\mu^2$-SGD.\n\n2. The authors show the error term is upper bounded by $\\frac{1}{t}$ with $\\mu^2$-SGD. Moreover, they obtain an upper bound for the excess loss that gives a noise independent optimal learning rate, which results in a wider range of optimal choice for learning rate compared to SGD. \n\n3. By adding Optimistic-OGD, the authors propse an accelerated version of $\\mu^2$-SGD--$\\mu^2$-ExtraSGD. The optimal learning rate in this case is constant in time so the optimal rate can be obtained without doing time-varying learning rate. \n\n4. The authors have experiments in both convex and non-convex setting to verify their results. Experiments show $\\mu^2$-SGD and $\\mu^2$-ExtraSGD are stable w.r.t different learning rates while other methods don't. Experiments are done thoroughly with details given in the appendix." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Combining two existing methods, the authors propose an accelerated SGD mechanism that is stable w.r.t. learning rate. The authors provide rigorous justifications for their claims and supporting experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main contribution of this work--$\\mu^2$-SGD and $\\mu^2$-ExtraSGD--comes by combining two existing works. The proofs for main theorems are quite traditional. One could argue this lacks novelty, although I personally found the results interesting.\n\n2. In the numerical experiments (non-convex setting), It seems $\\mu^2$-SGD and $\\mu^2$-ExtraSGD only make a difference when the learning rate is far away from normal choices. More precisely, Figure 2 shows all methods are quite similar for $\\eta\\leq1$. $\\mu^2$-SGD and $\\mu^2$-ExtraSGD are better only when $\\eta >1$, which is not a typical choice for learning rate anyway. I am not sure how learning rate stability is appreciated by the community. \n\n3. By allowing a wider range of learning rates, one could argue that with this method, we don’t need to tune the learning rate anymore, like we could set it to be 1. However, $\\mu^2$-SGD and $\\mu^2$-ExtraSGD are not hyperparameter free, i.e. they introduce weights $\\alpha_t$ and Corrected Momentum weights $\\beta_t$, which are chosen to be $t$ and $\\frac{1}{t}$ in the experiments. I wish the authors could give more insights into this topic. If the cost of reducing one hyperparameter is introducing two hyperparameters, I am not sure it’s worth it. \n\n4. The experiments are done for different learning rates, but I found the plots were hard to read when the accuracies/losses got too close. Presenting results with an additional table could help in these cases." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1.\tCould the author provide more comparisons between the proposed algorithm and parameter-free algorithms?\n2.\tCould the author include additional experiments on the impact of noise on the algorithm?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Considering algorithms that work in both noisy and noise-free conditions helps improve the robustness of the algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel gradient estimator for stochastic convex optimization, combining momentum-based techniques. Using this estimator, the authors develop robust SGD-style algorithms that achieve optimal convergence rates in both noiseless and noisy settings, maintaining stable performance over a wide range of learning rates." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe structure of this paper makes it difficult to follow, especially in terms of understanding the novelty of the algorithms.\n2.\tThe experimental results presented do not convincingly demonstrate the superiority of the proposed algorithm. The choice of learning rates between 10 and 1000 is unconventional and the hyperparameter settings for STORM are not clearly defined. Furthermore, the generalization benefits mentioned by the author in line 52 has no experimental support.\n3.\tThe novelty of this paper appears to be limited. For example, STORM was originally designed for a non-convex setting, and one of the main contributions of this paper seems to be the redesign of STORM's parameters for a convex setting.\n4.\tMany fully parameter-free algorithms[1,2] have appeared recently, which do not require knowledge of the smoothing constant $L$ and can achieve the same convergence rate. In contrast, this paper still relies on smoothing constants to determine the learning rate. Under fully parameter-free conditions, even a wide range of learning rate options appears to lose its significance.\n\n[1] Khaled A, Jin C. Tuning-Free Stochastic Optimization[J]. arXiv preprint arXiv:2402.07793, 2024.\n[2] Ivgi M, Hinder O, Carmon Y. Dog is sgd’s best friend: A parameter-free dynamic step size schedule[C]//International Conference on Machine Learning. PMLR, 2023: 14465-14499." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Stochastic convex optimization methods that allows easy tuning similarly to (noiseless) Gradient Descent" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024do,\ntitle={Do Stochastic, Feel Noiseless: Stable Stochastic Optimization via a Double Momentum Mechanism},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zCZnEXF3bN},\nnote={under review}\n}" }, "abstract": { "value": "Optimization methods are crucial to the success of machine learning, with Stochastic Gradient Descent (SGD) serving as a foundational algorithm for training models. However, SGD is often sensitive to the choice of the learning rate, which necessitates extensive hyperparameter tuning. In this work, we introduce a new variant of SGD that brings enhanced stability in two key aspects. First, our method allows the use of the same fixed learning rate to attain optimal convergence rates regardless of the noise magnitude, eliminating the need to adjust learning rates between noiseless and noisy settings. Second, our approach achieves these optimal rates over a wide range of learning rates, significantly reducing sensitivity compared to standard SGD, which requires precise learning rate selection.\nOur key innovation is a novel gradient estimator based on a double-momentum mechanism that combines two recent momentum-based techniques. Utilizing this estimator, we design both standard and accelerated algorithms that are robust to the choice of learning rate. Specifically, our methods attain optimal convergence rates in both noiseless and noisy stochastic convex optimization scenarios without the need for learning rate decay or fine-tuning. We also prove that our approach maintains optimal performance across a wide spectrum of learning rates, underscoring its stability and practicality. Empirical studies further validate the robustness and enhanced stability of our approach." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "online convex optimization", "stochastic convex optimization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/0daa0519287b4212f1139087d1d1fdf00f685183.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/53e552f2aed0303eb9fb47136c26f0579d17842e.zip" }, "title": { "value": "Do Stochastic, Feel Noiseless: Stable Stochastic Optimization via a Double Momentum Mechanism" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Could you use fast random subsampling methods, such as random Fast JL transform?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The idea of updating the iterate only in a subspace by rotations is not yet well explored and practically useful generalization of the random Riemannian coordinate descent. The paper provides extensive convergence analysis for the methods (although I haven't checked the proofs in the appendices in great detail). The paper is also well written and clearly explains the main concepts. There are multiple numerical experiments in the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a randomized Riemannian submanifold method for solving optimization problems constrained to the set of $n\\times p$ orthogonal frames, also called the Stiefel manifold. The main innovation of the algorithm is to randomly select an $r$-dimensional subspace, in which the current iterate is rotated in a way that locally decreases the objective. In this way, the method can be seen as a Riemmanian analogue to the random block coordinate descent for Euclidean optimization. \n\nThere are two sampling distributions considered for the random selection of the $r$-dimensional subspace: Haar uniform sampling on the orthogonal group and random selection of indices (referred to in the paper as a random permutation matrix). Both of these have the complexity of $nr^2$ although the latter can be computed in a more efficient way due to not needing to perform QR decomposition.\n\nThe paper provides global convergence analysis to critical points and local to a local minima (under PL-ineq.) in expectation and also in high-probability. The method can be also extended/analyzed to the stochastic variant when the objective gradients are random with bounded variance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "### Writing comments\nAlthough the paper is well written overall, I have some remarks on formulation of certain ideas:\n* Line 161: the authors state that the random submanifold is defined by the random orthogonal matrix $P_k$. This is not true strictly speaking, in fact, I think it is only the first $r$-columns of the matrix $P_k$ that define the subspace for the rotation.\n* Adding on the remark above: Isn't it then enough to generate only a random orthogonal frame from $St(n,r)$ that defines the subspace?\n* Lines 172 - 178: This part describes one of the the main technical observations: that it is possible to do a local approximation of the $\\tilde F(Y)$ by computing retraction around an identity matrix. Can this be theoretically shown that this is a local approximation? It is not immediately clear to me why this is the case?\n* Eq. 4: there is a mistake/typo after the first equality\n* Line 198: this is very minor remark, but does the order of the first $r$ indices in the sampled permutation matrix matter? In this sense, wouldn't it be sufficient to sample $r$ indices without replacement instead of the permutation? Practically this might be the same so it might not matter. \n\n### References on sketching\nThis method seems to be related to random subspace sketching. For example a very recent work of (Shustin & Avron, 2024; https://www.jmlr.org/papers/volume25/21-1022/21-1022.pdf) considers random sketching algorithm for fast optimization over orthogonal constraints. This would be good to discuss in the paper (if indeed related), and possibly also to compare against in the numerical experiments. \n\n### Numerical experiments\nThe paper provides numerous numerical experiments but I would like to see the study of dependency on the choice of the random subspace dimension $r$. In the first experiment, the orthogonal procrustes, I am missing the information of the choice of $r$ altogether. I would also like to see a discussion on the computational cost of the random Haar uniform sampling of O(n) as I imagine this will be computationally complex operation. \n\nOverall, I like the main paper idea, but I am missing more direct comparison with sketching methods in Riemnannian setting as well as the explanation why we can expect (4) to hold as a good first order approximation to minimizing $\\tilde F$. I would also like to see the numerical section improved with more extensive discussion on the choice of $r$ and how it relates to the per-iteration cost." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Theorem 1. How to ensure $X_k$ remains inside $U$ under $X_{k_0} \\in U$? The stochasticity induced by sampling seems problematic. Purely assuming all $X_k \\in U$ may be too strong. \n2. Remark 3. The retraction is at cost of $\\mathcal{O}(np^2)$, while calculating the Euclidean gradient is easy to be $\\mathcal{O}(n^2p)$ flops, as seen in problems like PCA. Why does the Riemannian gradient descent become impractical?\n3. Theorem 2. Is it possible to design a verison of submanifold method that has exact convergence rather than high probability.\n4. Figure 2. Many Stiefel maniofold applications are with small $p$. Could you also add some numerical experiments on small $p$ to see the comparsions?\n5. Figure 5. The improvement on the accuracy appears marginal. The early-stage superior performance might also be achieved by using a larger step size." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed randomized submanifold method includes several coordinate descent methods as specific examples and allows flexibility in setting the size of the submanifold.\n2. Convergence is established in both expectation and high-probability settings, with clear derivations provided." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a randomized submanifold algorithm for solving optimization problems with orthogonality constraints. The authors provide convergence guarantees for the proposed algorithm and conduct numerical experiments. Theoretical and numerical comparisons with existing literature are also presented." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The algorithm does not achieve exact convergence, potentially due to the stochasticity in subspace selection. Is there a way to ensure exact convergence, for example, by adding restrictions on the sampling process?\n2. It appears that the total computational complexity shows no improvement compared to Riemannian gradient descent and may even be higher when using orthogonal sampling. Additionally, the algorithm only converges to a neighborhood rather than a first-order stationary point.\n3. The numerical comparisons are insufficient. It would be helpful to see the effects of different submanifold sizes $r$ and small $p$ values." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. As cited in line 88 of the manuscript, the literature [1] proposed a randomized method, RSSM, which also updates on a low-dimensional random submanifold. Could you compare it with the method in this work in terms of computational cost per iteration? Additionally, is there any relationship between RSSM and RSDM?\n2. In line 382 of the manuscript, it states that analysis for Theorem 3 can be easily extended to mini-batch gradient descent. Can the authors explain this point briefly?\n3. In the experiments, the parameter $p$ is consistently close to $n$, for which the proposed RSDM outperforms other methods. It would be beneficial to demonstrate at what percentage of $n$ the value of $p$ needs to reach before RSDM begins to show its advantage.\n4. In Section 7.4, the numerical results of the deep learning task seem confusing, where the accuracy hovers around $40\\\\%$ to classify CIFAR10. Additionally, only the tendency of accuracy is presented; including the training/test loss would be more illustrative.\n\nI would like to increase my grade if the concerns are well addressed.\n\n[1] Randomized submanifold subgradient method for optimization over Stiefel manifolds. 2024" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The proposed RSDM is novel, which updates on a low-dimensional random submanifold instead of the full Stiefel manifold. In this way, it reduces the computational costs in each iteration, contributing to large-scale optimization.\n2. The paper provides a comprehensive theoretical analysis, including general nonconvex optimization problems, nonconvex functions satisfying the PL condition, and the results in stochastic settings. The discussion about the trade-off between efficiency and convergence is interesting.\n3. The effectiveness and efficiency of the proposed algorithm are validated on an array of numerical experiments, and the source code is available." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a randomized Riemannian submanifold descent method (RSDM) to address large-scale optimization with orthogonality constraints. Specifically, it mitigates the computational costs associated with the retraction in each iteration by updating on a low-dimensional random submanifold rather than the full Stiefel manifold. Two sampling strategies, orthogonal and permutation sampling, are provided. Theoretical analysis guarantees convergence, and extensive experiments demonstrate the effectiveness and efficiency of RSDM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The contribution seems limited in the scenario when $p=\\Omega(n)$. However, numerous applications typically involve the setting $p\\ll n$. Specifically, as discussed in Remark 3, the total complexity of the standard Riemannian gradient descent is $O(np^2\\epsilon^{-2})$, while the proposed method costs $O(n^3\\epsilon^{-2})$, which appears as a theoretical gap." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. In Line 69 and Line 92, the authors say that the complexity of non-standard linear algebra reduces to O(r^3) from O(nr^2). I understand O(r^3) comes from the cost of retraction operation. However, the per-iteration cost is still O(npr) or O(nr^2) (see Section 4.1 and Section 4.2). It would be better to clarify these in the statement.\n2. In line 142 and line 146, it is not suitable to use “denote … as …”. Besides, in line 155, what is the meaning of $X^*$?\n3. It can be better to explicitly write down the expression grad F_k(I_n) arising in (4) using the Euclidean gradient. Although it is very simple, it can help readers understand lines 212-213.\n4. In Line 222, why does the gradient computation only require O(nr^2)?\n5. Could the authors add more discussions on the parallel implementation?\n6. Could the authors get rid of the assumptions in Line 304? \n7. In line 484 of Section 7.3, it should be Figure 4 instead of Figure 2. In addition, if possible, could the authors also compare their RSDM with RSSM in (Cheung-Wang-Yue-So ‘24) in Line 577?\n8. In Line 467-470, Can the phenomenon of switching from sublinear rate to linear rate be explained by the error bound property (or the Riemannian PL property)? See Theorem 1 and Theorem 2 in “Liu H, So A M C, Wu W. Quadratic optimization with orthogonality constraint: explicit Łojasiewicz exponent and linear convergence of retraction-based line-search and stochastic variance-reduced gradient methods[J]. Mathematical Programming, 2019, 178(1): 215-262.” The authors can add some discussion if this is related.\n9. The update of RSDM is similar to the algorithm in \"A Block Coordinate Descent Method for Nonsmooth Composite Optimization\nunder Orthogonality Constraints\" (https://arxiv.org/pdf/2304.03641). For example, the formula in Line 215 is similar to (12) in the reference. Could the authors compare RSDM with the algorithm in the reference and highlight the differences more explicitly?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. This paper proposes a novel random submanifold descent method (with two sample strategies), which enables low-cost retractions and low per-iteration complexity. \n2. The authors provide a thorough convergence analysis under the non-convex setting, the local Riemannian PL setting, and the stochastic setting.\n3. The proposed algorithm outperforms other existing methods in various problems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper deals with large-scale optimization problems with orthogonality constraints. To overcome the computational bottleneck of the retraction operator, the authors propose a novel \"subspace\" technique that updates variables on random submanifolds, reducing per-iteration complexity. They introduce two strategies for selecting these submanifolds and analyze the convergence of their methods for both general nonconvex functions and those with local Riemannian PL conditions. Besides, the approach is also applicable to certain quotient manifolds. Experimental results demonstrate the method's effectiveness across various problems." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In line 87, the authors mentioned the poor parallelization of existing methods. However, they did not say too much about why RSDM can be implemented in parallel. In line 93, the authors say “allowing parallel computation”. Adding more discussion and exploration in this aspect can greatly strengthen this paper. For example, elaborating more on the parallel implementation of RSDM or even showing the speedup compared to the sequential implementation in the experiment part would be helpful.\n2. We know from Definition 1 that the Riemannian PL holds locally. In Line 304 of Theorem 1, it says “Suppose k0 large enough such that $X_{k0} \\in U$”. I understand that this assumption is because the iterates may not always stay in the local region due to algorithmic randomness. Is it possible to get rid of such an assumption? This seems difficult to me. Similar things happen in other theorems." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024efficient,\ntitle={Efficient optimization with orthogonality constraint: a randomized Riemannian submanifold method},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zCncHdGsOa},\nnote={under review}\n}" }, "abstract": { "value": "Optimization with orthogonality constraints frequently arise in various fields such as machine learning, signal processing and computer vision. Riemannian optimization offers a powerful framework for solving these problems by equipping the constraint set with a Riemannian manifold structure and performing optimization intrinsically on the manifold. This approach typically involves computing a search direction in the tangent space and updating variables via a retraction operation. However, as the size of the variables increases, the computational cost of the retraction can become prohibitively high, limiting the applicability of Riemannian optimization to large-scale problems. To address this challenge and enhance scalability, we propose a novel approach that restricts each update on a random submanifold, thereby significantly reducing the per-iteration complexity. We introduce two sampling strategies for selecting the random submanifold and theoretically analyze the convergence of the proposed method. We provide convergence results for general nonconvex functions and functions that satisfy Riemannian Polyak–Łojasiewicz condition as well as for stochastic optimization settings. Extensive experiments verify the benefits of the proposed method, showcasing its effectiveness across a wide variety of problem instances." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Oprimization", "Orthogonality constraint", "Riemannian optimization", "Stiefel manifold" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/344ec539980c9f2d76d94044fdfcf6c1c815ef9e.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/14fc03e4fdc748559eb4d12180a073724be0a254.zip" }, "title": { "value": "Efficient optimization with orthogonality constraint: a randomized Riemannian submanifold method" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See Weaknesses section. Additionally,\n- The range of the y axis for the four plots on the right in Figure 5 are missing. Are these also from 0 to 1?\n- How long does training take for training on 1B Kinetix environments? Would be good to see if training on such a large number of environments itself would be a bottleneck for learning generalist agents." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Kinetix provides 66 hand-designed levels while having the option to edit tasks with a graphical editor or to randomly generate more levels with rejection sampling. \n- The unified goal and dynamics within all environments encourage policies to have physical reasoning capabilities instead of merely memorizing the solution for some particular task, which is a valuable objective for researchers to pursue. \n- Kinetix provides a way to generate unlimited environments and tasks with a unified goal, objects, and dynamics, which could be of interest to multiple research communities like generalist RL policy learning, meta-learning, world modeling, spatial understanding and physics reasoning, and so on." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Kinetix, a new 2D simulated benchmark designed for training generalist agents with capabilities in fine-grained motor control, navigation, planning, and physical reasoning. The benchmark is built on a novel hardware-accelerated physics engine called Jax2D. Kinetix enables the procedural generation of a vast amount of environments using simple shapes, joints, and thrusters, allowing for tasks that include robot locomotion, object manipulation, simple video games, and classic reinforcement learning scenarios. Each environment shares a unified objective: “make the green shape touch the blue shape without touching the red shape,” enabling zero-shot generalization to new environments within the same distribution. Experimental results show that policies trained across a wide range of environments generalize better to unseen tasks, and fine-tuning these generalist policies on new tasks yields improved performance over training from scratch. This work contributes to the development of generalist RL agents and open-ended physics-based control tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper notes that as the generated environments increase in complexity, they may become unsolvable, which could contribute to the lower performance observed in the Large-level environments. If so, how does this impact the usability and interpretability of the benchmark results? To what extent does this affect the performance results reported in Figure 3?\n- It is unclear whether the proposed benchmark supports visual observations, which are essential for training generalist policies and building agents that can operate in real-world settings.\n- Although Kinetix can generate a vast range of environments, it is unclear how this benchmark would generalize to tasks or environments outside of its defined task distribution." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Kinetix enables a wide distribution of morphologies and initial environment scene configurations, but it doesn’t deviate beyond the single unified goal. How can it be expanded to also cover a wide distribution of goals and potentially even task specifications?\n\n2. Does Kinetix support parallel pixel rendering, such as for vectorized image-based experiments?\n\n3. Is SFL only choosing between the S, M, and L levels?\n\n4. Are the inputs into the policy purely one-hots? (One-hot encoding for each of the polygons, thrusters, and shapes.)\n\n5. Is the (x, y) 2D position of each polygon an input into the network? It would seem that positional embeddings not of the ordering of polygons, but their actual spatial positions, would matter a great deal in this problem.\n\n6. In section 4, authors write that Kinetix is a deterministic setting (thus satisfying one of the conditions for using SFL). How is Kinetix deterministic, given that environments are randomized?\n\n7. How many environments were used during training, and how many environments were held-out for zero-shot evaluation?\n\n8. What was the generalist policy in Figure 5 trained on? A distribution of environments over all 4 tasks at their hard level?\n\n9. Say we train an agent only trained on L levels. How does it perform on held-out levels of a different difficulty (M and S)?\n\n10. Heatmaps were an interesting way to convey the agent’s performance. Why not simply graph the agent’s x distance from the goal, with x-axis being the training iterations?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Introduces a physics engine that provides “almost entirely dynamically specified” scenes, where environments with different robot morphologies can be vmap-ed and run in parallel, which is not doable with prior Jax-based sim frameworks like Brax.\n\n2. Paper is clearly written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Jax2D physics engine, which is a reimplementation of Box2D but written in Jax, and introduce the Kinetix environment on top of Jax2D. Kinetix allows for procedurally generated open-ended environments with different robot morphologies. Authors create a self-attention-based policy and demonstrate performance zero-shot, with pretraining, and with finetuning on the target environments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. All environments in benchmark must fall under the goal of making green shape touch blue shape without touching red shape. This seems to mainly constrain the problem to single-step tasks, where the reward of minimizing the distance from green to blue always incentivizes progress. Was this unified goal constraint purposefully imposed by design, or was it a constraint of Jax implementation, where the reward function for all environments must be the same to be parallelizable?\n\n2. Authors emphasized that parallelism and speed were big advantages of Jax2D. Since it is a reimplementation of Box2D, and this is a critical contribution of the paper, what are the performance gain metrics over Box2D?\n\n3. Experiments were on multi-discrete action space with binary rewards. However, it would strengthen the argument of the paper to do experiments on more of the important features of Kinetix, such as pixel-based observations and continuous action space.\n\n4. The state representation of the policy is very specific to the Kinetix environment suite and not very generalizable to other 2D RL problems. For instance, each entity is encoded separately and there is no scene-level encoding that is passed in as observation for the policy. Often, it is essential for a policy to understand the entire scene when predicting an action.\n\n5. There were no supplementary materials submitted, which would have been a good opportunity to show video rollouts of the trained agent in action.\n\n6. Experiments were mainly limited to the improvement of finetuned policies over pretrained and task-specific, trained-from-scratch policies. However, I would have liked to see more experiments that provide additional insights beyond “finetuning is mostly good” and “zero-shot mostly doesn’t work.” For instance, using Kinetix for lifelong learning, transfer learning, and cross-embodiment learning.\n\n7. Abstract sentence seems like an oversell, given the results. “Our trained agent exhibits strong physical reasoning capabilities, being able to zero-shot solve unseen human-designed environments.” Most would also disagree with the 2D learned behaviors as “strong physical reasoning capabilities.”\n\n8. Minor: I think the wrong citation was provided for MJX in Section 7 (that work seems to be Mujoco).\n\n9. Minor: Experiments would benefit from some comparison to prior approaches/architectures, though this is less important given this is mainly a systems/framework paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- For the environment generator, it is mentioned that there may exist unsolvable levels which automatic curriculum methods can filter out. Could you clarify what was done here?\n- How does the choice of algorithm affect the performance in your benchmark. Do you anticipate releasing a dataset of transitions from Kinetics which can be used for offline 2 online RL?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Provide a highly efficient 2D rigid-body physics engine, leveraging JAX for scalable computation, with speedups of up to 30x when training an RL agent, allowing for \n- The learnt agent is highly effective at Zero-Shot transfer in the S and M levels that are held out, indicating the efficacy of pre-training on a wide set of procedural generation tasks. Additionally, show faster convergence/higher performance with this initialization\n- Have interpretable/handmade levels to understand the performance on different sizes/difficulties of tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides a framework for procedurally generated 2D physics-based tasks to learn an agent for physical control that can effectively transfer to tasks that involve physical control. They provide a hardware-accelerated physics engine that allows for cheap/efficient simulation to generate a mixed-quality pre-training dataset for online RL. The authors additionally provide human interpretable handmade levels in Kinetix to understand the type of tasks and interactions the agent must do. The authors show the efficacy of both zero-shot transfer to new tasks and the agent when fine-tuned on a new task. The authors evaluate their agent on classic RL environments such as Cartpole and video games like Pinball." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The JAX2D environment seems to be somewhat limited in its expressivities, modeling only 4 unique entities, which may not transfer to a wide set of domains/tasks outside of the ones studied. \n- The task/reward function seems to be fixed across all environments to collide the green and blue shaped objects, while avoiding red shapes. Additional reward shaping seems to be needed for effective training, leading to some limited applicability of generating this data at scale for any set of tasks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "They are mentioned in the Weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well-written, organized, and straightforward.\n- Extensive testing across various task complexities validates its robustness in diverse 2D environments.\n- This paper has strong potential to serve as a valuable benchmark for future research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Kinetix, a 2D physics-based RL environment aimed at enhancing generalization in RL agents. Leveraging the simulation, the agent is pre-trained for billions of steps, enabling zero-shot evaluation on novel tasks. Fine-tuning further improves performance, surpassing traditional RL methods. The approach integrates a transformer architecture with PPO as the core RL algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Real World Tasks:** While this paper provides a strong foundation in 2D simulations, expanding its scope to assess the agent’s adaptability to real-world tasks, such as 3D simulations or complex dynamics as seen in [1,2], would enhance its practical relevance. Bridging this gap could amplify the study’s contributions, offering broader insights into real-world generalization and scalability.\n\n- **Filtering out:** The authors mention that trivial and unsolvable levels are filtered out. What quantitative metrics were used to determine this filtering.\n\n- **Generalizability:** The claims of generalizability might be overstated given that the tasks remain in controlled simulations. Could the authors clarify the expected limitations of deploying such an agent in real-world scenarios with unpredictable environmental factors?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Training with reinforcement learning on a vast open-ended distribution of physics-based tasks leads to an agent that can zero-shot solve human-designed problems." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024kinetix,\ntitle={Kinetix: Investigating the Training of General Agents through Open-Ended Physics-Based Control Tasks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zCxGCdzreM},\nnote={under review}\n}" }, "abstract": { "value": "While large models trained with self-supervised learning on offline datasets have shown remarkable capabilities in text and image domains, achieving the same generalisation for agents that act in sequential decision problems remains an open challenge.\nIn this work, we take a step towards this goal by procedurally generating tens of millions of 2D physics-based tasks and using these to train a general reinforcement learning (RL) agent for physical control.\nTo this end, we introduce Kinetix: an open-ended space of physics-based RL environments that can represent tasks ranging from robotic locomotion and grasping to video games and classic RL environments, all within a unified framework.\nKinetix makes use of our novel hardware-accelerated physics engine Jax2D that allows us to cheaply simulate billions of environment steps during training.\nOur trained agent exhibits strong physical reasoning capabilities, being able to zero-shot solve unseen human-designed environments. Furthermore, fine-tuning this general agent on tasks of interest shows significantly stronger performance than training an RL agent *tabula rasa*. This includes solving some environments that standard RL training completely fails at.\nWe believe this demonstrates the feasibility of large scale, mixed-quality pre-training for online RL and we hope that Kinetix will serve as a useful framework to investigate this further.\nWe open-source Jax2D, Kinetix, and our final model weights." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "reinforcement learning", "open-endedness", "unsupervised environment design", "automatic curriculum learning", "benchmark" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/5e42f925c25f98147717c4f0ee2a220bd7d4caa0.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Kinetix: Investigating the Training of General Agents through Open-Ended Physics-Based Control Tasks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Refer to the content in the Weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "(1) The writing and organization of this paper are very clear.\n\n(2) This paper is intriguing because it is based on the idea of enhancing the binding capability of the current ligand by considering the binding positions of other ligands that target the same protein." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a molecular docking framework called GROUPBIND, which enhances the binding capability of a ligand to a target protein pocket by leveraging other ligands that bind to the same pocket. \nThe framework introduces message padding among groups of ligands and a triangle attention module for protein-ligand pairs. Experimental results validate that GROUPBIND improves docking performance based on diffusion models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) Why are the results of DIFFDOCK in Table 1 worse than those in the original paper, and it seems that the bold text annotations might be inaccurate?\n\n(2) From Figure 1, we can see that molecules binding to the same pocket indeed have similar structures, but how many pockets in the dataset exhibit this situation? Is there any statistical data on the number of pockets and the corresponding similar ligands in the PDBBind dataset? \n\n(3) During inference, when searching the database for ligands similar to the query ligand, how many entries in the test set can retrieve similar ligands? If similar ligands cannot be retrieved, does that mean the model becomes ineffective?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1.\tThe definition of \"ligand ground graph\" needs explicit formalization. The current lack of clarity makes understanding this crucial concept challenging. For example, when referring “noisy group ligands” in line 214, it is difficult to understand this concept.\n2.\tThe references to tables and figures are confusing. Avoid using \"Section\" to denote figures, tables, and section of manuscript at the same time. Use \"Figure\" when referring to images (e.g., Figure 1, Figure 2) and \"Table\" for tabular data (e.g., Table 1, Table 2). \n3.\tA thorough proofread is necessary to correct spelling and grammatical errors. For instance, \"beyound\" on line 101 should be corrected to \"beyond.\" A comprehensive review of the entire text is recommended.\n4.\tFigure 7's inclusion of Figures 3, 4, 5, and 6. For better visual organization and clarity, these figures should be presented as subfigures within a single figure (e.g., Figure 7a, 7b, 7c, and 7d). This allows for easier comparison and a more streamlined presentation.\n5.\tLacking explanation of \"NG\" in Figure 6. While the meaning of \"NG\" in Figure 6 might be discernible from the main text, its meaning should be explicitly stated. This ensures clarity and avoids requiring readers to search through the text for an explanation. A concise definition or clarification of \"NG\" within the caption is crucial." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The idea of leveraging similar binding poses among ligands targeting the same protein is intriguing and biologically relevant.\n2. The experimental results suggest that incorporating augmented ligands improves docking performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces GroupBind, a blind rigid docking method predicated on the biochemical observation that ligands binding to the same target protein often adopt similar poses. GroupBind employs an interaction layer for a group of ligands and a triangle attention module to embed protein-ligand and group-ligand pairs. Performance is evaluated on the PDBbind dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The core idea is similar to MCS (Maximum Common Substructure) docking [1, 2], which assume that ligands with similar substructures exhibit similar docking poses. GroupBind, however, assumes all ligands share similar docking poses. Figure 1 depicts highly similar ligands with similar docking structures. Conversely, [3] (Figure 4) illustrates cases where structurally distinct ligands adopt distinct poses. A statistical comparison quantifying the difference between the MCS docking assumption and GroupBind's assumption is warranted.\n\n2. Comparing GroupBind-Ref against blind docking methods like DiffDock is unfair. DiffDock performs blind docking, whereas GroupBind-Ref utilizes prior knowledge of the binding pocket, making it a site-specific docking method. This introduces a significant advantage for GroupBind-Ref.\n\n3. The evaluation should include more baselines such as FABind [5] and FABind+ [6], for which source code is available. Expanding the evaluation to include datasets like PoseBuster would assess GroupBind's ability to predict physically plausible structures.\n\n4. The reported top-1 docking success rate for DiffDock (32.4% with 40 samples in Table 1) appears considerably lower than previously \nreported results (38.2% in [4] and 36.0% in [5]). This discrepancy requires clarification.\n\n5. The clarity of the writing could be improved. Specific points of confusion are detailed below in the Questions section.\n\n[1] A High-Quality Data Set of Protein-Ligand Binding Interactions Via Comparative Complex Structure Mod, 2024\n\n[2] FitDock: protein-ligand docking by template fitting, 2022\n\n[3] DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking, 2024\n\n[4] DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking, 2022\n\n[5] FABind: Fast and Accurate Protein-Ligand Binding, 2023\n\n[6] FABind+: Enhancing Molecular Docking through Improved Pocket Prediction and Pose Generation, 2024" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1.\tThe authors should account for the effect of protein similarity on the results by performing redundancy removal on the test set proteins that are either duplicated or highly similar to those in the training or validation sets. Tools like MMseqs or other alternatives could be used for this process.\n2.\tThe authors need to explain how the similarity between ligands in the group ligand set impacts the results.\n3.\tHow would the results of GROUPBIND change if tested on the latest PoseBuster (version 1 and version 2)?\n4.\tIn Figure 4 on page 8, does the success rate refer to Top 1 or Top 40? What do “SG” and “AG” specifically mean? These details should be clarified in the figure caption.\n5.\tWhat does Figure 5 on page 8 illustrate, and where is it referenced in the article?\n6.\tIn Figure 6 on page 8, does the success rate refer to Top 1 or Top 40? The meanings of “NG”, “SG”, and “AG” should also be clarified in the figure caption.\n7.\tOn line 457 on page 9, the percentage 36.3% is mentioned twice, which could cause confusion and should be clarified." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The study proposes a new molecu lar docking framework to simultaneously consider multiple ligands docking to a protein." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a deep learning model for molecular docking, which improves data utilization and model quality by docking multiple molecules to the same protein. It further enhances the connections between similar atoms across different molecules and the same protein amino acids using a triangular perceptual network. Through this multi-molecule docking approach, the model surpasses existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, this paper offers a certain level of contribution, but the experimental section requires clearer descriptions and further discussion. For the figures in the paper, the authors should provide detailed captions, including explanations of the methods used." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "How does your method compare to ComBind in a head-to-head comparison? Answering this would significantly increase my enthusiasm for the paper as, in addition to providing a more relevant baseline, it would involve testing using ligands without known structure as augmentation ligands (assuming the ComBind evaluation framework is used)." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Particularly during the lead optimization phase, it is reasonable to expect that there are a number of known ligands for the target protein and using this information to improve docking performance could help in the discovery of better ligands.\n\nA reasonable approach for message passing across ligands with triangle attention is described. This might provide a more general template for other tasks where output are represented as graphs and there is a known consistency bias.\n\nThis approach does not require the accessory ligands to have a known structure at inference time, which is a realistic scenario.\n\nInformative ablation studies are performed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper describes an approach for leveraging the insight that similar ligands that bind to the same protein target are expected to binding similarly. A method for attending between ligands as part of a diffusion docking process is described and the results convincingly show the benefit of this approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The contributions are inappropriate as written. This is not the first time the concept that multiple similar molecules can be used to enhance molecular docking has been described. ComBind (Paggi et al, 2021; cited in the paper) does exactly that. There are other methods that use this insight in different ways (e.g. selecting poses from ensemble docking). The contributions should be qualified that this is the first end-to-end deep neural network approach to molecular docking that uses this insight.\n\nThe results only compare to single-ligand docking. Comparing to ComBind (or OpenComBind if lacking a Schroedinger license) would be more relevant. Can the diffusion model approach make better use of the similarity bias than previous methods? This question goes unanswered.\n\nNot going beyond the PDBbind to identify alternative augmentation ligands (of which there are many, as structures aren't needed for inference) is a missed opportunity that weakens the paper as with the current evaluation framework many ligands can't be put into groups. Replicating the ComBind evaluation would provide predefined groups of ligands while also making it possible directly compare to a conventional approach.\n\nThere are issues with using a time split to assess generalizability, but as this is the same split used with DiffDock it is appropriate to use for the comparisons performed here.\n\nThe overloading of \"k\" to mean two different things in Fig 2 is confusing.\n\nI found equations (4) and (5) confusing due to the overloading of z - the text says these two separate values (presumably equivalent to AF2 incoming and outgoing triangle attention edges) will be stacked, while the equations say they are summed.\n\n\"and since the C-C bond length is about 1.5A.\"\n\nTable 2 is apparently showing Top-10 results, but this fact is only stated in the main text. Why not replicate the reporting in Table 1? \"Med.\" isn't defined, but is presumably median RMSD.\n\n4.4: \"Section\" is used when \"Figure\" is meant. \"ifself\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024group,\ntitle={Group Ligands Docking to Protein Pockets},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zDC3iCBxJb},\nnote={under review}\n}" }, "abstract": { "value": "Molecular docking is a key task in computational biology that has attracted increasing interest from the machine learning community. While existing methods have achieved success, they generally treat each protein-ligand pair in isolation. Inspired by the biochemical observation that ligands binding to the same target protein tend to adopt similar poses, we propose \\textsc{GroupBind}, a novel molecular docking framework that simultaneously considers multiple ligands docking to a protein. This is achieved by introducing an interaction layer for the group of ligands and a triangle attention module for embedding protein-ligand and group-ligand pairs. By integrating our approach with diffusion based docking model, we set a new state-of-the-art performance on the PDBBind blind docking benchmark, demonstrating the effectiveness of our paradigm in enhancing molecular docking accuracy." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "molecular docking", "ai4science" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a720afd306b2b62a4a50993134d384d54f5022ea.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Group Ligands Docking to Protein Pockets" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The authors attempted to improve the state-of-the-art last-iterate convergence rate for general monotone game and their time-varying variants --- the existing best bounds are $T^{-1/2}$ for strongly monotone games by Lin et al. (2021), and $T^{-1/6}$ for linear games by Cai et al. (2023)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper design an algorithm for monotone games with bandit feedback that has last-iterate convergence guarantee. The convergence rate is $T^{-1/4}$ for monotone games, and $T^{-1/6}$ for linear games." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I have concern about the soundness of the results. At the first sight, there is a strange gap between the results for general monotone game and the linear game in Table 1: for general monotone game the bound is $T^{-1/4}$, and for linear game the bound is $T^{-1/6}$. However, linear game is a special case of monotone game. The author claimed that for linear games, one cannot find a convex function $p$ that satisfies the definitions in Line 217. However, for any linear game, one can always artificially add a perturbation function $\\epsilon ||x||^2$ to make it strictly monotone. Applying the $T^{-1/4}$ bound for general monotone game, and bounding the error due to perturbation, one should be able to get a bound of $T^{-1/4}+\\epsilon$ for linear games. Then taking $\\epsilon\\to 0$, it seems we should also get a $T^{-1/4}$ bound also for linear games. \n\n- It seems the main gap lies in Theorem 5.1. In the theorem, although the right-hand side of the bound is $T^{-1/4}$, the left-hand side is not the usual Euclidean distance to the equilibrium, but the distance induced by the $p$ function defined in Line 217. This distance could be arbitrarily smaller than the Euclidean distance if $p$ is close to linear. Translating this distance back to Euclidean distance will just make this bound vacuous. Therefore, I think the claim of the paper is flawed and it does not seem to make improvement over prior work." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "A couple of questions/issues for the authors:\n\n1. In the third example, it is claimed that splittable routing games are monotone. Can the authors justify this claim? It is not obvious to me.\n2. Before Proposition 5.1, the authors write that \"We have the following proposition which shows that the social welfare converges to optimal welfare on average\". This is not correct, the dynamics only approximate the optimal welfare (depending on the smoothness parameters)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper studies an important and well-studied problem in the intersection of optimization and game theory, and makes a number of concrete contributions that improve over the prior state-of-the-art. As described above, the paper obtains the best-known rates for monotone games and time-varying games under bandit feedback, which is certainly an important contribution. Further, the techniques employed in the paper are also quite non-trivial, and rely on using two different regularizers in an interesting way. While most prior results examine last-iterate convergence in the full-feedback model, the more realistic bandit model introduces several technical challenges. I believe that the results of the paper are well within the scope of ICLR, and will be well-received from the community on learning in games. The most related papers have been adequately discussed. All claims appear to be sound; I did not find any notable issue in the approach. The writing overall is of high quality." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the last-iterate convergence of no-regret dynamics in monotone games under bandit feedback. In particular, they provide a new algorithm that attains the best-known rate of convergence of $T^{-1/4}$ and $T^{-1/2}$ under strong monotonicity, while guaranteeing at the same time the no-regret property. This is the first convergence rate for uncoupled dynamics in monotone games under bandit feedback. Further, for time-varying games, the convergence rate improves over the prior state-of-the-art." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In terms of weaknesses, it appears that the proof mostly relies on existing techniques, although the way those techniques are used seems to have new aspects. I would encourage the authors to highlight more the technical challenges and the technical contributions compared to prior work. There is also one issue that really confuses me: Section 5.3 is about the special case of linear cost functions; how is possible that the rate is worse in that special case? For example, matrix games are clearly monotone, so why doesn't the rate of $T^{-1/4}$ apply in that case?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- A big concern is the potential issue induced by the non-uniqueness of the Nash equilibrium for monotone games. In Theorem 5.1, convergence is measured by the distance of the actions to x_i^*. Does this statement hold for each x_i^*? Should the convergence be measured using the distance of the actions to the set of Nash equilibria? \n\n- The concept of 'uncoupled' is quite confused. In games, each player's payoff depends on the actions of other agents, which suggests that their dynamics should be inherently coupled. However, a clear description of strongly uncoupled dynamics is not given.\n\n- The assumptions required for each theorem are not explicitly stated, making it difficult for readers to follow directly from the theorem statements.\n\n- The presentation of Algorithm 1 should be improved. For example, agents should first sample z and then play the perturbed action. Besides, each agent should receive feedback after all agents play their actions. Please see Algorithm 2 in [Lin 2021]." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper provides a theoretical analysis on several cases, including the convergence rate for monotone games (both in expectation and with high probability), social welfare, linear cost, and the time-varying case." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the convergence of mirror-descent algorithm in general monotone games and strongly monotone games. Several theoretical results are given. The time-varying monotone games are also considered. Overall I find the analyzed problem not new and the convergence rate is not surprised." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Many convergence rates in this paper fail to improve the existing result. The advantage of the proposed algorithm is unclear. The simulation results are also weak to demonstrate the advantage of the proposed algorithm. \n\n- There is a large room for the simulation section to be improved. Please add comparison of Algorithm 1 to, at least, the method in [Lin 2021]. Besides, the examples used in simulation are simple.\n\n- The technical contributions are unclear. The techniques used, including the ellipsoidal gradient estimator, are similar to those found in existing literature." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Can the $O(T^{-1/4})$ last-iterate convergence rate be considered as the $O(T^{3/4})$ regret for bandit convex optimization?\n\n2. For bandit optimization with convex and smooth functions, the state-of-the-art regret is $O(T^{2/3})$ by [1], which seems to correspond to a last-iterate rate of $O(T^{-1/3})$ in the setup of smooth monotone games. Is it possible to improve the current $O(T^{-1/4})$ to $O(T^{-1/3})$?\n\n3. For BCO problems with smooth or strongly convex functions, FTRL seems to be more popular than OMD [1,2]. Is it possible to represent Algorithm 1 using an FTRL-based update rule? Besides, are FTRL and OMD equivalent in the setup of BCO or bandit smooth monotone games?\n\n4. The rates in Table 2 are not last-iterate results. First, the guarantees in Duvocelle et al. (2023) and Yan et al. (2023) are represented in a regret-type rate. Second, the rate in Theorem 6.2 is actually not last-iterate since the performance measure is $\\frac{1}{T} \\sum_t \\cdots$. I suggest that the authors could correct this issue in the next version.\n\n5. It seems that the rates authors mentioned in Duvocelle et al. (2023) and Yan et al. (2023) do not need smoothness? I did not check their results very carefully, and I suggest that the authors could conduct a careful double check about whether the above two works need smoothness. If not, these works seem to be not comparable since Duvocelle et al. (2023) and Yan et al. (2023) focused on strongly monotone games without smoothness while this paper considers monotone smooth games.\n\n6. It should also be made clear that Duvocelle et al. (2023) and Yan et al. (2023) used tracking error $\\\\|x_t - x^*\\\\|^2$ as the performance measure, which is different from the measure used in this paper. Thus again I think these works are actually not comparable. Besides, if I am wrong about some statements, it is welcome that the authors could correct me.\n\n6. In Line 287, the authors said that \"We show that Algorithm 1 converges to the Nash equilibrium in monotone, strongly monotone, and linear games.\" It should be made clear that the results for different kinds of games require different parameter configurations. \n\n7. Could the authors provide some explanations about the validness of the permanence measure $\\sum_i D_p(x_i^*, x_i^{T+1})$? Why this measure is a valid measure of the distance of the algorithm's output to the Nash equilibrium? The gap using Bregman divergence heavily relies on how convex the function $p(\\cdot)$ is. If $p(\\cdot)$ is nearly a linear function, this measure seems not to be a good enough performance measure, from my point of view, intuitively. I guess that the authors use this quantity as the performance measure mainly due to technical reasons, as I said in the 'Weaknesses' part?\n\n\n\nReferences:\n\n[1] Improved Regret Guarantees for Online Smooth Convex Optimization with Bandit Feedback, AISTATS 2011\n\n[2] Bandit Convex Optimization: Towards Tight Bounds, NIPS. 2014" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Extending the studies from strongly monotone games to generally monotone ones is meaningful. The proposed algorithm where a new regularizer is used is novel. The authors have also proved a high-probability result for the global convergence rate. And they have also applied their results to the social welfare to validate the importance of their results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of learning in a general monotone game with bandit feedback, with strongly uncoupled dynamics. For smooth cost functions, the authors achieved an $O(T^{-1/4})$ last-iterate convergence rate. For strongly monotone games, the obtained last-iterate convergence matches the state-of-the-art rate $O(T^{-1/2})$. The proposed algorithm uses the standard gradient estimator for BCO, but updates with a slightly different OMD-based algorithm with a different regularizer. When each player runs the same algorithm, the individual regret can also be guaranteed with $O(T^{3/4})$ regret for generally monotone games and $O(T^{1/2})$ for strongly monotone games. Finally, the authors further extended the results to time-varying games, considering two cases: (1) the time-varying game sequence gradually converges to a static game, and (2) the game sequence does not converge, and the performance measure relies on some non-stationarity measures depicting how the game sequence changes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The core of achieving the last-iterate convergence is inserting a regularizer $p(\\cdot)$ into the OMD update. Concretely, in the OMD update, the authors imported an additional additive $\\eta (t+1) D_p(x_i, x_i^t)$ into the OMD update. And the core analysis is shown in Lemma J.1 in the appendix. Specifically, by using $\\sum_i D_p(x_i^*, x_i^{T+1})$ as the performance measure, the $\\eta (t+1) D_p(x_i, x_i^t)$ in the OMD update can be extracted and moved to the left-hand side of the three-point equality of Bregman divergence, dividing both sides by $O(T)$ yields the performance measure on the left-hand side and a term like $(Regret)/T$ on the right-hand side. Consequently, it remains to analyze the regret guarantee of the proposed method and dividing it by $T$ yields the final last-iterate convergence rate. \n\nThis idea is novel. However, the key question is how we can access such function $p(\\cdot)$. As stated by the authors on Page 5, the function $p(\\cdot)$ has to satisfy the condition that $c_i(x_i, x_{-i}) - \\kappa p(x_i)$ is convex. To compute such a valid $p(\\cdot)$, the algorithm must have the complete information of the original cost function $c_i$, which is not permitted by the learning protocol since $c_i$ is unknown. Otherwise, if $p(\\cdot)$ is given by the learning problem, the condition of $c_i(x_i, x_{-i}) - \\kappa p(x_i)$ being convex should be an assumption that should be put in Section 3 along with Assumption 3.1. However, I am more inclined to believe that this should not be an assumption since $p(\\cdot)$ is something that appears in the algorithm, while assumptions are actually statements about the problem-dependent and algorithm-independent quantities. Can the authors provide some further explanations on this issue?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024uncoupled,\ntitle={Uncoupled and Convergent Learning in Monotone Games under Bandit Feedback},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zDJNUDprhW},\nnote={under review}\n}" }, "abstract": { "value": "We study the problem of no-regret learning algorithms for general monotone and smooth games and their last-iterate convergence properties. Specifically, we investigate the problem under bandit feedback and strongly uncoupled dynamics, which allows modular development of the multi-player system that applies to a wide range of real applications. We propose a mirror-descent-based algorithm, which converges in $O(T^{-1/4})$ and is also no-regret. The result is achieved by a dedicated use of two regularizations and the analysis of the fixed point thereof. The convergence rate is further improved to $O(T^{-1/2})$ in the case of strongly monotone games.\nMotivated by practical tasks where the game evolves over time, the algorithm is extended to time-varying monotone games. We provide the first non-asymptotic result in converging monotone games and give improved results for equilibrium tracking games." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "online learning", "game" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/857d1fb1acf132c5af4c0a30d12cca3aceb5e12d.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/79325c8cd9a5c1a25a462615f29cd1cf9b5fd903.pdf" }, "title": { "value": "Uncoupled and Convergent Learning in Monotone Games under Bandit Feedback" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Kindly refer to [Weaknesses]." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The idea of using depth-warped images as guidance for novel view synthesis is reasonable.\n* It is interesting to see that the temporal consistent video diffusion model can be effectively reformulated to achieve geometrical consistent NVS in a training-free manner.\n* Experiments on several challenging settings, including 360-degree NVS from a single view, verify the significance of the introduced method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces a diffusion model-based approach to achieve novel view synthesis. In particular, it leverages the depth-warped views as guidance to achieve adaptative modulation. Experiments on single-view images, multi-view images, and monocular video input-based novel view synthesis showcase the efficacy of the introduced methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Accessing the geometry accuracy. For the 360-degree case, e.g., the truck, it would be better to apply mesh reconstruction on the rendered views, similar to Fig. 5(b) in latentSplat [Wewer et al. ECCV 2024]. The reconstructed mesh will provide a clearer understanding of how well the rendered views maintain correct geometry.\n\n* Pixel-aligned metrics. For the NVS task, it would be better to report comparisons with state-of-the-part methods regarding pixel-aligned metrics, e.g., PSNR and SSIM. \n\n* Discussion with feed-forward 3DGS models. It might be interesting to see comparisons with detailed analysis between the introduced methods and those feed-forward 3DGS models, e.g., pixelSplat [Charatan et al., CVPR 2024], MVSplat [Chen et al., ECCV 2024]. And it would be better to consider adding these methods to the related work for better coverage of recent NVS works." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "For dynamic scene comparison, it is said \"For monocular video-based NVS, we downloaded nine videos from YouTube, each comprising \nframes and capturing complex scenes in both urban and natural settings.\" \n\nWhy not just following the dynamic nerf settings? They have well aligned ground-truth for measuring the reconstruction performance.\nGeneration metrics like FID are not that reliable.\n\nSoma example datasets are HyperNerf, DyCheck (https://hangg7.com/dycheck/)" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed approach is entirely training-free, meaning it directly leverages pre-trained large video diffusion models without requiring additional fine-tuning or retraining. This feature not only reduces computational demands but also makes it adaptable to a wide range of applications where time or resources for training may be limited. The flexibility of using pre-trained models enhances its practicality, allowing users to apply this method to various scenes and tasks with minimal setup.\n\n2. The generated videos maintain high visual fidelity, delivering smooth results. This quality stems from the adaptive modulation of the diffusion process, which effectively incorporates scene details and structures from the given views, ensuring that outputs are kind of realistic.\n\n3. The method is underpinned by theoretical modeling, which guides its adaptive modulation strategy. By iteratively adjusting the score function with scene priors and analyzing estimation error boundaries, the approach achieves both controlled and adaptive modulation of the diffusion process." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a NVS method that leverages large pre-trained video diffusion models without additional training. The approach adaptively modulates the diffusion sampling process using input views to produce high-quality results from single or multiple views of static scenes or dynamic videos. Theoretical modeling is used to iteratively adjust the score function based on scene priors, enhancing control over the diffusion process. The modulation adapts to view pose and diffusion steps." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The comparison between this method and NeRF-based methods is fundamentally imbalanced. NeRF techniques incorporate an underlying 3D structure, enabling them to render any view with predictable performance, as the 3D structure informs which views are feasible and which are not. In contrast, the proposed method lacks an explicit 3D representation, limiting its view synthesis capabilities to specific views with no guarantee of consistent performance. This distinction is significant, as NeRF's inherent 3D information allows interpretable, reliable results across views, whereas this method’s output reliability is less predictable and may vary based on input views.\n\n2. To achieve a fair comparison between the proposed method and NeRF-based techniques, the authors should first reconstruct a 3D model from the output video of this method and then re-render the scene from that reconstructed model. This process would allow for a direct assessment of both methods’ rendering consistency and quality, ensuring that comparisons consider the 3D structure NeRF inherently leverages. Also, reconstruction error like PSNR should be reported.\n\n3. The video results of the proposed method exhibit visible flickering artifacts, which could substantially affect reconstruction quality and consistency. A deeper analysis is needed to assess how these artifacts impact overall reconstruction accuracy and to identify potential mitigation strategies. This might include tuning reconstruction parameters to minimize flickering, which would help improve the method’s output stability and robustness, especially for applications sensitive to temporal consistency.\n\n4. A major contribution is the derivation of the parameter $\\lambda$ in Section 4.2, which aims to minimize the estimation error upper bound in Equation 15. However, a gap remains between this upper bound and the actual estimation error represented in the left side of Equation 15. To strengthen the theoretical foundation, the authors should provide a more comprehensive analysis of how reducing the upper bound affects the actual estimation error. This could be achieved through statistical analysis and empirical evidence showing how well the method reduces estimation error in practice, thereby validating the theoretical assumptions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Line 415 mentions that the proposed method can achieve 360-degree NVS. Would it be possible to include a comparison with ZeroNVS [7] to better demonstrate its effectiveness?\n\n\n[7] Sargent, Kyle, et al. \"ZeroNVS: Zero-Shot 360-Degree View Synthesis from a Single Image.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2024." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed adaptive modulation of the score function in the diffusion process is novel.\n2. The proposed method achieves better results in various scenarios compared to baselines.\n3. The authors provide the code with an anonymous link, ensuring the applicability of the results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. This paper proposes a novel view synthesis pipeline without any training. The pipeline can take single or multiple views of static scenes or monocular videos of dynamic scenes as input.\n2. This paper modulates the score function with the warped input views to control the video diffusion process and generate visually pleasing results. They achieve the modulation in an adaptive fashion based on the view pose and the number of diffusion steps.\n3. They conduct extensive results on both static and dynamic scenes and show promising results with both evaluation numbers and visualizations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My primary concerns are with the references and experimental details:\n\n1. Some key references on diffusion-based NVS are missing [1,2,3,4,5,6,7]. Among these, [3] specifically focuses on scenes and has released its code. Is there a particular reason it was not included in the comparison?\n2. How is the synthesized view pose calculated in this paper? In Line 364, it states that 'current depth estimation algorithms struggle to derive absolute depth from a single view or monocular video, resulting in a scale gap between the synthesized and ground truth images.' When calculating pose error, does the proposed method account for this scale gap?\n\n**Minor Points:**\n\n1. In Table 1, it is stated that MotionCtrl [Wang et al., 2023b] and 3D-aware [Xiang et al., 2023] do not require training. However, as I understand, they do require fine-tuning.\n\n[1] Wu, Rundi, et al. \"Reconfusion: 3d reconstruction with diffusion priors.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2024.\n\n[2] Watson, Daniel, et al. \"Novel View Synthesis with Diffusion Models.\" *The Eleventh International Conference on Learning Representations*.\n\n[3] Yu, Jason J., et al. \"Long-term photometric consistent novel view synthesis with diffusion models.\" *Proceedings of the IEEE/CVF International Conference on Computer Vision*. 2023.\n\n[4] Cai, Shengqu, et al. \"Diffdreamer: Towards consistent unsupervised single-view scene extrapolation with conditional diffusion models.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.\n\n[5] Tseng, Hung-Yu, et al. \"Consistent view synthesis with pose-guided diffusion models.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2023.\n\n[6] Chan, Eric R., et al. \"Generative novel view synthesis with 3d-aware diffusion models.\" *Proceedings of the IEEE/CVF International Conference on Computer Vision*. 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The work is complete and results seem to be good quantitively." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a training free novel view synthesis paradigm based on video diffusion model. Specificcally, warped depth maps are utilized and certain sampling methods are proposed to ensure high quality NVS." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The work is complete, but the novelty and the contribution are not strong enough to meet the quality of ICLR. A good training-free is not that appealing after previous works like MotionCtrl, and it highly relys on the capability of SVD. It is hard to know about it generalizability around various video generation methods. \n2. It is recommended to show \"Directly guided sampling\" and \"Posterior sampling\" clearly in the view of practical implementation. Some illustration would be appreciated.\n3. Some equations should be displayed more clearly. For instance in eq.8, I(P_0) should not be related to the pose index i, but in eq.9 it is related to i. Maybe in eq.9 the I(P_0) should be revised as I(P_i)?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The proposed method sets the number of inference step to 100, which is quite large. Do other diffusion-based baselines also use the same number of inference step?\n2. The paper mentions the camera trajectory is estimated (L.360). Could you provide more details on how the pose metrics (ATE, RPE) are computed?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper presents a training-free method that enables a pre-trained diffusion model to generate image sequences based on given camera trajectories.\n2. The qualitative videos demonstrate that the proposed method generates plausible outputs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a training-free sampling method that generates image sequences conditioned on input camera trajectories. The input view is warped onto the target views, serving as pseudo-GT variables. To condition the reverse process, the paper re-computes the predicted mean of a reverse step with an interpolation between the predicted mean of the current noisy latent and the warped samples. It further explores calculating the optimal interpolation weight and two guidance methods (replacement and gradient). Both quantitative result and qualitative results show the proposed method can generate smooth, high-fidelity image sequences." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper relies on an off-the-shelf depth estimation network to warp the input image(s), which may be prone to scale inconsistencies across frames. Since the reverse process is largely dependent on the warped images, depth estimation errors could propagate to the final outputs.\n2. The paper does not explain how occluded regions are handled in depth-warping. Handling the occluded regions can be particularly challenging when the camera pose variation is large, as $\\tilde{\\mathbf{\\mu}}_{t, \\mathbf{p}_i}$ in Eq.12 may produce degenerate outputs. \n3. The camera trajectories in most qualitative results are limited to relatively small variations or 360-degree circular poses, where scale ambiguity from the depth estimation network can largely be ignored. This raises robustness of the proposed method when the camera variations are large. The paper could benefit from including more examples with larger camera trajectories.\n4. The toy experiment in Fig.2 (a) shows that the $\\mathcal{E}_D$ decreases with the diffusion reverse process. This brings another question on how the predicted means converge to the GT image (loss=0) where it is very unlikely to have sampled $\\mathbf{X}_t$ that would have led to the desired GT. I believe more details of the experiment can help understanding the paper better.\n5. The paper makes some assumptions to compute the optimal interpolation weight $\\lambda(t, \\mathbf{p}_i)$ in Sec.4.2. However, the paper does not present ablation study on the choice of the weight (except when $\\lambda(t, \\mathbf{p}_i) = \\infty$, which also shows comparable results). To validate the choice of the weight schedule, the paper could present comparisons to other naive techniques (e.g., linear, constant, exponential). Additionally, the interpolation weight depends on a set of hyperparameters $\\{ v_1, v_2, v_3 \\}$ which may require engineering effort to tune on new scenes. \n6. While the paper shows promising quantitative and qualitative results, the number of scenes used for evaluation is insufficient to validate the method's effectiveness and differs from previous state-of-the-art methods." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Our method adaptively modulates score functions of a pre-trained video diffusion to synthesize impressive novel views." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024zeroshot,\ntitle={Zero-shot Novel View Synthesis via Adaptive Modulating Video Diffusion Process},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zDJf7fvdid},\nnote={under review}\n}" }, "abstract": { "value": "By harnessing the potent generative capabilities of pre-trained large video diffusion models, we propose a new novel view synthesis paradigm that operates \\textit{without} the need for training. The proposed method adaptively modulates the diffusion sampling process with the given views to enable the creation of visually pleasing results from single or multiple views of static scenes or monocular videos of dynamic scenes. Specifically, built upon our theoretical modeling, we iteratively modulate the score function with the given scene priors represented with warped input views to control the video diffusion process. Moreover, by theoretically exploring the boundary of the estimation error, we achieve the modulation in an adaptive fashion according to the view pose and the number of diffusion steps. Extensive evaluations on both static and dynamic scenes substantiate the significant superiority of our method over state-of-the-art methods both quantitatively and qualitatively. The source code can be found on the anonymous webpage: https://github.com/PAPERID5494/VD_NVS. We also refer reviewers to the Supplementary Material for the video demo." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Diffusion sampling", "Novel view synthesis" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/7961b157c2c6f279fff19dfd8a51b528b7355e0f.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/2a56548e78db69ff5908b5167a1328dabf0e99a7.zip" }, "title": { "value": "Zero-shot Novel View Synthesis via Adaptive Modulating Video Diffusion Process" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Question: Please clarify the generalizability of your approach. Do you expect the rule learning method to be sensitive to data errors that arise naturally in real-world collection, such as uncalibrated sensors or aspect/focus variation in images, or frame rate in video? Consider the following comment, if that helps clarify the objective of the request.\n\nI am not familiar with the UCR datasets, and I don't think you provided a link, nor did the paper you cite have a currently valid link; I found data and descriptions here: https://www.cs.ucr.edu/~eamonn/time_series_data/. Some datasets seem to have a small degree of possible coordinate sensitivity, such as the spectrogram data, while other would vary significantly with scale, placement of the signal within a collection window, and image orientation. (This is also true of MNIST). Please review the table below, and help me understand the sensitivity of your method to data collection errors. This gets at the issue of generalizability / fragility of the method. In the table, the first column has the dataset name, the second column records whether or not your method was the top score in your reported results, the third records a (possibly incorrect) sense of how sensitive the data collection is to getting all the coordinates right, and the fourth is an explanation of the assessment. \n\n| Dataset | NeurRL | Coordinate Sensitivity | Rationale for sensitivity assessment |\n|---------|-------------|----------------------|-------------|\n| Coffee | Top | Low | Based on spectrographic measurements which are inherently coordinate-independent. The frequency patterns would be consistent regardless of measurement setup, as long as calibration is maintained. |\n| ECG | -- | Medium | While heart rhythms have characteristic shapes, the absolute voltage measurements depend on electrode placement. However, the relative patterns are fairly robust across different setups. |\n| Gun Point | -- | High | Heavily dependent on camera angle, distance calibration, and the defined origin point for measuring hand position. The distinction between gun/no-gun relies on precise spatial coordinates. |\n| ItalyPow.Dem | Top | Low | Power demand measurements are absolute values independent of coordinate systems. The readings represent actual power consumption regardless of measurement setup. |\n| Lightning2 | -- | Low (?) | While electromagnetic measurements depend on sensor placement, the characteristic patterns of lightning types are relatively robust across different setups. |\n| CBF | -- | Low | As a synthetic dataset designed to test pattern recognition, the shapes (Cylinder-Bell-Funnel) are meaningful in their relative values rather than absolute coordinates. |\n| Face Four | Top | High | Face profile traces are highly dependent on camera angle, distance, and orientation. Changes in perspective would significantly alter the time series. |\n| Lightning7 | Top | Low (?) | Similar to Lightning2, with moderate dependence on sensor placement but relatively robust patterns. |\n| OSU Leaf | Top | High | Leaf contour measurements depend heavily on orientation, scale, and starting point of the trace. Different coordinate systems would produce very different time series. |\n| Trace | -- | Low | As synthetic nuclear instrument data, the patterns represent relative changes that are meaningful independent of absolute coordinate systems. |\n| WordsSyn | Top | High | Pen trajectories are highly dependent on writing orientation, scale, and starting position. The coordinate system directly affects the recorded patterns. |\n| OliveOil | Top | Low | Spectroscopic measurements are independent of coordinate systems. The chemical signatures would be consistent across different spectrometers (after calibration). |\n| StarLightCurves | Top | Low | Brightness measurements are relative values that remain meaningful regardless of the telescope's exact positioning (assuming proper astronomical calibration). |\n\nQuestion 2: Can you explain \"leakage\" more clearly? The coordinate dependence an example above creates an implicit bias in the data, for instance limited field-of-view and imaging system orientation in MNIST and handwriting digitization; is this an example of \"label leakage\"?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality: This provides a new and instructive implementation to address the raw-to-symbolic data problem, while avoiding the data leakage issue.\n\nQuality and Clarity: The paper explains and illustrates the method in a clear and compelling way.\n\nSignificance: This appears to be an important contribution to the problem of addressing the problem of learning symbolic rules about raw sequential data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces Neural Rule Learner (NeurRL), a system designed to learn interpretable rules from sequential data, such as from time-series sensor data (like spectrograms) and serialized higher-dimensional data (such as images).\n\nThe system uses symbolization functions to map raw data into a predetermined framework of pattern and region predicates, allowing discrete raw data sequences to be represented as logical atoms that can be used in rule learning. In the supplied reference implementation, patterns are learned automatically through a differentiable k-means clustering process, while regions are predefined subdivisions of the sequence. The combination of learned patterns and predefined regions allows the system to express discovered features as logical rules using this symbolic vocabulary.\n\nThe framework uses two types of symbolization functions: body symbolization functions that map discrete sequences to symbolic concepts, and head symbolization functions that map inputs to a set of atoms. The system avoids the \"label leakage\" problem by not requiring pre-trained neural networks to map raw inputs to symbolic labels. The entire pipeline is differentiable and can be trained end-to-end, learning directly from raw data without needing intermediate symbolic representations.\n\nThe implementation uses a deep rule-learning module with multiple dense layers, and rules are evaluated using precision and recall metrics. The reference implementation includes an autoencoder component for learning concentrated representations of subsequences.\nExperimental validation was conducted on both time series data (UCR datasets) and image data (MNIST). Learning on UCR data-subsets highlights low-data or data-efficient learning methods. The reference implementation achieves comparable or better classification accuracy compared to baseline methods, and provides interpretable rules. The generated rules are human-readable and capture patterns in the data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The approach may require significant data invariance (orientation of images, sampling regularity). Many important datasets have these properties, although image data often does not. See questions for further clarification.\n\nAs a demonstration of the advantage of this approach as an interpretable and explainable method, it would be helpful to have the interpretation of rules on each dataset from UCR discussed in more detail. How interpretable are the rules in terms of the data patterns? Are there particular datasets or data types for which rules emerge that provide insight or match intuition?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Differentiable ILP frameworks ($\\partial$ ILP and $\\alpha$ ILP) typically suffer from their intense memory consumption due to the tensor encoding. Does NeurRL share the same problem? \n\nFor other questions, please refer to the weakness section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper tackles an important topic in the field of differentiable ILP, specifically the need for pre-trained models to ground raw input into symbols for effective rule learning. The approach of combining differentiable k-means and VAE with gradient-based rule learners is noteworthy, even though the concept is relatively straightforward.\n\nThe manuscript is well-written, with a high-quality presentation overall. It clearly outlines the research question and effectively conveys its core idea.\n\nThis work is a valuable contribution to the neuro-symbolic research community, paving the way for developing more robust systems capable of learning with fewer priors on raw input." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new differentiable Inductive Logic Programming framework where symbolic rules are learned via gradients from sequential data. To overcome the bottleneck of previous models that require pre-trained neural networks to perceive raw input, the authors propose Neural Rule Learner (NeurRL), which combines VAE and differentiable k-means to the differentiable rule learning approach. The resulting framework can learn classification rules that specify regions and corresponding patterns to be classified as positive. It is demonstrated in the experiments using simple datasets, including synthetic, UCR, and MNIST datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My primary concern is that it remains unclear whether the main claim is adequately supported by the methods and experiments sections. The paper asserts:\n\n> (Line 42-45) Learning logic programs from raw data is hindered by the label leakage problem common in neuro-symbolic research (Topan et al., 2021): This leakage occurs when labels of ground objects are introduced for inducing rules (Evans & Grefenstette, 2018; Shindo et al., 2023).\n\nand\n\n> (Line 82-85) In our work, we do not require a pre-trained large-scale neural network to map raw input data to symbolic labels or atoms. Instead, we design an end-to-end learning framework to derive rules directly from raw numeric data, such as sequence and image data.\n\nIf this claim holds, the proposed system should be trainable end-to-end, grounding symbols in meaningful ways such that rule learners can generate rules based on them. However, the paper does not make it clear how this grounding is accomplished.\n\nDoes NeurRL perform grounding on the obtained clusters? Can the rule learners manage rules involving multiple clusters by understanding the meaning of each cluster? If not, I am skeptical that the proposed framework fully addresses the bottleneck identified in the introduction, namely the reliance on perception models in prior rule learning frameworks like $\\partial$ ILP and $\\alpha$ ILP.\n\nWithout grounding, the method would be confined to the rule format (Eq. 5), as the system would lack the means to learn from non-grounded symbols, limiting its applicability to other domains.\n\nWhile I appreciate the method's ability to learn from small datasets, it is unclear how it would scale to larger, more complex datasets. Given that neural models typically require large datasets, future development would benefit from the capacity to handle such data at scale to better integrate neural networks. It would be beneficial to provide runtime comparisons on different dataset sizes, or to discuss potential challenges and solutions for scaling the approach to larger datasets. How does the model's performance change with increasing dataset complexity and size? If scalability to large-scale data is an issue, this limitation should be discussed somewhere in the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "In addition to discuss the above mentioned weaknesses, please find below some additional questions: \n\n1. The proposed solution introduces several hyperparameters, including parameters for data pre-processing (such as subsequence length and the number of temporal/spatial slots), fixed biases for the rule learner’s neurons, temperature parameters for the rule learner’s weights, and weights for the loss functions. How are these hyperparameters selected, and how sensitive is the model’s performance to their values? A sensitivity analysis could clarify this, and it should be feasible given the small size of the datasets. \n2. The post-hoc strategy achieves lower predictive accuracy compared to the full neural solution (see Table 1), with drops of up to 15 percentage points in some cases. This underscores the heuristic nature of the proposed strategy. Could you provide an explanation for this phenomenon? Additionally, what can be said about the faithfulness of the extracted rules in relation to the learned model’s weights? \n3. Why is neuro-symbolic learning necessary for the chosen MNIST task, and what is the rationale for this choice? The task could be solved by relying on superficial image cues (such as identifying regions that are either black or white), as illustrated by the qualitative demonstration in Figure 4, without needing to understand the underlying semantics of the digits. Therefore, a purely neural approach might suffice for this task. It would be interesting to assess the model’s capability to recognize digit classes within arithmetic tasks or, more broadly, in tasks defined by programs over integer sequences (e.g., sequences of even or odd numbers, counting). \n4. Are the codes learnable in the differentiable clustering layer and how do you ensure that meaningful clusters can be learnt, without experiencing any form of collapse? Please refer for instance to the work in [1].\n\n**References**\n\n[1] Sansone. The Triad Of Failure Modes and a Possible Way Out. In NeurIPS SSL Workshop 2023" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is overall clear **Clarity**. However the text should be polished to rewrite/rearrange words in some sentences and correct the several typos present throughout. Please refer the MINOR paragraph below for a non-exhaustive list.\n2. The considered problem is relevant and timely **Relevance**.\n3. Code is available, but no further check has been performed **Code availability**.\n\nMINOR \\\nL.109 -> $\\neg\\alpha_2$ \\\nL.129 -> a substitution or an interpretation \\\nL.151 -> is ordered real-valued observations with the dimension one \\\nL.159 -> discrete -> discretize \\\nL.161 -> closest \\\nL.195-197 -> could you rephrase the sentence? \\\nL.273 -> the its \\\nL.325-327 -> could you rephrase the sentence? \\\nL.381 -> synthetic" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a supervised neurosymbolic model for time series data, structured as a cascade of three main building blocks. First, an autoencoder extracts an embedding representation from raw data. Next, a differentiable clustering module maps the embedding to a symbolic representation, followed by a fuzzy neural rule learner that maps the symbolic representation to the target. The overall training objective includes three loss functions, one for each building block: a reconstruction error for the autoencoder, a k-means objective for the differentiable clustering, and a supervised cross-entropy loss for the rule learner.\nTraining occurs in two stages, beginning with the pre-training of the autoencoder, followed by the fine-tuning of the entire model. Additionally, a post-hoc heuristic is introduced to extract rules in propositional logic form from the weights learned by the rule learner. Experiments are conducted on a series of small datasets from UCR, comparing the model to existing neural and kernel-based classification baselines for time series data, with results demonstrating superior predictive accuracy. Lastly, a qualitative experiment on MNIST data is provided." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper lacks adequate context within the existing neuro-symbolic learning literature, making its contribution appear incremental in terms of novelty **Quality**. For example, neuro-symbolic (NeSy) autoencoders have been previously explored (see [1,2]), and differentiable clustering has also been investigated (refer to [3,4]). Additionally, the rule learner shows a strong resemblance to $\\delta$ILP. At a minimum, a discussion of related work should be included to clarify the paper's novelty and contribution. An experimental comparison with existing methods could also help demonstrate the advantages of the proposed approach. \n2. The paper lacks soundness due to several incorrect and overly ambitious claims that are not adequately supported **Soundness**. First, the paper states that this work is the first to enable joint training of the neural network and rule learner without pre-training. However, this claim is inaccurate, as the proposed training still requires a two-stage process. Additionally, contrary to what is mentioned, the work in [5] already achieves joint learning of a neural network and a rule learner. An experimental comparison on the datasets introduced in [5] would be valuable. Furthermore, the paper suggests that the proposed solution can uncover underlying rules without identifying conditions for failure. However, depending on the level of supervision and the complexity of the symbolic task, the neural learner may learn incorrect mappings from raw data to symbols. The work in [3,4] presents a basic sequential setting to analyze this issue, demonstrating that while representation learning (through differentiable clustering) is necessary to address the problem, it is not sufficient. Similar observations are subsequently reported in [6,7], where the authors label this issue as \"reasoning shortcuts.\" \n3. There is very limited discussion about the details of the experimental methodology, raising serious concerns regarding reproducibility **Reproducibility**\n4. Experiments are conducted on small datasets with relatively simple symbolic tasks **Quality**. I encourage the authors to consider more traditional and more complex experimental settings for neuro-symbolic learning, including the arithmetic tasks on handwritten digits or the datasets in [5].\n\n**References**\n\n[1] Misino, Marra, Sansone. VAEL: Bridging Variational Autoencoders and Probabilistic Logic Programming. In NeurIPS 2022 \\\n[2] Zhan, Sun, Kennedy, Yue, Chauduri. Unsupervised Learning of Neurosymbolic Encoders. In TMLR 2022 \\\n[3] Sansone, Manhaeve. Learning Symbolic Representations Through Joint GEnerative and DIscriminative Training. In ICLR NeSy-GeMs Workshop 2023 \\\n[4] Sansone, Manhaeve. Learning Symbolic Representations Through Joint GEnerative and DIscriminative Training. IJCAI KBCG Workshop 2023 \\\n[5] Evans et al. Making Sense of Sensory Input. In Artificial Intelligence 2021 \\\n[6] Marconato et al. Neuro-Symbolic Continual Learning: Knowledge, Reasoning Shortcuts and Concept Rehearsal. In ICML 2023 \\\n[7] Marconato, Teso, Vergari, Passerini. Not All Neuro-Symbolic Concepts Are Created Equal: Analysis and Mitigation of Reasoning Shortcuts. In NeurIPS 2023" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Is there a way to reduce the number of hyperparameters, such as determining the number of categories K adaptively during training? If this is not feasible, please at least provide guidance on how to select these hyperparameters and show how different choices impact the experimental results.\n2. Please compare your work with more related studies (see Weakness 4)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper formally establishes the ILP learning task from raw data, utilizing the interpretation transition framework of ILP to guide the learning process.\n2. Experimental results demonstrate that NeurRL shows a notable improvement over previous methods in terms of effectiveness.\n3. The paper's overall presentation is clear, making the concepts and methodologies easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel differentiable Inductive Logic Programming model called NeurRL, aimed at solving the problem of learning symbolic rules from raw data. Unlike previous methods that rely on pre-trained neural networks to extract labels for symbolic learning, NeurRL combines a differentiable clustering module and a deep neural network learning module, enabling end-to-end rule learning from raw data such as time series and images without leaking label information." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The model has a large number of hyperparameters (especially the number of categories K, subsequence length l, and the number of regions P), which may negatively affect reproducibility and generalizability.\n2. The experimental details are insufficient, lacking information on how hyperparameters were chosen and whether tuning was done separately for different experiments. Moreover, it is unclear if the results reported are the best performance or averaged.\n3. The learned predicates lack high-level abstraction, instead offering post-hoc explanations for neighboring pixels. This limits the generalizability and interpretability of the induced rules.\n4. Although previous works did not formally define the task of extracting rules from raw data, some have achieved similar goals with comparable architectures (e.g., autoencoder + discriminator)[1, 2], or going further in technical sophistication[3] . This paper lacks a discussion of and comparison with such works.\n5. The model does not consider translation invariance in the data, meaning some patterns should only need to be present rather than bound to a specific region. This limitation may hinder the model's performance in tasks where the position of the target is irrelevant.\n\n[1] Azzolin, S., Longa, A., Barbiero, P., Liò, P., & Passerini, A. (2022). Global explainability of gnns via logic combination of learned concepts. arXiv preprint arXiv:2210.07147.\n\n[2] Walter, N. P., Fischer, J., & Vreeken, J. (2024, March). Finding interpretable class-specific patterns through efficient neural search. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 8, pp. 9062-9070).\n\n[3] Wang, B., Li, L., Nakashima, Y., & Nagahara, H. (2023). Learning bottleneck concepts in image classification. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition (pp. 10962-10971)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024endtoend,\ntitle={End-to-End Rule Induction from Raw Sequence Inputs},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zDjHOsSQxd},\nnote={under review}\n}" }, "abstract": { "value": "Rule learning-based models are widely used in highly interpretable scenarios for their transparent structures. Inductive logic programming (ILP) is a form of machine learning that induces rules from facts and keeps the interpretability. Differentiable ILP models enhance their learning ability in a robust and scalable manner with the advantages of neural networks. However, most differentiable ILP methods learn from symbolic datasets. Learning from raw data needs an ILP model to tackle the symbol grounding problem: The inability to map continuous inputs to symbolic variables without explicit supervision. In this work, we incorporate a self-supervised differentiable clustering model and a novel differentiable ILP model to learn from raw data in an end-to-end way without leaking the labels. The learned rules describe the raw data with its features. We demonstrate that our method learns generalized rules from time series and images intuitively and precisely." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Neuro-Symbolic Methods", "Interpretability", "Inductive Logic Programming" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/b84668af5d4361522cf2d3378c0df9454a23f197.pdf" }, "presentation": null, "primary_area": { "value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/3cda9565406c36f0e6adc83ceecdac75e2986f5e.pdf" }, "title": { "value": "End-to-End Rule Induction from Raw Sequence Inputs" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. Are there specific conditions or types of tasks where width expansion or chain-of-thought reasoning may still be advantageous compared to log-depth scaling?\n\n2. Do you have plans to empirically validate these theoretical results on real-world datasets? If so, what metrics or benchmarks would you prioritize to assess the practical benefits of log-depth scaling?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. **Innovative Depth Scaling Insight**: This paper shifts the focus from fixed-depth transformers to a dynamic, log-depth scaling, addressing some limitations of traditional models in handling extended contexts. This perspective broadens our understanding of transformers’ potential expressivity.\n\n2. **Rigorous Theoretical Foundation**: The paper provides clear mathematical proofs and complexity analyses that validate the advantages of log-depth scaling for specific tasks, particularly regular language recognition and graph connectivity. This rigor strengthens the work’s contributions to understanding transformers’ computational capacity.\n\n3. **Comprehensive Comparison with Other Scaling Approaches**: The authors examine depth scaling in comparison to width scaling and chain-of-thought methods, demonstrating that logarithmic depth growth is more computationally efficient and effective for reasoning tasks, especially for state tracking and pathfinding in bounded contexts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores the computational expressivity of transformers when their depth scales logarithmically with input context length. Prior analyses typically assume a fixed transformer depth, which limits the model’s ability to solve certain tasks as context length increases. This work, however, argues that transformers can still solve certain problems up to a bounded input length, even if they cannot handle them for arbitrarily large inputs. By scaling depth logarithmically with the input length, transformers can effectively simulate finite automata and solve graph connectivity problems, which are critical for tasks involving multi-step reasoning and state tracking. These findings suggest that only a modest, logarithmic increase in depth is required to address such tasks, offering a path for efficient model scaling and highlighting the benefits of dynamic depth adjustments over simply expanding model width or employing chain-of-thought reasoning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited Consideration of Complementary Methods**: The paper promotes depth scaling over width or chain-of-thought expansion but could benefit from a more nuanced discussion of scenarios where those methods may still offer advantages or may complement log-depth scaling.\n\n2. **Lack of Experimental Validation**: Although the theoretical findings are compelling, the paper would be stronger with empirical experiments demonstrating the practical performance of log-depth transformers on real-world tasks and quantitatively comparing their efficiency and effectiveness with other models." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Q1. Given the two differences from the standard transformer architecture and the fact that Lemma 1 holds because of the masked pre-norm, do you think that the results hold for standard transformer architecture? \n\nQ2. Would it be possible to add a more informal introduction to let other people read and fully understand the work without looking at prior work? \n\nQ3. Did you also see a similar law for other architectures (like https://arxiv.org/abs/2405.06394 or other constant transformer-like model)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "S1. Authors showed that looped transformer (non-constant depth architecture) can solve problems more efficiently than CoT or adding width. This is quite interesting for theoretical transformer foundation, although the impact is pretty limited for current constant-depth architecture, given it's not clear and easy to scale large training for looped transformer. \n\nS2. The findings could inform more efficient model scaling strategies, maybe combining the effect of CoT with log n depth.\n\nS3. The paper is well written and all relevant lemmas are well explained." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work shows that two reasoning problems where standard transformer struggled to solve can be solved by a looped transformer. The work also shows that modular arithmetic with small modulus can be solved with the same architecture. Authors finally showed that scaling depth is more efficient than width or adding CoT intermediate steps." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. Major weakness is that the theoretical results lack experimental validation on the tasks author suggested, limiting real world applicability. Also, dismissing CoT with a couple of lines seems unfair, given the potential showed by CoT on arithmetic and graph related problems, above all.\n\nW2. Given the memory management issue, implementing dynamical depth is not clear and also it's not clear if a looped transformer can be indeed __trained__ to solve the two mentioned problems. \n\nW3. The paper is very intricate and a bit more explanation may be needed to understand this work. Also the work is not really self contained. \n\nW4. Although theorem 3 and 4 are the core part of the authors work, they are not well justified." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I would appreciate it if you could address the weaknesses I highlighted earlier. Specifically, I am interested in the following:\n\n- Could you provide additional clarification on your contributions regarding the first question, particularly on how you contribute to understanding the limits of the expressiveness of fixed-depth transformers with bounded inputs?\n- Could you offer a more detailed and rigorous explanation of the concepts discussed in Section 2.2?\n- Am I correct in identifying a potential confusion between decoder-only and encoder-only models, or is there a detail I might have overlooked?\n- What is the precise method of masking you employ within your model?\n\nI'm keen to resolve these issues as I believe your work holds significant value. However, the points mentioned above currently obscure its impact." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The three areas that the paper aims to contribute to are both relevant and intriguing. Particularly, \nthe assumption of bounded input lengths, which is closer to practical settings, deserves thorough investigation.\n- Setting aside the notational and definitional ambiguities discussed below, the paper supports all \ntheoretical claims with proofs, which appear to be sufficient.\n- The related work, specifically regarding the expressive capabilities of different transformer models, \nis well chosen. I concur that this paper contributes significantly to these domains." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the expressive capacities of transformers where \nthe depth grows logarithmically with the length of the inputs. Specifically, the paper seeks to \ncontribute to three questions:\n- Can fixed-depth transformers handle \"hard\" problems when restricted to bounded input lengths?\n- Does logarithmic depth scaling offer greater expressive abilities compared to a fixed depth?\n- What practical advantages do log-depth transformers or input-dependent depth offer?\n\nInformally, the paper's primary theoretical contributions regarding the \"universal transformer\" model, \nwhich is characterised by a basic encoder architecture with certain layers repeated a number of times \ndependent on the input length, are as follows:\n- (Theorem 1) For each regular language $L$, there exists a universal transformer, with logarithmic repetition, \nthat recognises $L$.\n- (Theorem 2) There exists a universal transformer, with logarithmic repetition, that decides whether, \ngiven a graph $G$ and vertices $s$ and $t$, there is a path from $s$ to $t$.\n- (Theorems 3 & 4) Placing different classes of transformers (fixed depth or CoT) in $\\text{TC}^0$." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "To begin, I will summarize the main issue I encounter with this paper. \nAs mentioned earlier, the contributions the paper aims to provide are intriguing, and I also think they are partially achieved. \nHowever, the current form of the paper makes it challenging to grasp the significance of these contributions and how they compare to existing results. This difficulty arises primarily from a lack of clear definitions, confusing informal statements, and a disconnect between the theoretical contributions and their proposed implications. \n\nI will go into detail below, loosely ordering the weaknesses from most to least significant:\n- I am not convinced that the paper sufficiently addresses the first and third contributions it aims to make. The main results focus on transformers with dynamic depth, leaving me unclear about the implications for fixed-depth models. Lines 76 to 83 seem to mention this in the context of an experiment, but it is only briefly referenced and lacks further development throughout the paper. Additionally, the explanation of the third contribution in Lines 86 to 9, which states, \"scaling depth logarithmically as a computational resource more efficiently expands the expressive power of transformers compared to scaling width...\", appears to be an overinterpretation based solely on the results of Theorems 1 & 2 versus Theorem 3. If this claim is to be maintained, the paper needs to provide more detailed justification for such a bold statement, which is currently absent.\n- The definition of the (s, r, t)-universal transformer model lacks clarity and detail. While I understand the authors' intent to be succinct, even those familiar with related research might struggle to comprehend the model's specifics. A thorough understanding is crucial for comparing the expressiveness of different transformer models. Here are some points related to the definition in Section 2:\n - There seems to be a confusion in terminology regarding the \"semantics\" of universal transformers, which are described as sequence-to-value functions. However, on line 103, they are referred to as a decoder, which is typically associated with sequence-to-sequence transformations.\n - The definition of masking within this model is unclear. The phrase \"... add a learned mask vector that can select specific dimensions of the residual stream ...\" is ambiguous and needs clarification. This issue is particularly important because, for example, in Lemma 1, the term \"...causally masked attention...\" is mentioned, but its meaning is not well defined.\n - The notation $L^l$ is ambiguous. Does it signify layer $l$ or the function it performs? Additionally, the notation $r$-layer $(l-s) \\mod r$ is not defined\n - The definition of averaging-hard attention using a limits construction and an exponential function diverges from the common use of hardmax in related works cited within the paper. Are there differences that justify this choice?\n - Section 2.2 appears disjointed and complicates the definition. The first sentence raises multiple questions, particularly on the necessity of \"memory management.\"\n- The presentation of Theorem 2 is somewhat misleading. The problem addressed is not the \"connectivity problem\" in the traditional sense, which involves deciding whether all nodes are interconnected by some path. Instead, it is the \"reachability problem,\" which focuses on deciding if there is a path connecting specific nodes.\n- The role of Theorem 4 is not clearly articulated. The theorem is referenced with \n(Anonymous, p.c.), which is somewhat unconventional. Is this result established within this paper? \nIt seems there is a proof provided in the appendix, so the rationale for this citation is unclear. \nIf the theorem is not original to this work, clarification is necessary regarding its inclusion.\n- The abstract begins with the statement: \"Most analysis of transformer expressivity treats the depth (number of layers) of a model \nas a fixed constant, ...\". I find this statement problematic. It seems that the authors are referring to research focusing on formal language theory, which aims to understand which classes of languages specific classes of transformers can recognize. \nWhile individual transformer models have a fixed depth, this is right, the class of models does not need to be constrained in this way. This is an example of an informal assertion that leads to confusion. \n- Lines 44 to 46 include the statement, \"This is analogous to how regular expressions cannot express all context-free languages, but one can write regular expressions that capture fragments of a context-free language.\" The paper uses this analogy to imply that transformers generally perform effectively on shorter input lengths. This statement is somewhat perplexing because the point about regular expressions is quite basic; if there is additional nuance or significance, such statements require further clarification." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What are the other configurations (e.g., model width, maximum number of attention heads, etc.) for the Transformers in the main theorems, specifically Theorems 1 and 2? \n\n2. Can you provide more clarification of your results in relation to prior work, such as [1]? (Refer to weakness 3)\n\n3. Can you conduct experiments on the two tasks considered in the paper, varying the model depth, model width, and the number of attention heads to empirically evaluate how these factors affect performance?\n\n[1] William Merrill and Ashish Sabharwal. The parallelism tradeoff: Limitations of log-precision transformers. https://arxiv.org/abs/2207.00729." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors utilize saturated attention rather than hard attention, aligning more closely with practical applications. Additionally, they consider the effects of layer normalization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates log-depth Transformers, demonstrating that their capabilities exceed those of constant-depth Transformers ($\\mathsf{TC}^0$) and constant-depth Transformers with $O(\\log n)$ chain-of-thought steps. Through theoretical constructions, the authors show that a log-depth Transformer can effectively solve the regular language recognition problem and the graph connectivity problem, both of which fall outside the expressive power of constant-depth Transformers with polynomial size." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation for log-depth Transformers lacks sufficient conviction. Given the bounded context assumption in line 50, it would be more appropriate to treat all elements as constant. If bounded context and model sizes are to be discussed carefully, as indicated in lines 76-78 regarding the maximum problem size of graph connectivity with fixed depth $d$, the coefficient is crucial, while the result of $2^{O(d)}$ remains unclear.\n\n2. The results presented lack clarity. In the main theorems (Theorems 1 and 2), only the model depth of $O(\\log n)$ is specified, omitting other critical configurations such as the required model width and the number of attention heads.\n\n3. Additional clarification is needed regarding the results in relation to prior work. Specifically, Lemma 3 of [1] demonstrates that a Transformer with depth $2d$ maps $\\langle C,x\\rangle$ to the circuit value $C(x)$, where $C$ is a threshold circuit with depth $d$. Since regular language recognition and the graph connectivity problem both belong to $\\mathsf{TC}^1$, this lemma could already suggest comparable results in this paper.\n\n4. The absence of experimental results to validate the theoretical findings is a significant gap. Empirical verification of the solid lines in Figure 1 is also essential.\n\n[1] William Merrill and Ashish Sabharwal. The parallelism tradeoff: Limitations of log-precision transformers. https://arxiv.org/abs/2207.00729." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We provide the first expressivity analysis of transformers that accounts for model depth and explains how transformers might use depth to successfully solve problems on bounded context lengths that they otherwise cannot solve." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Little Depth Goes a Long Way: the Expressive Power of Log-Depth Transformers},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zDze7VtB5C},\nnote={under review}\n}" }, "abstract": { "value": "Most analysis of transformer expressivity treats the depth (number of layers) of a model as a fixed constant, and analyzes the kinds of problems such models can solve across inputs of unbounded length. In practice, however, the context length of a trained transformer model is bounded. Thus, a more pragmatic question is: *What kinds of computation can a transformer perform on inputs of bounded length?* We formalize this by studying highly uniform transformers where the depth can grow minimally with context length. In this regime, we show that transformers with depth $O(\\log C)$ can, in fact, compute solutions to two important problems for inputs bounded by some max context length $C$, namely *simulating finite automata*, which relates to the ability to track state, and *graph connectivity*, which underlies multi-step reasoning. Notably, both of these problems have previously been proven to be asymptotically beyond the reach of fixed depth transformers under standard complexity conjectures, yet empirically transformer models can successfully track state and perform multi-hop reasoning on short contexts. Our novel analysis thus explains how transformer models may rely on depth to feasibly solve problems up to bounded context that they cannot solve over long contexts. It makes actionable suggestions for practitioners as to how to minimally scale the depth of a transformer to support reasoning over long contexts, and also argues for dynamically unrolling depth as a more effective way of adding compute compared to increasing model dimension or adding a short chain of thought." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "transformer", "expressivity", "limits", "bounded context", "circuits" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/196a1bf71ae37a3b4b5b6ef0331c4110dea635a3.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "A Little Depth Goes a Long Way: the Expressive Power of Log-Depth Transformers" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The writing quality of the article is subpar, and the paper appears to lack certain crucial elements: there are inaccuracies in the referencing of figures, particularly early in the overview section; there is a notable absence of analysis pertaining to experimental and visualization outcomes; the paper does not include any ablation studies; and there is a scarcity of citations and comparative work. Does this suggest that the work is yet to be considered complete?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ The method effectively leverages supervised contrastive learning to enhance feature separability across multiple classes.\n+ The use of a dimensionality reduction adaptor helps reduce feature complexity while preserving relevant information for anomaly detection, making the model more efficient and focused on essential patterns." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a Contrastive Learning-based multi-class Anomaly Detection (CLAD) method designed to enhance multi-class anomaly detection in industrial contexts. The proposed method combines dimensionality reduction and supervised contrastive learning to generate and differentiate between normal and anomalous samples across various classes. It uses a two-stage training process involving initial discriminative feature learning and subsequent fine-tuning, significantly improving detection accuracy on benchmark datasets like MVTec and VisA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ The author has made an error in the plotting of Figure 2, as the elements within the figure are not aligned properly.\n+ The paper lacks ablation studies and could be expanded to include a deeper analysis of how each training loss term contributes to the final model performance.\n+ The layout of tables and text in the experimental section of the article is not rigorous, with excessive spacing.\n+ The model structure in the article is simple and lacks innovation.\n+ The paper has a limited number of citations and lacks a thorough comparison with relevant existing work." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1.What does $p_m$ U (0, 1) mean in Line 209? What is $x_i′$ in Line 212?\n\n2.What is $\\mu$ in Eq.8? Where is the equation for calculating anomaly score $s_i$ in Line 323? You did not mention it in the methodology.\n\n3.Is the batch size B in contrastive learning equal to the number of multi-class categories? How do you balance the computational cost and training efficiency when dealing with a large number of categories?\n\n4.See weakness above." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "It proposed a framework for robust multi-class anomaly detection." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a framework for robust multi-class anomaly detection. It first generates representative anomaly samples for each class by adapting existing anomaly generation techniques. A supervised contrastive learning strategy is then introduced to construct feature distributions based on patch-level class labels, allowing the model to capture fine-grained distinctions between classes. The method achieves state-of-the-art results on various benchmarks, demonstrating both its effectiveness and efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The paper's writing quality is below standard, with excessive notation, numerous typos, and poor organization, all of which significantly hinder readability.\n\n2.The paper presents only the main results and lacks additional experiments, such as ablation studies, sensitivity analyses, etc., which are essential for thoroughly evaluating the robustness and effectiveness of the proposed approach.\n\n3.Overall, I believe this is an incomplete work that requires further development and additional experiments to fully validate the findings. It should be withdrawn in its current version." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Related work is lacking. In total 10 papers are cited but the anomaly detection field is very broad. Even the competing methods that the proposed method is compared with in the evaluation are not cited. \n\nAt times poorly written - Section 3.3 is an example. The first few sentences are very difficult to understand. Also lines 231 to 247 could be rewritten to improve clarity since it is filled with typos and unclear notation.\n\nL323 – The anomalous examples should be higher than sigma, the normal samples should be lower than minus sigma according to equations 21 and 21. Incorrectly stated in the text if I am not mistaken, although this is a minor detail.\n\nThe evaluation section is severely lacking. An ablation study evaluating design choices such as the use of the GLAS-based anomaly generation method, weighted loss, use of the contrastive objective, use of L_var etc." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- Good performance on the multi-class setting. \n- The proposed method is generally well described." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed method utilizes the anomalies generated following the GLAS method and incorporating additional losses and training techniques to improve the results. For training, the contrastive losses, hard negative mining, and a variance regularization term are used. The method is evaluated on the MVTec AD and the VisA datasets. The proposed method achieves excellent results on both datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lacking related work (only 10 papers are cited from the very active field of anomaly detection)\n- Severly lacking evaluation section. No ablation study. One of the contributions listed in the introduction is a thorough evaluation of model backbones and configurations which is not fulfilled in the rest of the paper.\n\nWhile the proposed method achieves solid performance on the multi-class setting, the evaluation section is severely lacking. The design choices, such as the additional loss terms, are not properly evaluated in an ablation study which makes it impossible for the reader to extract important insights and understand the contribution of individual components. This can not be overseen and would require a rewrite of a significant part of the paper to include so I believe this alone is grounds for rejection." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. The paper lacks ablation studies, making the overall structure incomplete.\n2. In lines 012-013 of the abstract, the authors claim to address the high computational complexity of existing models with their proposed framework. However, there are no subsequent experiments to support or validate this claim.\n3. The experimental results do not show a significant advantage over existing multi-class anomaly detection methods. It is recommended to validate the method's effectiveness on more datasets.\n4. Since there is already a contrastive learning-based anomaly detection method, ReConPatch, the authors should compare their method with ReConPatch in the experiments." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The authors propose a multi-class anomaly detection method based on contrastive learning. They introduce a two-stage training framework to train the Adapter and Discriminator separately." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a multi-class anomaly detection method based on contrastive learning. They introduce a two-stage training framework to train the Adapter and Discriminator separately. The effectiveness of the method is validated on the anomaly detection metrics of the MVTec-AD and VisA datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "See Questions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "NA" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed method achieves better anomaly detection performance than the selected alternatives." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to enhance existing embedding-based anomaly detection methods to multi-class anomaly detection through contrastive learning. There exists several critical problems thus I vote to reject this paper." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Related works are not comprehensive enough, which makes this manuscript less convincing. The authors categorize anomaly detection methods into reconstruction-based and embedding-based methods, ignoring lots of related works like memory-bank-based, knowledge-distillation-based, etc.\n\nAlso, lots of works lack proper references, like the author mentions DINO without refer to the corresponding paper.\n\nThis paper also lacks innovations. The authors utilize several existing techniques like gradient ascending from GLASS to construct anomalies in the feature space.\n\nGLASS is quite an important baseline for this paper, but the authors didn’t compare the proposed method to GLASS.\n\nThere are even no ablation studies." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A Contrastive Learning based Method for Multi-Class Anomaly Detection" }, "_bibtex": { "value": "@misc{\nguo2024clad,\ntitle={{CLAD}: A Contrastive Learning based Method for Multi-Class Anomaly Detection},\nauthor={Runtang Guo and Hong Li and Tongfei Chen and Yuguang Yang and Peng Zhou and Linlin Yang and Guodong Guo and Baochang Zhang},\nyear={2024},\nurl={https://openreview.net/forum?id=zE4mL85zgg}\n}" }, "abstract": { "value": "Anomaly detection is crucial yet challenging in industrial production, especially in multi-class scenarios. Existing high-performance unsupervised methods often suffer from low efficiency and high model complexity. While lightweight discriminator-based detectors have been proposed, they are typically designed for single-class detection and exhibit significant performance degradation when extended to multi-class tasks. To address these limitations, we propose a novel Contrastive Learning-based multi-class Anomaly Detection (CLAD) method.\nOur approach first encodes multi-class normal images to generate normal samples in the feature space, then synthesizes anomalous samples in this encoded space. We then employ an Adapter network to compress the samples and leverage contrastive learning to effectively cluster normal and anomalous samples across multiple classes. Finally, a discriminator network is used for anomaly classification and identification. By leveraging anomaly sample generation and a two-stage training process, our framework achieves state-of-the-art performance on the MVTec and VisA datasets under the discriminator-based paradigm. Our key contributions include a novel framework for multi-class anomaly detection, efficient sample generation techniques, and a comprehensive evaluation of model configurations." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Runtang_Guo1", "~Hong_Li4", "~Tongfei_Chen5", "~Yuguang_Yang1", "~Peng_Zhou14", "~Linlin_Yang1", "~Guodong_Guo1", "~Baochang_Zhang1" ] }, "authors": { "value": [ "Runtang Guo", "Hong Li", "Tongfei Chen", "Yuguang Yang", "Peng Zhou", "Linlin Yang", "Guodong Guo", "Baochang Zhang" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Industrial anomaly detection", "Multi-class anomaly detection", "Contrastive Learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "guo|clad_a_contrastive_learning_based_method_for_multiclass_anomaly_detection" }, "pdf": { "value": "/pdf/79e08af65187e5e7b968b9a2522ab72bd67be97f.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "CLAD: A Contrastive Learning based Method for Multi-Class Anomaly Detection" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "Authors seemed to give LLMs way more trust while processing data engineering on computer systems, involving unrestricted LLMs in such a field is not a good idea, I would suggest more ethical review for potential security concerns." }, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The multi-chunk strategy as the authors described in the paper, is a well-engineered strategy for building large-scale data RAG systems, however, there's not a universal way for chunking the dataset or documents, what specific method did you use to \"breaking the data into manageable sections while maintaining context across chunks.\"\n\n2. Since this is somehow like an end-to-end pipeline for dataset search on the web, would the authors mind sharing with me the overall latency of searching DataSEA? For example, suppose you are searching for the COCO dataset, regardless of the dataset download overhead (since it is related to your network bandwidth). What is the estimated waiting time I'm going to expect?\n\n3. One particular thing I'm also interested in is what kind of specific challenge did authors faced in this work. It's never been a settled one for processing data in the wild of free forms, there must be some different tasks remaining before and after this work. The authors did not, at least I did not, fully discuss the challenges for LLMs to process data. Understanding these points would also help understand the contribution proposed in this work.\n\n4. I'm curious about how authors process those huge datasets, e.g., datasets with 10GB+ single files, they can be binary-formatted or something else, making it extremely hard for simple general LLMs to do anything about it." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This work wrapped previous work in automated dataset discovery and analysis using the power of LLMs. DataSEA here is a fully automated framework for many data engineering applications, which is a creative combination of existing tools and engineering results. On the other hand, this work is a good start toward many automated information management systems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on data management using current techniques of LLMs, specifically, it uses first LLMs to provide a fully integrated solution for dataset discovery, evaluation, and custom analysis using large language models. It introduced the trilogy of automated dataset processing in the paper, Search, Evaluation (by LLMs), and Analyze (also by LLMs), which is an endorsable innovation in processing web data based on LLM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I'm concerned about the feasibility of this work in real applications, it seems the data processing itself is only somewhat related to tasks like document management with web search discovery, but lacks enough contribution. A comparative analysis with current or previous relevant methods could highlight DataSEA's strengths for other researchers to improve it and provide clearer insights into its performance and feasibility.\n\n2. On the other hand, the type of data this work can support is also very limited. As pointed out in their work, it is unable to process databases, making the work have less contribution in real applications, since nearly all the existing proven-valuable data is stored in different purpose-designed databases.\n\n3. This work heavily leaned on existing tools such as search engines and LLMs, which limits the contribution and novelty of the work. From my perspective, this paper uses LLMs and Google Search as a web crawler application combined with a few HTML tricks. It neither discussed the reliability of this framework using two existing tools nor evaluated it. In this paper, authors employ the LLMs and let them do their prompt-given jobs, if so, might not be a good research paper, but a good technical engineering application. It would be good if authors could showcase a few pieces of innovation beyond just combining existing tools, another good point would be elaborating your novel contribution at the beginning of the paper.\n\n4. The scope of this work seems too large and underestimates the complexity of actual web data processing. Real such data engineering means getting hands dirty from all-sourced data to efficiently store data, I'm conservative about whether LLMs are able to handle various sources, types, sizes, structures, and containers of data. I'm also curious about how authors process those huge datasets, e.g., datasets with 10GB+ single files, it would help to provide more detailed examples or some case studies showing how your system handles complex, real-world data processing scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "DataSEA offers an agentic framework to test how LLMs can help the data processing pipeline.\n\nThe paper is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents DataSEA, an automated framework for dataset processing that leverages LLMs to streamline dataset discovery, evaluation, and analysis. DataSEA consists of three core modules—Search, Evaluate, and Analyze—that autonomously handle tasks from locating datasets on the web to analyzing and generating code for custom visualization. The system aims to significantly reduce the manual labor associated with data preparation, allowing users to input a dataset name and receive automated support for finding, organizing, and analyzing data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Real-world Relevance of the Framework**: The paper suggests that LLM agents are capable of downloading datasets, finding metadata, and locating relevant websites independently. However, major dataset platforms (e.g., Hugging Face, Kaggle) now offer easy-to-use APIs for downloading datasets and metadata. This raises the question: why test if an LLM agent can handle these tasks from scratch, rather than prompting it with relevant API code? It would be beneficial for the authors to discuss the practical use cases for DataSEA given existing tools.\n- **Lack of Benchmark Focus**: The benchmark includes three components, each targeting different tasks. The search module retrieves dataset-related information, the evaluation module retrieves related literature and generates metadata, and the analysis module handles dataset downloading and visualization. However, it’s unclear what the unique challenges are for each module, which task presents the most significant bottleneck, and where the main challenges lie.\n- **Dataset Size**: The evaluation dataset is relatively small, with only 100 datasets tested.\n- The prompts in the appendix are poorly formatted and truncated due to page limits." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What is the model size used for Llama 3 in line 325, i.e., 8B, 70B? \n2. Can you reveal more experiment details based on the domain area of datasets? The author's team discussed the limitations of the system on biological fields so far.\n3. Can you describe a running example of how this system works? As the reviwer can't infer much detail from the current paper. Search: what would the optional details be? Extract: what would the custom property be? Analyze: user requirements?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1: The DataSEA system cooperates with multiple LLMs responsible for three core modules. Prompt engineering and a multi-chunk strategy are widely applied.\nS2: The paper claims to be the first to provide a fully integrated solution for dataset discovery, evaluation, and custom analysis using LLMs. By automating these processes, DataSEA has the potential to significantly impact data-driven research by reducing the time spent on preliminary data handling tasks.\nS3: This paper presents its objectives, the problem it aims to solve, and the solutions it proposes in a clear manner. Meanwhile, the author team has designed a bunch of experiments to support the system's robustness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a holistic and broad system, DataSEA, for dataset processing, including searching, evaluation, and analysis. This system is built on top of a set of LLMs driven by curated prompts. The author's team has conducted a series of experiments to evaluate the performance in terms of different processing speeds, and the result turned out to be promising. Code implementations are publicly available." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Grand but Sketchy Framework: The paper presents DataSEA as a comprehensive solution for dataset preprocessing, which is an ambitious goal. However, the description of the search, evaluation, and analysis modules, while innovative, may lack in-depth observation. It's not quite convincing to me whether the system is actually useful in wild dataset processing. For instance, line 330 quotes that it takes about 3-5 mins for high-speed mode. Also, the whole system employes LLM heavily in every stages as well as integration of Google product APIs, will it leads to a high cost?\n\n- Missing baseline: Comparative analysis is crucial for establishing the advancement of the system over existing ones. The paper could be improved by providing a section (or figure) that directly compares DataSEA's outcomes with those of other tools or frameworks mentioned in the related work section. If there are no direct competitors, the authors should explain why this is the case and how DataSEA's approach is fundamentally different or more effective.\n\n- Poor Presentation: The figures look conceptual but lack illustration with any examples. To strengthen this part of the paper, the authors could Include case studies or use cases that demonstrate the system's capabilities in real-world scenarios. In particular to content writing, the author team should amend properly in revision:\n1) Line 99 mentioned the system is powered by a set of LLMs. But model versions seem not aligning with the experiment settings in line:324. \n2) Line 158-163, the summary part appears to be suitable for the ending part of the introduction section, not the related work.\n3) Improve all in-line references and labeling of every figure or table.\n4) Wrong Table-of-Content in Appendix section." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Q1. What is the formal problem definition? Does the proposed solution address all aspects of dataset search, or are there limitations in scope?\n\nQ2. What are the specific scopes of data evaluation and analysis in the system? Does the framework aim to address all types of data evaluation and analysis problems, or are there certain boundaries or limitations?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "This paper tries to automate the entire dataset handling pipeline, from discovery to evaluation and analysis, significantly reducing manual labor." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces DataSEA (Search, Evaluate, Analyze), a fully automated framework for dataset processing using large language models (LLMs). Its goal is to streamline the data handling pipeline, reducing the manual effort involved in dataset discovery, preparation, and analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Key Weaknesses:\n\nW1. Lack of Formal Problem Definition:\n\nThe scope of \"dataset discovery\" and \"dataset processing\" is not clearly defined in the paper. While the authors claim that \"DataSEA is among the first to provide a fully integrated solution for dataset discovery, evaluation, and custom analysis using large language models,\" it remains unclear what specific types of datasets are targeted and how the boundaries of these processes are drawn. A formal definition of the problem and its components is essential to frame the solution appropriately.\n\nWhat types of datasets does DataSEA aim to handle?\nWhat are the specific steps involved in dataset discovery, evaluation, and analysis?\nWhat are the inputs and outputs at each stage of the pipeline?\n\nFor example, whether the user can search the dataset like \"COVID-19 cases in CA in the last season of 2022?\" or \"Give me datasets about house price is LA in 2023.\"\n\nW2. Simplified Approach to Dataset Discovery:\n\nThe dataset discovery aspect could be seen as limited, given that existing tools like ChatGPT for Google Search or Google Dataset Search (and many others, see (https://anaconda.cloud/useful-sites-finding-datasets)) could be easily integrated to achieve similar outcomes. The paper does not explore how DataSEA improves upon or distinguishes itself from these existing dataset discovery engines, raising questions about its novelty in this area. Therefore, it is suggested to:\n\n- Provide a comparison table showing the capabilities of DataSEA versus existing tools like Google Dataset Search and ChatGPT.\n- Explain any unique features or improvements DataSEA offers over these existing solutions.\n- Discuss why integrating existing tools was not chosen as an approach, if applicable.\n\nW3. Ambiguity in Types of Evaluation:\n\nThe different types of evaluations performed by DataSEA are not well-defined. While the paper suggests that GPT can handle these evaluations, it lacks a formal breakdown or categorization of the evaluation types, leaving the reader unclear about the specific contributions and effectiveness of the system.\n\nCan I ask questions like \"how many data errors exist in the searched dataset.\"\n\nW4. Undefined Scope of Custom Analysis:\n\nThe notion of \"custom analysis\" remains vague and insufficiently described. It is unclear what kind of analyses the system can perform and how flexible or adaptable these are to various research scenarios. Without clear examples or a formal definition, the custom analysis capability is hard to evaluate or differentiate from standard LLM tasks.\n\nCan I ask questions like \"what will the stock price trend next month?\"\n\nW5. Lack of Clear Comparisons with Baseline Methods:\n\nAlthough there are many LLM-powered methods available for automated evaluation and analysis (e.g., using prompts with GPT models), the paper does not provide a comparison with state-of-the-art baselines. Without experimental results that benchmark DataSEA against these existing solutions, it is difficult to assess its effectiveness or performance improvements. The authors could compare with\n\n- Standard LLM prompting approaches for dataset analysis,\n- Existing automated data analysis tools, or\n- Manual expert analysis as a human baseline\n\nExperimentation Weaknesses:\n\nW6. Missing Comparisons with LLM-Based Methods:\n\nThe experiments do not compare DataSEA’s performance with existing LLM-powered methods that could achieve similar results through effective prompting. This omission weakens the claims of novelty and superiority, as there is no evidence of how DataSEA outperforms such methods. \n\nIt is acceptable to include supplementary materials in the appendix. However, the formal problem definition and the proposed solutions must be clearly presented in the main body of the paper, rather than relying on simple descriptions" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A fully automatic system leveraging large language models to streamline dataset acquisition, metadata extraction, and preliminary analysis, enhancing research efficiency and data exploration." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024datasea,\ntitle={{DATASEA} - {AN} {AUTOMATIC} {FRAMEWORK} {FOR} {COMPREHENSIVE} {DATASET} {PROCESSING} {USING} {LARGE} {LANGUAGE} {MODELS}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zEPYCDaJae},\nnote={under review}\n}" }, "abstract": { "value": "In the era of data-driven decision-making, efficiently acquiring and analyzing\ndiverse datasets is critical for accelerating research and innovation. Yet, traditional manual approaches to dataset discovery, preparation, and exploration\nremain inefficient and cumbersome, especially as the scale and complexity of\ndatasets continue to expand. These challenges create major roadblocks, slowing down the pace of progress and reducing the capacity for data-driven breakthroughs. To address these challenges, we introduce DataSEA (Search, Evaluate, Analyze), a fully automated system for comprehensive dataset processing, leveraging large language models (LLMs) to streamline the data handling\npipeline. DataSEA autonomously searches for dataset sources, retrieves and organizes evaluation metadata, and generates custom scripts to load and analyze\ndata based on user input. Users can provide just a dataset name, and DataSEA\nwill handle the entire preparation process. While fully automated, minimal user\ninteraction can further enhance system accuracy and dataset handling specificity.\nWe evaluated DataSEA on datasets from distinct fields, demonstrating its robustness and efficiency in reducing the time and effort required for data preparation\nand exploration. By automating these foundational tasks, DataSEA empowers\nresearchers to allocate more time to in-depth analysis and hypothesis generation, ultimately accelerating the pace of innovation. The code is available at\nhttps://github.com/SingleView11/DataSEA." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Automated Data Processing", "LLM", "Data Pipeline Automation", "NLP", "Data Mining" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/9ce79fc7a36ed42e392aa531f929e0b809dee7a5.pdf" }, "presentation": null, "primary_area": { "value": "infrastructure, software libraries, hardware, systems, etc." }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/38c48d32357739f260192fe210b9fb07b7c51349.zip" }, "title": { "value": "DATASEA - AN AUTOMATIC FRAMEWORK FOR COMPREHENSIVE DATASET PROCESSING USING LARGE LANGUAGE MODELS" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "How does the runtime of GHRR compare to that of the vanilla Transformer? Given that vanilla attention has quadratic complexity, it would be good to know the complexity of GHRR relative to the standard self attention mechanism." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors draw useful connections between self-attention based Transformer models and Vector Symbolic Architectures, and use it to derive mathematical equivalence between QKV hypervectors and the self-attention mechanism.\n2. The authors use the insights to construct self-attention mechanisms for more complex data, proposing a GHRR-based Graph Transformer architecture for graph inputs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a Transformer encoder architecture based on Generalized Holographic Reduced Representations (GHRR), a Vector Symbolic Architecture (VSA) paradigm capable of implementing data structures including attention mechanisms. The proposed architecture uses binding-based positional encoding to encode sequential data, and also supports graph inputs. The architecture is evaluated on language modeling as well as node and graph classification tasks, showing better perplexity on language modeling tasks and comparable performance on vertex classification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The benchmarking of GHRR is limited; next token prediction perplexity is reported on two language datasets in Table 2 against a vanilla transformer, however the language modeling results would be strengthened by benchmarking on more recent NLP benchmarks and tasks, for instance the LAMBADA dataset for natural language understanding. Performance on a node classification and graph classification task is reported in Tables 3 and 4, however both experiments are missing baseline graph transformer models such as GPS Graph Transformer [1] and Graphormer [2]. Evaluating on more standard GNN benchmark datasets, such as ZINC and Open Graph Benchmark datasets, would also strengthen the empirical results of GHRR.\n\n1. Rampášek, Ladislav, et al. \"Recipe for a general, powerful, scalable graph transformer.\" Advances in Neural Information Processing Systems 35 (2022): 14501-14515.\n2. Ying, Chengxuan, et al. \"Do transformers really perform badly for graph representation?.\" Advances in neural information processing systems 34 (2021): 28877-28888." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Is the $\\delta$ similarity measure used in the model itself?\n- How were the hyperparameters for model training chosen?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper considers an interesting goal: implementing transformers using GHRR, which has rich algebraic structure\n- A clear and informative description of VSAs and GHRR is provided.\n- The proposed model is original; both for sequences and graphs.\n- Valuable experimental results are demonstrated" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new transformer architecture based on vector symbolic architectures (VSAs). The paper uses the Generalised Holographic Reduced Representation implementation of VSAs, which represents tokens as complex hypervectors ($\\mathbb{C}^{D\\times m \\times m}$), and models interactions between these through bundling and binding algebraic operators.\nThe paper implements a transformer for sequences, and a transformer for graphs using GHRR, and demonstrates experimental results on language modelling and vertex classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The presented results for the language modelling task use a single baseline, the vanilla transformer. For this model, the chosen embedding size is quite small, compared to those in the literature. It is also uncommon that the hidden dimension for the transformer is smaller than the model dimension.\n- No hyperparameter search is described for either task.\n- The presented results thus are not entirely informative about the model's performance.\n- While VSA's are known to have rich algebraic structure, the paper does not discuss this for the presented model. It would have been valuable to demonstrate whether, for example, semantically similar words are mapped to GHRR representations with high similarity as measured by $\\delta$ described on line 170. This would serve to motivate the architecture better.\n- Related to the above, it is not clear from line 242 whether the GHRR version of attention uses the similarity measure from the VSA, which would be most natural." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- unclear to me why there was a significant performance improvement over transformers if the attention mechanism is theoretically the same. Could you perhaps elaborate on the specific differences in the GHRR transformers that could contribute to the performance differences in the discussion/results? Ideally provide some analyses to causally investigate the important differences.\n- why does the VSA formulation help with graphs in a way that vanilla transformers cannot already do? Graphs can readily be encoded as tokens in a context." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- referred to bridge between transformers and hopfield networks\n- presented material in a digestible way even for those who have never heard of VSAs\n- demonstrate a formal, theoretical unification between VSAs and transformer attention\n- support theoretical claims with empirical evidence\n- show that their formulation can be readily used in another domain (graph classification)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present a theoretical bridge between Vector Symbolic Architectures and transformer attention through the specific VSA of Generalized Holographic Reduced Representations. They show how the transformer attention mechanism can be derived as a specific formulation of the GHRR with some slight relaxations. They then show that their derivation is correct by implementing it and comparing against vanilla transformers in a few settings. They also show how the VSA formulation more naturally extends to graph learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- as I have never encountered VSAs before, I'm unsure about impact/significance of VSAs in general. Does the broader ICLR audience care about VSAs? Why should they care?\n- difficult to understand the variants presented in Table 2. It would be nice if the differences between the variants was more clearly summarized closer to the point at which they are presented. Were they described in the methods?\n- would have appreciated a comparison of the graph and vertex classification results to vanilla transformers\n- generally difficult to understand some of the notation (probably due to my lack of exposure to VSAs)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1) Can you please provide evidence of what happens empirically when the model is used with a token context window n larger than m?\n\n2) Can you elaborate on the effect of relaxing unitarity of W?\n\n3) Can you provide experiments on widely accepted graph benchmarks, such as those on the Open Graph Benchmark (OGB)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The idea of deriving a more general framework for self-attention through VSA is interesting. In particular as it may enable a principled way to manipulate the inductive bias of the attentional module by operating on how the underlying vector space is generated and how positional information is embedded in the representation by the algebra operators. This makes the work, in principle, able to achieve some impact on the community.\n\n * The proposed approach, surprisingly, does not imply a substantially higher parameterization and computational costs (at least according to what the Authors state in the paper, as there is no specific experimental data supporting this claim)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work discusses how a specific implementation of Vector symbolic architectures (VSA) can be interpreted as a multi-head attention model for compound data encoding. The manuscript shows under which restrictions of the VSA model its computation map to self-attention with fixed context window. It further provides a discussion on how to define specialized positional encodings to process sequential and graph structured data. Experiments are provided on simple next token prediction, node classification and graph classification datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The work appears rather derivative when coming down to the learning methodology. The model underlying the proposed GHRR Transformer derives largely from Yeung et al 2024, aside from the part which integrate positional encoding information in the neural representation. Overall, the proposed equivalence of GHRR to self-attention boils down to a resemblance of the matrix operations involved in self-attention, when substantial constraints are imposed to the GHRR model. There is really no deep study and assessment of the relationship between the GHRR Transformer and self-attention, and of what are the consequences of some design choices and simplifications introduced in GHRR. Without a deeper insight on the key advantages introduced by the proposed GHRR-Transformer/Self-attention equivalence, this may appear rather empty and mostly a technical exercise. The equivalence itself hinges on quite strong simplifying assumptions, which are mentioned but whose impact is not discussed in depth. For instance, GHRR assumes a fixed context window: if such assumption is relaxed, then the dimensions of the embedding becomes entangled, loosing the very motivation for having introduced an holographic representation in a first instance. This aspect is mentioned in the paper, but only marginally, while it seems a major limitation of the approach.\n\n* The work appears derivative also with respect to the contribution on graph processing. The encoding of graph data in the proposed GHRR Transformer is heavily based on Poduval et al, 2022. The paper is not clear about what novel insight is being provided on graph encoding by the proposed approach against the one of GraphHD. It would be helpful if the Authors could elaborate more on what are the novelties of the proposed approach agains GraphHD. The technical discussion on the properties of the GHRR Transformer against what is the state of the art in graph NNs, remain on a shallow level. For instance, if I got this right, the disentanglement properties of the proposed Holographic embedding should allow to memorize exact information about inter-nodes relationships in the vertex encoding. This may be relevant with respect to the literature discussion about how to design graph NNs capable of capturing long range node relationships, surpassing limitations such as oversquashing. The work appears not well positioned with respect to relevant literature in this respect.\n\n * The work, at some point, relaxes the assumption on W being unitary. My understanding is that such an assumption in needed to preserve the holographic nature of the embeddings. It cannot be relaxed without a proper discussion of how this affects the properties of the model (the discussion should be both theoretical and empirical in this sense). Taking assumptions which contradict the very fundamental reasons for having introduced the holographic approach in the first instance, reduces the soundness of the contribution.\n\n * The empirical analysis is very limited in scope, depth and reproducibility. Little details are provided as concerns the experimental setup and no reference to code is given (neither public anonimyzed nor attached to the submission as supplementary). It would be helpful if the Authors can provide additional details to facilitate reproducibility, including choice of optimizers, hyperparameters, as well as to gain a deeper insight into the computational charateristics of the approach, such as its computational costs and parameterization in the experiments.\n\n* The experiments on sequential data are too limited: only a single 2-dataset experiment with simple next token prediction tasks is provided. If the approach is put forward as an holographic equivalent of Transformers (in year 2024), then one would expect to see experiments on how the approach can be used to at least match a Transformer on proper language modeling tasks. The experiments with graph data are quite poor. There is no reference baseline model from literature (I would have expected a comparison with at least the most popular Graph Transformer models). The datasets used in the experiments are not widely recognized benchmarks by the graph NN community and this does not allow to compare the proposed model against the relevant related literature. The dataset on graph classification cannot be considered a proper graph benchmark: deciding between a fully connected and a non-fully connected graph does not require a model with the ability to capture complex structured relationships in a graph. It suffices a model which can count the number of ones in the adjacency matrix." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024structureaware,\ntitle={Structure-aware Attention based on Vector Symbolic Architectures},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zET0Zg71WT},\nnote={under review}\n}" }, "abstract": { "value": "The introduction of the Transformer has brought about a revolution in AI. Central to the success of the Transformer architecture is the self-attention mechanism, enabling context dependence and long-range dependencies between tokens. Recent work has drawn an equivalence between Hopfield networks, a kind of associative memory model, and Transformers. In this work, we leverage this bridge, using Vector Symbolic Architectures (VSA), a brain-inspired computational paradigm capable of representing and implementing data structures, including associative memory models, to define a broad class of attention mechanisms catered for complex data types. In particular, we use Generalized Holographic Reduced Representations (GHRR), an implementation of a VSA, as the foundation for our proposed class of attention mechanisms. We show that GHRR is capable of implementing attention and design a GHRR Transformer encoder architecture based on the demonstrated mathematical equivalence. We propose a new kind of binding-based positional encoding based on methods used in VSAs for encoding sequential information. We extend the attention mechanism in our architecture to support graphs, inspired by techniques used in VSAs to encode graph representations. We evaluate the GHRR Transformer on language modeling, vertex classification, and graph classification tasks. Results suggest that our approach provides benefits in language modeling and graph classification tasks compared to baseline models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "transformers", "attention", "vector symbolic architectures", "neurosymbolic ai", "hyperdimensional computing" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/df207c160499bd6f96e1619821cf6a7056f97dce.pdf" }, "presentation": null, "primary_area": { "value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Structure-aware Attention based on Vector Symbolic Architectures" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to \"Weaknesses.\"" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper proposes a new cycle-consistent diffusion-based framework for virtual try-on, named CycleVTON.\n\n2. This paper introduces two complementary networks in CycleVTON. Alongside the conventional try-on network, we innovatively incorporate a clothing extraction network that extracts clothing from the human body and normalizes it to a frontal view.\n\n3. This paper proposes a cycle-consistency optimization strategy, which aligns the generated dressed human with real images and the extracted clothing with its true frontal appearance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes CycleVTON, which is a virtual try-on framework that uses a cycle-consistent diffusion-based approach. It consists of two conjugated networks: a regular try-on network and a clothing extraction network. The clothing extraction network standardizes clothing into a front-facing format, allowing for alignment between generated and real images. This cycle-consistent optimization strategy enhances the retention of clothing textures and structures, ensuring realistic and accurate clothing generation. The conjugated network structure supports traditional virtual try-on as well as flexible clothing extraction and exchange between individuals. Experiments on VITON-HD demonstrate its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In the proposed architecture, the CLOTHING EXTRACTION NETWORK differs only in its function, yet it does not fundamentally diverge from the ReferenceNet used in previous methods, such as IDM-VTON.\n\n2. The proposed Cycle-Consistency Optimization Strategy is successful. However, there is a lack of validation for the effectiveness of its subprocesses and sub-losses, such as $L^_h$ and $L^_c$.\n\n3. Some of the data in Table 1 seem to differ from those in the original paper, why?\n\n4. Please provide the code to demonstrate reproducibility.\n\n5. The entire process is overly engineered, lacking rigorous theoretical validation in this type of conference." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Can the CEN generalize to new datasets (such as DeepFashion)? \nDoes the CEN affect the generalization capability of the virtual try-on?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-written and easy to follow.\nThe proposed method outperforms existing baseline approaches on the VTON-HD dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents CycleVTON, a framework for virtual try-on. CycleVTON consists of two main components: a virtual try-on network that synthesizes human images given a specific clothing style, and a clothing extraction network that generates clothing templates in a unified space. Experiments are conducted on VTON-HD to evaluate the performance of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The idea of extracting human feature and clothing feature in a cycle framework is not novel, i.e., a similar work CycleVTON: A Cycle Mapping Framework for Parser-Free Virtual Try-On [Du et al., AAAI 2024]. The paper did not cite or discuss the CycleVTON paper [Du et al.]. A comparison against CycleVTON [Du et al.] should be provided since the ideas of the two papers are similar.\n\nThe technical contribution is limited. The virtual try-on part (ReferenceNet and denoising UNet) is borrowed from Animate Anyone [Hu et al.], and the main contribution is the idea of a clothing extraction network, which, however, has appeared in a previously published paper CycleVTON [Du et al.]. The technical contributions should highlight the differences against existing works.\n \nNo discussion on the failure cases (such as particular clothing types and poses). What are the limitations of the method?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Overall, CycleVTON represents an innovative approach to virtual try-on, balancing accuracy, flexibility, and robustness, with areas for improvement in handling complex clothing variations. But the cycle-consistent idea is not new, and please provide a detailed discussion on the paper: https://openaccess.thecvf.com/content/CVPR2021/papers/Ge_Disentangled_Cycle_Consistency_for_Highly-Realistic_Virtual_Try-On_CVPR_2021_paper.pdf" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The cycle-consistency strategy greatly enhances the model's ability to preserve clothing textures and structures, resulting in more realistic try-on outcomes.\n\n2. The dual-network design enables various applications, such as clothing extraction and swapping between persons.\n\n3. By utilizing clear frontal views for supervision, CycleVTON enhances adaptability to clothing deformations and variations in human poses, improving both supervision and robustness.\n\n4. It demonstrates superior performance, as the model surpasses state-of-the-art (SOTA) methods on the VITON-HD benchmark, excelling in both qualitative and quantitative measures, and highlighting its effectiveness in virtual try-on." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a cycle-consistent and diffusion-based framework for try-on. It consists of two networks: a regular try-on network and a clothing extraction network. The latter standardizes the clothing into a frontal view, allowing for a more realistic and accurate representation. Its consistent strategy optimizes the alignment between generated try-on images and original human images, as well as between extracted clothing and its accurate frontal view. This framework enables traditional virtual try-ons and adaptable applications such as clothing extraction and swapping. Experiments on the VITON-HD benchmark confirm CycleVTON’s superior performance compared to existing models. However, challenges remain in handling significant deformations and changes in clothing texture or orientation. Furthermore, the cycle-consistent idea is not new, and please refer to the https://openaccess.thecvf.com/content/CVPR2021/papers/Ge_Disentangled_Cycle_Consistency_for_Highly-Realistic_Virtual_Try-On_CVPR_2021_paper.pdf" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Significant clothing deformations or angle changes lead to artifacts such as wrinkling or stretching, which limit the model's ability to accurately reproduce the original textures and patterns.\n\n2. It depends on the standardized front views for optimal performance, which may limit its effectiveness with certain types of clothing.\n\n3. The dual-network cycle-consistent approach may raise computational costs and complexity, potentially affecting its use in real-time or resource-limited environments.\n\n4. The cycle idea is very close to this paper, so the novelty needs to be clarified.\nhttps://openaccess.thecvf.com/content/CVPR2021/papers/Ge_Disentangled_Cycle_Consistency_for_Highly-Realistic_Virtual_Try-On_CVPR_2021_paper.pdf" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. L258: Does the term \"frontal-view posture information\" refer to a canonical pose, or can it be any pose that is frontal? Additionally, is it correct to assume that the generated clothing image will be in this pose?\n\n2. Regarding the clothing extraction model, the input includes a photo of the clothed human, an image of the clothing itself, and the pose information. If the image of the clothing is already available, what is the purpose of using this model to extract the clothing again? Please clarify if there is a misunderstanding in my interpretation.\n\n3. Is there any limitations to the types of poses the model can handle?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The cycle consistency optimization strategy enables the two models to be trained jointly, providing the capability for both virtual try-on and virtual take-off of clothing.\n\n2. The input for the clothing image is not limited to a frontal view; the model can accept variations in the clothing image, which makes it more practical for real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a 2D virtual try-on network capable of: 1) rendering an image of a human wearing a given piece of clothing based on an input human image and clothing image, and 2) extracting the image of clothing from a clothed human image. Existing methods have been utilized to extract information from the human image in task 1), such as pose, human parsing, and body mask. A cycle consistency optimization strategy is introduced to jointly optimize the two models for tasks 1) and 2), where the ground truth of model 1) can serve as input for model 2), and vice versa. This approach simplifies supervision and data acquisition while achieving a better representation of the overall appearance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The explanation of the methodology could benefit from clearer clarification, as it is difficult to understand which phrase refers to which \"image\" in the pipeline. For instance, in lines 261-267, adding notations and labeling components in Figure 2 would help avoid ambiguity. Readers may struggle to differentiate between the frontal-view image and the distorted image and add labels (e.g., A, B, C) to each image in Figure 2 and use these labels consistently in the text description will be helpful. Additionally, in Figure 2, using different clothing items for the try-on and take-off procedures would help reduce confusion, as presenting the same blue t-shirt makes the process harder to follow.\n\n2. The experiments were conducted on a dataset with professional fashion images, but in real-world virtual try-on applications, \"in the wild\" images are typically used. However, there is no experiment demonstrating how the proposed method performs on in-the-wild images. Could this suggest that the model may be overfitting to the fashion dataset and might struggle to adapt to more diverse, real-world images? the authors could conduct additional experiments on a dataset of non-professional, real-world images to demonstrate the model's generalization capabilities.\n\n3. The poses in the presented dataset appear to involve relatively limited movement. It would be beneficial to include examples or conduct additional experiments with more extreme or dynamic poses to demonstrate the model's robustness. \n\n4. L112 typo: ‘arping’ →’Warping’" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024cyclevton,\ntitle={Cycle{VTON}: Improving Diffusion-Based Virtual Try-On with Cycle-Consistent Training},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zEUDoD9cU9},\nnote={under review}\n}" }, "abstract": { "value": "We present CycleVTON, a cycle-consistent diffusion-based virtual try-on framework. Unlike existing methods that rely on a single try-on network, our model consists of two conjugated networks. In addition to the regular try-on network, we design a clothing extraction network that extracts the clothing worn by the person and standardizes it into a front-facing format. These two networks are symmetrical, enabling alignment between the generated dressed human and real images of dressed human, as well as between the extracted clothing and its front-facing ground truth. This cycle-consistent optimization strategy allows for enhanced retention of clothing textures and structures, ensuring a more realistic and accurate clothing generation in virtual try-on scenarios. Moreover, the conjugated network structure not only supports traditional virtual try-on but also allows flexible clothing extraction and clothing exchange between different individuals. The experiments on VITON-HD demonstrate the effectiveness of our approach." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "virtual try-on", "diffusion", "cycle-consistency" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/d6f430bab3aa7094ff65ffc55a4f38a94857a30c.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "CycleVTON: Improving Diffusion-Based Virtual Try-On with Cycle-Consistent Training" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "•\tIs it likely that the LLMs have been trained on much of the data of AIDBench? How much does this contribute to their performance?\n\n•\tIt’s okay not to have the formulas for standard metrics such as precision and recall. \n\n•\tPutting the random guess metrics from Table 3 in Table 4 / 5 would’ve made reading those tables much easier. \n\n•\tWill the dataset be released?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "•\tThe research question is interesting, namely whether LLMs can perform authorship identification tasks. \n\n•\tThe authors collect a new benchmark that could be useful for the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce AIDBench, a new benchmark to evaluate the ability of LLMs to perform authorship identification. The benchmark is made up of five domains and two evaluation settings: one-to-one author identification (authorship verification), and one-to-many author identification (authorship retrieval). Multiple LLMs are prompted, both with and without a topic-controlled prompt, and a RAG-based approach is proposed in cases where the length of the candidate texts exceed context length of the LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "•\tIt’s not clear whether a new benchmark was needed. There are many authorship verification and authorship retrieval datasets that could’ve been used to evaluate the attribution abilities of LLMs. This would’ve also made comparisons against established methods easier. \n* Authorship Verification (One-to-One) – https://pan.webis.de/clef23/pan23-web/author-identification.html (essays, emails, interviews, and speech transcriptions)\n - https://pan.webis.de/clef22/pan22-web/author-identification.html\n - https://pan.webis.de/clef21/pan21-web/author-identification.html\n\n* Authorship Retrieval (One-to-Many) – \n - Test dataset from - https://arxiv.org/abs/2105.07263\n\n•\tIt would’ve been better to evaluate with metrics that account for class imbalance, such as the F1 score. This would sharpen the results in Table 2. \n\n•\tThe RAG-based approach relies on semantic embeddings for it search. Given the nature of the task, it would’ve been more natural to rely upon stylistic embeddings such as the following:\n- https://aclanthology.org/2022.repl4nlp-1.26/\n- https://aclanthology.org/2021.emnlp-main.70/\n- https://arxiv.org/abs/2410.12757\n\n•\tAlthough AIDBench contains 5 domains, it seems that only two domains were evaluated on (in the main text), namely the Guardian and Research Papers. \n\n•\tIn the One-to-Many scenario, it’s not clear whether one can attribute a research paper to a single author, since they’re collaborative pieces of writing after all. I am not sure what to make of the One-to-Many results in this scenario. \n\n•\tThere are no baselines compared against. Some of the papers linked above, as well as previous submissions to the PAN CLEF Author Verification challenge would’ve been good baselines. \n\n•\tI found the description of the way the one-to-many subsets were created to be confusing. Lines 365-369." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In the one-to-many identification setting, the authors randomly sample a number of authors and compile all of their writings into a set. Does this approach sufficiently simulate real-world conditions? Could this method inadvertently bias the results by creating sets where all texts are from the same author, and if so, how could this be mitigated?\n\nIn section 2.4, there is a discussion about using cosine similarity for authorship identification. Should the paper provide more rigorous discussion on whether high cosine similarity actually correlates with authorship? Could high similarity merely indicate shared topics or content rather than distinct writing styles, and if so, how could this be addressed?\n\nDoes the paper sufficiently address the potential limitations posed by potential data leakage? What strategies could the authors adopt to mitigate or acknowledge this potential bias in their results?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The research shows that LLMs, particularly when enhanced by the RAG method, can outperform random chance in identifying authorship, demonstrating potential privacy risks in de-anonymizing texts in various systems.\n- The paper introduces a benchmark that includes a variety of datasets from different domains, allowing for a thorough evaluation of LLMs in authorship identification.\n- The authors address LLM limitations with long input by introducing a RAG-based method, which improves large-scale authorship identification tasks with large number of candidate authors and text.\n- It highlights the underexplored privacy risks posed by LLMs, bringing attention to the potential de-anonymization of authors, particularly in sensitive environments like peer reviews." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the AIDBench benchmark that includes a diverse range of datasets to systematically test LLMs’ authorship identification capabilities. The authors test LLMs using two tasks. To enhance LLM performance in authorship identification when texts are too lengthy, the paper introduces a RAG-based method to improve accuracy. This method involves first filtering candidate texts and retrieving the most relevant chunks to keep within the LLM’s context window limits." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper's methodology for one-to-many identification may have a potential problem. By randomly sampling a number of authors and placing all their writings into a set, there is a risk that all texts in the set come from the same author. This introduces a bias and undermines the validity of the experiment.\n- The paper uses identifying anonymous reviewers as one of the motivations, yet the dataset used for analysis consists of research papers rather than actual academic reviews. Consider modify the motivation part, as the two types of text differ substantially in terms of length, tone, and style. Reviews are typically shorter and more opinion-based, while research papers are formal and content-rich.\n- In Section 2.4, the paper doesn't clearly discuss whether high cosine similarity is a reliable indicator of authorship. Cosine similarity typically measures the overlap in content or topic rather than writing style. High similarity scores could simply reflect the fact that two texts discuss similar subjects rather than being written by the same author. \n- The paper should acknowledge the potential limitation of the benchmark dataset, particularly the likelihood that the data used in this study were also used to train LLMs.\n- Section 3.1 notes that the results in Table 2 lead to a different conclusion from prior work. However, this discrepancy may be due to the fact that the previous work employed a different evaluation metric (accuracy) and a different dataset." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "None" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The manuscript mentions that embedding models tend to focus more on semantic meanings than on linguistic characteristics. However, there might be instances where texts from the same author share similar linguistic styles but differ significantly in semantic content. How to ensure that the top-k candidates cover texts that belong to the target author in all scenarios?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors introduce the RAG approach to address the challenge of one-to-many authorship identification, which could handle large-scale text collections and improve the efficacy of identifying multiple texts by a single author across extensive datasets.\n2. The experiments conducted across various text datasets . By demonstrating the model's performance in different scenarios, the authors show the adaptability of the RAG method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new benchmark, AIDBench, to show the privacy risks posed by large language models (LLMs) in potentially compromising the anonymity of authors across various text formats. The Retrieval-Augmented Generation (RAG)-based method is proposed to help authorship identification for large-scale texts exceeding typical model context windows." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The abstract suggests that the aim of this study is to explore the privacy risks associated with the use of LLMs for recognizing the authorship of anonymous texts. However, the experimental section seems primarily focused on validating the authorship recognition capabilities of these models and proposing new method to enhance their efficiency. There might appear to be a gap as the experiments do not adequately assess the outlined privacy risks, nor do they propose potential mitigation strategies.\n2. While the the RAG method is proper for one-to-many authorship identification, providing a comparison with baseline models, particularly under conditions where text lengths do not exceed the model's context window, would enable a better evaluation of the RAG method’s effectiveness.\n3. The selection and filtering of datasets, such as the \"50 papers filter 10 papers\" scenario, might be crucial in evaluating the model's performance but are insufficiently described. A more detailed explanation of how these settings were chosen and optimized are expected.\n4. The results section lacks the further analysis of why certain models perform differently under various tests. For instance, the Claude-3.5-sonnet model significantly outperforms others in the task involving 5 authors and 50 papers. It would be insightful if the authors could discuss potential reasons for this model's superior performance and possible factors leading to failures in others." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce AIDBench, a benchmark for evaluating large language models' ability to identify the authorship of anonymous texts, highlighting new privacy risks" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024aidbench,\ntitle={{AIDB}ench: A benchmark for evaluating the authorship identification capability of large language models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zEm5nXxiXU},\nnote={under review}\n}" }, "abstract": { "value": "As large language models (LLMs) rapidly advance and integrate into daily life, the privacy risks they pose are attracting increasing attention. We focus on a specific privacy risk where LLMs may help identify the authorship of anonymous texts, which challenges the effectiveness of anonymity in real-world systems such as anonymous peer review systems. To investigate these risks, we present AIDBench, a new benchmark that incorporates several author identification datasets, including emails, blogs, reviews, articles, and research papers. AIDBench utilizes two evaluation methods: one-to-one authorship identification, which determines whether two texts are from the same author; and one-to-many authorship identification, which, given a query text and a list of candidate texts, identifies the candidate most likely written by the same author as the query text. We also introduce a Retrieval-Augmented Generation (RAG)-based method to enhance the large-scale authorship identification capabilities of LLMs, particularly when input lengths exceed the models' context windows, thereby establishing a new baseline for authorship identification using LLMs. Our experiments with AIDBench demonstrate that LLMs can correctly guess authorship at rates well above random chance, revealing new privacy risks posed by these powerful models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "large language models", "privacy", "authorship identification" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/8cab422fe32dd1a1ffce73245de25ce69257eaac.pdf" }, "presentation": null, "primary_area": { "value": "datasets and benchmarks" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "AIDBench: A benchmark for evaluating the authorship identification capability of large language models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I'm wondering about the setup of the ground truth data and novel view data in the proposed dataset:\n1. only 1-3 objects are dynamic in the synthesis space yet all the other 10-20 objects and backgrounds are still. Is it a reasonable ratio? if the still objects (including the simple background) are less challenging, will the error from the dynamic object be overwhelmed? Especially when the input camera poses go around the scene like a scanning, the still objects will get much benefit from this camera movement. It is recommended to justify this ratio of dynamic to static objects, perhaps based on real-world scenarios or existing datasets. Another idea is to conduct an ablation study varying this ratio to show its impact on the model's performance.\n\n2. In the appendix, it is described that the monocular input ground truth is the 100 frames from the first camera, and the novel view ground truth is the 100 cameras at time 0. Is the novel view evaluation only tested on time 0? Won't introduce bias? Please clarify if it has been considered to evaluate novel views at different time steps. Or at least discuss potential biases this evaluation method might introduce and how they might address or mitigate these biases." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper presents a significant advancement by proposing a batchwise optimization method for DUSt3R, enabling it to be used effectively for dynamic Gaussian Splatting (DGS) without relying on Structure from Motion (SfM) pipelines like COLMAP. This innovation is particularly valuable as it removes the dependency on pre-calibrated camera data, allowing the model to perform self-calibration even in scenes with extensive object and camera motion. By introducing this self-calibrating approach, the paper expands the application range of DGS, addressing a major gap in pose-free dynamic view synthesis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents SC-4DGS (Self-Calibrating 4D Gaussian Splatting), a novel approach for pose-free dynamic view synthesis from monocular videos. Current dynamic neural field models rely heavily on Structure from Motion (SfM) and static scene assumptions, which limit their applicability in real-world scenarios with significant motion. SC-4DGS addresses these limitations by jointly optimizing camera poses and dynamic Gaussian representations without requiring pre-calibrated camera data or static scenes.\n\nThe approach leverages DUSt3R, a geometric foundation model, to provide initial pose and point cloud estimations. It introduces batch-wise optimization and an extended motion representation tailored to DUSt3R’s capabilities, allowing the model to handle dense frames efficiently. SC-4DGS also incorporates several regularization terms to improve geometric accuracy in rendering.\n\nTo evaluate SC-4DGS, the authors introduce Kubric-MRig, a new benchmark dataset designed to test calibration and rendering performance in dynamic scenes with extensive object and camera motion. Experimental results show that SC-4DGS outperforms previous pose-free dynamic neural fields and achieves competitive results against state-of-the-art pose-free 3D neural fields." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Clarity of Figures and Tables: \nThe visual presentation of some figures and tables could be significantly improved for readability and clarity: Figure 1 is not illustrative and hard to follow even if combining the main content. At least it can add a legend or annotations that explain each variable clearly, making it challenging for readers to follow without detailed reference to the text. Adding a legend or in-figure labels could improve comprehension. Figure 4 would benefit from a horizontal arrangement of camera rotations, as it would make the sequence and rotation dynamics easier to interpret at a glance. Addressing these issues would make the paper more accessible, particularly in sections introducing new methods and experimental setups.\n\nExcessive Detail in Preliminary and Method Sections (Sections 3.1 and 3.3):\nSection 3.1 delves too deeply into the basics of 3D Gaussian Splatting, including eight detailed equations that do not play a critical role in subsequent sections. While introductory information is helpful, streamlining this part to focus only on the essentials would allow readers to concentrate on the novel contributions. Similarly, Section 3.3 introduces multiple variables in equations that are only referenced once, which adds cognitive load without enhancing understanding. Reducing or simplifying these equations, or moving some details to an appendix, could maintain the paper’s technical rigor while improving readability.\n\nDataset Design Limitations: \nThe Kubric-MRig dataset, while useful, is overly simplified in certain respects: The background lacks detail and depth variation, as it is primarily a simple ground plane. This simplification limits the model's ability to generalize to more complex real-world scenes where background intricacies affect depth and spatial perception. The dataset's objects are all synthetic geometric shapes or scanned objects with basic forms, which fall short of representing realistic, complex shapes such as human bodies, animals, or vehicles. Incorporating more diverse objects with varied textures and structures would enhance the dataset’s utility as a benchmark for dynamic view synthesis in practical applications. For example, including urban environments with buildings, natural scenes with vegetation, or indoor settings with furniture. For objects, adding articulated models of humans or animals, or complex mechanical objects like vehicles.\n\nLimited Novelty: \nWhile the batch-wise optimization approach for DUSt3R to enable its use in dynamic Gaussian Splatting is a valuable engineering contribution, the novelty is relatively incremental: The main contribution is an engineering improvement rather than a theoretical innovation, which may not meet the high standards of groundbreaking novelty expected at ICLR. \nOn the other hand, to strengthen the impact, the authors could discuss broader implications of this method, such as potential extensions or applications in other areas, or propose future directions that could build on this optimization. For example, it can be tested on other types of 3D reconstruction models including dynamic NeRFs which also rely on COLMAP, or if it has potential applications in fields like robotics or augmented reality." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Q1. The batchwise optimization makes joint optimization of pose and dynamic rendering possible, but is it capable of correcting pose estimation errors if DUSt3R provides a suboptimal initial pose estimation? In other words, how robust is the system to inaccuracies in the initial poses provided by DUSt3R?\nQ2. If SC-4DGS shares similar failure conditions with most SfM-based methods, it seems like the approach is merely combining a two-step process into one without addressing the inherent issue of pose estimation errors. In scenarios where ground truth poses are not available, what specific conditions would lead to failures in pose estimation and rendering for SfM-based methods, but would still allow SC-4DGS to correctly render dynamic scenes and estimate poses?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper presents a fully integrated pipeline that allows for pose-free dynamic Gaussian splatting rendering.\n\nBased on my understanding, this is more of a pipeline paper rather than a structure-based improvement. However, the batch-wise optimization introduces novelty by addressing the memory and computational limitations of DUSt3R, making the method scalable for dynamic scenes.\n\nA new dataset that addresses limitations that previous benchmark could not assess" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces SC-4DGS, a pose-free optimization pipeline for dynamic view synthesis by combining DUSt3R for initial self-calibration, DynMF for dynamic motion representation, and a novel batchwise optimization strategy. The method claims to eliminate the need for structure-from-motion and enables the joint optimization of camera poses and dynamic scene representations, achieving high-quality rendering for complex dynamic scenes in a monocular setting. It also introduces a new dataset, Kubric-MRig, to evaluate both camera calibration and novel view synthesis performance in dynamic scenes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The proposed method heavily depends on prior approaches such as DUSt3R for both pose estimation and dynamic rendering. As a result, any failure in DUSt3R directly causes the failure of the proposed method. Moreover, the paper lacks some critical comparative experiments. While it repeatedly claims that structure-from-motion (SfM) methods, like Colmap, often fail, leading to inaccurate pose estimation and rendering, it would be beneficial to include experiments that clearly show where SfM combined with standard 4DGS falls short and where SC-4DGS excels. This would more effectively demonstrate the advantages of the proposed approach.\n\nAdditionally, the paper feels somewhat incomplete and lacks polish in its writing. For example, in the ablation study on pose initialization strategies, the following statement is problematic: \"We explore four batch sampling strategies when initializing poses via DUSt3R: naive, sequential (SQ), strided batch (SB), and ours… SB uses the strided batch technique for optimization.\" This sentence is uninformative, as it merely reiterates the meaning of the abbreviation without adding value. Furthermore, the following sentence refers to \"Appendix 4.4\" for more details, but upon checking, there is no such appendix." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1.Based on my understanding, RoDyNeRF uses an explicit neural voxel radiance field in the section 3. The paper’s description of it as an implicit representation is inaccurate and could lead to misunderstandings.\n2.In Table 1, the terms “wide viewpoints” and “large motion” are used but lack clear quantitative measures. Providing specific metrics or thresholds for these terms, such as angular range for viewpoints or displacement magnitude for motion, would help readers better understand how your dataset outperform than others.\n3.The preliminary section in 3.1 is extensive. If this content is not the authors’ own contribution, a concise overview would be more effective, allowing readers to focus on the novel aspects of the paper.\n4. DUSt3R requires multi-view images for point cloud reconstruction, similar to COLMAP. What scenarios does DUSt3R handle that COLMAP cannot? Identifying and explaining these differences would clarify DUSt3R’s advantages and justify its use over COLMAP. The author could provide more results to show DUSt3R outperform COLMAP.\n5.If my understanding is incorrect, please let me know. Most objects in the dataset are rigid objects, meaning they do not undergo mesh deformations like humans or animals do during motion, which makes the dataset less challenging. The authors could include some non-rigid objects in the dataset or discussing why you don't do that in the paper." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is among the first to achieve pose-free 4D Gaussian Splatting, which provides valuable insights for future research in dynamic scene reconstruction.\n2.The paper provides highly detailed mathematical formulations, which effectively build the theoretical foundation of SC-4DGS. Also, this thorough mathematical foundation enhances clarity, helping readers understand both the model structure and the optimization processes involved.\n3.The introduction of the Kubric-MRig dataset fills a gap in DVS benchmarking by incorporating extensive camera and object motion alongside ground truth data for calibration and novel view synthesis, providing a robust dataset for dynamic scene evaluation.\n4.The SC-4DGS method demonstrates large improvements over previous dynamic neural field approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the limitations of traditional dynamic view synthesis (DVS) from monocular videos, which relies on Structure from Motion (SfM) and thus requires stationary scenes. To overcome this, the authors propose SC-4DGS, a pose-free optimization pipeline for dynamic Gaussian Splatting (3DGS) that eliminates the need for SfM through self-calibration using geometric priors from DUSt3R. SC-4DGS introduces batchwise optimization and an extended motion representation to jointly optimize camera poses and scene representations in dynamic settings. Additionally, the paper introduces Kubric-MRig, a benchmark dataset with photorealistic scenes featuring complex camera and object movements for evaluating calibration and rendering quality. Experiments show that SC-4DGS outperforms existing pose-free 4D neural fields and performs competitively with 3D neural fields, proving effective for complex, real-world dynamic scenes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The SC-4DGS model lacks originality, as it largely combines the existing DYNMF and DUSt3R models. The primary change appears to be the replacement of COLMAP with DUSt3R for point cloud generation, without significant structural improvements. Demonstrating unique contributions or optimizations beyond this combination would strengthen the paper’s impact.\n2. Although the paper highlights the high computational complexity of DUSt3R, it lacks specific training time comparisons. Quantitative results showing the time reduction achieved through the paper’s optimizations would help illustrate the efficiency of SC-4DGS.\n3.The experiments mainly compare SC-4DGS with RoDyNeRF under conditions without ground-truth (GT) camera poses, and the results do not surpass methods using GT. This limited comparison, with only RoDyNeRF as a baseline in the absence of GT, raises concerns about the reliability of the findings.\n4.The paper states that most methods use COLMAP to serve as GT camera poses and point cloud; however, COLMAP estimates are not true GT values but approximations. Since SC-4DGS replaces COLMAP with DUSt3R for point cloud and pose estimation, it’s essential to demonstrate the advantages of this change. An experiment comparing COLMAP’s and DUSt3R’s effectiveness in scenarios where COLMAP fails but DUSt3R succeeds would substantiate SC-4DGS’s advantage.\n5.Although SC-4DGS is designed for monocular video inputs, the Kubric-MRig dataset features wide variation among its 100 cameras, and rapid camera switching may lead the monocular sequence to resemble stereo video. This could create inconsistencies when using the dataset to evaluate the performance of a strictly monocular input model. Adjustments to the dataset or a careful clarification of how it aligns with the paper’s objectives would address these concerns." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Since not mentioned in the paper, did you use the G.T. camera intrinsic? Or optimizing the camera intrinsic? Please make it clear in the paper. \n2. The proposed method seems to highly rely on DUSt3R which is an existing work for static scenes. But in the method section, the author claimed they used DUSt3R for camera pose initialization, I agree that DUSt3R can be a good initialization in some small or regular motion scenes, but I am wondering whether DUSt3R can work well in some large movement datasets such as DAVIS[1], iPhone,...? I am expecting the authors can conduct more experiments on DAVIS and iPhone to show the effectiveness and robustness of their method.\n3. Can the authors provide some motion mask samples generated by them? Since I think the motion mask estimation of RoDynRF is not that robust and effective.\n\n[1] Pont-Tuset, Jordi, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. \"The 2017 davis challenge on video object segmentation.\" arXiv preprint arXiv:1704.00675 (2017)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors incorporate DUSt3R for a wonderful point cloud and camera pose initialization. \n2. The authors proposed a new dataset that can benefit the research in this field.\n3. The author provides experiments on both dynamic and static scenes." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is built on the existing work DUSt3R. Firstly, the authors implemented DUSt3R with their new way of batch-wise alignment for the point clouds and camera pose initialization. After the canonicalization of points, with the benefit of monocular depth estimation from DepthAnything and the motion mask estimation from RoDynRF, they introduce more regularization loss to jointly optimize the camera poses, and scene representations. In addition, they claimed to propose a more challenging dataset for DVS. However, I keep my doubts about the statement that the proposed dataset is challenging. Besides, their heavy reliance on the initialization of DUSt3R makes me doubt their contribution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. From my perspective, this paper lacks comparisons with proper baselines. For example, there are some accepted works like casualSAM[1], LEAP-VO[2], ParticlesfM[3], ..., which I think should be compared with. Or can the authors provide some reasons for why lack of them?\n2. I think the authors should include the comparisons with COLMAP, since although COLMAP is somehow theoretically designed for static scenes, I believe it does not completely fail on most of the datasets. The exhaustive matching step can help COLMAP get rid of some (dynamic) outliers, even though some noise exists.\n2. I do not agree that in table 1, the authors claimed that the iPhone dataset does not contain large motion (e.g. 'spin', 'apple', 'space-out', 'pillow', 'teddy', ...).\n3. From my perspective, the authors introduce a lot of regularization terms, however, for dynamic scene experiments, the authors only conducted experiments on one public dataset, which I think is not enough.\n4. The authors claimed that COLMAP is very time-consuming, but I do not see any time comparisons with COLMAP or the existing baselines.\n\n\n[1] Zhang, Zhoutong, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. \"Structure and motion from casual videos.\" In European Conference on Computer Vision, pp. 20-37. Cham: Springer Nature Switzerland, 2022.\n\n[2] Chen, Weirong, Le Chen, Rui Wang, and Marc Pollefeys. \"LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry.\" In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19844-19853. 2024. \n\n[3] Zhao, Wang, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. \"Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild.\" In European Conference on Computer Vision, pp. 523-542. Cham: Springer Nature Switzerland, 2022." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Pose-free 4D Gaussian Splatting for Causaully Captured Videos" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024towards,\ntitle={Towards Pose-Free Dynamic Neural Fields: Leveraging Geometric Foundation Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zFfZEQHUiv},\nnote={under review}\n}" }, "abstract": { "value": "Dynamic view synthesis (DVS) from monocular videos has remarkably advanced in recent years, achieving high-fidelity rendering with reduced computational costs. Despite these advancements, the optimization of dynamic neural fields still relies on traditional structure from motion (SfM), requiring that all objects remain stationary during scene capture. To address this limitation, we present \\textbf{SC-4DGS}, a pose-free optimization pipeline for dynamic Gaussian Splatting (GS) from monocular videos, which eliminates the need for SfM through self-calibration. Specifically, we jointly optimize dynamic Gaussian representations and camera poses by utilizing DUSt3R, enabling accurate calibration and rendering.\nFurthermore, we introduce a comprehensive benchmark, \\textbf{Kubric-MRig}, that includes extensive camera and object motions along with simultaneous multi-view captures. \nUnlike previous benchmarks for DVS, where ground truths for camera information are absent due to the difficulty of capturing multiple viewpoints simultaneously, it facilitates evaluating both calibration and rendering quality in dynamic scenes.\nExperimental results demonstrate that the proposed method outperforms previous pose-free dynamic neural fields and achieves competitive performance compared to existing pose-free 3D neural fields." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Neural Rendering", "Pose-free", "Gaussian Splatting", "Dynamic View Synthesis" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/c5f53331d12d457972a0c3ce3da4321f29f7c250.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Towards Pose-Free Dynamic Neural Fields: Leveraging Geometric Foundation Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Some additional questions and suggestions:\n\n-In figure 3, the SAC line seems to be flat? I am surprised this task cannot be learned at all by SAC, and that the best method only reaches a ~60% success rate. Perhaps the degree of sparsity is not clear to me from the task description. Perhaps that sparsity could be elaborated on somehow?\n\n-I suggest adjusting page spacing so that the critical algorithms 3 and 4 are on the same page, since they are connected and key to this paper's contribution\n\n-Figure 4 feels like supplemental material to me, since it's basically just validating hyperparameter choices.\n\n-The caption on figure 6 should say what the shaded regions represent.\n\n-Also in figure 6, it looks like the algorithms haven't clearly converged on most of the three tasks. 500k environment steps is not a very large number, perhaps it would be worthwhile to run these experiments for the commonly-used 1 million steps to make sure there's separation at convergence?\n\n-Why is it valuable to vary update to data ratios so much? Is there a reason this is a key hyperparameter for this algorithm? If so that should be explained in the paper. It's not clear how this relates to explore-exploit tradeoffs\n\n-For that matter, why is sigma (which appears to be a key hyperparameter) not varied? It seems like performance should be sensitive to that parameter since it controls algorithm switching (the core contribution)\n\n-Further, when and how much does KEA actually switch between the policies? This is the core phenomenon of this paper, so surely there should be some confirmation it's actually switching regularly and in a structured way, no?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper is generally well written, and tackles a valid problem in RL, how to gracefully handle the explore-exploit tradeoff. The core results are clear, and the resulting algorithm seems to outperform the baseline by a significant amount on the evaluation task suite. The approach is, to my knowledge, original, though other methods have tackled this problem before in various ways." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Keeping Exploration Alive (KEA), an algorithm that switches between exploration policies depending on the value of the intrinsic reward in order to reduce sub-optimal performance due to explore-exploit tradeoffs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The core issue I have with this paper is that there's very little exploration of the proposed algorithm beyond claims of superior performance. The core concept is simple (not a bad thing), but there's no real consideration given to what makes it good or useful. Seeing as the core contribution of this work is very simple, it should be straightforward to do deeper analysis and benchmarking, but this paper doesn't do much of that, and largely just concludes \"it works.\"\n\nAs a result, I can't help but feel this paper is just sort of insubstantial and doesn't contribute that much to the field. It proposes an idea that is novel, but straightforward as an extension of past work, and doesn't do anything with the idea beyond show that it outperforms the baseline. If accepted as is, I'm not sure how much this will actually move the topic of exploration forward.\n\nAs such, I'm inclined to recommend rejection, but would like to see an expanded version of this work submitted to a future conference, because I do like the core idea of dynamically switching between explore and exploit policies for off-policy RL." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "-How sensitive is the model to the switching threshold parameter? \n\n- Was this threshold hand-tuned or optimized in some manner?\n\n- What is the computational impact of using two agents in terms of memory and processing requirements?\n\n- Does the model perform well in highly dynamic environments with frequently changing goals?\n\n- Can the method applied to other RL algorithms, beyond SAC?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The method is easy to understand and contributes to alleviating an important issue in the field of reinforcement learning. \n\n- Proposes a unique combination of SAC and a co-behavior agent with a dynamic switching mechanism.\n\n- Shows substantial performance gains in benchmarks like 2D navigation tasks and robot control.\n\n- Paper well written and easy to follow" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces KEA (Keeping Exploration Alive), a framework designed to enhance exploration efficiency in reinforcement learning, specifically for sparse reward environments. The core contribution of the paper is a novel method that combines the Soft Actor-Critic (SAC) approach with a co-behavior agent, designed to improve exploration by proactively coordinating multiple strategies. This coordination is facilitated by a dynamic switching mechanism, which alternates control between the SAC agent, which includes a curiosity-driven exploration method, and the co-behavior agent based on the novelty of the state being explored. This mechanism enables KEA to maintain high stochasticity in high-novelty regions and avoid premature convergence, thereby improving both learning efficiency and robustness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper should discuss how the switching mechanism threshold is tuned and its effect on the performance. This appears to be a key mechanism but somehow ablation studies on this component are limited. \n\n- The evaluation lacks comparison with papers that address a similar issue, such as [Never give up: Learning directed exploration strategies, Badia et al., 2020]. I would suggest adding stronger baselines to validate the proposed method.\n\n- The paper’s proposed method, KEA, is designed specifically with SAC in mind, leveraging SAC’s ability to handle exploration-exploitation trade-offs in off-policy settings. However, the authors do not discuss how KEA might generalize to other off-policy algorithms. \n\n- Finally, one of my main concerns comes from the switching mechanism, which utilizes a fixed threshold to decide when to switch between the SAC agent with curiosity-driven exploration and the co-behavior agent. While this approach appears effective in simulations, using a fixed threshold might lack the flexibility needed for different environments or tasks, especially those with varying reward sparsity or novelty patterns. This inflexibility could limit KEA’s adaptability, potentially leading to less effective exploration in more complex or dynamic scenarios" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In Figure 2 (right-upper part), when the intrinsic reward is higher than a threshold, KEA switches to the co-behavior agent $\\pi^B$. While this region is novel (i.e., high intrinsic reward), meaning the agent has not been well trained here. Why does the non-selected agent $\\pi^{SAC}$ show an uneven distribution? Generally, for unfamiliar samples (those that have been rarely trained), the policy tends to output a relatively uniform distribution.\n\n2. In Section 3.2, how is UTD set for both $\\pi^{SAC}$ and $\\pi^B$ respectively? In the UTD experiment, how is the batch size set? Since batch size has an impact on UTD performance, more details would be helpful.\n\n3. The reference paper [4] presents a framework similar to KEA, where an additional agent is trained, and a switching control mechanism is designed. I encourage the authors to discuss this relevant work and make comparisons.\n\n[4] Mguni, David, et al. \"Learning to shape rewards using a game of two partners.\" Proceedings of the AAAI Conference on Artificial Intelligence. 2023.\n\nI would like to increase the score if these concerns are addressed." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The approach presented in this paper is straightforward and easy to implement. It effectively combines two exploration strategies (curiosity-driven exploration and stochastic policy sampling). Experimental results demonstrate improved performance compared to approaches that rely on a single exploration strategy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a switching mechanism to cooridnate exploration behaviors from two agents, an SAC agent with curiosity-based exploration strategy, and a co-behavior agent with only stochastic policy to draw actions. The proposed algorithm used the novelty intrinsic reward to choose which agent to output action by comaparing with a threshold. Therefore, the proposed algorithm can actively corredinate two different exploration strategies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The underlying assumption for the switching criterion is not intuitive. In Equation (4) and the following explanation (from around Line 239 to Line 245), it is stated that for low intrinsic reward regions (i.e., low novelty), KEA switches to curiosity-based exploration to encourage the agent to explore novel areas. Conversely, for high novelty regions, KEA switches to a more stochastic sampling policy. However, it would make sense if it was the other way around, as high intrinsic rewards usually indicate that the area is novel and worth exploring further, making it more logical to apply a strong exploration strategy. Conversely, low intrinsic reward regions (low novelty) might be better suited for stochastic sampling. I recommend that this assumption be supported with more quantitative analysis or additional comparative experiments.\n\n2. The paper conflates curiosity-based exploration and novelty-based intrinsic reward exploration. They are generally considered two distinct approaches for encouraging exploration. While this doesn't undermine the core idea of the paper, clearer distinctions between the two should be made, and the terminology used should be more precise. Including more related representative works would also help.\n\n3. The experimental section can be improved. First, I suggest that the authors include additional baselines that focus specifically on exploration, beyond just the backbone algorithms. Since KEA’s core idea is to combine two exploration strategies, only comparing it against SAC, RND, and NovelD seems more like an ablation study (with only one agent in KEA involved). Representative SOTA works like [1][2][3] would offer better context for evaluating KEA’s performance. Additionally, for KEA itself, it would be valuable to explore other factors, such as how different values for the switching criterion (hyperparameter $\\delta$) affect the model's switching behavior, how many times were $\\pi^{SAC}$ and $\\pi^B$ activated in training, etc.\n\n4. The writing and expression can be improved, for example: (1) Lines 199-203 and Lines 128-130 are repeated verbatim, and (2) paragraphs like Lines 308-315 and Lines 404-411 could be better presented in table format for clearer understanding.\n\n[1] Devidze, Rati, Parameswaran Kamalaruban, and Adish Singla. \"Exploration-guided reward shaping for reinforcement learning under sparse rewards.\" Advances in Neural Information Processing Systems (2022).\n\n[2] Trott, Alexander, et al. \"Keeping your distance: Solving sparse reward tasks using self-balancing shaped rewards.\" Advances in Neural Information Processing Systems (2019).\n\n[3] Hong, Zhang-Wei, et al. \"Diversity-driven exploration strategy for deep reinforcement learning.\" Advances in neural information processing systems (2018)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In Fig. 1, I didn't understand why \"high intrinsic rewards can cause premature exploitation of novel regions\". Could you further elaborate this sentence?\n2. The statement of \"Soft Actor-Critic (SAC) is known to be sensitive to the scale of rewards\" should also be property justified either with a small experiment or some relevant literatures.\n3. How is the co-behavior policy actually trained?\n4. Does high stochasity always help with collecting diverse data from the environment? Methods such as novelD may encourage agent to try new actions to reach new states. This would consequently create more diverse data. But after introducing the high-variance co-behaviour policy, the agent may fall back to random policy and thus, collect repeated transitions.\n5. Could you justify why RND is used for futher analysis not NovelD?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper addresses an important research questions. Applying Off-policy to tasks with sparse rewards is always tricky and existing curiosity-based methods are not working out of the box.\n2. The paper is well motivated via a good visualization. The interplay between policy entropy and state novelty is interesting to look into.\n3. The proposed algorithm is simple to implement with a minimal extra computing budget. The switching mechanism is clearly written.\n4. The ablation study of various UTD ratios is also interesting. \n5. Relevant literature is good summarised." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Reinforcement Learning tasks with sparse rewards are always hard to solve. Curiosity-based approaches are derived to facilitate the exploration. However, Soft Actor-Critic (SAC) does not perform well even with curiosity-based exploration. To address this issue, this paper introduces a new exploration strategy KEA (Keeping Exploration Alive), which combines entropy-based exploration and curiosity-based exploration strategies by introducing a simple switching mechanism between two policies with different variance. This mechanism decides which policy to deploy based on the intrinsic rewards.\n\nFor experiments, the authors concluded a simple 2D discrete navigation tasks and analysed the learned policy’s entropy and novelty. Results show that their method can improve performance or learning speed for NovelD and RND. They also tried the methods on a set of sparse continuous control tasks. KEA outperforms baseline algorithms. An ablation study of different UTD (Update-to-Data), KEA is more consistent than other baseline methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The scope of the paper is rather limited as the authors only tackle Soft Actor-Critic (SAC) in their paper. This paper can be extended to other off-policy algorithms as they all share the issues of premature convergence of the exploration and the influence from different UTD ratios.\n2. The experiments are also not very convincing. First, the proposed 2D navigation task is too simple and the performance improvement is also not significant. Second, NovelD and RND both are used as baseline algorithms and NovelD has achieved better performance compared to RND. Yet, the latter experiments are all about RND. Thirdly, in the continuous tasks, the three tasks are not trained for enough time. The performance imporvement is also limited.\n3. The presentation/clarity of the paper should also be improved. For example, I didn't fully get Fig. 1's layout in the beginning. Some sentences are also written in a vague way without further explaination (See questions section). The training of the co-behavior policy is not mathematically defined." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "KEA improves exploration in sparse reward environments by proactively coordinating exploration strategies when combining SAC with curiosity-based methods, maintaining exploration-exploitation balance, and substantially improving learning efficiency." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024kea,\ntitle={{KEA}: Keeping Exploration Alive by Proactively Coordinating Exploration Strategies in Curiosity-driven Exploration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zG2vcC1l1f},\nnote={under review}\n}" }, "abstract": { "value": "In continuous control tasks, Soft Actor-Critic (SAC) has achieved notable success by balancing exploration and exploitation. However, SAC struggles in sparse reward environments, where infrequent rewards hinder efficient exploration. While curiosity-driven exploration methods help address this issue by encouraging the agent to explore novel states, they introduce challenges, such as the difficulty of setting an optimal reward scale and managing the interaction between curiosity-based exploration and SAC’s stochastic policy. These complexities often lead to inefficient exploration or premature convergence and make balancing exploration-exploitation challenging. In this paper, we propose KEA (Keeping Exploration Alive) to tackle the inefficiencies in balancing the exploration-exploitation trade-off when combining SAC with curiosity-based methods. KEA introduces an additional co-behavior agent that works alongside SAC and a switching mechanism to facilitate proactive coordination between exploration strategies from the co-behavior agent and the SAC agent with curiosity-based exploration. This coordination allows the agent to maintain stochasticity in high-novelty regions, preventing premature convergence and enhancing exploration efficiency. We first analyze the difficulty of balancing exploration-exploitation when combining SAC with curiosity-based methods in a 2D grid environment. We then evaluate KEA on sparse reward control tasks from the DeepMind Control Suite and compare against two state-of-the-art curiosity-based exploration baselines — Random Network Distillation (RND) and NovelD. KEA improves episodic rewards by up to 119% over RND and 28% over NovelD, significantly improving learning efficiency and robustness in sparse reward environments." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Reinforcement Learning", "Curiosity-based Exploration", "Sparse Reward", "Soft Actor-Critic" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/b1b6117e136b5d1e4be59749d158a39856a55871.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "KEA: Keeping Exploration Alive by Proactively Coordinating Exploration Strategies in Curiosity-driven Exploration" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Significance of position-weighted mean pooling: How would the model perform with these designs? Would the proposed position-weighted mean pooling produce significantly better results?\n2. Effects of concatenated multiple images: Would this affect VLM performance, particularly when handling a larger number of retrieved images (e.g., more than five)?\n3. Efficiency of VLM-based retrieval: Would using the VLM significantly increase retrieval time? How would the number of retrieved images affect retrieval time?\n4. Clarity on accuracy measurement: How is accuracy measured? Is an additional LLM used to evaluate the textual responses?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. [Originality] The proposed method is promising. With the VLMs showing increasing capabilities in understanding images, encoding images with VLMs is a natural improvement from traditional text-based retrieval methods.\n2. [Quality] The paper presents relatively comprehensive experiments, demonstrating advantages in both retrieval and generation capabilities of the proposed methods.\n3. [Significance] The proposed method clearly outperforms various baselines.\n4. [Clarity] The paper is generally well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes VisRAG, a multimodal retrieval-augmented generation pipeline. Different from traditional RAG methods, which rely on text parsing to retrieve information from visual content, this paper utilizes a VLM-based retriever and generator to navigate information relevant to the query and generate responses. The VisRAG design consists of two main stages: (1) Retrieval: Given a query, the VLM retrieves a set of relevant images from the dataset leveraging the cosine similarity between the query embedding and image embeddings; (2) Generation: the VLM uses a combination of retrieved images and the query to produce responses. VisRAG is compared with a series of baselines and show advantages in both retrieval and generation capabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major concerns:\n1. In the paper, when retrieving relevant images, the authors adopt position-weighted mean pooling over the last-layer VLM hidden states. This choice seems heuristic-based. Although the paper points out that VLMs are with causal attention, potentially emphasizing weights on later tokens, after multiple VLM layers, we could expect information to propagate across most tokens. In fact, previous works often use the leading token or simply learn a linear layer from all tokens. How would the model perform with these designs? Would the proposed position-weighted mean pooling produce significantly better results?\n2. When using multiple images to generate responses, one method proposed in the paper is image concatenation, which is performed horizontally. This could change the resolution of the final image. Would this affect VLM performance, particularly when handling a larger number of retrieved images (e.g., more than five)?\n3. The paper does not report retrieval efficiency compared to text-based methods. Would using the VLM significantly increase retrieval time? How would the number of retrieved images affect retrieval time?\n\nMinor comments:\n1. When measuring the performance of generation, how is accuracy measured? To me, this is particularly unclear for synthetic data, where ground truth answers and queries are generated by GPT-4. Is an additional LLM used to evaluate the textual responses?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Improve text-only RAG by establishing a vision-language model (VLM)-based RAG pipeline.\n2. The experiment is thorough under the designed scenario.\n3. The authors have promised to make the data and code open source." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces VisRAG to address the limitations of text-only RAG systems by incorporating vision-language models (VLMs). VisRAG enables the processing of documents as images. By training the VisRAG retriever with open-source and synthetic data and exploring various generation methods, experimental results show VisRAG's superiority in both retrieval and generation tasks. The authors also plan to make their code and data publicly available." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of novelty: For retrieval, it uses the dual-encoder-style and just changes them to VLM embeddings. The position-weighted mean pooling is also off-the-shelf. For generation, they only try a few straightforward strategies.\n2. Lack of design for document understanding: There is a lack of design considerations specifically for document understanding. The proposed method is not tailored for document understanding scenarios. Alternatively, why not validate the method on non-document (common image) datasets?\n3. Confusion regarding the application: Why does a person always need the assistance of document images when asking general questions? For example, in Line 884, Why would a person directly ask a question like 'Where were the two magnesium containing papers made at?' without reference to any document? If he already has this document as a reference, why does he still need to do a retrieval process at first?\n4. Why only experiment with text-only or image-only features, not image+text(+layout)?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. In the comparison process, Colpali and Vis RAG-Ret should be fine-tuned based on the same data, rather than using the pre-trained model to compare directly. Please add more comparing results.\n2.The article needs to provide more comparison results with the current best models, such as GPT o1, Qwen and other leading models that can directly input image for VQA task." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "In this paper, the authors propose an image-based RAG method named VisRAG, and construct a new dataset to verify the relevant effects. This paper does a lot of comparative experiments to verify its innovation. A large number of state-of-the-art methods have been combined and proven to be effective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focus on the research of vision-language model (VLM)-based RAG\nPipeline and propose a novel method named VisRAG. It broadens the external data processed by RAG from text to image. To achieve the goal, the authors introduce two vision-optimized modules: VisRAG-Ret and VisRAG-Gen. This VisRAG-Ret directly utilizes the images without extracted textual content and establish the map between the query and the documents in the latent space. The VisRAG-Gen leverage the help of the existing LLMs and VLMs for generating the answers. The image processing technology is integrated to realize the efficient generation of multiple inputs. Moreover, a new dataset is built combining a vision question answering (VQA) dataset and synthetic data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The method in this paper is mainly based on the retrieval of image data, which is basically consistent with the text retrieval framework, and lacks innovation. In addition, similar to the comparative model Colpali's innovation, the VisRAG is fine-tuned with a new dataset and new VLM which is not sufficient to express its innovation in multimodality area." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See \"weaknesses\"" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper provides a clear and detailed description of the VisRAG framework., which makes it easy to understand.\n2. The experimental results on document-related VQA tasks are promising, suggesting that the proposed method could be a practical solution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes VisRAG, a model that leverages vision-language models (VLMs) for retrieval-augmented generation in multi-modal documents. The framework comprises two components: VisRAG-Ret for retrieval and VisRAG-Gen for generation, utilizing a data construction method that combines visual question answering (VQA) datasets with synthetic data. Experiments demonstrate that VisRAG outperforms traditional RAG in both retrieval and generation tasks, exhibiting improved training data efficiency and generalization capabilities. It achieves a significant relative increase in end-to-end accuracy, highlighting its potential to replace text-based RAG for multi-modal documents." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The rationale behind the proposed VisRAG-Ret is unclear. If no OCR information is included, how can the target image be retrieved, especially for answers that require precise text content? From my perspective, a pooling-based representation feature does not offer sufficient information.\n\n2. The experimental setup is not sufficiently convincing, indicating that the VQA task alone cannot fully demonstrate the method's effectiveness in the retrieval task. In other words, employing such a complex pipeline to address the VQA task is akin to using a sledgehammer to crack a nut.\n\n3. The methods discussed are insufficient. To the best of my knowledge, there are several document-specific MLLMs, such as mlug-docowl, ureader, and text monkey, among others. While these works may not be directly related to RAG, they should at least be mentioned and analysized.\n\n4. In my view, the comparison is unfair, as both Minicpm and Siglip can not only perform VQA tasks but also handle other tasks related to document image understanding and content perception." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024visrag,\ntitle={Vis{RAG}: Vision-based Retrieval-augmented Generation on Multi-modality Documents},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zG459X3Xge},\nnote={under review}\n}" }, "abstract": { "value": "Retrieval-augmented generation (RAG) is an effective technique that enables large language models (LLMs) to utilize external knowledge sources for generation. However, current RAG systems are solely based on text, rendering it impossible to utilize vision information like layout and images that play crucial roles in real-world multi-modality documents. In this paper, we introduce VisRAG, which tackles this issue by establishing a vision-language model (VLM)-based RAG pipeline. In this pipeline, instead of first parsing the document to obtain text, the document is directly embedded using a VLM as an image and then retrieved to enhance the generation of a VLM. Compared to traditional text-based RAG, VisRAG maximizes the retention and utilization of the data information in the original documents, eliminating the information loss introduced during the parsing process. We collect both open-source and synthetic data to train the retriever in VisRAG and explore a variety of generation methods. Experiments demonstrate that VisRAG outperforms traditional RAG in both the retrieval and generation stages, achieving a 25–39% end-to-end performance gain over traditional textbased RAG pipeline. Further analysis reveals that VisRAG is effective in utilizing training data and demonstrates strong generalization capability, positioning it as a promising solution for RAG on multi-modality documents. Our code and data will be made publicly available." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Retrieval-augmented Generation", "Vision-language Models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/c07a4d6b9ca632b81e1e9e93876976a97cdc1c2a.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. in Table 1: the baselines with max token 512 generate very high hallucination (>30%). Is it aligned with user's experience? At least I feel Gemini and ChatGPT's numbers should be much better than that.\n\n2. I feel the proposed method can be used for any LLM based method. Why does the title claim it only for multimodal LLMs?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Simple method which is very easy to use\n2. Good insights from eigenspectrum variance\n3. Constantly improve the decoding methods of InstructBLIP, MiniGPT-4, LLaVA-1.5, and Shikra." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the role of anchor tokens in causing hallucinations in LLMs, and proposed Dynamic Token Propagation Mechanism (TAME) based on eigenspectrum variance of the attention weight. The results show that TAM can be integrated with existing decoding method including Beam/VCD/ICD/SID and improve the performance consistently on InstructBLIP, MiniGPT-4, LLaVA-1.5, and Shikra." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I hope the paper could explore additional approaches to reduce hallucinations, such as Retrieval-Augmented Generation (RAG), Reinforcement Learning from Human Feedback (RLHF), improvements in factuality metrics, and high-quality, data-based instruction tuning.\n\n2. The paper currently uses the a few simple eval sets like MS COCO. I would suggest incorporating more challenging benchmarks for a more rigorous assessment.\n\n3. The writing could benefit from refinement. For instance, many citations are not in the correct format.\n\n4. The paper only reports metrics for hallucination. But it is not clear the new predictions do better or worse in other metrics. Sometimes less hallucination may lead to less details or not as friendly to users to read." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1.\tThis paper provides a solid theoretical framework by analyzing the relationship between LVLM hallucinations and attention patterns.\n2.\tThis paper introduces a novel plug-and-play decoding strategy that dynamically adjusts the eigenspectrum variance of attention weights to mitigate hallucinations without adding inference time.\n3.\tExtensive experiments show that TAME improves hallucination metrics across multiple MLLMs.\n4.\tTAME can be integrated into various decoding strategies without requiring additional training or data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the causes of hallucinations in LVLMs through the attention weight matrix of anchor tokens. Then this paper demonstrates the propagation pattern of anchor tokens by the eigenspectrum of the attention weight matrix. The authors further propose a versatile plug-and-play decoding strategy named Dynamic Token Propagation Mechanism (TAME) to reduce the over-propagation of anchor tokens through dynamically intervening in the eigenspectrum variance. Extensive experiments show that TAME improves hallucination metrics across multiple MLLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tAlthough this paper provides an additional theoretical framework, the impact of anchor tokens on the LVLM hallucination was first proposed by OPERA [1].\n2.\tTAME primarily addresses object hallucinations but may not be as effective for other hallucination types, such as incorrect attributes or relations (e.g., limited performance improvement on the MME benchmark). Experiments on more types of hallucination should be conducted.\n3.\tIt would be interesting to study applying TAME to some layers rather than all layers through experiments.\n\n[1] OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The motivations are strong, with in-depth analysis and derivation of the mechanisms behind hallucinations in MLLMs.\n2. The proposed method, TAME, is a simple and effective plug-and-play decoding strategy.\n3. The experiments are thorough, validating both the causes of hallucinations mentioned in the paper and the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "MLLMs encounter hallucination issues, where the generated text does not match the provided visual content. Previous methods have attempted to mitigate hallucinations by designing special decoding strategies, such as penalizing summary tokens, but they lack an analysis of the relationship between the model’s hallucinations and the summary mechanism. In this paper, a study and analysis of the causes of hallucinations in MLLMs are presented. A general decoding strategy, TAME, is proposed to reduce the excessive propagation of anchor tokens by dynamically intervening in the variance of feature spectra. Experiments demonstrate the correlation between feature spectra and hallucinations in MLLMs, as well as the effectiveness of TAME." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The introduction mentions summary tokens but does not define them, leaving unclear the distinction between summary tokens and anchor tokens.\n2. There are some typos in the text that need to be checked:\n 1. Line 301: \"Conversly\" -> \"Conversely\"\n 2. Line 333: \"7,3\" -> \"7.3\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In the qualitative results, it appears that the proposed method tends to make the model’s output shorter, which is expected, as the goal is to reduce hallucinated content. However, from another perspective, could this lead to insufficient perception of the image by the model, potentially missing some details? If so, how could we measure and balance this trade-off?\n\n2. The example in Figure 6 is somewhat unclear, as it’s difficult for reviewers to determine which parts of the image or which specific patches correspond to the enhanced visual tokens. Could the authors provide a visualization of the relevant visual tokens mapped onto the image?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The paper is well-written and clear motivated. The authors begin by introducing hallucinations and related work, gradually diving deeper by focusing on the phenomena of token information flow and information aggregation, ultimately leading to the proposal of TAME. \n\n2. This paper is built on a detailed and clear theoretical foundation. Although the proposed method appears simple, the step-by-step theoretical derivation strongly supports it. \n\n3. The proposed method is simple-yet-effective, which obviously surpasses all baselines on CHAIR, POPE and GPT-4 assisted hallucination evaluation. Meanwhile, the method introduces no additional computation overhead, which is superior to existing contrastive decoding based methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new decoding method to address hallucination issues in multimodal large language models (MLLMs). The authors identify that excessive propagation of “anchor tokens”, influenced by polarized variance in the attention mechanism’s eigenspectrum, contributes to hallucinations by sidelining crucial visual information. To mitigate this, the paper introduces the Dynamic Token Propagation Mechanism (TAME), a plug-and-play decoding strategy that adjusts the eigenspectrum variance in the attention weights, limiting over-propagation without adding to inference time. Extensive experiments demonstrate that TAME effectively reduces hallucinations across multiple MLLMs and achieves great improvements over existing methods like Greedy Decoding and Visual Contrastive Decoding (VCD). The method is simple-yet-effective and easy to implement." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some of the color schemes in the figures, such as Figures 4 and 5, are prone to causing confusion. It is recommended to use colors with higher contrast and more distinct differentiation.\n\n2. The paper begins its analysis with the phenomena of token information flow and token information aggregation, gradually developing a method to mitigate hallucinations. However, it does not yet discuss the underlying causes or triggers of these phenomena, only relying on empirical analysis. Nevertheless, the analytical approach used in the paper is still worth learning from.\n\n3. The paper provides limited explanation of the proposed method. Although the method is simple (just one single formula), a smoother transition from the theoretical derivation is needed, along with a clear description of its application scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024intervening,\ntitle={Intervening Anchor Token: Decoding Strategy in Alleviating Hallucinations for {MLLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zGb4WgCW5i},\nnote={under review}\n}" }, "abstract": { "value": "Multimodal large language models (MLLMs) offer a powerful mechanism for interpreting visual information. However, they often suffer from hallucinations, which impede the real-world usage of these models. Existing methods attempt to alleviate this issue by designing special decoding strategies that penalize the summary tokens. However, these methods lack analysis of the relationship between hallucination and summarization mechanism of LLMs. Interestingly, we find that penalizing summary tokens is not necessary: merely intervening the query-key parameters variance, without costing extra inference time, still alleviates hallucinations. Specifically, we explore the causes of hallucinations by analyzing localized self-attention patterns called ``anchor\" tokens and define the attention localization degree of the model as token propagation probabilities. Our analysis reveals that over-propagation of anchor tokens occurs when the distribution of eigenvalues of the query and key matrices has a non-zero mean and a polarized variance, leading to excessive dependence on anchor tokens while neglecting vision information and describes the image content with hallucination. Based on the observation, we propose a versatile plug-and-play decoding strategy, Dynamic Token Propagation Mechanism (TAME), to alleviate excessive propagation by dynamically intervening the eigenspectrum variance of the attention weight, thereby alleviating hallucinations without relying on complex decoding strategies. Extensive experiments reveal a correlation between the eigenspectrum and hallucinations across various MLLMs, and show that TAME reduces the percentage of hallucinated objects." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multimodal large language models; Hallucination" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/7262be39349b4cb875590a81dab3fe80a0578c09.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Intervening Anchor Token: Decoding Strategy in Alleviating Hallucinations for MLLMs" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weakness above. In addition, \n\n1. Do authors plan to open-source the implementation?\n\n2. There is also a line of work similar to token healing and seems working well in both code and text. Could the authors study it and see whether the token alignment method could mitigate the issue especially in single-model completion tasks? \nToken Alignment via Character Matching for Subword Completion https://aclanthology.org/2024.findings-acl.929.pdf" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The authors attempted to address a critical problem on tokenization for LLMs. They are the first to introduce tokenization bias and its effects on model behavior. They present a theory of it and introeduced efficient implementation of the algorithm.\n\nThe authors also studied the practical impact of the proposed solution through both code generation task (FIM, speficially) and model emsambles. Results indicate that this training-free approach is working well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper reveals a fundamental issue in language models regarding tokenization bias - where tokenized and byte-level models, despite being statistically equivalent, produce different predictive distributions. The authors introduce the Byte-Token Representation Lemma, providing a framework to convert tokenized language models into statistically equivalent token-free ones without additional training. They demonstrate practical applications in fill-in-the-middle tasks and model ensembles, achieving significant improvements: 18% on FIM coding benchmarks and up to 3.7% on ensemble tasks. The paper provides both theoretical foundations and practical implementations through efficient algorithms for cover encoding search and next-byte sampling." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the current results are great, it'd be better if authors can also cover tasks other than FIM that suffer from this issue, e.g. complete a story in natural language or generic code completion task that doesn't require FIM.\n\nThe authors discussed computational overhead but doesn't thoroughly analyze the latency impact especially in a practical setting. It'd be better if authors can study these as part of the experiments authors conducted.\n\nThe interaction between the proposed byte-level prediction method and popular sampling techniques (like top-p, top-k) have not been explored enough, especially in the ensemble context." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* How is the proposed method compatible with other tokenization algorithms besides BPE and MPE?\n* What do the right two sub-figures mean in Figure 6?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* This paper defines the tokenization bias phenomenon, which could cause performance issues in practical scenarios. \n* They evaluated the proposed method on FIM code completion task and mode ensemble task. \n* The proposed byte-level prediction method considers both the accuracy and efficiency in recovering the unbiased byte-level distribution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the impact of tokenization on the performance of language models, particularly highlighting a phenomenon called \"tokenization bias.\" This bias comes from the discrepancies between the predictive distributions of tokenized LMs and their statistically equivalent byte-level distribution when predicting the next byte or token. In this paper, they introduce the Byte-Token Representation Lemma (BTR Lemma) to map between the token and byte domain, which allows the computation of exact byte-level probabilities without requiring extra modification of the model. This method identifies \"cover encodings\" for a given byte string which are all possible valid tokenizations of a string that \"optimally\" contain the string in the prompt. Experiments on FIM code completion task and model ensemble task show that the proposed method improves model performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* This method is very similar to the ACL 2024 paper 'Token Alignment via Character Matching for Subword Completion' which addresses the same issue and proposes more evaluation scenarios. However, in this paper, they didn’t compare with the token alignment method. \n\n* Based on the results shown in Figure 6, the improvement of the proposed method on the FIM evaluation tasks in sampling settings (pass@10) over the token-level with token healing method is very minor. For example, from 84.0 to 84.1. Why is that? \n\n* The evaluation scenarios are limited. Only two types of tasks are covered. How’s the performance of the method on general completion tasks without the FIM setting? Does the tokenization bias also have an impact on natural language completion tasks?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I am curious what do authors think about the generalizability of their methods to non-deterministic tokenization or in applications that involve mixed token vocabularies ? Could some experimental results be shown to further strengthen the paper ? \n\nHow do authors explain the not-so-strong gains with ensebling (Table 1 and Table 2). It was not clear how the models are being ensembled together? I would have also liked to see other combination strategies as a baseline e.g. consensus voting or ranking over diverse samples from different LMs. Are these experimental results available ? \n\nIn order to truly appreciate the need for this research, its important to know just how much diverseity there exists in tokenization vocabulary across different LLMs today. To that end, I would like to see some sort of distribution characterizing the popularity of tokenization and distribution of vocabulary sizes across various LLMs. If the world is already coverging to one tokenization and the vocabulary of tokens homogenizes over time, then what practical advantages does this paper bring ? We should debate whether or not this is true and I would like to see authors providing their rebuttal." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "What I really liked about this paper is that it very systematically investigates the phenomenon of tokenization bias. It develops a byte level prediction method and overcomes the bias. The applicability of this method and the findings in the paper have implications in areas such as coding where fill in the middle is crucial as well as in model ensembling. The idea of combining ensembles of foundation models at the logit layer is very appealing. Back in the day when there were no tokens and LMs were being built on whole words, ensembling using to be the norm in any practical LM based applications, such as speech recognition. With varying tokenization powering different LLMs today, the ability to freely combine any LLM with each other became restrictive. This paper proposes byte level efficient sampling algorithms and in conjuction with the byte-token representation lemma and the associated theoretical foundation, now allows the community to confidently combine predictions from any LLMs. Tokenization allows the LMs to overcome closed vocabulary limitation and from my perspective, this paper further allows us to overcome distinct vocabulary challenges associated with model ensembling." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides a thorough investigation into the phenomenon of tokenization bias in language models, proposing a byte-level prediction method to mitigate this bias. The approach has promising implications, especially for tasks like fill-in-the-middle (FIM) in coding and for model ensembling. The idea of combining language models with distinct tokenizations at the logit level recalls older ensemble methods, commonly used before tokenization became mainstream in modern LLMs. The proposed method, underpinned by the Byte-Token Representation Lemma and efficient sampling algorithms, allows for effective ensemble models by bypassing tokenization restrictions, facilitating seamless integration across diverse language models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the results are promising for fill in the middle (in coding) tasks, I would have liked to see experimentation on other tasks to ascertain this method's generalizability to a wider variety of tasks, espically those that require context senstitive token predictions. Computer use, a new usecase demonstrated by companies such as Anthropic, could be a good use case that the authors could consider as an example. \n\nAn area that authors could delve further into is more analysis on the tokenization bias in multi step or long context tasks. \n\nIn order to truly appreciate the need for this research, its important to know just how much diversity there exists in tokenization vocabulary across different LLMs today." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In the caption for Figure 2, should the probability of sampling “A” on the LHS be $(1 - \\alpha)$? It states the probability is $\\alpha$.\n\nI would be willing to change my opinion based on the responses to the weaknesses!" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The characterization of the tokenization bias phenomenon, with examples / and the synthetic example with a 3rd order Markov chain, is very interesting and useful to the community\n- The paper is well-written and studies an important problem" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Tokenization is a common pre-processing step, which shortens input sequences by mapping multiple bytes to discrete tokens. \nPrior work finds: Even for unlimited data and compute, tokenized models can achieve better cross-entropy loss vs. untokenized models\nOne hypothesis for why tokenization helps is that it allows models to handle longer contexts at less compute\nMain problem: there are certain combinations that cannot be represented in token space, limiting the model’s abilities. This is harmful esp. for prompts that end mid-token.\nApproach: this work presents a way to predict when tokenization is getting in the way and converts tokenized LMs into statistically equivalent byte-level models" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The code infilling setting is an interesting example of where this work could be useful. However, it is not clear why the prefix would stop in the middle of a token/word or why new tokens would arise at inference time. Why are the <PRE>, <MID>, and <SUF> tokens needed? Why is the study restricted to two prompting formats/are others possible? Since tokenization schemes are lossless, it seems like any prefix should be representable with the set of available tokens at training time. Further explanation of the evaluation setting and why it is realistic would be helpful.\n- Several prior methods can ensemble the predictions from models that do not share a vocabulary, or ensemble multiple predictions from a single model (https://arxiv.org/abs/2203.11171, https://arxiv.org/abs/2210.02441, https://arxiv.org/abs/2307.11031). It is not clear whether the ensembling method proposed in this paper would outperform these methods . The motivation that byte-representations help ensembling is not fully examined due to the lack of these comparisons.\n- It would be useful to provide decoding efficiency metrics corresponding to each decoding approach in Figure 6 since efficiency is a motivation in Section 4." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We describe an algorithmic process to turn any tokenized LLM into its statistically equivalent byte-level LLM. Applications; FIM task + model ensembles." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024exact,\ntitle={Exact Byte-Level Probabilities from Tokenized Language Models for {FIM}-Tasks and Model Ensembles},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zGej22CBnS},\nnote={under review}\n}" }, "abstract": { "value": "Tokenization is associated with many poorly understood shortcomings in language models (LMs), yet remains an important component for long sequence scaling purposes. This work studies how tokenization impacts model performance by analyzing and comparing the stochastic behavior of tokenized models with their byte-level, or token-free, counterparts. We discover that, even when the two models are statistically equivalent, their predictive distributions over the next byte can be substantially different, a phenomenon we term as ``tokenization bias''. To fully characterize this phenomenon, we introduce the Byte-Token Representation Lemma, a framework that establishes a mapping between the learned token distribution and its equivalent byte-level distribution. From this result, we develop a next-byte sampling algorithm that eliminates tokenization bias without requiring further training or optimization. In other words, this enables zero-shot conversion of tokenized LMs into statistically equivalent token-free ones. We demonstrate its broad applicability with two use cases: fill-in-the-middle (FIM) tasks and model ensembles. In FIM tasks where input prompts may terminate mid-token, leading to out-of-distribution tokenization, our method mitigates performance degradation and achieves an approximately 18\\% improvement in FIM coding benchmarks, consistently outperforming the standard token healing fix. For model ensembles where each model employs a distinct vocabulary, our approach enables seamless integration, resulting in improved performance (up to 3.7\\%) over individual models across various standard baselines in reasoning, knowledge, and coding." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Language models", "Tokenization", "Probability", "Sampling" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/4c8d5ca9ac45632e40720c72bcc911ca29f7b78e.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could you explain the mechanism behind cross-bias generalization, e.g., why does training on sycophancy bias help reduce pattern matching biases, or have you identified common features among biases that show better generalization?\n\nThe paper demonstrates BCT works but doesn't fully explain why. Could you provide analysis of what the model is actually learning during BCT?\n\nRegarding paraphrase requirements, what is the minimum number of paraphrases needed for effective BCT? What is the relationship between number of paraphrases and performance?\n\nIs it possible to experiment with some methods of balancing bias reduction with instruction following?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper addresses an important problem in AI systems - the tendency of language models to generate biased reasoning without acknowledging the influence of biasing features. The empirical results show some promise, with reported reductions in certain types of biases. The experimental setup is generally well-documented, facilitating reproducibility, and the authors are commendably transparent about some of the method's limitations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Bias-Augmented Consistency Training (BCT), presenting a novel unsupervised approach to reducing biased reasoning in large language models. The work reframes the challenge of biased reasoning as a consistency problem, training models to maintain coherent reasoning patterns across prompts with and without biasing features. Through extensive evaluation the authors demonstrate BCT's effectiveness in trained bias and held-out biases." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the presented approach to bias mitigation is interesting, several weaknesses limit its potential contribution. \n\n- The paper lacks a clear theoretical explanation for the effectiveness of BCT, making it difficult to predict its applicability to different biases. While this is understandably a common challenge for LLMs, the resource-intensive nature of BCT, requiring numerous paraphrases for each bias type, raises concerns about its scalability and practicality. Additionally, the paper does not adequately address the potential challenges in generating high-quality paraphrases at scale.\n\n- The inconsistent performance of BCT across different bias types raises questions about its reliability and generalizability. Furthermore, the observed degradation in instruction-following capabilities suggests potential unintended consequences, highlighting the need for a more thorough investigation of potential negative impacts. Perhaps its worthwhile to be more comprehensive in more instruction-following tasks.\n\n- The experimental evaluation is limited in scope, focusing primarily on multiple-choice questions and bias reduction metrics. This narrow focus raises concerns about the generalizability of the findings to open-ended tasks and real-world applications. The paper's claim of successful generalization to unseen biases lacks a convincing analysis of the underlying mechanisms.\n\n- The paper does not adequately position its contribution within the broader context of existing debiasing approaches. While BCT demonstrates some empirical success, its limitations and potential drawbacks raise questions about its overall significance and potential advantages over alternative methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Are there results for other LLMs beyond GPT-3.5T?\n- Can you formalize what the BCT objective is?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Originality: The paper addresses an important problem of doing biased/unbiased reasoning answers to questions that are potentially biased. \n- Quality/Clarity: The paper is clearly written and the structure is simple and clear to follow. \n- Significance: The paper provides some results that show that, under standard bias metrics, the proposed approach could help generate unbiased answers in the context of biased questions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for fine-tuning of large language models by using the concept of biased-consistency training (BCT). The aim is to reduce biased reasoning in chain-of-thought by incorporating a suite testing nine forms of biased reasoning under the (BCT) unsupervised fine-tuning method. The method is tested on question answering tasks run with GPT-3.5T." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- From Figure 2, I understand the main contribution of the paper is the BCT fine-tuning objective. However, there is no comprehensive explanation of what it entails, the assumptions, limitations of the proposed approach. Furthermore, Section 2 which is supposed to be the methodology part delves more around existing approaches, whilst leaving a single small paragraph for the explanation of the method.\n- The experiments are limited to GPT-3.5T. Therefore, there is no demostration of how generalizable the proposed method is to LLMs in general.\n- Beyond Figure 3, there is no other place in the experiments where the results of the proposed method in question answering tasks are clearly explained, with clear examples that show how this proposed method produces unbiased/biased answers." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- What is the BCT training objective? From the description in Section 2.2, is it simply adding both biased and unbiased responses when fine-tuning the model, or the objective actually changed?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The studied problem is important: faithfulness in CoT reasoning is relatively under-explored and it's important to ensure the models to produce unbiased reasoning to users.\n\n- The results on generalization to other types of biased reasoning is very interesting, and demonstrates the usefulness of the proposed method.\n\n- The authors provided fairly comprehensive analysis on models' behavior when trained with BCT, especially the observation where BCT reduces coherent biased reasoning is kind of interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new method called bias-augmented consistency training to improve faithfulness in CoT Reasoning. Specifically, the paper shows that training on certain bias types can help reduce biases in other types of biased reasoning behaviors." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Novelty: overall the proposed method is similar to existing literature on reducing bias in ML/NLP models with data augmentation or counterfactual augmentation, hence the novelty of the work is a bit lacking. The authors should provide more discussion on how this work is different from existing literature.\n\n- For the generalization behavior, currently the main experiments are done on Sycophancy examples and generalization is shown on other bias types. It would be more useful to show if this trend holds or not (and to what extent) regardless of which bias type the model is being trained on. Some of the experiments are provided in Appendix E but the results are not very detailed. In Figure 5, why would multi-bias training perform on-par and sometimes even worse than single-bias? More discussion on this would be useful.\n\nMinor:\n- Many of the important results are included in the appendix, it would be more useful to re-organize the experiment section a bit and show all the major results in the main text. For example, the main results are only shown over one model: gpt-3.5-turbo-0613. The authors provided some additional experiments on LLaMa3-8B in the appendix, but it would be more useful to bring those experiments to the main text to show that the method generalizes across models. Also, is there a study on what model sizes are required for BCT to be effective?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. How does the instruction-following data influence the effectiveness of BCT?\n2. Will BCT affect the model's calibration?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The introduced bias-augmented consistency training is simple yet effective. Its unsupervised nature suggests broad applicability and scalability.\n2. The authors presents strong empirical results showing generalization across bias types and tasks.\n3. The paper is very well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces bias-augmented consistency training (BCT), an unsupervised fine-tuning method to reduce biased reasoning in chain-of-thought (CoT) language model outputs. The authors demonstrate that BCT significantly reduces biased reasoning by training models to give consistent reasoning across prompts with and without biasing features. The key contribution is showing that training on a single bias type generalizes well to reducing other forms of biased reasoning and to new tasks, without requiring labeled data. The authors also validate that BCT minimally impacts model performance on standard tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My main concern with the proposed BCT is its potential negative impact on few-shot performance or, more broadly, in-context learning. While Appendix I shows some evidence for this, I'm not fully convinced, as Figure 9 only includes TruthfulQA results. I recommend including few-shot and few-shot CoT results on additional benchmarks, such as MMLU, to provide a more comprehensive assessment.\n2. Section 5.2 demonstrates that BCT reduces coherent biased reasoning, highlighting a trade-off between CoT faithfulness and sensitivity to biases. This important finding warrants further discussion and analysis.\n3. Some experimental design choices lack clear justification. For instance, the rationale behind measuring task and bias generalization simultaneously, as well as the selection of training data mixture proportions, could be better explained." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce an unsupervised consistency training method that effectively reduces unfaithful biased reasoning in language models, even on held-out forms of bias." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024biasaugmented,\ntitle={Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zGvwENuzPU},\nnote={under review}\n}" }, "abstract": { "value": "While chain-of-thought prompting (CoT) has the potential to improve the explainability of language model reasoning, it can systematically misrepresent the factors influencing models' behavior--for example, rationalizing answers in line with a user's opinion without mentioning this bias. To mitigate this biased reasoning problem, we introduce bias-augmented consistency training (BCT), an unsupervised fine-tuning scheme that trains models to give consistent reasoning across prompts with and without biasing features. We construct a suite testing nine forms of biased reasoning on seven question-answering tasks, and find that applying BCT to GPT-3.5-Turbo with one bias reduces the rate of biased reasoning by 86% on held-out tasks. Moreover, this model generalizes to other forms of bias, reducing biased reasoning on held-out biases by an average of 37%. As BCT generalizes to held-out biases and does not require gold labels, this method may hold promise for reducing biased reasoning from as-of-yet unknown biases and on tasks where ground truth reasoning is unavailable." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Chain-of-thought prompting", "Explainability", "Generalization", "Reasoning", "Bias" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/9ffabdefe6da242b2a44477ade75d4bec4533f7d.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Is the algorithm scalable? It would be valuable to discuss if this approach can handle high-dimensional datasets effectively and whether its complexity scales polynomially or exponentially with the number of variables. \n\n2. Can you discuss the relation between the proposed method with other types of causal discovery algorithms, such as (1) constraints-based algorithms such as PC algorithm etc., and (2) continuous optimization based algorithms such as the one in https://arxiv.org/abs/1803.01422, which provides scalability? \n\n3. What are the effects of small sample sizes on SkewScore’s performance? It would be great if authors can discuss the convergence guarantee of the algorithm in terms of the size of samples. \n\n4. Can the proposed algorithm provide confidence intervals or uncertainty quantification in measuring the skewness or determining causal directionality?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. An example in Section 5 provides a clear, practical illustration of the algorithm, connecting theoretical concepts with application. \n\n2. This paper provides a well-structured and self-contained summary of prior works in causal discovery algorithms based on functional causal models, allowing readers to clearly follow the evolution for learning causal graphs. As a reviewer, this summary enables me to track the advancements in the field and understand the recent progress. \n\n3. The core idea behind the proposed method is simple and intuitive, based on the asymmetry of score skewness to identify causal directions. This approach is conceptually accessible and computationally efficient solutions. \n\n4. The experiments provide strong empirical evidences of the paper’s contributions, with a wide range of scenarios that demonstrate the robustness of the proposed framework in heteroscedastic settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides a novel causal discovery algorithm that leverages skewness to determine causal direction when heteroscedastic noise (HN) is present." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper provides correctness guarantees only for the two-variable case, where skewness in the score function can reliably distinguish causal directions. However, for settings with more than two variables, there is no formal guarantee of correctness for Algorithm 1. \n\n2. Current empirical results seem insufficient in showcasing its performance in high-dimensional scenarios. In other words, it remains unclear whether SkewScore scales effectively as the number of variables grows significantly. \n\n3. While the authors claim robustness in the presence of latent confounders, this robustness appears to be limited to simple bivariate cases. In multivariate settings, SkewScore can still produce incorrect causal graphs when latent confounders are present. For instance, in the causal structure $X_1 \\rightarrow X_2 \\rightarrow X_3$ with a latent confounder $L$ influencing both $X_1$ and $X_3$ ($X_1 \\leftarrow L \\rightarrow X_3$), the method would infer the structure $X_1 \\rightarrow X_2 \\rightarrow X_3$ along with a spurious direct link $X_1 \\rightarrow X_3$. This incorrect edge arises because the latent confounder $L$ induces a conditional dependence between $X_1$ and $X_3$, which SkewScore interprets as a direct causal link. As a result, while the method can handle some latent confounding, it does not offer sufficient mechanisms to detect and mitigate confounding in more complex structures, limiting its application in practical scenarios where latent variables are common." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Is it possible to provide identification results such as 'the skew score is larger under the wrong causal direction' without symmetric noise assumption?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The notation is well-organized and the paper is easy to follow overall." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes a causal discovery method for a class of restricted structural causal models with heteroscedastic symmetric noise. The identification of the causal graph is established using a score that quantifies the skewness of the score function. A causal discovery algorithm is then proposed based on the skewness score." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Fact 2 is incorrect. Let X follow univariate exponential distribution with parameter $\\lambda$, then $E[\\frac{d \\text{log} p(X)} { dx}] = -\\lambda$. Its proof is also incorrect. \n\n2. The proof of Theorem 1 heavily relies on the assumption of symmetric noise. Without this symmetry, skewness no longer provides useful implications for the causal direction, as the proof of Theorem 1 breaks down. \n\n3. Definition 1 needs detailed discussions. As stated, \"this measure captures the asymmetry...,\" it would be helpful to demonstrate that it is zero when the distribution is symmetric." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How is Theorem 1 applied to establish the causal direction between variables $X$ and $Y$ in the experiments? Equations (4) and (5) represent the population mean, while in experiments, we can only obtain the sample mean.\n2. The performance of SkewScore in terms of accuracy, as shown in Figure 3, is not consistently superior to other methods. Specifically, SkewScore outperforms others in the case of GP-sig Student's t in Figure 3(a), and in the cases of GP-sig Gaussian and GP-sig Student's t in Figure 3(b). What are the computation times associated with the different methods?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The introduction of the SkewScore, which is leveraged to identify causal structures without requiring the extraction of exogenous noise.\n2. A theoretical extension of the criterion to multivariate cases, along with empirical validations showing its robustness, particularly in scenarios involving latent confounders." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel approach to uncovering causal relationships in data with heteroscedastic noise, which challenges traditional causal discovery models. The authors propose the use of a skewness-based criterion derived from the score of data distributions to identify causal directions, establishing that this measurement is zero in the causal direction but non-zero in the anti-causal direction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "See questions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "### On Proposition 5\nCould you elaborate on why, “when $\\sigma(x)$ is not constant (i.e., in the heteroscedastic model), $c \\neq 0$ in most cases”? More broadly, could you comment on the condition $c=0$? This characterization could be crucial as it defines a class of distributions $p(u, x)$.\n\nProposition 5 considers $f$ in a finite-dimensional linear space, specifically polynomials. Could neural networks with hidden layers be considered also (since Example 7 doesn’t)?\n\nIt’s interesting that $f$ is required to lie in a “large finite-dimensional space.” I would assume that if $f$ were in an infinite-dimensional functional space, the space should instead be “small.” Could you share your thoughts on this apparent contrast?\n\n### Additional Comments\nCould similar theoretical results be proven for standard skewness?\n\nAssumption 4: Typically, “regularity” refers to the properties of a well-behaved functional space. In Assumption 4, condition 1) is indeed a regularity condition, but it’s unclear if condition 2) qualifies as such. Additionally, could you explain why condition 2) is considered mild?\n\nMinor: There is a duplicate bibliography entry for Mityagin (2015)." }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper provides a neat and effective solution to a critical problem. The connection between score skewness and causal direction is both theoretically grounded and practically effective.\n\nThe extensions to multivariate cases and the potential applicability under hidden confounding are valuable contributions.\n\nThe theoretical analysis appears rigorous (though I haven’t verified every detail).\n\nThe paper is clearly written, with helpful elements like the two properties of the score function, the motivation of score skewness by analogy to standard skewness, and well-designed figures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method to identify causal directions in heteroscedastic symmetric noise models (HSNMs) using the skewness of the score function, which is the gradient of the log density. The skewness considered here is analogous to standard skewness but is distinct, offering computational advantages. Importantly, the paper proves that the score's skewness is zero only in the causal direction. This idea is later generalized to the multivariate case. The proposed method also shows robustness in the presence of hidden confounders, and a theoretical analysis is provided for an \"additive\" confounder in the bivariate case. The method demonstrates strong empirical performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The generality of Assumption 1, which defines the applicable HSNMs, could be explored further, although the paper presents sufficient results for a conference submission.\n\nThe method's applicability under hidden confounding might be limited to special confounding structures." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. Which ER graph model is used in the experiments, $G(d, p)$ or $G(d, m)$? \n2. Can you explain in more detail why $c\\neq0$ in general in Remark 6?\n3. What is the runtime for the order learning and independent tests separately?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The method is novel and can be potentially applied to causal discovery with latent confounders.\n2. The theoretical justification of Assumption 1 seems sound, and the experiments showed that the proposed method works particularly well on sparse HSNMs.\n3. The author extensively evaluated the method when the assumptions are violated, e.g. non-symmetric noise distribution, and latent confounder in the multivariate setting, and proved its effectiveness and robustness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the author proposed a novel causal learning algorithm for heteroscedastic symmetric noise models (HSNMs), where the noise term of a variable $X_i$ can be parameterized by $\\sigma(pa(X_i))N_{X_i}$ with symmetric noise $N_{X_i}$. Specifically, the author showed that the skewness of the score function (SkewScore) in the $X_i$-coordinate is zero if and only if $X_i$ is a leaf node under certain assumptions. Therefore, the topological order can be identified by iteratively removing the variables with the least SkewScore. Conditional independence tests are then applied to determine the edges. Its validity for the HSNM with latent confounder is also discussed. Extensive experiments showed that the proposed method is effective on HSNMs and is robust to assumption violations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The author mainly used ER1 to evaluate the methods on HSNMs without latent confounders, which is relatively sparse.\n2. I think the justification of Equation (6) was a bit missing. Proposition 5 justified that Assumption 1 is mild, and therefore Equation (5) in the bivariate setting holds. However, there is no such explanation for Equation (6). Can you provide some concrete examples where Equation (6) is reduced to some known results, or simplified by specific function and noise classes?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Skewness-Based Criterion for Addressing Heteroscedastic Noise in Causal Discovery},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zGzs5SIwT8},\nnote={under review}\n}" }, "abstract": { "value": "Real-world data often violates the equal-variance assumption (homoscedasticity), making it essential to account for heteroscedastic noise in causal discovery. In this work, we explore heteroscedastic symmetric noise models (HSNMs), where the effect $Y$ is modeled as $Y = f(X) + \\sigma(X)N$, with $X$ as the cause and $N$ as independent noise following a symmetric distribution. We introduce a novel criterion for identifying HSNMs based on the skewness of the score (i.e., the gradient of the log density) of the data distribution. This criterion establishes a computationally tractable measurement that is zero in the causal direction but nonzero in the anticausal direction, enabling the causal direction discovery. We extend this skewness-based criterion to the multivariate setting and propose \\texttt{SkewScore}, an algorithm that handles heteroscedastic noise without requiring the extraction of exogenous noise. We also conduct a case study on the robustness of \\texttt{SkewScore} in a bivariate model with a latent confounder, providing theoretical insights into its performance. Empirical studies further validate the effectiveness of the proposed method." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Causal Discovery", "Heteroscedastic Noise", "Score Matching" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/09f74a4c1ee9476c978d3330a568927cec8ff225.pdf" }, "presentation": null, "primary_area": { "value": "causal reasoning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "A Skewness-Based Criterion for Addressing Heteroscedastic Noise in Causal Discovery" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why retraining is not required needs further explanation and study?\n2. Why was the 9 qubit agent not trained till convergence?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Upto 60% shorter circuit depths are obtained.\n2. Generalized to more than 2 qubits without re-training." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a zero-shot quantum state preparation approach. RL has been used.\nA moving goal post reward function has been designed. Training is performed on less than 0.0001% of the\nstate space and then generalized." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Proof of Prop. 2 of convergence is not completely convincing. Conditions \nunder which it is valid need rigorous statement.\n2. No guarantee that the potential function will be strictly convex. No consideration\nis given to this.\n3. It is not clear why retraining is not required. How general is this is not stated." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- see above.\n\n- HSH gate in the gateset:\n - What is the rationale behind including HSH gate in the gateset $G$ (action space)? Because in principle the agent should learn the symmetry on $S$ w.r.t. to $X$ and $Z$ operations by itself. I believe this inductive bias is introduced during the learning process while the whole point of using RL is not to have any kind of bias.\n\n - Also HSH gate has three single qubit gates. In Table 1, do you count it as one gate for all the methods? Do you use the same gateset for all the methods for a fair comparison?\n\n - Can you provide an ablation study with and without HSH gate to see how the number of gates differ between these two settings for your method?\n\n- Why is there no circuit size numbers for $9$-qubit DRL (local gates) in Table 1?\n\n- In Fig. 3(d), $n=9$, can you please explain the behaviour of the circuit size for depth $t=1, 2$ signficantly larger for your method compared to other two baselines?\n\n- How does the algorithm perform for the QSP task when the target states which are complex and physically relevant but are not classically simulable?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper does a reasonable job introducing the quantum state preparation task and, from the presented difficulties of the domain, manages to derive its adaptations to standard RL employed to tackle this domain. The adaptations presented are sound and straight-forward. The results give a nice guideline for future research in that direction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors employ reinforcement learning (RL) to the task of quantum state preparation (QSP) on physically relevant states. Additionally, their method enables zero-shot preparation for any target state when the system size is fixed. \nNumerical results for systems with up to $9$ qubits show that their method generates circuits up to $60$% shorter than the baseline methods evaluated in the study." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The presentation of the paper is somewhat convoluted and challenging to follow; clearer descriptions would be helpful.\n - The contributions could be better highlighted, such as the introduction of the moving goalpost reward (MGR) for the QSP task.\n - The methodology and results, presented through plots, are introduced quite late in the paper. Providing clearer scenario descriptions of the experiments would improve readability.\n - A clearer contrast with related work would be helpful.\n - Figure legends could be enlarged for improved readability.\n - To aid understanding of the MGR function, a figure illustrating its workings would be valuable.\n\n- The paper appears to misinterpret the QSP definition by assuming access to the target state $|\\psi\\rangle$. Typically, in QSP, one has access only to a succinct (classical) description of the target state, not the circuit that implements it. However, in this approach, the authors start each episode with $|\\psi\\rangle$ as the initial state and then find a circuit to prepare $|\\mathbf{0}\\rangle^{\\otimes n}$ for zero-shot state preparation. This approach may not be physically feasible in a lab; if the circuit for $|\\psi\\rangle$ is known, the rationale for using RL (or any method) to find a circuit that prepares $|\\mathbf{0}\\rangle^{\\otimes n}$ and then inverts it is unclear. It might be argued that the given $|\\psi\\rangle$ circuit is too deep or not NISQ-friendly, necessitating a more compact architecture, though this would still be costly in a real lab setting.\n\n- Insufficient comparison with relevant state-of-the-art methods for the quantum state preparation task:\n - RL methods: [1-4]\n - ML methods: [5,6]\n - SAT-based methods: [7,8]\n\n- The algorithm's performance under common noise models, such as state preparation and measurement (SPAM), bit flip, phase flip, depolarizing noise, and thermal relaxation is not evaluated. These noise sources are standard in quantum computing.\n\n- Several claims in the paper lack theoretical or numerical support. Ablation studies providing numerical evidence would help support these claims:\n - Line 306-308: \"However, the cumulative reward obtained ..... terminating the episode.\"\n - Line 803-806: \"Finally, for the linear-connectivity agents, .... faster with this term.\"\n\n- The paper lacks open-source code and a reproducibility statement.\n\n\n[1] Fosel, Thomas, et al. \"Quantum circuit optimization with deep reinforcement learning\", arXiv:2103.07585 (2021)\\\n[2] Patel, Yash J., et al. \"Curriculum reinforcement learning for quantum architecture search under hardware errors\", ICLR (2024)\\\n[3] Zen, Remmy, et al. \"Quantum Circuit Discovery for Fault-Tolerant Logical State Preparation with Reinforcement Learning\", arXiv:2402.17761 (2024)\\\n[4] Kremer, David, et al. \"Practical and efficient quantum circuit synthesis and transpiling with Reinforcement\nLearning\", arXiv:2405.13196 (2024)\\\n[5] Wang, Hanrui, et al. \"Quantumnas: Noise-adaptive search for robust quantum circuits.\" 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA)\\\n[6] Wu, Wenjie, et al. \"Quantumdarts: differentiable quantum architecture search for variational quantum algorithms\", ICML (2023)\\\n[7] Peham, Tom, et al. \"Depth-Optimal Synthesis of Clifford Circuits with SAT Solvers\", QCE (2023)\\\n[8] Schneider, Sarah, et al. \"A SAT Encoding for Optimal Clifford Circuit Synthesis\", ASPDAC (2023)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The questions are included in Weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors introduce a novel reward function that is inspired by the intrinsic nature of the QSP problem, integrating it into the RL framework to enhance performance in terms of circuit size reduction compared to existing algorithms.\n2. The proposed algorithm demonstrates zero-shot generalization, significantly decreasing computational complexity compared to traditional RL algorithms, which require retraining for each new target state.\n3. The authors provide a theoretical analysis of the generalization capabilities of the proposed algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study presents a zero-shot reinforcement learning (RL) algorithm that incorporates a novel reward function specifically designed for quantum state preparation (QSP). The authors concentrate on stabilizer state preparation tasks involving up to 9 qubits. Notably, the proposed algorithm demonstrates strong generalization capabilities on unseen data, utilizing a training dataset that is 0.001% smaller than the total state space. Furthermore, the quantum circuits generated by this algorithm achieve a circuit size reduction of over 60% compared to existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As someone who is not well-versed in RL algorithms, I found it somewhat challenging to follow the overall algorithms framework. The authors present the general RL algorithms separately before introducing their specific formulation for QSP tasks along with the novel reward function. It would enhance clarity if the authors summarized the algorithm in a table format.\n2. The authors primarily use the total number of gates in the circuit, including both single-qubit and two-qubit gates, as the main evaluation metric for circuit size. However, two-qubit gates are more challenging to implement in practical devices and are more susceptible to noise. It would be beneficial to separately compare the counts of single-qubit and two-qubit gates in the circuits generated by the proposed algorithm against those produced by existing algorithms.\n3. Typos: Line 256: \"a the target\"." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The term \"Zero-shot\" is ambiguous. Typically, \"zero-shot\" refers to training without class labels. In the manuscript, the authors seem to be referring to the unified target trick, which should be clarified.\n\n2. In the introduction, the authors claim that their RL method is applicable to \"arbitrary states.\" However, the states studied are specifically stabilizer states. The authors should either correct this claim or provide results for additional categories of states.\n\n3. How are the experimental results related to the theoretical lower bound of generalization? The authors should consider displaying the success probability achieved during their experiments.\n\n4, What are the definitions of $q$ and $N$ in Lines 517-518? Additionally, the last equation should be numbered for clarity." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The problem studied is highly relevant to the quantum computing community. The manuscript presented a RL algorithm to efficiently prepare stabilizer state, which is valuable for future research.\n\n2. The numerical results is promising compared to conventional algiorthm, even with connectivity constraints, showing the potential utility on near-term quantum computers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript studied the problem of quantum state preparation (QSP). While theoretical lower bounds for the required operations have been established, efficiently approximating quantum states on near-term quantum computers remains an open question. The authors proposed a reinforcement learning (RL) algorithm for QSP, focusing on the preparation of stabilizer states— a set of states of significant practical interest—using Clifford gates. The authors introduced a unified target state by replacing $|\\psi\\rangle$ with $|0\\rangle$ and approximating backwards, referred as \"Zero-shot\" in the manuscript. Additionally, the authors proposed a moving-goalpost reward (MGR) function that aligns the maximum cumulative reward with the highest final fidelity. On the experimental side, the authors conducted experiments with up to 9 qubits, with and without connectivity restrictions, achieving better performance in terms of gate count compared to referenced algorithms. Theoretically, they proved a loose lower bound on the probability of generalization success." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The nanuscript lack discussing about relevant literatures. The \"Zero-shot\" trick is not new; amany existing works have adopted it in reinforcement learning for quantum circuit synthesis and compiling, with QSP being only a subset [1,2]. \n\n[1] Zhang, Yuan-Hang, et al. \"Topological quantum compiling with reinforcement learning.\" Physical Review Letters 125.17 (2020): 170501.\n[2] Qiuhao, Chen, et al. \"Efficient and practical quantum compiler towards multi-qubit systems with deep reinforcement learning.\" Quantum Science and Technology (2024).\n\n2. The claim in the abstract that \"To our knowledge, this is the first work to prepare arbitrary stabilizer states on more than two qubits without re-training\" is questionable. A recent work [3] also proposed a reinforcement learning algorithm for Clifford circuit synthesis.\n\n[3] Kremer, David, et al. \"Practical and efficient quantum circuit synthesis and transpiling with Reinforcement Learning.\" arXiv preprint arXiv:2405.13196 (2024).\n\n3. The authors established a loose lower bound for the probability of generalization success (as indicated in the last equation of the main text). In Table 2, the probability decreases with qubit size, raising concerns about the generalizability of the proposed algorithm. Additionally, in Line 517, the authors stated, \"We did not train the 9-qubit agent to convergence...,\" which raises further concerns about the trainability of the proposed RL algorithm." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the \"Weaknesses\" above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This manuscript introduces a novel novel reward function with provable guarantees, which adds a level of theoretical robustness to the approach.\n- The proposed method's success on stabilizer states, which are crucial for quantum error correction and other quantum information processes, suggests its applicability.\n- The paper demonstrates that their method generates circuits that are up to 60% shorter than those produced by existing algorithms in some cases, which is an improvement in efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript presents a deep reinforcement learning approach for quantum state preparation (QSP) . The authors design a novel reward function with guarantees that significantly scales beyond previous works, allowing for generalization to unseen states. They demonstrate their method on stabilizer states up to nine qubits, achieving state-of-the-art performance by preparing target states with varying degrees of entanglement." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I believe the term \"zero-shot\" in the title of this paper is misleading.Zero-shot learning (ZSL) is a problem setup in deep learning where, at test time, a learner observes samples from classes which were **not** observed during training, and needs to predict the class that they belong to. However, in the context of this paper, training is conducted on a set $\\mathcal S$ of $n$-qubit states. From my perspective, this setup aligns more with traditional supervised learning rather than what is commonly referred to as \"zero-shot\" learning.\n- The authors claim that \"The key idea that allows us to obtain zero-shot inference is the following: a sequence of actions starting in state |ψ⟩ leading to the state |0⟩ yields the inverse of a circuit preparing |ψ⟩ starting from |0⟩\". However, I believe this concept is not original to this paper. For example, a similar idea was proposed in [1]. Yet, the authors have failed to cite it.\n- I doubt the authors overstated their results. It's important to distinguish between the preparation of arbitrary quantum states and the preparation of arbitrary stabilizer states. The latter is a more specialized and arguably less challenging problem compared to the general case of arbitrary state preparation. Stabilizer states, due to their specific properties and structure, may be more amenable to efficient preparation methods, which might not directly translate to the broader and more complex task of preparing any arbitrary quantum state.\n\n[1] Huang, Hsin-Yuan, et al. \"Learning shallow quantum circuits.\" *Proceedings of the 56th Annual ACM Symposium on Theory of Computing*. 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- This max { } function notation is confusing to me, L323 (e.g.) uses max{a : b}, L324 uses max{a | b}? How is this supposed to be read and understood?\n- Why was n=9 picked as the benchmark and not something higher, since stabilizer circuits can be so efficiently computed? L517 mentions that the he 9-qubit agent was not trained to convergence, why is that?\n- How were the observation and action choices implemented in detail?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The ideas of this paper are presented understandable and mostly concise, the format is clean and the structure is logical. The background is explained very extensively. While I think the application of RL for stabilizer states is has fundamental issues, see below, I find the concept of reverse construction for a fixed end-state an interesting idea. I also think the extra effort for the study of entangling entropy and the attempt at a generalization bound, albeit quite loosely estimated, deserve praise and should be more common in empirical QRL papers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work utilizes a reinforcement learning PPO agent to solve the quantum state preparation problem. The key idea involves a potential based reward that is based on goalposts - earlier fidelity penalties are weighted less to encourage exploration and the preparation of partially suboptimal but intermediately necessary states. Training is conducted on stabilizer circuits, which can be completely classically simulated and only involve the H, CNOT and phase gate variants, the placement of which are used as actions. To reduce the set of possible end-target states the RL agent trains to reverse the state preparation, starting from the target state towards the |0> state. After training a one-shot model is evaluated on a study of random target states and a set of states relevant for quantum error correction is shown to yield circuits preparing valid states with fewer gates than analytical models. An analysis on the entangling entropy is given empirically and a upper bound is approximated via Monte-Carlo estimations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Nonetheless, I think this paper has three major issues:\n\n- While fundamental explanations are generally good, the first 5 1/2 pages are simply explanation of fundamental QC / RL basics (excluding Figure 2). The first original section starts at the bottom of page 6, which is not efficient use of space. Much of Chapter 2 could be formulated much more concise or moved to an appendix. In general the information and result density is rather light for a A* publication with 10 pages. On the other hand the actual RL algorithm realization, and the training details should be made much more clear. From what I understand the observation is the whole stabilizer tableau (flattened probably?) and the actions are discrete choices over the set of all applicable gate-actions - or gate-target pairs for two-qubit gates. I am not clear if the addition of g_i in L343 is included somehow, or how often the actions are chosen. Once for each qubit for every layer? Some of the training details like hyperparameters are in the appendix, but are critical information for a self-contained paper and should be moved to the main-paper. (E.g. the training episodes, the max-T curriculum for increasing n, the fidelity relaxation and the fact that the otherwise-case of the reward was not even used as described in the paper.)\n- The goalpost potential-reward as part of the contribution feels quite over-engineered and I’d argue this work is lacking a critical evaluation in the RL context. Similarly, while the reverse generation thought is interesting, without a comparison to the forward-generation training (zero to target) I find it hard to judge any significance. A study of the generation with a simpler fidelity based reward as is found in the literature (even the ones cited here, Kölle et al., Gabor et al., Wu et al.) would have helped, as would a test with a simpler, fixed step-based regularizer instead of the goalpost. A direct comparison of ‘backward’ vs. ‘forward’ generation training would also help support the claims made here.\n- Finally, I think the RL application for stabilizer circuits in particular has fundamental issues. While they are indeed used in QEC, the point and the reason for that is, that they are completely and very efficiently classically simulated. (The abstract of Aaronson & Gottesman mentions thousands of qubits.) In other words, any efficiency gain w.r.t. circuit size, generalization ability or else, could also be classically brute-forced. Without application to circuits with unsimulatable quantum-effects (including rotation gates, e.g.) efficient one-shot generalization is an interesting insight, but lacking overall impact in terms of QC relevancy.\n---\nMinor Issues:\n- L256 Typo: parameters: [a] the target …\n- Inconsistent use of ‘Fig.’ and ‘Figure’ references." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Reinforcement learning framework for zero-shot quantum state preparation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024train,\ntitle={Train once and generalize: Zero-shot quantum state preparation with {RL}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zHeHIIFQVF},\nnote={under review}\n}" }, "abstract": { "value": "Quantum state preparation is an essential cornerstone of quantum information science and quantum algorithms. Notwithstanding worst-case hardness results, designing efficient and scalable methods for approximate state preparation on near-term quantum devices remains a significant challenge. In this work, we present a deep reinforcement learning approach to quantum state preparation that allows for the zero-shot preparation of any state at a fixed system size. We scale significantly beyond previous works by designing a novel reward function with provable guarantees. In our experiments on stabilizer states up to nine qubits, we achieve generalization to unseen states by training on less than $10^{-3}$\\% of the state space. We prepare target states with varying degrees of entanglement content and obtain insight into the quantum dynamics generated by our trained agent. Benchmarking shows our model produces stabilizer circuits up to $60$\\% shorter than existing algorithms, setting a new state of the art. To our knowledge, this is the first work to prepare arbitrary stabilizer states on more than two qubits without re-training." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Quantum State Preparation", "Deep Reinforcement Learning", "Zero-shot Inference", "Off-the-shelf Algorithms", "Generalization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/52b4326078a6c1e06c5148280b1d8e049997093b.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/114d5149e0368650799a76966c9d81f6b95be3c5.zip" }, "title": { "value": "Train once and generalize: Zero-shot quantum state preparation with RL" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- In the preliminatries, 3.1., the formulation is bit awkward; if the function $f_{\\theta_t}$ is a function $\\mathbb{R}^d \\rightarrow \\mathbb{R}$, then the output will be a single scalar, but in the equations the l2-norm calculation is applied … I also checked the reference paper (Xu et al., (2021)), but it is slightly different. \n\n- The overall experiments on Kendall’s tau seems to be not high (Fig.2,Fig.4). (Relatively high, but not significant in absolute terms). Can the authors explain this in further details?\n\n- The cost analysis is done with the GPU search day metric. How is the cost calculated? Under which GPU?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is easy to read through, well presented.\n- The paper proposes relatively strong method for the proxy of searching adversarially robust architectures.\n- The paper includes numerous validations including performance under white-box, black-box attacks, datasets, numerous datasets, and ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper targets a lightweight proxy to assess the adversarial robustness of Neural Architecture Search (NAS) networks. The proposed proxy is represented as the product of two terms (Eq. 4). The first term, based on the intuition that adversarial and natural accuracy are reversely correlated, is introduced to consider adversarial accuracy and is experimentally validated. The second term is proposed through a theoretical approximation using the Neural Tangent Kernel. The proposed proxy is experimentally validated." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. To read through, I felt that the paper has to focus on two points : \n- Is the proposed method superior in terms of cost-efficiency? (zero-cost)\n- Is the proposed method superior in terms of robust accuracy?\n- For now, RQ1 and RQ2 looks a bit similar(line 255-256), before looking at the experimental results.\n\n2. I feel the cost-efficiency related explanation is slightly short. The authors could include detailed explanations for (line 183-185.) such as , the reason the authors chose to iterate samples over generating adversarial samples itself and why is it efficient." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "While the contribution in terms of speed reduction is significant for the proposed method, my concern is the increase in adversarial robustness as compared to the baselines is very minimal in the majority of the cases." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The reduction in the search cost is very significant. \n- The paper is well-written with typos. \n- From Table 3, the transferability results are good. \n- A novel dataset will be released by the authors, which will help advance research in the robust NAS direction. \n- Comparison with the latest state-of-the-art methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a zero-shot proxy that requires no training and depends only on the initial neural network weights to find the robust architecture. Using this zero-shot proxy reduces the neural architecture search speed drastically. Experiments are performed on multiple datasets with varying resolution and number of classes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the search cost is reduced significantly, the increase in the adversarial robustness is minimal (less than 1% in most cases for Table 1). A similar observation can be derived from the results in Table 4. \n- Minor typos: \"8/2550\" instead of \"8/255\" on line 301, \"We\" instead of \"we\" on line 63." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to Weaknesses Section" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ The preliminaries on the Neural Tangent Kernel (NTK) and input loss landscape are clearly presented, making the methodology easy to understand. All symbols and equations are well-defined. \n\n+ A robust theoretical justification for the proposed zero-shot proxy is provided by establishing the corresponding loss as an upper bound for adversarial loss.\n\n+ The efficiency of the search process using their technique is clearly demonstrated in the experimental section.\n\n+ The authors introduce a novel dataset called Tiny-RobustBench, which may prove valuable to the research community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a zero-shot proxy applied directly to the initial weights of a deep neural network (DNN), serving as an upper bound for adversarial loss. This approach enhances the adversarial robustness of the network without requiring adversarial examples. Theoretical justification for the zero-shot proxy is provided based on the Neural Tangent Kernel (NTK) and the input loss landscape. Experimental results demonstrate a 20x speedup over state-of-the-art robust neural architecture search (NAS) methods, with no loss in performance against black-box and white-box attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- In the theoretical analysis section, $\\lambda_{\\min}(\\hat{\\Theta_{\\theta_o}})$ is approximated based on empirical observations to $-\\lambda_{\\min}(\\Theta_{\\theta_o})$, which weakens the theoretical foundation of the work. I wonder if the authors could establish the relationship between the two NTKs in strict mathematical terms, such as showing that the approximation holds true within a specific margin of error with a certain probability. Currently, the analysis relies on an empirically driven approximation, which undermines its theoretical robustness.\n- The performance of the black-box and white-box attacks is demonstrated using a single dataset (CIFAR-10) in Tables 1 and 2. The authors should consider providing performance comparisons across multiple datasets for the chosen baselines in both tables. Additionally, Table 2 should clearly present a comparison of search efficiency.\n- I am curious about the rationale for selecting only specific models (ResNet-18 and PDARTS) for comparison in Table 3. Would it be possible to include models from Table 1 to evaluate their transferability in Table 3?\n- There is inadequate justification for why their approach exhibits better transferability in Table 3. It would be beneficial to provide justification grounded in their theoretical contributions, particularly highlighting the unique advantages that their proposed techniques offer to enhance transferability.\n- The authors should discuss the assumptions made in their analysis and their validity within the context of the study. For instance, when transforming from Equation 1 to Equation 3, it would be helpful to clarify the assumptions involved and how they apply in their setting. Specifically, I am uncertain whether the assumption of infinite-width DNN parameters is valid; if it is not, what implications does that have for this paper? I encourage the authors to include these details, even if they are standard, as doing so would enhance the paper's readability and facilitate the assessment process." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024zerocost,\ntitle={Zero-cost Proxy for Adversarial Robustness Evaluation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zHf7hOfeer},\nnote={under review}\n}" }, "abstract": { "value": "Deep neural networks (DNNs) easily cause security issues due to the lack of adversarial robustness. An emerging research topic for this problem is to design adversarially robust architectures via neural architecture search (NAS), i.e., robust NAS. However, robust NAS needs to train numerous DNNs for robustness estimation, making the search process prohibitively expensive. In this paper, we propose a zero-cost proxy to evaluate the adversarial robustness without training. Specifically, the proposed zero-cost proxy formulates the upper bound of adversarial loss, which can directly reflect the adversarial robustness. The formulation involves only the initialized weights of DNNs, thus the training process is no longer needed. Moreover, we theoretically justify the validity of the proposed proxy based on the theory of neural tangent kernel and input loss landscape. Experimental results show that the proposed zero-cost proxy can bring more than $20\\times$ speedup compared with the state-of-the-art robust NAS methods, while the searched architecture has superior robustness and transferability under white-box and black-box attacks. Furthermore, compared with the state-of-the-art zero-cost proxies, the calculation of the proposed method has the strongest correlation with adversarial robustness. Our source code is available at https://anonymous.4open.science/r/ZCP-05B6." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Neural architecture search", "adversarial robustness", "zero-cost proxy" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/6617f2a0a717046cf4003582bdfec08791e443dc.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Zero-cost Proxy for Adversarial Robustness Evaluation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Is there any harder QA benchmark that involves logic or multi-step reasoning suitable to evaluate the proposed method?\n\n- How do current LLMs w/wo CoT perform in the SimpleLogic dataset?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The proposed method is effective. The experiments show the proposed method trained with rule labels can obtain nearly 100% accuracy on the OOD splits, which largely improves the previous methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on multi-step reasoning tasks that require a model to predict whether a statement is true given unification rules and facts in natural language. The authors first define several rule templates that contain different numbers of unification terms, then train a Transformer and cross-attention module to fill the term slots with entities and their features from a sentence. Lastly, the extracted rule is input to a neural theorem prover to obtain the proof and the truth prediction. The author uses three types of supervision: label supervision, proof supervision, and rule supervision. Experiments show training with rule supervision can obtain 96.7% accuracy on the OOD test split." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The evaluation dataset and the compared method are too simple. Since the paper claims current TLMs have limits in reasoning, it should compare with recent SOTA LLMs and show their incapability in a complicated benchmark.\n\n- The comparison is unfair. The proposed method with proof and label supervision achieves inferior performance to baseline TLM and only obtains nearly 100% accuracy with rule-level supervision. A more appropriate baseline should also have these labels. For example, an LLM trained to generate text in \"fun :- happy kind\" format and predict the label with CoT." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How does the approach work on tasks other than SimpleLogic?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper presents multiple training objectives and studies which impact performance.\n- An efficient and hardware-aware algorithm has been proposed for proof search.\n- The authors show evaluations of the SimpleLogic dataset and generalization capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce Proof Search Augmented Language Models (PSALMs), a differentiable proof search module combined with a transformer. The authors propose an efficient hardware-aware method for proof search (at further depths than prior works) and pruning and performing ablations to identify the strengths of granular rule supervision." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There are limited empirical evaluations other than SimpleLogic.\n- More investigation could be conducted into the scaling of the proof search." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Could you explain the reasons why loss L_{rule} is minimized much smaller than the other two? It seems to me it shows that many entries in matrix T are zeroes.\n\n- As a follow-up question, since the loss terms can be minimized to a different scale, should different weights improve the performance when combining them such as the results in Tables 1 and 2?\n\n- Section 3.1 states that \"Encoding rules independently prevents the TLM from “shortcutting” the NTP,\" could you provide corresponding empirical evidence and quantify the impact?\n\n- Figure 3 shows the range of proof score is different from the case on the left and the case on the right. Could you explain how the larger range for L_{rule} affect its generalization and separation between positive and negative ones?\n\n- Could you provide the key reasons underlying the big differences between the last two rows of Table 1 in terms of OOD accuracy and OOD soundness?\n\n- I assume the vanilla TLM model is the DeBerta model with 435M parameters. However, there are manyother larger LLMs developed. Could you provide a baseline or baselines using state of the art LLMs?\n\n- Could you explain how the proposed method overcomes the exploding computational cost compared to other methods? As far as I could understand, the pruning and parallelization do not change the nature of exponential growth.\n\n- Could you comment on the complexity of the proposed system with respect to the depth? In addition to depths 5-6, could you provide results for depths 7, 10, and 20?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well written and the key components are motivated and explained well. The experimental results seem to be systematic and convincing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a PSALM model that combines a transformer-based language model with a proof search system. Compared to the NTP introduced by Rocktaschel & Riedel in 2017, it uses pruning and parallel execution to improve scaling and throughputs. Using a new loss term, the system provides significant improvements over a vanilla model for (limited) out-of-distrbution generalization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The novelty introduced in the paper seems limited. Pruning is well known; while it improves scaling, in the worst case, the complexity remains the same. Also the improvements over the vanilla model do not reflect the improvements over the state of the art transformer models.\n\nThe experimental results are not fully explained. For example, the row of PSALM L_{rule} in Table 1is not explained and no comments are provided in the paper even though it is a key result." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why are language models such as Llama not used in the experiments?\n2. If language models are used, do we still need the complex procedure to make it work?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This paper proposes a new method to improve the model reasoning abilities by semantics proof search. The experimental results on the SimpleLogic task are nice." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Summary:\n\nThis paper applies an encoder-only model to represent the rule and statement in the proof. Then, the Neural Theorem Prover (NTP) utilizes the representation to perform a backward-chaning search of proofs and sort them.\n\nContributions:\n\nThis paper proposes a new method to improve the model reasoning abilities by semantics proof search. The experimental results on the SimpleLogic task are nice." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiment section is insufficient. First, only one dataset named SimpleLogic is used in the experiment. Second, the author only uses one model in the experiment. Last but not least, there are some work aimed to improve the proof reasoning abilities. However, I have not seen the gap between those work and the method proposed by the authors.\n\n2. The terminology in this article is not to the standard. For instance, we typically refer \"transformer\" to encoder-decoder architecture models instead of encoder-only models." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We add differentiable proof search to transformers to improve generalization across problem distributions" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024proof,\ntitle={Proof Search Augmented Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zIPFFhowuM},\nnote={under review}\n}" }, "abstract": { "value": "Transformer language models (TLMs) exhibit an impressively general range of capabilities. A growing body of work aims to harness these models for complex reasoning problems expressed in natural language. However, recent theoretical and empirical results have revealed limits to the algorithmic generalization of TLM reasoning. Transformers trained to solve deduction problems from one distribution fail to solve instances of the same problem class drawn from other distributions. We propose to improve the systematic reasoning capabilities of TLMs via a differentiable proof search module, yielding proof-search augmented language models (PSALMs).\nIn a PSALM, a Transformer is responsible for predicting rule and statement representations for a neural theorem prover (NTP). The NTP performs a backward-chaining search over proofs, scoring them based on a soft unification operation. Our principal challenge is to train models to reason without also learning spurious features.\nOur results show that rule-level supervision allows PSALMs to successfully generalize across problem distributions in deduction tasks where vanilla transformers fail to learn systematic behavior. We also find we only need label supervision to adapt PSALMs to more natural text." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "reasoning", "transformers", "neural theorem proving", "neural network architectures", "differentiable algorithms" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a40e2dcd942299ee758c8eb9676946dfe2867784.pdf" }, "presentation": null, "primary_area": { "value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Proof Search Augmented Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "A different paper was used to cite GenZProt in Section 5.1, instead of (Yang & Gomez-Bombarelli, 2023)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The model shows good structural accuracy and chemical validity.\nThe usage of latent diffusion for backmapping is new. Discretization of the local geometry for backmapping has been previously explored (Chennakesavalu & Rotskoff, 2024) using rotamer libraries, but the learnable discretization (via VQ-VAE) proposed by this paper is original and reasonable.\n\n(Chennakesavalu & Rotskoff, 2024) Data-efficient generation of protein conformational ensembles with backbone-to-side chain transformers, J. Phys. Chem. B 2024, 128, 9, 2114–2123" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new framework for backmapping, called Latent Diffusion Backmapping (LDB). This method uses a VQ-VAE to discretize local all-atom structures into a codebook latent space. The diffusion model is then trained on this latent space distribution, so that during inference, the latent variables are generated conditioned on the CG structure and then mapped back to all-atom structures via the VQ-VAE decoder." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I think the modeling choices like discretization and latent diffusion make sense and are meaningful, the evidence of their advantages is not convincing, primarily due to the lack of evaluation on diversity or distribution matching between the ground truth and sampled conformations.\nIn GenZProt paper, diversity and the Earth Mover’s Distance (EMD) scores are reported in Appendix A, and there are qualitative analysis of diversity/distribution matching in Figure 8. In DiAMoNDBack paper, the diversity score is reported with the comparison with GenZProt in Table 1 and 2, and there are both quantitative (in JSD) and qualitative analysis of the generated distributions in Figure 5.\nSince the paper is tacking the ‘backmapping’ problem, which aims to reconstruct the fine-grained conformational ensemble conditioned on the CG structure, rather than protein side chain packing (PSCP), which might suffice to find a single probable fine-grained structure, some evaluation on the generated ensemble’s diversity and distribution matching is required.\nMoreover, the paper claims “Diffusion leverages stochastic noise, which allows for exploration across diverse conformations while maintaining structure validity. This is evident in the RMSD and Clash metrics, where our diffusion-based model consistently achieves better results. (line 485-505)“. The structure validity claim is well-supported by RMSD, GED, Clash, Interaction, GDR scores, but there is no evidence for the conformational diversity claim.\nI have a concern that the discretized latent space would be bad for generating diverse structures despite being better at generating chemically valid structures. So it would be great if the authors can show that the discretization does not harm the diversity. If the authors provide diversity and distribution matching evaluations, I am willing to consider increasing the rating." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "* \"Residues with fewer than 13 heavy atoms are excluded.\" Most amino acids have fewer than 13 heavy atoms. Could the authors clarify?\n* How are the internal coordinates defined exactly? Why are there 13 bond lengths, bond angles, and dihedral angles?\n* \"The network processes three inputs: node coordinates, atom types, and an initial noise vector.\" Why are atom types still involved at the latent diffusion stage?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The proposed latent diffusion approach has not been previously developed for the backmapping task. The approach is reasonable and appears to be effective.\n* By encoding the all-atom coordinates into node-level latents in a graph, the dimensionality of the generative modeling task is reduced.\n* The encoding and decoding of the latents from internal coordinates helps ensures that decoded structures are chemically valid." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper develops a model for backmapping coarse-grained protein structure. The model architecture consists of two parts (1) a VQ-VAE is trained to map all-atom structures into a discrete latent associated with each amino acid residue (2) a GNN is trained to generate these discrete latents from the coarse grained structure via diffusion. The method is assessed on PED, ATLAS, and PDB structures and is shown to outperform recent methods GenZProt and DiamondBack." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Originality**\n* Although the approach is reasonable and well-motivated, it amounts to a relatively straightforward application of currently popular ideas. Latent diffusion and discrete tokenization have been well explored in many related contexts, including protein structure. There are not any aspects of the model design that seem particularly surprising or insightful.\n* To improve on this axis, the paper should present nonobvious and/or nontrivial conceptual development that motivates the methodology.\n\n**Quality**\n* The experimental validations are brief and are of average quality. Given that the method is generative, i.e., produces a backmapped distribution, it is surprising that there are no distributional accuracy metrics. Chemical validity is also a rather low bar that can be passed by any heuristic reconstruction + relaxation. Some other metrics introduced by the authors are not well justified or explored.\n* To improve on this axis, the paper should present evaluations that go beyond than existing works in this space. Some suggestions include, e.g., energies, distributional similarities, or metrics used by the PDB in validating atomic models. Since there is a smaller body of ML work on this task, case studies or similar experiments would better make the case that the improvements are meaningful (and not just incremental).\n\n**Clarity**\n* The paper's structure is a little odd, placing a lot of emphasis on architectural and training details but deferring dataset and evaluation details to the appendix. I would recommend swapping this because the task is relatively unfamiliar to the ML audience.\n* The paper contains many overly verbose \"filler phrases\" that repeat similar points and do not add much to the exposition. Some examples of what I mean (there are many, many more):\n\n> Such dynamics lead to distinctive conformational variations, which contribute to the diverse functions\nof proteins and are of great significance for maintaining normal features in vital organisms. (What are \"distinctive variations\", \"normal features\", or \"vital organisms\"?)\n\n> Despite challenges posed by the scarcity and uneven distribution of protein conformation data, we\nemployed a Vector Quantized Variational Autoencoder. (What is an \"uneven distribution\"? Why does scarcity contraindicate a VQ-VAE, and why did you proceed with one anyways?)\n \nThese are just rhetorical questions --- my point is that the authors do not need to say these things if there is not a clear reason to say them. In general my sense is that the introduction could be trimmed by an entire page without loss of meaning.\n\n* To improve on this axis, the paper should be significantly rewritten for concision and clarity, and should focus on explaining things less likely to be familiar to the audience.\n\n**Significance**\n\nIn sum, although the work does technically advance the state of the art on this task, in my opinion the significance of the results and novelty of the method do not quite meet the bar for ICLR." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "NA" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- LDB effectively addresses the challenges of limited exploration in conformational space and accurately reconstructs all-atom 3D protein structures from coarse-grained representations. \n- The use of discrete latent representations simplifies the diffusion process and improves overall efficiency, while also enhancing structural accuracy and chemical fidelity. \n- The evaluation of LDB on multiple diverse protein datasets demonstrates its state-of-the-art performance and versatility, highlighting its potential for practical applications in computational biology." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Latent Diffusion Backmapping (LDB), a novel approach for backmapping coarse-grained protein structures. LDB combines discrete latent encoding with diffusion to address challenges in reconstructing atomistic conformations. Results on multiple datasets demonstrate its state-of-the-art performance and versatility." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The presentation of the method is not very clear. Feel free to correct me if there are any misunderstandings.\n\n- In line 243, an SE(3)-equivariant GNN is used to output bond lengths, angles, etc. Does this indicate that an equivariant network generate an invariant representation? In line 284, the input of ProteinMPNN is CG models. Why does a CG model contain the atom type? \n- In line 286, why do you add noise to the CG model? In my opinion, it serves as a condition into the network, like the BERT in text-to-image models. But in Stable Diffusion, we never add noise to text embeddings.\n- In line 295, how do you parameterize the epsilon-theta? Is it the ProteinMPNN?\n\nMore ablation studies are required. Since a main contribution of this study is latent discrete diffusion, it would be interesting to see what would happen if the diffusion models are trained in the coordinate space and how the vocab size and hidden size affect the result." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "My major questions are above. Additional questions.\n\n* Line 168. What is n_f? What are \"element types of the protein\"? Why are the amino acids real valued instead of integer or categorical?\n\n* Line 177. There are no details with how the internal coordinates are obtained. This is another reason why this work is not reproducible due to the lack of algorithmic details.\n\n* Line 185. The work claims to generate all-atom structures but then this states only the C-alpha coordinates are used. So which is it?\n\n* Line 253. The authors point to Jing et al 2022 to reconstruct full-atom structure but Jing et al 2022 is torsional diffusion for small molecule conformations. How does this relate to the protein structures here?\n\n* Line 273. Where is your evidence of 3D diffusion doing poorly? We know SOTA methods like AlphaFold3 [1] and RFdiffusion [2] work very well in 3D.\n\n* Why was AlphaFlow not compared to?\n\n[1] https://www.nature.com/articles/s41586-024-07487-w\n\n[2] https://www.nature.com/articles/s41586-023-06415-8" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "* Discrete latent diffusion approach is novel to modeling protein conformations.\n* LDB outperforms Genzprot and DiAMoNDBack on most metrics across the three datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Latent Diffusion Backmapping (LDB) is a generative model approach for modeling protein conformations. A discrete latent space is learned over all-atom protein structures with a VQ-VAE followed by a diffusion model to sample protein conformations. The work claims state-of-the-art (SOTA) results on three protein datasets which I do not agree with. LDB does not achieve SOTA on all metrics and the metrics are confusing as to whether they are relevant to evaluating protein ensembles." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have several major concerns.\n\n* After having read the paper twice (including the appendix), I still do not understand the method. The problem definition in section 3 is to sample from p(x|X,V) where x is the coordinates and (X, V) is the coarse grained (CG) representation. Then the method in section 4 proceeds to describe a VQ-VAE for compressing protein structures then a graph latent diffusion model to sample discrete latents. Which part is sampling x and which part is sampling (X, V)? The methods section is poorly written with many details left unanswered.\n\n* The paper claims several times that \"diffusion in high-dimensional spaces complicates the multi-step denoising process\" but provides no evidence of this. In fact, if the latent dimensionality is 36 then isn't this higher dimension than the 3 dimensional coordinates? I would have liked to see a direct comparison or some references but none are provided.\n\n* I do not understand the metrics. If the benchmarks are protein conformation datasets then there should be distributional metrics such as in AlphaFlow [1]. For instance, what does it mean for computing RMSD to PED00055 when it has 20-140? Do you take the min RMSD of the model's samples to the frames? Each metric described in the appendix lacks details on how to calculate them.\n\n* The paper in its current form is not reproducible. There just lacks too much details on the model, metrics, baselines and ablations (how was VAE-difusion and VQ-VAE-flow implemented?). Why was AlphaFlow not compared to?\n\n\n[1] https://arxiv.org/abs/2402.04845" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024the,\ntitle={The Latent Road to Atoms: Backmapping Coarse-grained Protein Structures with Latent Diffusion},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zIqLQVBxdd},\nnote={under review}\n}" }, "abstract": { "value": "Coarse-grained molecular dynamics simulations offer computational efficiency for exploring protein conformational ensembles and thermodynamic properties.\nThough coarse representations enable large-scale simulations across extended temporal and spatial ranges, the sacrifice of atomic-level details limits their utility in tasks such as ligand docking and protein-protein interaction prediction.\nBackmapping, the process of reconstructing all-atom structures from coarse-grained representations, is crucial for recovering these fine details.\nWhile recent machine learning methods have made strides in protein structure generation, challenges persist in reconstructing diverse atomistic conformations that maintain geometric accuracy and chemical validity.\nIn this paper, we present Latent Diffusion Backmapping (LDB), a novel approach leveraging denoising diffusion within latent space to address these challenges. \nBy combining discrete latent encoding with diffusion, LDB bypasses the need for equivariant and internal coordinate manipulation, significantly simplifying the training and sampling processes as well as facilitating better and wider exploration in configuration space. \nWe evaluate LDB’s state-of-the-art performance on three distinct protein datasets, demonstrating its ability to efficiently reconstruct structures with high structural accuracy and chemical validity.\nMoreover, LDB shows exceptional versatility in capturing diverse protein ensembles, highlighting its capability to explore intricate conformational spaces. \nOur results position LDB as a powerful and scalable approach for backmapping, effectively bridging the gap between CG simulations and atomic-level analyses in computational biology." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Protein Structure Reconstruction", "Latent Diffusion", "Discrete Protein Representations" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/b2e524b09b9b83663885af44578dabf01f1b50de.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "The Latent Road to Atoms: Backmapping Coarse-grained Protein Structures with Latent Diffusion" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Please refer to the Weaknesses:\n\nThe most important questions are regarding the hypothesis (W1-3), missing related work (W7), and the originality (W5).\n\n2. Here are additional questions:\n\n1) Isn’t the inference cost for graph-diffusion inference much higher compared to bottom-up or top-down approaches? \n\n2) Which metric is the most important in the evaluation (L317-323)?\n\n3) Could incorrect predictions also propagate through graph diffusion and negatively impact inference?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Hierarchical classification is a significant problem.\n- The inference method using graph diffusion is interesting. \n- Various ablation studies were provided to demonstrate the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a training and inference method to address hierarchical classification, based on the assumption of hierarchical manifolds. First, a graph diffusion-based approach is introduced, utilizing predictions at all levels, which differs from traditional bottom-up/top-down inference methods. Next, hierarchical multi-modal contrastive finetuning is proposed. The approach demonstrated good performance on the iNat18/21 dataset, and various ablation studies were conducted to validate the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Clarity**\n\n1. Figure 1 is difficult to understand. Please provide more explanation. Also, is it a visualization of actual features? If so, how was it created? If it’s intended as an intuitive example, not actual data, what evidence supports this specific representation? \n\n2. While the hypothesis of hierarchical manifolds serves as a key motivation for the proposed method (Lines 165-169 & Figure 1), it is difficult to find strong evidence supporting this hypothesis. Specifically, why do pretrained features lie on a hierarchical manifold? I believe the structure of the feature space is significantly influenced by the training process, including factors such as the choice of loss function, model architecture, and data labeling. These factors can lead to feature spaces that do not strictly conform to hierarchical manifolds. This is why some previous works focus on learning or embedding hierarchical structures within the latent space [1, 2]. Furthermore, hierarchical relationships in real-world data may not always be represented by nested structures. As such, this assumption appears somewhat oversimplified. If the intention was to convey that the data does not inherently lie on a hierarchical manifold but must be trained to conform to one, it would be beneficial to clarify this in the writing.\n\n3. Since the hypothesis regarding hierarchical manifolds is not clear, it is difficult to understand how graph-diffusion inference can be leveraged even in the absence of HMCF (Figure 1(a)). In other words, what is the relationship between graph-diffusion inference and hierarchical manifold? Isn’t it possible to conduct graph diffusion-based inference using the relationships among all labels, even without the discussion about the hierarchical manifold?\n\n4. Some sections of the writing appear unorganized. For example, in Line 84, the abbreviation \"HMCF\" is introduced before it is defined. Additionally, in Line 188,Is “14,036” referring to the number of test samples or the total number of hierarchical labels? Making these points clearer would enhance the conciseness of the writing.\n\n\n\n**Originality** \n\n5. The hierarchical contrastive learning method that utilizes labels at all levels has been proposed previously [3], so its novelty here is somewhat limited. It would be helpful to clarify how the proposed approach differs from [3]. \n\n**Quality** \n\n6. Relatedly, there are many additional experiments in the Appendix that are not referenced in the main text (except for Table 5). If these experiments are relevant and essential to this method, it would be helpful to mention them in the main text so that readers can find them. \n\n7. Some important works seem to be missing from the Related Work section. In addition to top-down and bottom-up approaches, there are hierarchical classification studies that use multi-branch architectures to predict all levels simultaneously based on the complete label hierarchy [4, 5]. A comparison with these approaches would be helpful. These studies aim to address the question raised in Lines 163–164: “Can the predictions across different levels in the category hierarchy mutually reinforce each other to improve overall accuracy?” and are driven by similar motivations.\n\n\n8. As a minor point, citations would be easier to read if placed in parentheses.\n\n[1] HIER: Metric Learning Beyond Class Labels via Hierarchical Regularization, (CVPR, 2023) \n[2] Learning Structured Representations by Embedding Class Hierarchy, (ICLR, 2023) \n[3] Use All The Labels: A Hierarchical Multi-Label Contrastive Learning Framework (CVPR, 2022) \n[4] Your “Flamingo” is My “Bird”: Fine-Grained, or Not (CVPR, 2021) \n[5] Label Relation Graphs Enhanced Hierarchical Residual Network for Hierarchical Multi-Granularity Classification (CVPR, 2022)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The paper uses a generic prompt structure (“a photo of a {class}”) for training the VLM with HMCF. Did you test alternative prompt formulations?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper introduces a novel perspective on hierarchical classification by leveraging manifold learning through a graph diffusion-based inference approach. The combination of hierarchical multi-modal contrastive fine-tuning (HMCF) and graph diffusion is a creative." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the challenge of classifying images within a hierarchical labels, which prioritizes correct coarse labels when fine labels are difficult to predict accurately. The authors argue that standard fine-tuning approaches, which typically optimize models on fine classes using cross-entropy loss, may not fully leverage hierarchical structures. They propose the use of a graph diffusion-based inference strategy. The approach adjusts posterior probabilities across hierarchical levels, which differs from traditional top-down and bottom-up methods by treating the hierarchical taxonomy as a graph." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper could improve by including a more comparison with alternative hierarchical inference strategies beyond traditional top-down and bottom-up methods. \n2. the method depends on vision language models to fully leverage the hierarchical structure.\n3. There is not enough details on the graph diffusion process, and it should give more intuitive explanation of the mechanism." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* What is the difference between differentiable diffusion (DD) and diffusion (D)? Since the explanation for DD is in section 3.2, the diffusion-only case needs to be explained. \n* Based on the results in Tables 1 and 2, DD hardly affects AP and metrics for hierarchical classification, although it improves leaf classification. Why is this the case? \n* Based on the results in Table 3, most of the improvements are due to adding MCF on leaf classes and a minor improvement by adding hierarchy. Can you explain why this is the case? Since there is a discussion in the paper that fine-tuning just on the leaf level is not effective. \n\nSuggestions:\n\n* Adding more discussion to connect the two methods would be great.\n* It would be nice to show detailed performance based on each class to see the method's effectiveness in cases where distinguishing the hierarchical classes is more challenging." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper is well written.\n* The idea of hierarchical classification is exciting and important, especially in the application.\n* The inference strategy based on graph diffusion is novel for this task." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a hierarchical multi-modal contrastive for fine-tuning pre-trained vision language models. The authors claim the method can capture hierarchy more efficiently than top-down and bottom-up approaches. Furthermore, they introduce a graph diffusion algorithm for inference motivated by a hierarchical manifold. In the end, they compare the performance of their inference and fine-tuning strategy across different tasks and datasets with other methods and show the superiority of their method compared to them." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The discussion from lines 171-176 is confusing and not accurate since it is about the importance of incorporating the hierarchical nature of data in training and inference. Yet one of the paper's main contributions is an inference method to improve performance.\n* The paper lacks cohesion. It proposes two methods that are hardly connected. One is an inference method based on graph diffusion, and the other is a fine-tuning method for VLMs based on contrastive loss. \n* The numbers in all Tables don't have confidence intervals, so it is hard to grasp how significant the differences are. The authors should include confidence intervals or standard deviations from multiple runs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- [Methodology] It is not clear how sections 3.2 (inference), and 3.3 (fine-tuning) come together in practice. Are these both applied? Do you first finetune using HMCF, and then apply the graph diffusion? Please elaborate this, perhaps in a pseudo code, algorithmic view. \n- How do you justify this swing of outperformance between D and DD across performance eval tables? Wouldn't it be possible to have one approach (a hybrid) that works best in most/all scenarios? If not, when to use which method? Please elaborate on this." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper reads reasonably well, and the narrative is coherent. \n- The experimental setup in comprehensive. \n- The results look promising compared to the considered baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a combination of a contrastive fine-tuning technique and a graph diffusion inference strategy to improve the downstream performance of hierarchical classification models. This is investigated across two main benchmarks and several settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- [Narrative] Some key points are repeatedly mentioned over and over again throughout the paper (eg. top-down, bottom-up, why they work and don't and why one outperforms the other and even why the proposed approach is better). Overall, the paper can be simplified/shortened to go straight to the point. I understand these are core points to take home, but stating them so many times might not be helpful nor efficient. While on the same topic, several grammatical errors and typos can be found throughout the paper, please make sure you give it another good proof-read. \n- [Relevance] Most baselines and references upon which the problem definition and its significance are built date back to 2 to 5 years ago. Whether hierarchical classification is still such an important problem can be questioned in that light. My suggestion would be to definitely look for more recent literature and baselines that accentuate on the relevance and importance of the problem. \n- [Proof] I'd suggest adding a proof that walks the reader from eq. (2) to (4). It is not that straightforward. \n- [General] I feel using \"graph diffusion\" might be confusing as most recently there is a new branch in literature that focuses on using Diffusion models (in computer vision) over graphs. I'm not sure though how this could be addressed, please give it a thought." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024hierarchical,\ntitle={Hierarchical Classification via Diffusion on Manifolds},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zJbwrk1DHc},\nnote={under review}\n}" }, "abstract": { "value": "Hierarchical classification, the problem of classifying images according to a predefined hierarchical taxonomy, has practical significance owing to the principle of ``making better mistakes'', i.e., better to predict correct coarse labels than incorrect fine labels. Yet, it is insufficiently studied in literature, presumably because simply finetuning a pretrained deep neural network using the cross-entropy loss on leaf classes already leads to good performance w.r.t not only the popular top-1 accuracy but also hierarchical metrics. Despite the empirical effectiveness of finetuning pretrained models, we argue that hierarchical classification could be better addressed by explicitly regularizing finetuning w.r.t the predefined hierarchical taxonomy. Intuitively, with a pretrained model, data lies in hierarchical manifolds in the feature space. Hence, we propose a hierarchical multimodal contrastive finetuning method to leverage taxonomic hierarchy to finetune a pretrained model for better hierarchical classification. Moreover, the hierarchical manifolds motivate a graph diffusion-based method to adjust posteriors at hierarchical levels altogether in inference. This distinguishes our method from the existing ones, including top-down approaches (using coarse-class predictions to adjust fine-class predictions) and bottom-up approaches (processing fine-class predictions towards coarse-label predictions). We validate our method on two large-scale datasets, iNat18 and iNat21. Extensive experiments demonstrate that our method significantly outperforms prior arts w.r.t both top-1 accuracy and established hierarchical metrics, thanks to our new multi-modal hierarchical contrastive training and graph-diffusion-based inference." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "hierarchical classification", "graph diffusion" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/5ce3a44c17a014dbee2eed3e2a62e05ee900cbab.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Hierarchical Classification via Diffusion on Manifolds" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "What is the final choice of the batch size for PPO? Only candidates 1024, 2048, 4096, and 8192 are reported." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Strengths:\n1. Estimating the on-policy gradient requires on-policy samples. Because we do not have access to the true gradient, using the target policy to draw on-policy samples to estimate the true gradient has an estimation error. This error is caused by under-sampled or over-sampled data. Encouraging under-sampled on-policy samples reduces the sampling error and improves gradient estimation. This paper is the first to apply this idea to the PPO algorithm.\n2. This paper proposes an algorithm to learn the behavior policy to encourage data in PPO algorithms." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem that on-policy sampling may have sample error throughout the sampling process. Such sampling errors could be remedied by sampling using a different policy. This paper applied this idea to PPO by remediating the sampling error inside each minibatch. Specifically, this paper divides the minibatch of the PPO into more fine-grained nano batches and adjusts the sampling policy based on sampled nano batches to encourage under-sampled samples. This paper claims their results are better than PPO." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses:\n1. This paper adjusts the batch size of PPO to \"1024,2048,4096,8192\" which is much larger than the original batch size (64 samples) of PPO. This paper makes this change because they need to learn a behavior policy to encourage under-sampled data at each update step. In other words, they run a mini-PPO to encourage under-sampled data by finding a behavior policy. This drastic increase in batch size is not well-justified. In fact, it is confusing that their PPO algorithm still achieves the comparable performance of the original PPO with at least 16 times less update due to the batch size increase. I am not sure if the original PPO paper has such large room to change the batch size to 16 times larger while achieving the same performance under the same hyperparameters.\n2. This paper is not well-written. The length of this paper can be greatly reduced. Currently, this paper has too many unnecessary examples and sentences. The theoretical inside of this paper is also weak. I think a good improvement direction is to discuss the relationship between the sampling error and the number of samples and how this paper can remediate the sampling error from the theoretical perspective. Some new interesting findings might also be discovered during this process.\n3. This paper also introduces many fragile hyperparameters. For example, the learning rate for its proposed behavior policy ranges from 10^{-3} to 10^{-5} for an ADAM optimizer. Such a large difference in learning rate is not common for the ADAM optimizer. Moreover, they also require KL cutoff and regularizer coefficient inside the mini-PPO to encourage under-sampled data, which are two more dimensions of hyperparameters." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. As a follow-up insight, there may be cases where we can improve the sampling error in one state, but this will cause worse sampling errors down the road because we will be forced to go into specific states. Wouldn't it make more sense to push this sampling error into the reward instead, and form a suitable policy? if you think this is not an issue, it would be helpful if you can explain why.\n\n2. Something doesn't sit right with me regarding equation 6 and its motivation. If we are only aiming to improve the sampling, in the discrete case we wouldn't care about the distance from the target policy. If we are aiming to also improve control, then most algorithms prevent policy changing too much to improve stability in which case we don't need this additional term (for example in PPO). So is this specifically for the continuous case? It does not make a lot of sense to me, at least according to the examples for the discrete case. I would be happy if you could elaborate on this point. \n\n3. Why did you chose the form given in Equation 5? why not its log or difference? Some explanation, ideally guided by theory would be helpful (it is the importance sampling term, widely used in off-policy learning, so it should make some sense using it in some variation)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The sampling errors problem the authors tackle is interesting, there has not been sufficient discussion on this topic in the RL community. \n\n2. The presentation and writing is generally clear. \n\n3. The sampling errors problem can arise in many real-world setups. \n\n4. The proposes solution is simple in a good way." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors tackle the problem of sampling errors from insufficient data such that the empirical data has some discrepancy with desired policy. The authors propose a method to adaptively adjust the policy to resolve this discrepancy. They show their method works in several test cases including control and continuous action space." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiments and motivation the authors provided are not accommodating for the problem. When is this really a problem in the real-world? In most standard examples there is ample time to correct the sampling errors with more data and simulation (this happens with a not a lot of samples in practice). Perhaps in cases of non-stationarity of the system itself this \"quicker\" adaptation become crucial? or maybe for very large state\\action-spaces this is more of a problem. \n\nI propose the authors will try to distill where they're problem is really important and design experiments correspondingly, or at least motivate more strongly towards these cases.\n\n2. Even in a quick look, Equation (5) can be rewritten much simpler to be -max(g, 1-eps). I'm not sure forcing it to look like PPO helps with insight, but it is confusing and unnecessary. That is, unless you formulate it similarly to the advantage.\n\n3. Some questions were raised for me regarding the soundness of the proposed solution, see in the questions section." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I have only one question:\n\nIn Zhong et al., the policy is updated at every step in an online fashion: to me, this makes a lot of sense, since we want constantly to \"correct\" the sampling distribution. Ideally, one wants to employ the method you describe in Proposition 1. \n\nHowever, if I understand well, you propose first to use the target policy to collect the samples and, afterward, to generate a policy that \"corrects\" the previous behavior. However, such a policy is not well defined: it depends on how many samples will be drawn from such a policy. For example, suppose that the target policy is the distribution $\\pi(a=0|s) = 0.5$ and $p(a=1|s)=0.5$. Suppose then that the samples drawn from this distribution are:\n\n0, 0, 0, 0, 0.\n\nThe policy that you want to learn should then put much more probability density on 1 to compensate, but if we sample hundreds of samples with this new policy, then the dataset will be again \"unbalanced\". In my understanding, the idea of Zhong et al. to continuously update the policy is to avoid this issue.\n\nWhy did you propose this \"batch update\" rather than an \"online update\"?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality\n-------------\n\nThe idea proposed is novel. Although the method proposed until page 5 is from Zong's paper, and the application to policy gradient estimation seems somewhat trivial, I think it is still valuable and necessary.\n\nQuality\n----------\n\nThe quality of the paper is really good. The authors explain the problem and the main idea very well. The method is sounds and directly addresses the problem presented. The experiments are well-designed, explained, and commented on. The results are equipped with good statistical significance, and the appendix reports the hyperparameters necessary for reproducibility. \n\nClarity\n--------\n\nThe paper is well-written, and the overall idea is exposed very clearly.\n\nSignificance\n-----------------\n\nIt is hard to assess the significance of this paper: 1) it heavily relies on a previous publication, and 2) it does not seem that the method improves so much over the baselines. However, I think the core idea is significant as it can serve as a source of inspiration for other RL problems, as MC is a core technique in RL to estimate many different quantities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the usual problem of the high-sample complexity of on-policy policy gradient algorithms by refining a very interesting idea proposed by Zhong et al. in 2022. In that paper, the authors observe that when the number of samples is scarce, the samples might resemble the on-policy distribution poorly, which would be equivalent, in a sense, to dealing with an off-policy distribution. In more simple terms, a low number of samples usually cause a high variance policy gradient estimation. \n\nZhong et al. observe that on-policy samples do not necessarily need to come from the on-policy distribution; for example, the samples\n\nx_i = 0, 1, 1, 0 \n\ncan come from p(x=1) = 0.5 and p(x=0)=0.5, but also from q(x=1) = 0.4 and q(x=0) = 0.6. \nSimilarly, a batch of samples that look very off-policy might actually have been generated from the on-policy distribution, for example\n\nx_i = 1, 1, 1, 1, \n\nmight have been generated from the policy $p(x=0) = 0.5$, $p(x=1) = 0.5$. \nMonte-Carlo estimation (i.e., $E[X] \\approx 1/n \\sum_i X_i$) has higher variance exactly due to the issue described above, when the number of samples $n$ is low, while the variance disappears for $n\\to\\infty$, making it a consistent estimator.\n\nThe variance is the main problem affecting MC estimators and, thus, policy gradient algorithms as well (which are all MC estimators, as they attempt to estimate a respected value using $1/n \\sum_i X_i$).\n\nThe variance of MC estimators can be ideally reduced if small batches of samples seem to better adhere to the target distribution. This can be achieved by breaking the i.i.d sampling assumption. For example, when the target distribution is $p(x=0) = 0.5$, $p(x=1) = 0.5$, One can choose a deterministic policy that alternates 0s and 1s. Such policy makes the MC estimator strictly more efficient. In this sense, **an off-policy distribution (that violates the i.i.d assumption) can generate better samples \"on-policy\" samples than the target distribution.**\n\nZhong et al. in 2022 propose to learn a policy that aims to lower the variance of the policy gradient estimators by generating samples that adhere better to the target policy distribution.\n\nThe contribution of this ICLR 2025 paper is to use the idea of Zhong et al. for on-policy policy gradient estimation (and policy improvement), while in the original publication, the idea was only proposed for policy evaluation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Contribution\n-----------------\n\nThe paper's contribution starts on page 5: the author focuses on presenting Zhong et al.'s paper there. The core contribution is the application of Zhon's idea for PG estimation and the use of PPO clipping and KL divergence to prevent too aggressive updates. \n\nWhile I think it is necessary that the authors devote that space to expose Zhong et al.'s idea, I think that it might not be clear to the reader that **that is not** the main core idea of the paper. \n\nClarity\n--------\n\nWhile the paper is well written overall, I do not always like the choice of words and how the problem is described. \nI find talking about on-policy \"data\" vs. \"sampling\" confusing. A sample is not per se on or off-policy. When we speak about \"on-policy\" and \"off-policy\" in estimation, we refer to the generating process (i.e., the distribution) that generates the samples. A sample X_i can be generated by an off-policy distribution $\\beta$ and simultaneously very unlikely w.r.t. $\\beta$ and very likely w.r.t. the target distribution $\\pi$, but that does not make the sample on-policy (or off-policy). Policy gradient estimation is simply off-policy when the distribution that is used to generate the samples is off-policy. \n\nPairwise, I find speaking of \"empirical policy\" as in Proposition 1 confusing. One thing is samples, and one thing is the distribution. What can be told, in general, is that if $n > m$, then $Var[1/n \\sum_i^n X_i] < Var[1/m \\sum_i^m X_i]$, where X_i are samples coming from the same distribution. Now, it would be possible to introduce a new distribution (as done in Proposition 1), and show that the variance of the batch of samples decreases even faster in $n$, when compared to the original distribution. \n\nOverall, I am conscious that much of the \"nomenclature\" is borrowed from Zhong et al., thus the authors might have preferred to keep consistency with that. \n\nMinor Comments\n-----------------------\n\nEquation 1 is unclear as it is written: s_0 comes from a different distribution than s_1, which comes from a different distribution than s_2, and so on. Writing $s \\sim d_\\pi$ gives the impression that all states are drawn from the same distribution. Usually, there are two ways to express the policy's return,\n\n1. \n\n$$\nJ(\\theta) = \\mathbb{E}_{s \\sim d^\\gamma_\\pi, a \\sim \\pi}\\left[r(s, a)\\right]\n$$\n\n where $d^\\gamma_\\pi$ is the discounted state-action visitation (see Nota and Thomas \"Is policy gradient a gradient?\" and Sutton et al. \"Policy gradient with function approximation\")\n\n2. \n\n$$\nJ(\\theta) = \\mathbb{E}\\left[\\sum_{t=0}^\\infty \\gamma^t r(s_t, a_t)\\right]\n$$\n\nwhere $s_0 \\sim d_0, a_t \\sim \\pi(\\cdot | s_t), s_{t+1} \\sim p(\\cdot | s_t, a_t)$.\n\nEquation 2 and 4, should be with the discounted state visitation (see Nota and Thomas \"Is policy gradient a gradient?\" and Sutton et al. \"Policy gradient with function approximation\" :D )\n\nReferences\n---------------\n\nNota, C., & Thomas, P. S. (2019). Is the policy gradient a gradient?. arXiv preprint arXiv:1906.07073.\n\nSutton, R. S., McAllester, D., Singh, S., & Mansour, Y. (1999). Policy gradient methods for reinforcement learning with function approximation. Advances in neural information processing systems, 12. (**see page 3, definition of $d^\\pi$**)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "**[Q1]** What do the authors believe are the main benefits of PROPS relative to existing methods that address data or computational efficiency, including: \n- Methods that reuse data by making off-policy corrections (references in lines 122-124): These algorithms appear to be the most similar to PROPS, and some of them introduce similar levels of added complexity / runtime when compared to on-policy algorithms. However, they are not compared against in the experiments.\n- Off-policy RL algorithms: These methods have been shown to be extremely data efficient and can even be used to train complex tasks directly from real-world data (e.g., [1]). They have also been shown to significantly outperform on-policy methods on some of the MuJoCo tasks used in this paper.\n- Massively parallel on-policy implementations: When a simulator is available, it has been shown that massively parallel implementations of on-policy algorithms such as PPO can result in very efficient RL training in terms of wall-clock time (e.g., [2]). \n\n**[Q2] Experiments**\n- The use of different hyperparameter values across each algorithm makes the comparison of PROPS and PPO difficult to interpret. I appreciate the value in the best vs. best comparisons presented in the paper, but how does PROPS perform when using the same hyperparameter values as PPO (and only tunes the PROPS-specific hyperparameters)?\n- It is not obvious to me that reducing sampling error will necessarily lead to improved performance. Given the non-convex objective of policy gradient methods, noisy updates could even be helpful to avoid / escape local optima. Did you perform any experimental analysis on this connection (e.g., comparing PPO for the same number of total updates but varying batch sizes)? A comparison of PPO and PROPS for $b=1$ (with the same PPO hyperparameters) would also provide some insight into this question.\n- It seems that “on-policy sampling” refers to PPO-Buffer in the sampling error figures, which does not seem like a fair comparison because PPO-Buffer does not attempt to correct for the use of off-policy data at all. What is the sampling error of standard on-policy PPO (and how does it compare to PROPS for $b=1$ with the same PPO hyperparameters)? Are similar trends observed?\n\n**Minor:**\n- In (1), the combination of an expectation over the visitation distribution and a summation over time inside the expectation is a bit strange. When summing over time inside the expectation, the expectation is typically written w.r.t. trajectories rather than the visitation distribution.\n- Equation (5) can be written more directly as $\\min(-\\pi_{\\phi} / \\pi_{\\theta}, -(1-\\epsilon_{\\textnormal{PROPS}}))$. The more complicated structure currently being used is needed in PPO’s objective because advantages can be positive or negative, but this is not relevant in (5).\n- Why is IQM used as the performance metric? Typically the mean return is reported.\n- Typos: line 341: Samping -> Sampling, line 485: Figure 5b -> Figure 3b\n- Some of the legends appear inconsistent with other figures (Figure 12) / with the figure caption (Figure 15)\n\n---\n\n**References:**\n\n[1] Smith et al. Demonstrating a Walk in the Park: Learning to Walk in 20 Minutes With Model-Free Reinforcement Learning. In RSS 2023.\n\n[2] Rudin et al. Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning. In CoRL 2022." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- **[S1] Clear and well-written:** The paper is easy to read. The authors clearly present the interesting idea of on-policy data vs. on-policy sampling, as well as the proposed PROPS algorithm. \n- **[S2] Detailed, transparent presentation of experimental implementation:** The authors do a nice job of clearly describing the implementation details used in the experiments. This includes supporting analysis in the Appendix showing results across hyperparameter values." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Proximal Robust On-Policy Sampling (PROPS), an adaptive off-policy sampling method designed to reduce sampling error and improve data efficiency in on-policy policy gradient algorithms. The paper generalizes the adaptive off-policy sampling method ROS introduced in Zhong et al. (2022) by proposing regularization techniques to handle continuous action spaces, and incorporates PROPS into RL training with PPO. Experiments on GridWorld, other discrete-action tasks, and MuJoCo benchmarks compare PPO with PROPS sampling to standard on-policy PPO and a version of PPO that naively reuses recent data without off-policy corrections." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**[W1] Lack of novel theoretical analysis**\n- While Proposition 1 is valuable because it is not obvious that this sampling procedure converges to the correct state visitation distribution, it is a slight generalization of a result that already appears in Zhong et al. (2022). It is also restricted to finite states and actions, and is based on a different sampling procedure than what is proposed in Section 5. \n- Proposition 1 only analyzes what happens in the limit of infinite data, but the transient behavior is important in the context of policy gradient algorithms where a finite (and often limited) amount of data is collected between each policy update.\n- It seems that there would likely be an important interplay between the size of policy updates, the amount of data to reuse, and the size of behavior policy updates determined by the PROPS objective. Unfortunately, the paper does not perform any analysis to connect these components, and instead sets all hyperparameters using a hyperparameter sweep. \n\n**[W2] Experiments do not demonstrate convincing performance benefits**\n- The experimental results suggest that PROPS can achieve slightly better sampling error during RL training, but these improvements only lead to marginal performance benefits in continuous control MuJoCo tasks. \n- Experiments only compare against PPO and a version of PPO that naively reuses recent data (PPO-Buffer). There have been several works that improve data efficiency in a more principled way than PPO-Buffer that are similar to the goals of this work (see references in lines 122-124), but none of these are considered as baselines. Off-policy RL algorithms are also not included for comparison, which can significantly outperform on-policy or near on-policy methods in some of the tasks considered in this work.\n- Due to the regularization needed for stable performance, PROPS requires the behavior policy to remain close to the current policy (small $\\delta_{\\textnormal{PROPS}}$). The experimental results suggest that this prevents PROPS from reusing a significant amount of past data (as mentioned in lines 470-471 and 1181-1183), limiting the data efficiency that can be achieved.\n\n**[W3] Contribution is not clear relative to existing methods:** The paper does not clearly position its contribution relative to existing RL methods that also address data or computational efficiency. See [Q1] below." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a non-i.i.d., off-policy sampling method to produce data that more closely matches the expected on-policy data distribution than on-policy sampling can produce, thus improving the data efficiency of on-policy policy gradient algorithms." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024onpolicy,\ntitle={On-Policy Policy Gradient Reinforcement Learning Without On-Policy Sampling},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zJfOyS1YLW},\nnote={under review}\n}" }, "abstract": { "value": "On-policy reinforcement learning RL algorithms perform policy updates using i.i.d. trajectories collected by the current policy. However, after observing only a finite number of trajectories, on-policy sampling may produce data that fails to match the expected on-policy data distribution. This sampling error leads to noisy updates and data inefficient on-policy learning. Recent work in the policy evaluation setting has shown that non-i.i.d., off-policy sampling can produce data with lower sampling error than on-policy sampling can produce~\\citep{zhong2022robust}. Motivated by this observation, we introduce an adaptive, off-policy sampling method to improve the data efficiency of on-policy policy gradient algorithms. Our method, Proximal Robust On-Policy Sampling (PROPS) reduces sampling error by collecting data with a behavior policy that increases the probability of sampling actions that are under-sampled with respect to the current policy. Rather than discarding data from old policies -- as is commonly done in on-policy algorithms -- PROPS uses data collection to adjust the distribution of previously collected data to be approximately on-policy. We empirically evaluate PROPS on both continuous-action MuJoCo benchmark tasks as well discrete-action tasks and demonstrate that (1) PROPS decreases sampling error throughout training and (2) improves the data efficiency of on-policy policy gradient algorithms. Our work improves the RL community’s understanding of a nuance in the on-policy vs off-policy dichotomy: on-policy learning requires on-policy data, not on-policy sampling." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "reinforcement learning", "on-policy", "policy gradient", "data collection" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/b0860f2fd30166c07622fe9adc01e1c035fdebcc.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/f97b624cc2f76b646ab962b7d189665661bc29fd.zip" }, "title": { "value": "On-Policy Policy Gradient Reinforcement Learning Without On-Policy Sampling" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In Table 3, human-human agreement seems very low. Does this saturate the performances above 60% to some extent?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The authors core objective is important and compellingly achieved: create a lightweight model judge with high performance. Their use of a dynamic checklist, and normalized score predictions both show notable improvements in scoring, from their experiments.\n* The authors unpack the issues with existing lightweight model judges, related to their analyses, position bias and uncertainty, that could be informative to other solutions as well.\n* The comprehensively compare a series of model judges in different settings to understand how they can best be optimized. Their quantitative analysis and ablations are particularly informative." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose RocketEval, a system comprised of API calls and lightweight models, performing as a model judge. The system is designed to be minimally computationally expensive while achieving comparable human agreement scores to GPT4o on MT-Bench and WildBench. Their subsequent ablations and analysis inform how the system can achieve these efficiency gains." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Their explanation for lightweight model “analysis” being weak could be framed more conservatively. If I understand correctly, these experiments more show that lightweight models find classifying GPT4o reviews into scores as an easier task than predicting the score themselves. It wouldn’t necessarily naturally follow that a GPT4o checklist would be the obvious solution, but rather a GPT4o reasoning.\n* It isn’t immediately apparent why GPT4o is so much more expensive than the RocketEval (which includes GPT4o calls). This should be more closely explained. Is it because the reasoning is super long? If so, it would be beneficial to see harder baselines (in terms of efficiency), e.g. prompting GPT4o to produce shorter CoT explanations before its score prediction." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Could authors explain their framework's relation and difference with paper [1] ?\n- RocketEval can enable lightweight models the ability to judge large models, is this similar to the idea of weak-to-strong supervision? Please explain the relation and difference between your paper's ideas and weak-to-strong [2].\n- Conducting experiments solely on a total of 1104 data from MT-BENCH and WILDBENCH is not convincing enough. Could authors conduct experiments on more datasets?\n\n[1] Ribeiro, Marco Tulio, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. \"Beyond accuracy: Behavioral testing of NLP models with CheckList.\" arXiv preprint arXiv:2005.04118 (2020).\n\n[2] Burns, Collin, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen et al. \"Weak-to-strong generalization: Eliciting strong capabilities with weak supervision.\" arXiv preprint arXiv:2312.09390 (2023)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Originality: The paper presents a novel evaluation framework, RocketEval, which evaluates LLMs by grading checklist. This approach is distinct from existing methods like multiple-choice questions. However, I believe the idea of using checklist to evaluate NLP models was actually proposed by another paper [1] in 2020.\n- Quality: The analysis of the lightweight LLMs' abilities in section 2 provides valuable insights into the framework’s effectiveness.\n- Clarity: The paper is well-organized and clearly structured, with each section logically following the previous one.\n- Significance: The framework’s high agreement with human judgments and significant cost reduction make it a promising solution for large-scale LLM evaluations.\n\n\n[1] Ribeiro, Marco Tulio, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. \"Beyond accuracy: Behavioral testing of NLP models with CheckList.\" arXiv preprint arXiv:2005.04118 (2020)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes RocketEval, a novel framework for efficiently evaluating large language models (LLMs) using lightweight LLMs as judges. The framework addresses the limitations of existing automated evaluation methods, such as high costs and lack of interpretability, by using checklist grading. This paper trains lightweight LLMs to independently evaluate each checklist item, providing a multifaceted and unbiased judgment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper primarily focuses on evaluating responses to queries, as the checklist is generated by the input query. However, LLMs are used in a wide range of applications, including text summarization, and translation, where evaluation metrics based on queries may not be efficient since the input query may be very long.\n- Since authors employ GPT-4o as the checklist creator, while the paper argues that checklist creation is a one-time process, the cost of using a powerful LLM like GPT-4o for this task could be significant, especially for large-scale evaluations. Exploring more efficient methods for checklist creation would be beneficial.\n- Typo: Line 215: We select responses from 12 test models and compare the benchmark results when using *GPT-4o* and *Claude-3.5-Sonnet* as judges as strong baselines." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Clear Motivation: The author clearly explains the motivation of the paper in the introduction. I particularly agree with the point raised about the shortcomings of \"Fine-Tuned Judge Models,\" as these models often lack understanding of complex instructions (although the author did not provide proof for this claim).\n- Overall Well-Written: The paper is generally well-written.\n- Relevant Problem: The paper raises a highly relevant problem that urgently needs to be addressed in the current landscape of LLM evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces RocketEval, a lightweight and automated evaluation method for LLMs that addresses the high costs and privacy concerns of using powerful LLMs as judges. By reframing evaluation tasks into a Q&A format using instance-specific checklists, RocketEval improves the judgment accuracy of smaller models, overcoming issues like uncertainty and positional bias." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Regarding Section 2.4: While information entropy is a good metric, it is not very intuitive. I recommend that the author use self-consistency (sampling n times and calculating consistency) to demonstrate this uncertainty.\n- Limited Advantage Over GPT-4o-Mini: Although RocketEval shows a clear advantage over GPT-4o, its superiority over GPT-4o-mini is less significant. Additionally, the N=1000 setting is impractical, as it is rare for someone to test 1000 models simultaneously.\n- Missing Ablation Studies: The paper lacks ablation studies for two components: Independent Checklist Item Judgment and Normalized Probability.\n- Comparison with Other Models: I also suggest that the author provide comparisons with other Fine-Tuned Judge Models, such as the Prometheus series.\n- Minor Issues: In lines 289-290, the author mentions that reference responses and hand-crafted rubrics may not work well in all cases. Please provide examples to illustrate this. Figure 10 is confusing—why does the first row have three models, while the following subplots have only one model per row?\n\nOverall, I believe the problem and methods proposed in the paper are valuable, but there is room for improvement. I hope the author can address my concerns during the rebuttal stage." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "How does the method compare to other approaches, such as [1, 2], as well as fine-tuned judge LLMs?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- I enjoyed reading the paper, especially Section 2, as the problem is well-motivated and the charts effectively convey the key messages.\n- The idea is simple yet smart. By decomposing a single (potentially difficult) evaluation task into multiple simpler items in a checklist, lightweight LLMs can be leveraged to perform the evaluation. More importantly, checklist generation is largely a one-time effort, which could make LLM evaluation more economical in everyday LLM development environments.\n- The authors have thoroughly evaluated the proposed method from multiple perspectives, including agreement, cost, and a qualitative study." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose RocketEval to address the high costs, privacy concerns, and reproducibility challenges associated with using powerful LLMs like GPT-4o for evaluation purposes. RocketEval leverages lightweight LLMs combined with a multi-faceted checklist approach to improve judgment accuracy. Experimental results from benchmarks like MT-BENCH and WILDBENCH demonstrate that RocketEval achieves a high correlation with human evaluations, comparable to GPT-4o, while reducing costs by more than 50-fold." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some aspects of the design lack scientific rigor. For instance, the estimation of “Normalized Probability” does not appear to be justified. Why would it provide an unbiased (or even plausible) probability? I suggest simply referring to it as a “score” instead of “probability”.\n- Direct evaluation using lightweight LLMs is critiqued for high uncertainty and position bias. However, the authors fail to provide a comparison in this regard during the evaluation. Could this issue be addressed by bias mitigation algorithms [1, 2] at an even lower cost? What is the unique advantage offered by the checklist-based method compared to those approaches?\n- The authors mentioned the potential benefits of using prefix caching for cost savings. However, this is not reflected in Section 4.2, even though this feature is already available through OpenAI.\n- In Section 4.2, the authors quote the cost from `vast.ai`, a bidding-based crowd-sourced GPU rental platform. The quoted price ($0.8/hr) is the minimum price for renting an A100 card according to `https://vast.ai/pricing` (why not use the median price?). This might be unfair for several reasons. First, it is unlikely that $0.8/hr represents the regular price. More importantly, OpenAI offers production-level stability, which `vast.ai` does not. A fair comparison would be to quote the price from a public cloud provider like AWS.\n\n[1] Large language models are not fair evaluators\n\n[2] Split and merge: Aligning position biases in large language model based evaluators" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A simple, replicable,interpretable, and accurate automated evaluation method that uses lightweight LLMs as judges to efficiently assess various scenarios and questions." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024rocketeval,\ntitle={RocketEval: Efficient automated {LLM} evaluation via grading checklist},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zJjzNj6QUe},\nnote={under review}\n}" }, "abstract": { "value": "Evaluating large language models (LLMs) in diverse and challenging scenarios is essential to align them with human preferences. To mitigate the prohibitive costs associated with human evaluations, utilizing a powerful LLM as a judge has emerged as a favored approach. Nevertheless, this methodology encounters several challenges, including substantial expenses, concerns regarding privacy and security, and reproducibility. In this paper, we propose a straightforward, replicable, and accurate automated evaluation method by leveraging a lightweight LLM as the judge, named RocketEval. Initially, we identify that the performance disparity between lightweight and powerful LLMs in evaluation tasks primarily stems from their ability to conduct comprehensive analyses, which is not easily enhanced through techniques such as chain-of-thought reasoning. By reframing the evaluation task as a multi-faceted Q\\&A using an instance-specific checklist, we demonstrate that the limited judgment accuracy of lightweight LLMs is largely attributes to high uncertainty and positional bias. To address these challenges, we introduce an automated evaluation process grounded in checklist grading, which is designed to accommodate a variety of scenarios and questions. This process encompasses the creation of checklists, the grading of these checklists by lightweight LLMs, and the reweighting of checklist items to align with the supervised annotations. Our experiments carried out on the automated evaluation benchmarks, MT-Bench and WildBench datasets, reveal that RocketEval, when using $\\textit{Gemma-2-2B}$ as the judge, achieves a high correlation (0.965) with human preferences, which is comparable to $\\textit{GPT-4o}$. Moreover, RocketEval provides a cost reduction exceeding 50-fold for large-scale evaluation and comparison scenarios." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "automated evaluation", "large language models", "natural language processing" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1253b5d0c183f9b66e90566f9e5cfdf3d1c6d306.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/4f6c277c069d7d67c2cadabc8c18e84a843f662f.zip" }, "title": { "value": "RocketEval: Efficient automated LLM evaluation via grading checklist" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Have the authors thought about ways to handle observed instances of hallucinated tags? And currently, how are these affecting retrieval accuracy?\n2. Was a quantitative evaluation considered for the quality and relevance of extracted tags maybe like a comparison with human annotated tags?\n\nPlease also refer to weaknesses for other questions." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The paper introduces an approach to use foundational models to extract modality-specific tags for both videos and text, this is a novel approach to enhance cross-model alignment.\n2. Extensive experiments are conducted on three diverse datasets which cover a wide range of video retrieval tasks, comparison of MAC-VR against a strong baseline with SOTA methods and different inference strategies are documented.\n3. The authors have documented extensive ablation studies for validating the contribution of each components such as the effect of different numbers of tags, foundation models and architectural choices on retrieval performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. The authors in the paper introduce Modality Auxiliary Concepts for Video Retrieval, a novel approach that aims at improving alignment between video and text for effective video retrieval.\n2. This work provides a framework for extracting and utilizing modality-specific tags using foundational models from video and text and an alignment loss is introduced to align modality-specific auxiliary concepts with visual and textual latent concepts. \n3. In the paper they discuss how the architecture is tested with various inference strategies achieving competitive results across three benchmark datasets and ablation studies further validate the impact of each component." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper does not cover the quality and the diversity of the tags extracted. An evaluation of tag relevance and coverage could enhance the contribution.\n2. The paper has limited analysis of the individual contributions of video tags and text tags within the MAC-VR framework. Understanding the contribution of modality-specific tags to overall alignment can maybe help to identify the potential areas for improvement." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. The novelty of the paper is insufficient. It only uses LLMs to obtain the tag information of text and video to enhance the performance of the model. In addition, based on the introduction of additional data and large model knowledge, our model still lags far behind the state-of-the-art methods including T-MASS as shown in Table 3.\n2. The performance of the proposed method in this paper lags significantly behind the state-of-the-art (SOTA) techniques. Upon examining Table 3, it is evident that the proposed MAC-VR method falls behind T-MASS in terms of R@5 and R@10 for the MSRVTT dataset. For DiDeMo, MAC-VR also underperforms T-MASS across R@1, R@5, R@10, and MeanR metrics. The disparity in performance is notably substantial. Also, for ActivityNet, MAC-VR also falls behind T-MASS at R@1 and R@5 by a large margin.\n3. The logic in L53 is strange. The authors explain the effectiveness of different inference strategies which seems to have nothing to do with the auxiliary concept loss function described later.\n4. Some important papers need to be referenced and compared, including:\n[1] Clip-vip: Adapting pretrained image-text model to video-language representation alignment. ICLR 2023\n[2] Uatvr: Uncertainty-adaptive text-video retrieval. ICCV 2023\n5. The writing of the paper is not professional and needs further improvement. The symbols of this paper should be consistent, such as QB in L157 vs \\textit{QB} in L158, maybe $\\in$in L140, et al.\n6. The second term in Eq. 5 should also be preceded by a weight parameter to control it, and the selection of its parameters should be verified through experiments.\n7. The authors do not define the K in Eq. 4. Besides, the K in Eq. 4 seems to be different from the K in L323, so I suggest the authors use different symbols to distinguish them." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper introduces Modality Auxiliary Concepts for video retrieval, a novel approach that leverages modality-specific tags to enhance video retrieval\n2. The authors propose to extract modality-specific tags from foundational VLMs and LLMs to augment the video and text modalities.\n3. The authors propose a new Alignment Loss to better align and distinguish these learnt latent concepts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Modality Auxiliary Concepts for video retrieval, a novel approach that leverages modality-specific tags to enhance video retrieval. However, the novelty of the paper is insufficient. It only uses a large model to obtain the tag information of text and video to enhance the performance of the model. In addition, based on the introduction of additional data and large model knowledge, our model still lags far behind the state-of-the-art methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty of the paper is insufficient. It only uses LLMs to obtain the tag information of text and video to enhance the performance of the model. In addition, based on the introduction of additional data and large model knowledge, our model still lags far behind the state-of-the-art methods including T-MASS as shown in Table 3.\n2. The performance of the proposed method in this paper lags significantly behind the state-of-the-art (SOTA) techniques. Upon examining Table 3, it is evident that the proposed MAC-VR method falls behind T-MASS in terms of R@5 and R@10 for the MSRVTT dataset. For DiDeMo, MAC-VR also underperforms T-MASS across R@1, R@5, R@10, and MeanR metrics. The disparity in performance is notably substantial. Also, for ActivityNet, MAC-VR also falls behind T-MASS at R@1 and R@5 by a large margin.\n3. The logic in L53 is strange. The authors explain the effectiveness of different inference strategies which seems to have nothing to do with the auxiliary concept loss function described later.\n4. Some important papers need to be referenced and compared, including:\n[1] Clip-vip: Adapting pretrained image-text model to video-language representation alignment. ICLR 2023\n[2] Uatvr: Uncertainty-adaptive text-video retrieval. ICCV 2023\n5. The writing of the paper is not professional and needs further improvement. The symbols of this paper should be consistent, such as QB in L157 vs \\textit{QB} in L158, maybe $\\in$in L140, et al.\n6. The second term in Eq. 5 should also be preceded by a weight parameter to control it, and the selection of its parameters should be verified through experiments.\n7. The authors do not define the K in Eq. 4. Besides, the K in Eq. 4 seems to be different from the K in L323, so I suggest the authors use different symbols to distinguish them." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1、1. Please refer to Figure 3 and answer this question: During model testing, are the features $T_{i}$ generated by the Text encoder used as input to the T-CVE model?\n2、In line 362, it is mentioned that K=8. I believe that the choice of K is crucial, and I recommend an ablation study to investigate the impact of different values of K." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper proposes a method called Modality Auxiliary Concepts to enhance video retrieval performance, utilizing large models to generate Visual/Textual Tags that help align visual and textual concepts. The approach is also logically clear and well-structured." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method called Modality Auxiliary Concepts to enhance video retrieval performance, utilizing large models to generate Visual/Textual Tags that help align visual and textual concepts. I hold a positive view of this research." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1、In line 300, for the proposed alignment loss $L_{A}$, please provide the principle or the formula for further clarification.\n2、2. I suggest formatting the tables in the style of three-line tables for improved aesthetics. In line 472, $L_{L_{A}}$ should be changed to $L_{A}$." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "none" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "none" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper conducted thorough experiments to demonstrate the effectiveness of the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper utilizes auxiliary concepts to improve the alignment of visual and textual latent concepts, enabling the distinction between each concept. It also introduces an Alignment Loss to strengthen the alignment between visual and textual latent concepts. Experimental results demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper contains numerous typographical errors; please review it carefully. For example, on line 142, it reads \"into a shared ddimensional latent embedding\"; on line 236, it says \"Alignment Loss LA and\"; on line 314, it mentions \"Flikr videos\"; and on line 317, it states \"on the vall split\". Figure 3 has low resolution. There are also missing bold elements in Table 2. The authors are requested to carefully proofread their paper. The spelling in Figure 5 also contains errors, such as \"enjoyoing\". The capitalization of the table labels throughout the paper is not consistent.\n2.The LA loss proposed by the author is not clearly stated in the paper, and it seems to lack innovation.\n3. The method proposed by the authors leverages the capabilities of VLM and LLM to generate tags for auxiliary alignment. However, as the authors' visualization results and experimental data analysis suggest, this kind of annotation largely generates confusing information that is detrimental to semantic alignment. This does not sound reasonable. Moreover, the authors still heavily rely on the alignment capabilities of the latent space itself, and the experiments do not reflect the effect of using only tags. I speculate that the effect obtained by using only tags is not good. The authors could further demonstrate the rationality and effectiveness of the proposed method through additional experiments or other means. It is difficult to confirm the effectiveness of the method from the current experimental data because I have observed that some results on DiDeMo and ActivityNet Captions have actually declined. This may be due to overfitting to the first dataset, rather than an improvement brought about by the method itself." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024leveraging,\ntitle={Leveraging Modality Tags for Enhanced Cross-Modal Video Retrieval},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zKFUNRH0hN},\nnote={under review}\n}" }, "abstract": { "value": "Video retrieval requires aligning visual content with corresponding natural language descriptions. In this paper, we introduce Modality Auxiliary Concepts for Video Retrieval (MAC-VR), a novel approach that leverages modality-specific tags---automatically extracted from foundation models---to enhance video retrieval.\nPrevious works have proposed to emulate human reasoning by introducing latent concepts derived from the features of a video and its corresponding caption. Building on these efforts to align latent concepts across both modalities, we propose learning auxiliary concepts from modality-specific tags. \nWe introduce these auxiliary concepts to improve the alignment of visual and textual latent concepts, and so be able to distinguish each concept from the other.\nTo strengthen the alignment between visual and textual latent concepts—where a set of visual concepts matches a corresponding set of textual concepts—we introduce an Alignment Loss. This loss aligns the proposed auxiliary concepts with the modalities' latent concepts, enhancing the model's ability to accurately match videos with their appropriate captions. \nWe conduct extensive experiments on three diverse datasets: MSR-VTT, DiDeMo, and ActivityNet Captions. The experimental results consistently demonstrate that modality-specific tags significantly improve cross-modal alignment, achieving performance comparable to current state-of-the-art methods." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Video Understanding" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/5893040ac6e26a674c55010c980ff52c3aa1756f.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Leveraging Modality Tags for Enhanced Cross-Modal Video Retrieval" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Your method depends on DPM-Solver to generate the trajectories. Do you think this affects results in any way? What other dependencies does your method have that might affect results?\n2. Do the hyperparameters for DPM-Solver affect the results drastically? Why did you pick the current hyperparams?\n3. Do you think a larger amount of timesteps might be needed to get satisfactory results? Is this compatible with your method?\n4. Skip baselines seem conceptually weak. The one-step baseline is guaranteed to fail given results on toy examples. Are there no better baselines - for example progressive distillation, that is used for diffusion models?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The work presents a method that leverages deterministic flow matching to create training data (from an AR model) for a one-step image generation model. When trained on this data this model is a distilled version of the original AR model. The idea of using determinstic flow matching to create the data is novel and seems like a good and innovative candidate idea to achieve this. The paper evaluates the claims on class-to-image generation on ImageNet and compares to simple baselines, achieving acceptable FID scores, that are high but not too high." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The method maps autoregressive model output distributions to a Gaussian distribution through flow matching, creating a deterministic transformation. The goal is to then learn this mapping with a neural network. This enables parallel token generation while preserving the conditional dependencies between tokens in the AR model's output distribution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "A. It seems like the FID increases, although seemingly acceptable in numerical terms, give rise to blurry and artifact-ridden images, many of which don't even preserve the structure of the class they are trying to generate (monkeys without eyes etc.). Also, another thing that undermines my confidence is that no images are shown in the main paper and instead shown in the appendix. At least some examples are shown.\n\nB. One big problem from (A) is that, since the paper's main premise is to distill an AR model into a one-shot model, good samples are required to demonstrate that this approach is viable. Given the current results we do not know if this approach can actually generate satisfactory images even when scaled or taken into another regime with larger data such as T2I.\n\nI think these are the main issues for me to be convinced of this approach. My question to authors are:\n- Do you expect the approach to become good enough to generate satisfactory images with different scale (data, model size)?\n- If so, what leads you to believe this?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The writing is somewhat unclear. It's not obvious whether $q_t$ represents tokens in the discrete codebooks or outcomes of flow matching. Algorithms 1-4 are much clearer than the main text, so refining the writing to enhance intuitiveness would benefit the readers.\n\n- The performance gap between few-step AR samplers and the baseline remains quite large.\n\n- Some baseline values, such as VAR/LlamaGen in Table 1, appear to be significantly lower than the original paper's reported results.\n\n- It would be interesting to explore how DD performs in text-conditioned image generation tasks.\n\n- The structure of DD closely resembles DMD in diffusion distillation, where a <noise, clean image> dataset is constructed; however, DMD isn't addressed in this paper. Adding a discussion on DMD would be beneficial. Additionally, transforming next-token prediction to AR-based token denoising at a high level seems conceptually similar to MAR (Li et al., 2024, cited). Including further discussion on MAR could also be helpful." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The authors proposed a new framework to make few-step AR distillation possible. \n- The conversion from next image token prediction to next (set of) image token denoising is a very smart design. The method naturally combines the best of worlds in diffusion models / flow matching and autoregressive image modeling.\n- The performance gain against baseline few-step samplers is huge." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Distilled Decoding (DD), a novel method to reduce the sampling steps of autoregressive image generation models. The approach reframes the traditional next-token or scale prediction process into a denoising procedure guided by flow matching. Starting from fully noisy tokens, DD employs an autoregressive process to map these tokens into image tokens in a flexible number of sampling steps. Experiments with the LlamaGen and VAR models demonstrate that DD significantly outperforms conventional few-step sampling baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The authors propose a novel framework enabling few-step autoregressive (AR) distillation.\n\n- Converting next-image-token prediction to next-image-token denoising is an ingenious design choice, seamlessly integrating the strengths of both diffusion models and autoregressive image modeling.\n\n- The performance improvement over baseline few-step samplers is substantial." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "My questions lean more toward potential future design possibilities, as the explanations provided in the paper are relatively clear.\n\nQ1: The approach seems very similar to a variant of the Consistency Model (CM) applied to the AR domain, replacing the diffusion model in CM with an AR model. What are the advantages of this substitution? Line 319 mentions that, compared to diffusion flow matching, DD “has an easy way to jump back to any point in the trajectory.” This is noted as a characteristic, but what practical benefits does it provide? Additionally, the ODE sampling trajectory in CM is also theoretically fixed and reversible. If my understanding is incorrect, please correct me.\n\nQ2: Is there a way for this method to be trained from scratch, similar to Consistency Training?\n\nQ3: Currently, the method requires a dataset size of 1.2M. How significantly does the data volume impact performance? Related distillation methods, such as DMD[1], use relatively smaller datasets. Is there a possibility of a similar approach to DMD2[2] that could discard the need for noise-data pairs?\n\nQ4: Q: In line 234, is $\\pi$ the same as z used later on? There may be a misalignment in the notation here.\n\n[1]. Yin T, Gharbi M, Zhang R, et al. One-step diffusion with distribution matching distillation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 6613-6623.\n[2]. Yin T, Gharbi M, Park T, et al. Improved Distribution Matching Distillation for Fast Image Synthesis[J]. arXiv preprint arXiv:2405.14867, 2024." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is concisely written with a clear line of thought. The approach of constructing a continuous embedding space in AR and matching it to a Gaussian distribution is particularly interesting. This construction effectively combines the discrete cross-entropy model from AR-based methods with the probability distribution strategies that have proven successful in diffusion models, allowing the concept of consistency distillation from diffusion to be successfully applied within the AR field. The comparisons are comprehensive, and the experiments are well-executed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper is concisely written with a clear line of thought. The approach of constructing a continuous embedding space in AR and matching it to a Gaussian distribution is particularly interesting. This construction effectively combines the discrete cross-entropy model from AR-based methods with the probability distribution strategies that have proven successful in diffusion models, allowing the concept of consistency distillation from diffusion to be successfully applied within the AR field. The comparisons are comprehensive, and the experiments are well-executed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper successfully applies the consistency distillation (CD) technique from diffusion models to the AR field. However, the current results still fall short compared to the pre-trained model." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In table 1, why VAR-DD with one step achieves better performance than VAR-DD with two steps?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well written and easy to follow.\n2. The presented idea of compressing AR models with Flow Matching is intersting. Given that autoregressive models tend to be slow, accelerating them is a crucial challenge. The efforts presented in this paper could provide valuable insights for the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a method called Distilled Decoding (DD) that distills autoregressive models into fewer steps. It employs flow matching to establish a deterministic mapping from Gaussian noise to the output distribution of the pretrained autoregressive models. Extensive experiments demonstrate that DD outperforms the baselines, accelerating autoregressive models with minimal performance degradation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lines 201-203 state that modeling the joint distribution of several tokens is impractical due to the vast possible space. However, it seems that the proposed method uses Flow Matching to learn the joint distribution of tokens (the full sequence in the one-step case). This is somewhat confusing—how does the proposed method tackle this issue?\n2. The image quality shows significant degradation (as indicated by both FID scores and the figures), raising concerns about the practicality of the proposed method.\n3. It would be beneficial to include experiments with text-conditioned LlamaGen to demonstrate that the method can be adapted to more complex scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024distilling,\ntitle={Distilling Auto-regressive Models into Few Steps 1: Image Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zKlFXV87Pp},\nnote={under review}\n}" }, "abstract": { "value": "Autoregressive (AR) models have recently achieved state-of-the-art performance in text and image generation. However, their primary limitation is slow generation speed due to the token-by-token process. We ask an ambitious question: can a pre-trained AR model be adapted to generate outputs in just one or two steps? If successful, this would significantly advance the development and deployment of AR models. We notice that existing works that attempt to speed up AR generation by generating multiple tokens at once fundamentally cannot capture the output distribution due to the conditional dependencies between tokens, limiting their effectiveness for few-step generation. To overcome this, we propose Distilled Decoding (DD), which leverages flow matching to create a deterministic mapping from Gaussian distribution to the output distribution of the pre-trained AR model. We then train a network to distill this mapping, enabling few-step generation. We evaluate DD on state-of-the-art image AR models and present promising results. For VAR, which requires 10-step generation (680 tokens), DD enables one-step generation (6.3$\\times$ speed-up), with an acceptable increase in FID from 4.19 to 10.65. Similarly, for LlamaGen, DD reduces generation from 256 steps to 1, achieving an 217.8$\\times$ speed-up with a comparable FID increase from 6.52 to 17.98. In both cases, baseline methods completely fail with FID scores $>$100. As the first work to demonstrate the possibility of one-step generation for image AR models, DD challenges the prevailing notion that AR models are inherently slow, and opens up new opportunities for efficient AR generation." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "image autoregressive models", "parallel decoding", "distillation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f66ca334bf5d720ab5a87d71ffa3b733e5818ec6.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Distilling Auto-regressive Models into Few Steps 1: Image Generation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What’s the relationship between agent features F_a, map features F_m and surrounding features F_o? Will map queries involved into the selection of interactive queries? \n2. Will the path of other objects be used to calculate the geometric prior in query selection? There may be some objects drive far form the ego path, which is closet to the path but should be pick out. While some objects drive close to the ego path when lane changing, which are far from the path but should pay more attention to.\n3. In Figure 4, why the ego query is updated by weighted object queries? Does that also mean the map queries will not affect the ego query feature?\n4. The introduction of DiFSD(Dense) confuses me about the contribution in sparsity. As you claimed that dense scene representation is computationally intensive, why the dense-based setting is even faster than your fully sparse design (14.8 FPS vs 10.7 FPS)? Meanwhile, I thinke the latency comparison of w/ and w/o interactive query selection will help to illustrate the efficiency of the module.\n5. The visualization of CIPV/CIPS in different scenes iteratively will help reader to understand the process of iterative motion planner. Will you provide some qualitative results?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This idea of iterative motion planner make sense, which is similar to TCP[1] which iteratively adjust attention with guidance of trajectory.\n2. The performance on nuScenes dataset outperforms SparseDrive especially in L2 error.\n\n[1] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: a simple yet strong baseline. NeurIPS, 2022." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed an iterative motion planner based on sparse paradigm, which find out interactive object queries into joint motion-planning module. In open-loop nuScenes dataset, the method outperforms previous sparse method in L2 error and efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The overall novelty of this paper is limited. It’s a good attempt to apply trajectory-guided attention based on previous sparse perception works. However, the presentation of this paper is not enough to demonstrate how it works and how many time it saves. \n2. The contribution in sparsity which aims to reduce the computation is confused, since the dense setting achives even faster inference speed. It looks that the dense representation obtains a better balance between performance and efficiency, while the results of dense setting are also used to calculate the improvements in abstract by the authors.\n3. The comparision between previous works in Table 1 is unfair. Specially, the UniAD and VAD utilize different evaluation protocol as mentioned in PARA-Drive [2]. It looks this paper follows the approach in VAD without transforming UniAD results, which will obtain lower L2 error and collision rate. Therefore, the claimed performance improvements to UniAD are incorrect. Meanwhile, the collision rate of SparseDrive shows large gap between reported in its paper, which deserve to re-check.\n4. The method is only evaluated in open-loop nuScenes dataset without close-loop simulation.\n\n[2] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why do you use bs=1 on A100 GPUs, is the model training consuming so much memory?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper draw inspiration from human driver, and makes the model learning to focus the closest target in path. The proposed Intention-Guided Geometric Attention is interesting. \n2. The experiments are extensive and detailed, supporting the effectiveness of proposed modules." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a fully sparse framework for end-to-end driving, it considers the driving habit of human driver to focus on closest target in path to improve the interaction modeling. Besides, it incorporates position-level diffusion and trajectory-level denoising to improve planning performance. The proposed method achieves SOTA performance on nuScenes dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. One main concern is the experiments are only conducted on open-loop benchmark nuScenes, which suffers from short-cut learning in [1]. Close-loop experiments is needed to prove the effectiveness of the whole model.\n2. I think the framework is built on SparseDrive[2] with optimized planning modules, however, the worst result for L2 in ablation study has already surpassed SparseDrive by a large margin, it is confusing where the good performance comes from.\n\n\n[1] Li, Zhiqi, et al. \"Is ego status all you need for open-loop end-to-end autonomous driving?.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. \n[2] Sun, Wenchao, et al. \"SparseDrive: End-to-End Autonomous Driving via Sparse Scene Representation.\" arXiv preprint arXiv:2405.19620 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "What is the performance of the model if the driving command never appears in the training set? (e.g., turning left and right, or the driving command is empty)" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed solution sounds solid in theory. Specially, the proposed geometric prior through intention-guided attention considers environment in a novel way. In addition, excluding computationally intensive dense scene representation learning and redundant environmental modeling indeed helps speed up the computation. \n2. The ablation study and supplementary material are helpful. Readers can get more information from the ablation study results.\n3. The analysis and discussion guide readers think more deeper about the model design and the performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose an ego-centric fully sparse paradigm for end-to-end self-driving. Specifically, the proposed solution mainly consists of sparse perception, hierarchical interaction and iterative motion planner. The model does not consist of any computationally intensive dense scene representation learning and redundant environmental modeling. Besides, the authors introduce a geometric prior through intention-guided attention, where the closest inpath vehicle/stationary are gradually picked out through ego-centric cross attention and selection. The experimental results show that the proposed solution achieves high accuracy in the self-driving planning task and runs faster." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It will be better if the authors could report some failure cases. Especially for those cases which the predictions are totally opposite toward the navigation command. \n2. It will be better if the authors could mark the symbols on Figures. In this way, readers can easily associate text context with Figures." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "none" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please clarify the superiority of the combination of existing modules, which makes the work in this paper with limited novelty." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "•\tThe paper proposes a novel autonomous driving paradigm, DiFSD, which adopts an ego-centric design and sparse representation, showcasing originality.\n•\tThe experimental results demonstrate DiFSD's superior performance on the nuScenes dataset, indicating high quality.\n•\tThe paper is well-structured and logically coherent, making it easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces DiFSD, an ego-centric fully sparse paradigm for autonomous driving. By focusing on key agents (CIPV/CIPS) and mimicking human driving behavior, DiFSD improves efficiency and performance through sparse perception, hierarchical interaction, and iterative motion planning, while modeling uncertainty with motion diffusion and trajectory denoising. Experiments on the nuScenes dataset show a 66% reduction in L2 error, 77% lower collision rate, and 8.2× faster efficiency compared to UniAD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "•\tThe innovation seems to be built upon combining existing modules, lacking in-depth theoretical analysis and explanation.\n•\tThe reference format is inconsistent, with varying capitalization and abbreviations for conference names. For example, \"IEEE/CVF conference on computer vision and pattern recognition\", \"IEEE/CVF Conference on Computer Vision and Pattern Recognition\", and \"IEEE conference on computer vision and pattern recognition\" are used interchangeably. Additionally, while ICRA and IROS are abbreviated, it is suggested that CVPR and ICCV also be abbreviated for consistency.\n•\tFig. 3 and Fig. 4 illustrate the dual interaction layer within the hierarchical interaction module and the planning optimization layer in the motion planner module, as well as the interactive score fusion process in the geometric-attended selection step, respectively. However, both figures are missing essential captions.\n•\tThe paper includes only two formulas: Formula 1 explains the interactive score fusion by combining attention, geometric, and classification scores, while Formula 2 describes the Loss Function. However, important sections such as Uncertainty Denoising lack necessary formulaic descriptions, which would provide more clarity and rigor to the proposed methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024difsd,\ntitle={Di{FSD}: Ego-Centric Fully Sparse Paradigm with Uncertainty Denoising and Iterative Refinement for Efficient Self-Driving},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zKoUV1wHRJ},\nnote={under review}\n}" }, "abstract": { "value": "Current end-to-end autonomous driving methods resort to unifying modular designs for various tasks (e.g. perception, prediction and planning). Although optimized in a planning-oriented spirit with a fully differentiable framework, existing end-to-end driving systems without ego-centric designs still suffer from unsatisfactory performance and inferior efficiency, owing to the rasterized scene representation learning and redundant information transmission. In this paper, we revisit the human driving behavior and propose an ego-centric fully sparse paradigm, named DiFSD, for end-to-end self-driving. Specifically, DiFSD mainly consists of sparse perception, hierarchical interaction and iterative motion planner. The sparse perception module performs detection, tracking and online mapping based on sparse representation of the driving scene. The hierarchical interaction module aims to select the Closest In-Path Vehicle / Stationary (CIPV / CIPS) from coarse to fine, benefiting from an additional geometric prior. As for the iterative motion planner, both selected interactive agents and ego-vehicle are considered for joint motion prediction, where the output multi-modal ego-trajectories are optimized in an iterative fashion. Besides, both position-level motion diffusion and trajectory-level planning denoising are introduced for uncertainty modeling, thus facilitating the training stability and convergence of the whole framework. Extensive experiments conducted on nuScenes dataset demonstrate the superior planning performance and great efficiency of DiFSD, which significantly reduces the average L2 error by 66% and collision rate by 77% than UniAD while achieves 8.2x faster running efficiency." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Autonomous Driving", "End-to-End Fully Sparse Paradigm", "Iterative Refinement", "Uncertainty Denoising" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/17076696cb8c0980e3147185ea46e34a4b69bcf7.pdf" }, "presentation": null, "primary_area": { "value": "applications to robotics, autonomy, planning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/1b8cef457a59dad7462e9fdd87769580e7c36d8f.pdf" }, "title": { "value": "DiFSD: Ego-Centric Fully Sparse Paradigm with Uncertainty Denoising and Iterative Refinement for Efficient Self-Driving" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the Weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors propose a post-hoc method for developing robustness to spurious correlations, which does not involve any additional training / heavy computational workload, which is a big positive in my opinion. The method operates by identifying specific dimensions in the embedding space that satisfy certain threshold criteria for some forms of spuriosity metrics proposed by the authors, dropping them from the data representations, and solving a logistic regression from this updated embedding space to the label space. None of these operations seem to be computationally expensive.\n\n2. The authors propose ways to measure the degree of spuriosity of a specific dimension, when additional unbiased samples both are and are not available. This is a useful contribution which may also be applicable in other scenarios beyond bias mitigation.\n\n3. The experimental results are strong. Performance improvements over state-of-the-art are quite significant on standard benchmarks, with minimal extra computational overhead." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an approach to mitigate biases learned through spurious correlations by erasing dimensions in the embedding space that are more likely to be associated with spurious features. The authors propose two metrics - namely consistency and evidence energy, for cases when additional unbiased data respectively are and are not available for training, which are used to identify spurious dimensions in the embedding space. Experiments reveal significant improvements over state-of-the-art bias mitigation approaches mainly in the absence of unbiased training data, but also to some extent when it is available." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The calculation of both consistency (eq. 4) and evidence energy (eq 7) are done independently for each feature dimension $i$. However, it is well known that input features do not necessarily map neatly onto independent dimensions in the feature space [a, b]. Given this, the feature erasure in eq. 9 should technically not be achieving what it claims to achieve. By turning full dimensions on/off, it could be - (i) partially turning off some core features that could be in superposition with a spurious feature; and (ii) not fully turning off spurious features, since it may be in superposition with some core feature dimension which is not turned off.\n\n2. In computing $C_{(ik)}$ in eq. 4, there would be significant disparity in the distribution of the features stemming from sampling bias, since the number of unbiased samples is typically << the number of biased samples in practical datasets, which itself could skew the value of $d$, not providing a faithful measure of consistency. How do the authors account for this imbalance when applying the distance metric $d$?\n\n3. Since the idea of EvA is essentially to remove dimensions from the embedding space that correspond to spurious features, an ablation for what constitutes the best approach for selecting the subspace to drop is necessary. EvA employs thresholding on consistency and evidence energy to select spurious subspaces. However, whether consistency and evidence energy are really the best metric to look at when it comes to performing this selection needs to be either theoretically or empirically established. For this, one empirical baseline could be comparison with random feature drop out in the embedding space. I would suggest the authors to explore this in further detail and consider developing such baselines to establish the optimality of their specific feature erasure mechanisms.\n\n4. Table 5 shows that the gains from EvA are more significant in the absence of extra unbiased training data. The authors should provide a discussion on why their method might have a stronger advantage over SOTA when unbiased training data is not present.\n\n5. When extra training data is not available, how does one perform the hyperparameter search for $\\epsilon$? Additionally, when training data is available, although a single pass of EvA is computationally inexpensive, the hyperparameter sweep over $\\epsilon$ may be time consuming. Discussing these, especially, the former case of selecting $\\epsilon$ in the absence of an unbiased train subset is necessary." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In Table 5, how many runs were performed to compute mean and std, and was the same erase ratio used in all of the EvA runs?\n\nSection 4.4 (and Fig. 3 c) discuss the effect of the erase ratio on the worst accuracy. How does this hyperparameter influence the mean accuracy?\n\nIn Table 7, mean acc and worst acc are reported for CelebA but unbiased/conflicting acc in Table 5. The results on Waterbirds suggest that your method results in a better worst acc but worse mean acc for the ResNet-50 than for the ResNet-18: Does this pattern hold on CelebA?\n\nGiven the computational efficiency and no requirement for unbiased data or annotations of the spurious correlations, it should be quite feasible to evaluate EvA-E on ImageNet scale spurious correlations benchmarks (e.g. [1],[2]). In these more realistic settings, the distinction between spurious and core features become more complex. Can you provide empirical evidence or arguments that your evidence energy measure can still be used for detection and mitigation in such settings?\n\n\n[1] Singla et al., Salient ImageNet: How to discover spurious features in Deep Learning?\n\n[2] Neuhaus et al., Spurious Features Everywhere -- Large-Scale Detection of Harmful Spurious Features in ImageNet" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- computational cost are lower compared to similar methods that do not require unbiased data\n- their mitigation outperforms other methods on several common spurious correlation benchmarks\n- for the variant including unbiased data, the introduced method requires less samples for the same accuracy\n- the procedure seems to be well motivated including theoretical analysis" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a method to detect and mitigate spurious correlations learned by an image classifier. They present two measures to identify channels in the penultimate layer of the model which correspond to such spurious correlations, one of them requiring an additional unbiased dataset (“consistency”) while the other can be computed on the the training data (“evidence energy”). After identifying a set of spurious channels, the corresponding weights in the last layer are set to zero and the remaining weights (of the last layer) are retrained." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- the experiments establishing the correlation between their two measures are based only on one class of one of the datasets\n- experiments are only based on small toy settings with strong spurious correlations\n- the method is strongly inspired by OOD methods, but apart from a brief mention in the conclusion, this is not discussed in the main paper \n\nminor: fontsize in Fig. 1, Tab. 3, Tab. 4 are too small, the definition of “energy” actually corresponds to the “free energy” which might be confusing" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Why spurious features often lead to high confidence predictions?\n- Can the proposed method used for other model architectures beyond convolutional neural networks?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Two practical methods, namely EvA-C and EvA-E, are proposed for detecting and mitigating spurious features in the penultimate activations, depending on the existence of an unbiased dataset. The two proposed methods are both efficient and effective in mitigating spurious correlations.\n\n- The two metrics originate from the spurious feature detection methods can be used to measure the spuriousness of features from the penultimate layer of a model. These metrics are also useful to validate the effectiveness of spurious correlation mitigation methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Existing methods on mitigating spurious correlations typically requires unbiased data and multiple rounds of retraining. The paper proposes a computational- and data-efficient pipeline that detects spurious features before the last linear layer. When there is no unbiased datasets, the contribution of each feature to the network’s prediction\nconfidence is exploited for detection. If the unbiased data is available, the consistency of penultimate activations between spurious and unbiased datasets is used for detecting spurious features. Through channel-based erasure and re-weighting on the final linear layer, reliance on spurious correlations can be effectively mitigated." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Contrary to the claim in Line 93, the performance of the proposed method is not the state-of-the-art in comparison with some methods [1,2]. For example, with ResNet50 as the backbone, EvA-E has a worst-group accuracy of 86.6% (Table 7), while [2] achieves 89.1%.\n\n- The unbiased dataset $\\mathcal{D}\\_{\\text{unbiased}}$ is first introduced in Line 156; however, it is unclear what $\\mathcal{D}\\_{\\text{unbiased}}$ can be considered as \"unbiased\". Based on the descriptions in Line 230 and Line 312, $\\mathcal{D}\\_{\\text{unbiased}}$ is selected from the validation set. In some datasets, such as CelebA, the validation set also has gender bias. If $\\mathcal{D}\\_{\\text{unbiased}}$ is defined as a set of group-balanced data where each class of samples distribute equally across different spurious features, then group annotations are required to obtain $\\mathcal{D}\\_{\\text{unbiased}}$, contrary to the claim in Line 50-51. Further clarification on this point would be beneficial.\n\n- In Eq. (4), why the unknown test distribution $\\phi_{\\text{test}}^{ik}$ can be approximated via $\\phi_{\\text{unbiased}}^{ik}$? In some datasets such as CelebA, the test set is also biased. It would be helpful to further clarify this point.\n\n[1] LaBonte et al., Towards last-layer retraining for group robustness with fewer annotations, NIPS, 2023.\\\n[2] Li et al., Bias Amplification Enhances Minority Group Performance, TMLR, 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the questions raised in the Weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The method is well-motivated, and I appreciate its simplicity (in terms of both understanding and computation).\n- Providing theoretical analysis adds a more principled foundation to the two measures.\n- There are experiments showing that the erased features are indeed the spurious ones on CMNIST, and the correlation between evidence energy and consistency, which further supports that the method works as expected." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes EVA, a method for addressing spurious correlations in datasets. It identifies spurious features using a consistency measure and the evidence energy measure in scenarios with and without unbiased datasets, respectively. Theoretical analyses are provided regarding the relationship between these two measures and a feature’s spuriousness. Experiments are conducted to demonstrate the method's effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Since the original DFR paper uses ResNet50 as the architecture, it would be a good sanity check to compare results in that same setting. The authors provided results for ResNet50 in Appendix E.2, but they mention that, to mimic real-world scenarios, they tuned hyperparameters based on Mean Accuracy instead of Worst Accuracy. This might be debatable, as the setting where DFR is applicable assumes an unbiased validation set. Additionally, I wonder if the setup is consistent between ResNet18 in the main paper and ResNet50 in the appendix for the Waterbirds dataset. Could you clarify whether hyperparameters for ResNet18 were tuned based on Mean Accuracy or Worst Accuracy?\n- Overall, since most existing papers evaluate their methods on ResNet50 while this paper primarily uses ResNet18 (this makes the numbers not comparable to those reported in other papers; in general, the numbers are much lower for all methods than those in the literature potentially due to the architecture), and some baselines (e.g., JTT, SSA [1], AFR [2], and [3]) are not included in the main table, Table 5, it is hard to compare the proposed method to the state of the art and to assess whether it truly stands out among existing techniques. For example, I wonder if the proposed method can outperform JTT on the Waterbirds dataset when ResNet50 is used, which has a reported worst-group accuracy of 86.7%, a relatively high value compared to the numbers in this paper obtained on ResNet18.\n- Perhaps a detailed discussion and comparison with [1] is necessary, given that it also achieves the same goal: lightweight reweighting of the last layer without group labels.\n\n\n[1] Qiu, Shikai, et al. \"Simple and fast group robustness by automatic feature reweighting.\" International Conference on Machine Learning. PMLR, 2023.\n\n[2] Nam, Junhyun, et al. \"Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation.\" arXiv preprint arXiv:2204.02070 (2022).\n\n[3] Liu, Sheng, et al. \"Avoiding spurious correlations via logit correction.\" arXiv preprint arXiv:2212.01433 (2022)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "EvA is a method that effectively handles spurious correlations in pretrained networks by explicitly erasing class-specific spurious connections, improving both data and compute efficiency." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024eva,\ntitle={EvA: Erasing Spurious Correlations with Activations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zKvrOOBouT},\nnote={under review}\n}" }, "abstract": { "value": "Spurious correlations often arise when models associate features strongly correlated with, but not causally related to, the label e.g. an image classifier associates bodies of water with ducks. To mitigate spurious correlations, existing methods focus on learning unbiased representation or incorporating additional information about the correlations during training. This work removes spurious correlations by ``**E**rasing **wi**th **A**ctivations'' (EvA). EvA learns class-specific spurious indicator on each channel for the fully connected layer of pretrained networks. By erasing spurious connections during re-weighting, EvA achieves state-of-the-art performance across diverse datasets (6.2\\% relative gain on BAR and achieves 4.1\\% on Waterbirds). For biased datasets without any information about the spurious correlations, EvA can outperform previous methods (4.8\\% relative gain on Waterbirds) with 6 orders of magnitude less compute, highlighting its data and computational efficiency." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "spurious correlation", "compute efficiency", "data efficiency" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/4e5c5685f1373f04de115d86cf61edd9d99bc3c6.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "EvA: Erasing Spurious Correlations with Activations" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See Weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The combination of long-tailed, semi-supervised, and open-world learning is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the limitations of existing open-world long-tailed semi-supervised learning (OLSSL) algorithms, which assume identical distributions between known and novel categories. It introduces a more realistic open-world long-tailed semi-supervised learning (ROLSSL) setting, where no assumption is made about the distribution relationship between known and novel classes. The authors propose a novel approach called dual-stage post-hoc logit adjustments, which dynamically adjusts predictive probabilities to mitigate category bias during learning. The method takes into account the frequency of samples, the number of categories, and the size of the data, improving performance by more selectively utilizing imbalanced unlabeled data. Experiments on CIFAR100 and ImageNet100 show some improvements in performance, demonstrating the effectiveness of the proposed method and setting a strong baseline for future research." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing of this paper is not very clear. For example, 'S/N Consistency' in Tab. 1, a key feature of the proposed setting, is never explained.\n2. Open-world Long-tailed Semi-supervised Learning seems quite similar to [*1], which is not discussed in related works. Therefore, the novelty of this paper may be weak.\n3. Experiments may be insufficient. Only some tiny datasets are considered. Large-scale benchmarks (e.g., ImageNet-LT or iNaturalist) should be included.\n\n\n[*1] SimPro: A Simple Probabilistic Framework Towards Realistic Long-Tailed Semi-Supervised Learning. ICML'24" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How to handle the scenarios where the number of new classes is unknow? It seems that the proposed method requires the prior class number, which might not realistic.\n2. Could the authors summarize the main novelty intuitively?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper is well-motivated and easy to follow.\n2. The proposed problem setting is realistic and reflects the long-tailed nature of the real world.\n3. This paper conducts comprehensive experiments to show the superiority of the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies a realistic open-world setting, namely open-world long-tailed semi-supervised learning (ROLSSL), where the distribution relationships between known and novel categories might be different. To solve this problem, the paper proposes dual-stage post-hoc logit adjustments. Specifically, it estimates the distributions of the unlabeled data and utilizes the estimations to re-adjust the logits. Comprehensive experiments validate the superiority of the method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The comparative methods are outdated, and the most recent comparative method of this paper is ECCV 2022. In the literature, OSSL is also referred to as generalized category discovery (GCD). Some recent GCD methods, including GCD [R1] and SimGCD [R2], should be included in Table 2 and 3. There are also some works [R3, R4] in GCD considering long-tailed scenarios, which are highly related to this work and should be compared.\n2. The novelty of this paper is limited. The basic loss functions $L_{pair}$ and $L_{reg}$ are well-known techniques in novel class discovery and GCD, while the logit adjustments are also prevalent methods in long-tailed classification. Simply applying model predictions in logit adjustments is naïve and might not provide great insights to the community.\n3. The method still requires strong priors of the ground truth new class numbers, which is unrealistic.\n4. Experiments with the ViT backbone should be conducted.\n5. The notations should be concise and clear.\n\nReferences:\n[R1]. Generalized Category Discovery. CVPR 2022.\n[R2]. Parametric Classification for Generalized Category Discovery: A Baseline Study. ICCV 2023.\n[R3]. Novel Class Discovery for Long-tailed Recognition. TMLR 2023.\n[R4]. Generalized Categories Discovery for Long-tailed Recognition. ICCV 2023, workshop." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Have you considered running experiments on iNaturalist? It has more classes than any of the benchmarks used in this paper and presents a more challenging fine-grained, long-tailed distribution." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The problem addressed by this paper is highly practical, as determining the distributional relationships between known and novel categories is indeed nearly impossible in real-world scenarios.\n\n2. The paper is well-written and easy to follow.\n\n3. Key experimental results demonstrate that the DPLA approach consistently outperforms the OpenLDN baseline in both known and novel class recognition, achieving up to a 50.1% improvement on datasets like CIFAR-100 and ImageNet-100. Authors further showcase the robustness and adaptability of the proposed method across varying levels of data imbalance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a approach called Dual-stage Post-hoc Logit Adjustments (DPLA) to tackle the challenges in Realistic Open-world Long-tailed Semi-supervised Learning (ROLSSL). DPLA is a two-stage logit adjustment technique designed to improve model performance on both frequent and infrequent (tail) classes. In the first stage, logits are adjusted based on the sample frequency and class count, which helps balance predictions across known classes. In the second stage, adjustments are dynamically refined using the predicted class distribution from unlabeled data, with a particular focus on emphasizing tail classes to counter bias towards dominant classes. This dual-stage approach enables better recognition in long-tailed, semi-supervised, and open-world settings, especially where labeled data is sparse and imbalanced. Key experimental results highlight that the DPLA approach consistently surpasses the OpenLDN baseline in both known and novel class recognition, achieving up to 50.1% improvement experiments on datasets such as CIFAR100 and ImageNet100. They also show the robustness and adaptability of the proposed method across different data imbalance conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. A key concern is the complexity of the proposed pipeline, which involves multiple stages and numerous hyper-parameters. With additional stages and hyper-parameters to tune, it becomes more challenging to generalize this pipeline to new problem settings and datasets. For instance, as shown in Tables 3 and 4, using non-optimal scaling factors can significantly impact performance on CIFAR-100 and ImageNet-100, despite both datasets having the same number of classes. Have you explored methods to automatically tune the hyperparameters or analyzed which parameters are most sensitive? \n\n2. Technical contributions. A primary contribution of this paper is the proposed logit adjustment method for pseudo-labels, which often exhibit class imbalance. However, the issue of naturally imbalanced pseudo-labels has been previously addressed by DebiasPL [1], which also applies a logit adjustment based technique to mitigate class imbalance in pseudo-labels, resulting in improved classification performance on long-tailed datasets. A discussion on the technical distinctions between this approach and DebiasPL would be beneficial.\n\n3. While this paper presents experiments on several benchmarks to demonstrate performance gains, most datasets used are small to medium-scale, such as ImageNet-100 or CIFAR. I believe it is essential to include experiments on larger datasets, such as iNaturalist or ImageNet-1k or ImageNet-22k, to better assess the generalizability of the proposed method. It will be nice if you could provide any potential challenges or modifications needed to apply your method to larger datasets.\n\n4. In addition to presenting results on novel and known subsets, I believe it is essential to include performance metrics for many-shot, medium-shot, and few-shot subsets. This would help clarify whether the current logit adjustment methods compromise performance on many-shot classes to improve results on few-shot classes—a common issue with most logit adjustment techniques. It would be valuable for the authors to discuss whether their proposed method faces this challenge and, if not, to explain why it avoids this problem. If it does, providing insights into potential strategies to address this limitation in future work would be beneficial. \n\n5. I may suggest authors to include a specific breakdown of performance metrics for these subsets in their results tables. Add a discussion section that explicitly addresses the trade-offs between performance on different shot categories and potential strategies to balance these trade-offs. In my opinion, while improving performance on few-shot classes is important, it is equally crucial to avoid jeopardizing performance on many-shot classes, as they are more prevalent in real-world applications. \n\n[1] Wang, Xudong, Zhirong Wu, Long Lian, and Stella X. Yu. \"Debiased learning from naturally imbalanced pseudo-labels.\" In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14647-14657. 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "This work may need more effort to refine its writing and the methodology sections need more explanations for equations as mentioned above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "SSL in a more realistic setting is proposed and explored in this work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper mainly focuses on the more challenging SSL setting - open-world long-tailed SSL (OLSSL). The authors propose a more realistic setting than the existing OLSSL: labeled samples are much less than the unlabeled ones. Based on the existing OLSSL method, this work proposes a dual-stage post-hoc logit adjustment method to calibrate and optimize the model. Several experiments are conducted for verification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Since readers may not be quite familiar with the development of open-world long-tail SSL, it is suggested that more background about the existing OLSSL algorithms be included in Section 3.2, especially for the definition of Eq(1).\n2. The dual-stage post-hoc logit adjustment section is not easy to read. Specifically, \n- why the aremax is used in Eq(2)? According to the Fig 1, it is expected to be a distribution adjustment. \n- What is the definition of F_{y_i^l} in Eq(3)? If the labeled samples are limited, will this sample frequency be biased? How C_base and S_base are defined? \n- What is the relationship of the equation in L244 and the Eq(2)? \n- for unlabeled samples, how the sample ratio \\pi_c^r in Eq(4) is obtained when the class of the unlabeled samples is unknown? How alpha and beta are defined? What is the meaning of (alpha-beta) in Eq(4)?\n3. Does the \\tau in Eq(7) denote the ones in Eq(2)? If so, why the operation become plus but not minus as in Eq(2)? \n4. The proposed method is argued to be a two-stage method, does it mean training labeled and unlabeled samples in different stages? However, according to Eq(8), the model seems optimized on labeled and unlabeled samples together.\n5. In Figure 1, the right side of the diagram shows a question mark (\"?\") symbol, does it represent the unlabeled samples? If so, why does the labeled branch also include a depiction of unlabeled samples? Is there a particular significance or reason for this choice in the illustration?\n6. The paper seems primarily focused on addressing the long-tailed problem while offering little innovation for the open-world problem.\n7. In the defined Realistic Open-World Long-Tailed Semi-Supervised Learning (ROLSSL) framework, it is assumed that the number of novel classes is known (L200). The proposed method leverages this prior knowledge for model calibration (Eq(3)). However, this assumption seems somewhat unrealistic, as it is uncommon to have prior knowledge of the exact number of novel classes present in the unlabeled samples." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The analysis of hyperparameters $\\tau_1$ and $\\tau_2$ are missing." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The studied problem in this paper is practical and underexplored. Most previous studies focus on long-tailed or open-world semi-supervised learning, without considering both.\n2. The proposed approach is simple and easy to implement. This paper improves previous methods by proposing a dual-stage post-hoc logit adjustment technique, which does not involve complex strategies or additional learnable parameters.\n3. Experiments on six datasets show the advantage of the proposed approach against previous methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to solve a complex task in semi-supervised learning, i.e., the long-tailed label distribution and the presence of novel classes in unlabeled data. The paper presents a simple solution for this task based on techniques such as logit adjustment, and entropy maximization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty of the proposed method seems limited. As stated in the *Strengths*, the proposed method is simple and easy to implement because it is pretty much built upon existing techniques. However, this also makes the contribution of the method minor.\n\n2. The rationale behind the method is not well-explained. It is unclear why the logit adjustment strategies for labeled and unlabeled data are different. Also, the design of Eq. (3) is not straightforward and needs in-depth understanding and justification. In Eq. (4), the impact of hyperparameters $\\alpha$ and $\\beta$ are not studied.\n\n3. The writing needs improvement. For instance, the paper should include the definition of $\\mathcal{L}\\_{pair}$ and $\\mathcal{L}\\_{reg}$ to be self-contained.\n\n4. It would be better if the paper could include real-world semi-supervised datasets which follow long-tailed label distribution and also include novel classes in unlabeled data.\n\n5. It seems that the logit adjustment strategy in open-world semi-supervised learning has been considered in [1].\n\n[1] Bridging the Gap: Learning Pace Synchronization for Open-World Semi-Supervised Learning. IJCAI 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nhe2024towards,\ntitle={Towards Realistic Long-tailed Semi-supervised Learning in an Open World},\nauthor={Yuanpeng He and Lijian Li and Xiancai Chen and Chi-Man Pun and Wenpin Jiao and Zhi Jin},\nyear={2024},\nurl={https://openreview.net/forum?id=zLHP6QDWYp}\n}" }, "abstract": { "value": "Open-world long-tailed semi-supervised learning (OLSSL) has increasingly attracted attention. However, existing OLSSL algorithms generally assume that the distributions between known and novel categories are nearly identical. Against this backdrop, we construct a more Realistic Open-world Long-tailed Semi-supervised Learning (ROLSSL) setting where there is no premise on the distribution relationships between known and novel categories. Furthermore, even within the known categories, the number of labeled samples is significantly smaller than that of the unlabeled samples, as acquiring valid annotations is often prohibitively costly in the real world. Under the proposed ROLSSL setting, we propose a simple yet potentially effective solution called dual-stage post-hoc logit adjustments. The proposed approach revisits the logit adjustment strategy by considering the relationships among the frequency of samples, the total number of categories, and the overall size of data. Then, it estimates the distribution of unlabeled data for both known and novel categories to dynamically readjust the corresponding predictive probabilities, effectively mitigating category bias during the learning of known and novel classes with more selective utilization of imbalanced unlabeled data. Extensive experiments on datasets such as CIFAR100 and ImageNet100 have demonstrated performance improvements of up to 50.1%, validating the superiority of our proposed method and establishing a strong baseline for this task. For further researches, the experimental code will be open soon." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Yuanpeng_He1", "~Lijian_Li1", "~Xiancai_Chen1", "~Chi-Man_Pun1", "~Wenpin_Jiao1", "~Zhi_Jin1" ] }, "authors": { "value": [ "Yuanpeng He", "Lijian Li", "Xiancai Chen", "Chi-Man Pun", "Wenpin Jiao", "Zhi Jin" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Open-world", "Realistic long-tailed semi-supervised learning", "Logit adjustment" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "he|towards_realistic_longtailed_semisupervised_learning_in_an_open_world" }, "pdf": { "value": "/pdf/2cd398bf25fc2602a15dc6780847e91c92d2059c.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Towards Realistic Long-tailed Semi-supervised Learning in an Open World" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the above Weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper uncovers an observation that the decomposed singular vectors and values naturally undertake the different types of degradation information, ascribing various restoration tasks into two groups, i.e., singular vector dominated and singular value dominated.\n2. Two operators are developed to favor the decomposed optimization of degraded singular vectors and values for various restoration tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Decomposition Ascribed Synergistic Learning (DASL), a method for unified image restoration that optimizes singular vectors and values for handling multiple degradations within a single model, and integrates seamlessly into existing restoration architectures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is a limited evaluation with recent AIR methods like PromptIR, GenLV, and the diffusion-based AIR techniques. Such comparisons ensure the research is aligned with the latest advancements in the field.\n2. Do the observations regarding the decomposition of singular vectors and singular values hold true in terms of other degradations (low-resolution, jepg compression, etc.) or compound degradations (noise+blur, rain+haze, etc.)? Whether the proposed method is available in multiple degradation combinations?\n3. How well do the decomposed singular vectors and singular values represent degradation information, and are they discriminative enough to distinguish between different degradations?\n4. Some typos like \", ,\"." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Are there specific references supporting the findings presented in Figures 1 and 2? Additionally, clarification is needed on the type of Singular Value Decomposition (SVD) employed in the analysis and whether any post-processing techniques were applied to produce the displayed images. It is noted that the deblurring case exhibits a relatively large area of damaged pixels, while the hazy and low-light image enhancement cases only show minor pixel degradation. What factors contribute to this discrepancy?\n\nThe authors conducted tests on underwater enhancement and sandstorm enhancement but did not include these results in Figure 1 & 2. Are the observations in Figure 1 commonly encountered across different datasets, including underwater and sandstorm? Moreover, Figure 2 is based on 100 images; Additional images should be tested to validate the findings. It is unclear whether only synthetic images were used to generate Figures 1 and 2, such as in the rainy case.\n\nThe assumptions regarding singular vectors capturing content information and spatial details, and singular values representing global statistical properties, are questionable. The authors should provide more supporting references for these claims. Furthermore, given that degradations are categorized into two groups, does the proposed method require prior knowledge of which group the degradation falls into before reconstruction?\n\nA clear diagram illustrating the entire blueprint of the proposed methods should be included to demonstrate how the SVEO and SVAO are integrated into the algorithm. The current description lacks clarity, particularly the statement about substituting convolution layers with SVEO.\n\nThe experiments primarily report PSNR results; however, additional metrics should also be considered to provide a more comprehensive evaluation of the proposed method's performance.\n\nThe rationale for using the L1 norm in L_dec should be clarified.\n\nFigure 5 does not effectively illustrate progressive component reconstruction as intended.\n\nTypos in the abstract: 'tasks into two groups, , '" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper addresses a significant gap in the field by proposing a unified approach to handling multiple image degradations, moving beyond the conventional practice of treating each degradation in isolation.\n\nThe use of singular value decomposition to categorize restoration tasks into singular vector dominated and singular value dominated groups provides a fresh perspective on the relationships between different types of degradations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript explored the restoration of multiple image degradations through a unified model. By employing singular value decomposition, the authors classified restoration tasks into singular vector dominated and singular value dominated categories. They introduced the Decomposition Ascribed Synergistic Learning (DASL) framework, which optimized the decomposed components to exploit relationships among various restoration tasks. The framework included two operators, Singular VEctor Operator (SVEO) and Singular VAlue Operator (SVAO), along with a supporting decomposition loss." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The manuscript raises several important questions regarding its methodology and results that require clarification.\n\nThe current description lacks clarity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "[Q1] As per my understanding, the orthogonal regularization loss and decomposition loss in formula (5) represent the reconstruction loss of singular vectors and singular values respectively, and different dominant factors are employed for various types of degradation. During the training process, the balanced weights are set as a fixed ratio. Will this setting negatively harm the mixed results?\n\n[Q2] In Figure 2, which baseline method did you use? It seems that the exchange reconstruction error for dehazing is very close. Thus, this figure can't convince me of the effectiveness of the proposed idea in dehazing task.\n\n[Q3] This work aims to use a single model to handle multiple kinds of degraded images with one degradation type. How does it perform on images that contain multi-types of degradation? What is the difference in performance between your model and a model that trains only with a single type of degradation?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "[S1] Lightweight networks have great application prospects at present.\n\n[S2] DASL reduces the computations of the baseline model and improves the effect." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper mainly proposes Decomposition Ascribed Synergistic Learning (DASL) to restore multiple image degradations by SVD and FFT. Specifically, diverse degradations are ascribed into two groups, singular vector-dominated degradations and singular value-dominated degradations. The proposed DASL dedicates the decomposed optimization of them respectively, rendering a more unified perspective to inherently utilize the potential partnership among diverse restoration tasks for ascribed synergistic learning. Two effective operators SVEO and SVAO have been developed to favor the decomposed optimization. Experimental results demonstrate that DASL improves restoration quality while reducing model parameters and accelerating inference speed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "[W1] The paper writing is very poor, containing many typos. Many sentences, paragraphs, and sections are difficult to understand. \n\n[W2] The proposed method applies the SVD in the FFT space and validates its effectiveness to some extent. However, it does not clearly explain the insight and why it works.\n\n[W3] More visualization results of DASL+baseline should be presented rather than only MPRNet. In addition, in Fig. 14-16, the superiority of the proposed method is hard to distinguish. \n\n[W4] The article presumes that the SVD and IDFT operate similarly in terms of signal formation, but does not show the Ablation Study of the restore results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Same as described in the weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is easy to understand and has a clear overall structure.\n2. Exchange of the singular value and the singular vector analysis is performed on the image restoration tasks of different degradation types, showing the role of singular values ​​and singular vectors in image restoration.\n3. The singular value operator and singular vector operator are proposed to make the existing methods lighter, with some improvement in computational complexity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The goal of this paper is to solve the image restoration task of multiple image degradation types. The paper explores the relationship between singular values and singular vectors between degraded images and clear images in different degradation types. Singular value-dominated and singular vector-dominated image restoration are ascribed. Singular vector operators and singular value operators are proposed and lightly integrated into the existing image restoration backbone. Experiments on multiple degradation types are conducted to verify the effectiveness of the method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Figure 2, the authors slightly show that the decomposed singular values ​​and singular vectors between different degradation types undertake different degradation information. Specifically, the exchange reconstruction error of Figure 2 (a) in the dehazing task is close to 50% for both singular value and singular vector, and the low-light and haze degradation in Figure 2 (b) are similar to other degradations such as rain and noise, while the blur and low-light in Figure 2 (c) are also very close. Therefore, I think that the authors don't explain the relationship between singular values ​​and singular vectors and degradation information well. The assumptions proposed are also untenable.\nThe authors could consider conducting an in-depth analysis in high-dimensional space or frequency space to analyze the impact of singular values ​​and singular vectors in terms of signal-to-noise ratio, frequency changes, etc. Even ablation analysis without using singular values ​​or singular vectors can be used instead of just statistical analysis of reconstruction errors.\n\n2. The experiments only include results on a single degradation task dataset, not on a mixed dataset. This does not meet the requirement of using a single model for a unified image restoration task. The authors could refer to the similar experimental setup in the work of Li et al. [1] to conduct experiments on mixed degradation datasets to demonstrate the effectiveness and robustness of the proposed DASL. \n\n3. The stability improvements brought by the loss curves of the training trajectories in Figure 6 and Figure 7 are also not clearly demonstrated for tasks such as denoise and derain, etc. The authors should include the corresponding intermediate feature map changes or any other visualization statistics that can assist in demonstrating the effectiveness of singular values ​​and singular vectors.\n\n[1] Li, Boyun, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. \"All-in-one image restoration for unknown corruption.\" In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 17452-17462. 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024decomposition,\ntitle={Decomposition Ascribed Synergistic Learning for Unified Image Restoration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zLaayPL8f0},\nnote={under review}\n}" }, "abstract": { "value": "Learning to restore multiple image degradations within a single model is quite beneficial for real-world applications. Nevertheless, existing works typically concentrate on regarding each degradation independently, while their relationship has been less comprehended to ensure the synergistic learning. To this end, we revisit the diverse degradations through the lens of singular value decomposition, with the observation that the decomposed singular vectors and singular values naturally undertake the different types of degradation information, dividing various restoration tasks into two groups, \\ie, singular vector dominated and singular value dominated. The above analysis renders a more unified perspective to ascribe diverse degradation connections, compared to previous task-level independent learning. The dedicated optimization of degraded singular vectors and singular values inherently utilizes the potential partnership among diverse restoration tasks, attributing to the Decomposition Ascribed Synergistic Learning (DASL). Specifically, DASL comprises two effective operators, namely, Singular VEctor Operator (SVEO) and Singular VAlue Operator (SVAO), to favor the decomposed optimization, which can be lightly integrated into existing image restoration backbone. Moreover, the congruous decomposition loss has been devised for auxiliary. Extensive experiments on five image restoration tasks demonstrate the effectiveness of our method." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Image Restoration", "Decomposition", "Orthogonality", "Signal formation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/6e6d558f084294c197e57b496bbf9b1af751c5fd.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Decomposition Ascribed Synergistic Learning for Unified Image Restoration" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "As in weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "-\tThe authors propose a two-stage distillation scheme that progressively distills the model to a few-step regime with reduced training costs.\n-\tThe proposed method achieves superior results with SDXL on COCO-5k generation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose TLCM (Training-efficient Latent Consistency Model) for accelerating text-to-image latent diffusion models in a data-free manner with a small amount of training time. The key innovation is a two-stage distillation process. A data-free multistep latent consistency distillation (MLCD) is proposed to accelerate the base model, followed by another improved data-free latent consistency distillation to ensure global consistency. The authors enhance the model's performance through several techniques including latent LPIPS for perceptual consistency, multi-dimensional preference learning, distribution matching, and adversarial learning. Their empirical results show that TLCM can generate high-quality images in just 2-8 inference steps." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-\tThe submission is basically a technical report on achieving the SOTA few-step image generation performances with data-free multi-step consistency distillation, which is a successful practical combination of [1] and MCM. The second distillation stage also uses a series of SOTA techniques including ADD, DMD, and preference tuning. There are very limited scientific insights or technical innovations.\n-\tThe current presentation is poor and needs substantial improvement. For instance, there is extensive and unnecessary usage of acronyms, making the submission so confusing and hard to read.\n-\tLimited empirical evaluation. The method is only validated on SDXL. How does it perform on other diffusion models with different architectures, e.g., Pixart?\n\nReferences\n\n1.\tKohler, Jonas, et al. \"Imagine flash: Accelerating emu diffusion models with backward distillation.\" arXiv preprint arXiv:2405.05224 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "no" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "see above" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Advantages:\n1. While straightforward, the idea shows some novelty\n2. The writing is acceptable, though the innovation isn't immediately apparent" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a direct combination of Multistep Consistency Models [1] and Latent Consistency Models [2], resulting in a Multistep Latent Consistency Model.\n\nAdvantages:\n1. While straightforward, the idea shows some novelty\n2. The writing is acceptable, though the innovation isn't immediately apparent\n\nDisadvantages:\n1. The experimental results appear counterintuitive: contrary to normal expectations where more sampling steps yield better results, this paper shows deteriorating FID scores with increased sampling steps\n2. The paper lacks directness - the core innovations are hard to identify, as the authors seem to attempt highlighting multiple contribution points\n3. The generated images predominantly show a cyberpunk style, lacking photorealism and overall quality\n\nTechnical Issues:\n1. Equation 1 appears incorrect - contains an extra x_t term\n Should be: dx_t = f(x_t, t)dt + g(t)dw_t\n\n2. Results Inconsistencies:\n - Tables 1 and 3 show FID increasing with more sampling steps, contradicting common understanding\n - Ablation study shows puzzling results: combining all techniques significantly worsens FID scores\n - Equations 11-13 look unnecessary, just traditional confrontational losses.\n\nWriting Concerns:\nThe paper lacks focus on core contributions. It is hard for me to distinguish core contributions.\n\nReferences:\n[1] Heek et al., \"Multistep consistency models,\" 2024\n[2] Luo et al., \"Latent consistency models,\" 2023\n[3] Song et al., \"Consistency models,\" 2023" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Disadvantages:\n1. The experimental results appear counterintuitive: contrary to normal expectations where more sampling steps yield better results, this paper shows deteriorating FID scores with increased sampling steps\n2. The paper lacks directness - the core innovations are hard to identify, as the authors seem to attempt highlighting multiple contribution points\n3. The generated images predominantly show a cyberpunk style, lacking photorealism and overall quality" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "(L144-145) This is a vague statement. What is the meaning of the effectiveness?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Interesting idea to use synthetic data for distillation.\n- Reasonable motivations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Existing methods for distilling the pretrained knowledge of Diffusion Models have several drawbacks; expensive training time, necessity of large scale real data, and image generation quality given a few steps to generate an image. To this end, the authors propose a data-free distillation pipeline named as Training Efficient Latent Consistency Model (TLCM). Briefly, the initial parameter of TLCM is obtained by the proposed first process, and the second process boosts global consistency of TLCM. The experiment results show that TLCM has benefits over the baselines for the short training time and the data-free mechanism." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Overview figure is very complex to understand. My suggestion is that it would be better if the right figure is decomposed into two parts; section 4.2 and section 4.3\n- Weak explanations on the experiment settings.\n - What is the meaning of each measure? \n - Is higher better or lower better? \n - How many samples are used? \n - What was the input to generate the samples to measure?\n - It sounds more reasonable if the proposed methods contain only IS or FID like [1]. The aim of the work is to distill the knowledge of the pretrained models and reduce the required sampling steps to some extent. Shouldn’t it enough If the generated images by TLCM (2-8 steps) are as realistic as DDIM? Please provide justifications why the other metrics are needed in detail. \n- Weak interpretations of the experiment results.\n - (L364-L365) The proposed methods use synthetic data to distill the pretrained knowledge while the baselines are using real data. How does the proposed method show better performance than the baselines? It sounds more reasonable if the baseline is a sort of upper bound.\n - FID of TLCM (2 steps) is better than FID of TLCM (8 steps). It also shows the pattern that the # of steps and FID are inversely proportional, which is not reasonable. \n - Prioritize what information is important for each table and emphasize those numbers.\n\n- Too many engineering techniques are applied for improving the marginal performance. They dilute the main point of the paper.. \n\nOverall, the proposed methods are interesting, but the results (FID) are not convincing enough.\n\n[1] PROGRESSIVE DISTILLATION FOR FAST SAMPLING OF DIFFUSION MODELS, ICLR’22" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- After the emergence of rectified flow transformer, I think these kinds of distillation methods should be re-assessed. Apart from the non-predictable nature of the original diffusion trajectory, we now can learn the straightened flow of latent space and the distillation must be a lot easier. Even in the paper of SD3, it says that if the model capacity is given enough, reducing timesteps into a small number sacrifices little performance. This is not a critical comment about this paper and I would recommend the authors to think about this aspect." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Explanation about existing methods have been given enough. It is easy to follow and the application of each component sounds reasonable.\n- Resulting images with only 4 steps look quite impressive.\n- Necessary ablation studies to make this method solid have been provided." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "- The authors propose a multi-step distribution matching diffusion model to generate samples within few steps. By dividing the whole iterations into several predefined steps, the consistency model can focus on matching distribution in narrower range.\n- Adversarial loss, latent-LPIPS loss, and human preference loss have also been adopted to boost the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Even though the results are quite impressive, this method is a combination of Multistep Consistency Models(MCM), Distribution Matching Distillation(DMD), and DiffusionGAN. It is hard to ignore novelty issue and this would be the main reason of my decision.\n- Too many notations made me really hard to follow this paper. It would be much better if some expressions can be trimmed.\n\nMinor issue\n- The caption of the 3rd row in Fig.3 should be modified." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Does the 70GPU-hour training time include the time spent on data generation? The overall cost of data generation is needed to justify L241 'with cheap cost in data-free manner.'" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper presents the method in a very intuitive way with very informative equations and figures. The overall writing quality of this paper is good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a method for latent diffusion model distillation. The core idea of this paper is mainly inspired by latent consistency distillation. The authors implemented some improvements based on the previous success of diffusion distillation, such as sparse timesteps. To prevent the cost of collecting real-world data, the proposed method is trained on generated data from the teacher model instead. A bag of independent techniques such as adversarial training are further included to help the model match the SoTA performance. The evaluations are performed on SDXL." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Novelty\n\nThis paper presents insufficient novelties. The entire method seems to be an ensemble of diffusion distillation techniques already proven effective. For example, this paper advocates the advantage of relying on no training data and uses synthesized data instead. However, this has already been proven useful in SD3-turbo [1], and learning from generated data is by no means a new idea under the context of model distillation. \n\n- Empirical evaluations\n\nThis paper is presented with insufficient evaluations. SDXL, which is a relatively old model, is the only one evaluated in the paper. And there are some important comparisons missing such as SD3-turbo, DMD, and instaflow. \n\nThe authors claim, 'We only report FID for reference and do not analyze it since FID on COCO is not reliable to evaluate text-to-image models.' I agree that FID is not the best way for diffusion evaluation. However, the values do reveal something about the proposed methods. I noticed that the FID values increase with the number of steps for the proposed method. In my understanding, this is a result of learning distillation from generated data instead of real data, as the modeled distribution deviates further from the real-world data distribution. \n\n- Supports to claims \n\nI find some claims in the paper require further support. For example, the authors claimed 'They (other distillation methods) need to perform long-time learning with a huge volume of real data.' I find it very hard to agree with this. According to my experience, methods such as LCM can be trained very fast with a small portion of training data. The authors are strongly encouraged to fill in the 'TH' value for LCM in Table 1 for a direct comparison. For the rest of the metrics, I don't see how the proposed method clearly outperforms the rest, except for IR, which is a score that the model is explicitly optimized for. \n\n\n[1] Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation, arxiv." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a novel Training-efficient Latent Consistency Model (TLCM) to tackle the challenges of expensive cost and the performance drop when sampling with few steps in large distilled latent diffusion models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tlcm,\ntitle={{TLCM}: Training- efficient Latent Consistency Model for Image Generation with 2-8 Steps},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zM92zziRtQ},\nnote={under review}\n}" }, "abstract": { "value": "Distilling latent diffusion models (LDMs) into ones that are fast to sample from is attracting growing research interest. However, the majority of existing methods face two critical challenges: \n1) They need to perform long-time learning with a huge volume of real data. \n2) They routinely lead to quality degradation for generation, especially in text-image alignment. \n\nThis paper proposes the novel Training-efficient Latent Consistency Model (TLCM) to overcome these challenges. \nOur method first fast accelerate LDMs via data-free multistep latent consistency distillation (MLCD), then data-free latent consistency distillation is proposed to guarantee the inter-segment consistency in MLCD at low cost. \nFurthermore, we introduce bags of techniques to enhance TLCM's performance at rare-step inference without any real data, e.g., distribution matching, adversarial learning, and preference learning. \nTLCM demonstrates a high level of flexibility by allowing for adjustment of sampling steps within the range of 2 to 8 while still producing competitive outputs compared to full-step approaches.\nAs its name suggests, TLCM excels in training efficiency in terms of both computational resources and data utilization.\nNotably, TLCM operates without reliance on a training dataset but instead employs synthetic data for the teacher itself during distillation. With just 70 training hours on an A100 GPU, a 3-step TLCM distilled from SDXL achieves an impressive CLIP Score of 33.68 and an Aesthetic Score of 5.97 on the MSCOCO-2017 5K benchmark, surpassing various accelerated models and even outperforming the teacher model in human preference metrics. \nWe also demonstrate the versatility of TLCMs in applications including controllable generation, image style transfer, and Chinese-to-image generation." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "latent diffusion model", "consistency model", "acceleration" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f3969c19aede76b6812cadcca514c3cbb300b7e1.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "TLCM: Training- efficient Latent Consistency Model for Image Generation with 2-8 Steps" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The description of a vector spectrum by summing the DFT appears a little non-standard (L285-7). Could the authors provide a reference for this approach (potentially outside the INR literature?)\n- In Appendix C it is mentioned that the for image regression the same image is resampled 10 times due to randomness of the model output. The same network should be deterministic in output - does the randomness occur due to re-sampling the network with reinitialisation? \n- The paper mentions that other activations are addressed L131. Have experiments on non-periodic functions (e.g. Gaussian) been performed? \n- Could the authors provide details on the compute required for Fresh on more complex signals like NeRF (L414-6 notes that this is negligible for images). \n- Table 5 records video results. It mentions 'Results for NeRF are provided for reference only as it is incompatible with FreSh'. Could the authors clarify this?\n - Could the authors clarify what is meant by 'direction-dependent frequency magnitudes' and its importance for video (L510)? \n\nMinor Presentation: \n- The presentation of Figure 2 could be improved (visually the figure is difficult to interpret without the caption / main text). Tweaking this figure (perhaps additional labeling) would greatly improve the clarity of the method and extend its impact.\n- The font in figures 12, 13, 14 is inconsistent \n- The text size in Table 6 is inconsistent / appears to have been scaled \n- Figure 5 - suggest changing 'Thousands to steps' to 'Steps ('000)' \n- Figure 1 could be improved (the PSNR labels are difficult to distinguish)\n- Left-hand quotation makes can be done in Latex using the ` character" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "There are a number of strengths in this paper: \n- The issue is well-motivated: selection of frequency is a common frustration when using SIREN implicit neural representations. A method for a-priori selection of frequency rather than needing to grid-search across trained networks will be useful for the community. \n- The method is conceptually simple and should be easy to implement / include in INR pipelines meaning it may be broadly applicable.\n- Experimental results extensively tested with results averaged across seeds for multiple modalities including image datasets (WikiArt, FFHQ, Chest X-Ray, etc), video, and neural radiance fields. \n- The authors have incorporated their method in multiple pipelines (Siren, Fourier, Hashgrid embeddings, Finer). Results broadly indicate that the method leads to a performance improvements within these pipelines. \n- The experiments are well ablated, with key components (e.g. spectrum size n) evaluated with multiple configurations and datasets\n- The paper is well written, logically structured, and clearly presented. Existing literature is reviewed well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a method for selecting the frequency hyperparameter involved in using periodic activations in implicit neural representations. Selection of the frequency hyperparameter is critical to INR performance, and authors often either select arbitrarily e.g. $\\omega=30$, or employ a costly grid-search. The authors propose selecting this hyperparameter based on the target signal frequencies, by minimising the Wasserstein distance between the target signal spectrum and the untrained network output spectrum. This is conducted by searching across candidate frequencies. The authors demonstrate performance improvements across images, neural radiance fields, and videos. Overall this is a nicely written paper which introduces a simple method with interest to the INR community." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are a few minor weaknesses in the paper. These principally relate to small aspects of presentation (e.g. Figure 2). I've listed these in the Questions below as they are largely minor issues rather than critical weaknesses. \n\nIn terms of technical weaknesses, while the method avoids a costly grid search across trained networks it still requires a grid search across candidate frequencies on untrained networks (this cost will be negligible in comparison). In addition, the method introduces an additional hyperparameter (selection of the spectrum cut-off). If I'm reading it correctly, this hyperparameter would still need a trained network to be evaluated (even though the authors note it transfers across signals). This could reduce some of the benefit of the method, especially if the frequency hyperparameter needs to be searched in combination with the frequency cut-off. Could the authors clarify this issue?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "> Similarly, Wire (Saragadam et al., 2023) uses very high frequencies and increases frequencies at each hidden layer, not just at the embedding layer. This makes it incompatible with FreSh\n\nI am curious if you can say more about these models (NeRF and Wire) which use very high frequencies. What purpose do these high frequency components serve if they are not reflected in the training data? Why don’t these models perform poorly as FreSh would predict?\n\nYou mention that your FreSh measurements are noisy due to random initialization, so you average across 10 initializations. [1] find that SIREN’s random initialization of the first layer has a significant effect on the final PSNR (Appendix A.6). I would be interested to see if your method could be further improved by using FreSh to select not only the hyperparameters, but also the random initialization to start from.\n\nIn the appendix you say that there is high variability in the model configurations selected by FreSh, even within a single dataset. But how much of this variability is random vs. actually important for performance? I would be curious to see how per-image FreSh compares to a dataset-wide FreSh: how much of the advantage of FreSh comes from optimizing for each individual image, and how much can be achieved just by making an appropriate hyperparameter choice for each dataset. \n\n[1] https://openreview.net/pdf?id=iKPC7N85Pf" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "Clear, simple, effective solution to a practical problem. Thorough demonstrations that it works. The paper is well-written and well-presented, easy to read and understand. I can imagine this technique being widely adopted for INR hyperparameter optimization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Hyperparameter optimization is a tricky but important problem for implicit neural representations (INRs), since different hyperparameters may be optimal for different signals, but hyperparameter sweeps can be expensive. The authors introduce a fast proxy for determining if an INR’s hyperparameters are well-suited to a particular training signal. Instead of directly optimizing the post-training loss, the authors propose to minimize the Wasserstein distance between frequency distributions of the training signal and the newly initialized INR’s output. This skips the step of actually training each INR, dramatically reducing the time needed to perform a hyperparameter sweep. This proxy appears to perform quite well, conferring most of the performance advantage of a traditional hyperparameter grid search. The authors evaluate their method across a range of image datasets, videos and NeRFs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Section 5 contains a few typos:\n\nSection 5, paragraph 2, “a trail and error approach”\n\nSection 5.1: You fell victim to LaTeX’s backwards quotation marks, one of the classic blunders\n\nOtherwise the paper is quite solid." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Typically, spectral bias means that neural networks tend to fit the high-frequency component of the target signal. Therefore, spectral bias can be reflected as different learning speeds for different frequency components. Although the advanced activation functions such as Sine improve the representation performance, MLPs with Sine still first fits the low-frequency parts as shown in some paper such as “Improved implicit neural representation with Fourier reparameterzied Training”. Inspired by this phenomenon, these methods might not change the bias; rather, it broadens the range of frequencies the model can represent. Under this perspective, Fresh might shift some special spectrum distribution that enjoys the fastest learning speeds of MLPs towards the target signals. So the detailed description of this special spectrum distribution or the validation of the existence of this special spectrum might be an interesting problem. This point will not affect my scoring. If authors could try to find this special spectrum and do more exploration about this problem, I think that it will further deepen our understanding of deep learning." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The method of this paper is fairly easy to understand and flow. This is the first hyperparameter selection method for implicit neural representation models based on the idea of frequency alignment. \n2. The computation cost of this method can be ignored. \n3. Through relatively extensive experiments and the experiments on \"decreasing the default embedding frequency\" in the supplementary materials, this paper well demonstrates that this method could adjust part hyperparameters to adjust the “preferred spectrum” of MLPs, achieve better representation of target signals." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for selecting the hyperparameters for current MLP-based INR models to improve the suboptimal representation performance of models with default hyperparameters. This method measures the Wasserstein distance between frequency distributions of initial-state models and the target signal and selects the hyperparameters that minimize the Wasserstein distance. This paper validates this method on several current INR models via 2D image approximation, 2D video fitting and neural radiance fields. Experimental results demonstrate that this method is effective to some extent." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The related work of this paper seems overly lengthy and overlooks some of the latest methods for spectral bias. For example, spectral bias can be overcame by special initialization and training dynamics adjustment, such as reparameterization and batch normalization. Moreover, as stated in the paper that Fresh is a simple initialization technique in line 024, Fresh should compare to the previous initialization method like “From Activation to Initialization: Scaling Insights for Optimizing Neural Fields”. \n\n2. There lacks a reliable theoretical or experimental link between the key observation (line 019) and Fresh (line 241，Table 2). More concretely, there need an explanation that the adjustment of embedding layer hyperparameters could affect the whole model's spectrum. A more quantitative expression would be better. The illustration in Fig. 1 seems to be vague. Clarification of this point would certainly raise my opinion of the paper.\n\n3. Some experimental results show that Fresh could not provide a better result even worse than baseline models, such as results with time input in Table 5 and in Table 6. Intuitively, Fresh should not obtain worse results. It might show that there are several unclear relationships. Solving the weakness 2. might help this issue.\n\n4. The experiment in Appendix A seems for demonstrating that weights of embedding layer could not be trained sufficiently by gradient descent. However, why is the optimization algorithm SGD? Meanwhile, is the index L2 norm （Eq. 8） used by the paper (720) meaningful? This experiment should be illustrated more clearly." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "n/a" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Why not just have these parameters be learned rather than grid searched (as I believe omega_0 can be learnable at least, I am not sure about the others)? \n- 15,000 steps seems very excessive to fit a single image, how much wall clock does this take on a gpu for good results? Other methods like instant-ngp and spder (above) are much faster.\n- In 5.2 in experiments why is NeRF not used as a baseline also as it is mentioned in the paper? I am not sure which one it is in Table 6." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well written, and introduces a novel method of initialization by fitting the frequency spectrum of the untrained model with the desired downstream task. The logic is sound, and this concept may apply to other fields." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors notice that the initial frequency spectrum of an untrained model’s output correlates strongly with the model’s eventual performance on a given target signal. They want to “pretrain” the untrained model output with the end signal through a hyperparameter search over initialization parameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Evidence\n- The experimental evidence is extremely weak in my opinion and does not demonstrate a significant increase that would warrant other researchers’ usage of this method. Most results show improvements of less than 1dB, which is essentially identical to the naked eye.\n- The results in Table 1 are concerning. After 15,000 steps the difference between the Base Siren model and using Fresh have virtually no difference (PSNRs of 31.18 and 31.81). \n- Table 3 - Why does FreSH have no improvement on Finer and Finer with k=0 and most samples? If it is only applicable to the SIREN and Fourier features model it would be way too limited to warrant an accept as these are several years old baselines.\n- Moreover in Table 6, overall the results are not impressive as most PSNRs increase by < 1dB (i.e. Finer has virtually no change)\n- Table 8, why does Fresh hurt the results when you include time? It seems as expected input for video.\n\nBaselines\n- Paper seems to have a lot of overlap with SPDER https://arxiv.org/pdf/2306.15242 (especially in theory, and that the untrained model output aligns with end reconstruction) and should compare the baseline against it.\n- Other baselines should be (Ramasinghe & Lucey, 2022) which suggests Gaussian activations. It is mentioned in the paper but not reported on to the best of my knowledge.\n- Instant NGP (HashGrid?) may also be used in the image baseline, as it excels in high-resolution images but is not included here for that.\n- Overall, SIREN seems to be a very weak baseline model. I understand that tuning the omega_0 parameter can improve results, but I don’t believe this to be as significant of a contribution on top of the existing model.\n\nComplexity\n- Although computationally cheap (utilizing random initialization and DFT), it adds much more complexity to the setup\n- For example, on the Wasserstein measurement setup, “To prevent this from affecting the selection process, we measure the Wasserstein distance 10 times and use its mean to select.” I feel this is too much of a hack to be reliably used by other researchers, but I may yield to what other reviewers say.\n- Authors could clarify what the parameter values represent in the first section of Experiments for ease of reading.\n\nImpact\n- It is stated the method is incompatible for “extremely high frequencies” (i.e. NeRF and Wire), which makes me doubt its generalizability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024fresh,\ntitle={FreSh: Frequency Shifting for Accelerated Neural Representation Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zMjjzXxS64},\nnote={under review}\n}" }, "abstract": { "value": "Implicit Neural Representations (INRs) have recently gained attention as a powerful approach for continuously representing signals such as images, videos, and 3D shapes using multilayer perceptrons (MLPs). However, MLPs are known to exhibit a low-frequency bias, limiting their ability to capture high-frequency details accurately. This limitation is typically addressed by incorporating high-frequency input embeddings or specialized activation layers. In this work, we demonstrate that these embeddings and activations are often configured with hyperparameters that perform well on average but are suboptimal for specific input signals under consideration, necessitating a costly grid search to identify optimal settings. Our key observation is that the initial frequency spectrum of an untrained model's output correlates strongly with the model's eventual performance on a given target signal. Leveraging this insight, we propose frequency shifting (or FreSh), a method that selects embedding hyperparameters to align the frequency spectrum of the model’s initial output with that of the target signal. We show that this simple initialization technique improves performance across various neural representation methods and tasks, achieving results comparable to extensive hyperparameter sweeps but with only marginal computational overhead compared to training a single model with default hyperparameters." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "spectral bias", "automatic hyperparameter selection", "implicit neural representation", "discrete fourier transform" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/6e20622979b5ee9c810123e9c5d4316d41f6fffe.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/056dfc49e2068f5943905e3f00585770577c2c4f.zip" }, "title": { "value": "FreSh: Frequency Shifting for Accelerated Neural Representation Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Thank you for submitting to ICLR 2025. The research problem is both interesting and timely. I have the following questions:\n\n- The differences between OpenCarbonEval and related works, such as CodeCarbon and Eco2AI, appear to be mainly focused on engineering efforts. For example, they could just refactor the code and adopt equation (1). Could the authors expand on their comparisons from a research perspective or discuss additional experimental results?\n\n- The accuracy of estimating f(t) needs improvement. Could the authors clarify how they derived the function, such as comparing with other possible functions?\n\n- Distributed deep learning systems could be complex. Models mapped onto the same class of GPUs can result in varying energy consumptions. Could the authors provide experimental results on how OpenCarbonEval accounts for differences in energy consumption when adjusting for data parallelism, tensor parallelism, pipeline parallelism, and data offloading? Additionally, how does the carbon footprint scale with the number of GPUs?\n\n- The authors have conducted a substantial number of experiments. Are there any key takeaways from these experimental results? For instance, how does the knowledge of CO₂ emissions impact future training strategies?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- Timely problem\n- Good motivation" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a tool named OpenCarbonEval to estimate energy consumption and carbon emissions during the training process of large ML models. The authors present a new formulation for estimating the training carbon footprint of various models and evaluate the effectiveness of their approach in comparison to related work." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Novelty\n- Soundness\n- Insufficient hardware details" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Address all the questions raised in the weakness points, mentioned in the following again:\n\n1.\tInadequate Justification for the α Parameter: The derivation of the α parameter lacks theoretical depth, as the paper does not substantiate the choice of logarithmic modeling. Providing empirical or theoretical evidence for using f(t)=ln(1+αt) would strengthen its validity; a comparison with alternative functions could clarify this choice.\n2.\tLimited Model Generalization: OpenCarbonEval does not convincingly show its ability to generalize across diverse ML tasks and architectures. The adaptability of the α parameter remains unclear, particularly for models outside the initial dataset. Additional validation across a wider range of model types by extending Table 1, 2 will reinforce its versatility. Detailed results for the validation of the method is required.\n3.\tLack of Explanation for Equations: The paper lacks the connection between equation (2) and equation (3), and also lacks the explanation of how the Lcomputation is used to estimate the energy consumption E. Moreover, the Clifelong needs to be elaborated in terms of how it is attained. \n4.\tComparison with results for LLMCarbon: Can the authors present the analysis of same models and hardware combinations presented in Table 4 in the LLMCarbon paper?\n5.\tJustification or Citation for Assumption: The assumption of 1-year GPU lifespan for the embodied carbon estimation lacks justification or citation from a reliable source. \n6.\tOverlooked Factors in Operational Carbon Calculation: OpenCarbonEval does not account for essential factors like Power Usage Effectiveness (PUE) in data centers, leading to potential underestimations of emissions. Including PUE in calculations would create a more realistic operational carbon estimate.\n7.\tSimplistic Treatment of Training Dynamics: OpenCarbonEval applies Little’s Law simplistically, assuming a steady state in training dynamics, which oversimplifies the training process. More practical grounding, perhaps through empirical evidence, would enhance applicability in ML contexts. LLMCarbon addresses this by using detailed hardware efficiency and optimal parallelism settings, providing a robust framework for accurately modeling training dynamics.\n8.\tEmbodied Carbon Calculation: OpenCarbonEval’s approach to embodied carbon appears oversimplified, lacking in-depth parameters that affect emissions, such as hardware-specific manufacturing and lifetime estimates. Moreover, the Clifelong needs to be elaborated in terms of how it is attained.\n\nOpenCarbonEval addresses a timely topic and proposes interesting methods for estimating emissions, but fundamental conceptual and methodological gaps needs to be clarified. Without rigorous validation, robust comparisons, and clearer theoretical grounding for key parameters, the method may not yet be practical for diverse ML scenarios. Addressing these weaknesses could make OpenCarbonEval a valuable contribution in the future." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1.\tRelevant Topic: The environmental impact of large ML models is an important concern, and OpenCarbonEval’s focus on a general framework for carbon footprint estimation. OpenCarbonEval showcases improved error rate in comparison to LLMCarbon across various large-scale ML models.\n\n2.\tMulti-Domain Scope: The method’s attempt to generalize across model types, hardware types, and tasks, potentially making it more versatile than existing carbon estimation among the estimation methods. \n\n3.\tDataset Creation: OpenCarbonEval contributes an open resource by curating a dataset of carbon emissions containing 110 records." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents OpenCarbonEval, an innovative approach for estimating the carbon footprint of training large ML models, with claims to improve prior models by incorporating hardware-specifications, embodied and operational carbon estimation, and dynamic power consumption. It introduces an α parameter to model dynamic power consumption and introduces an open-source dataset of 110 models across multiple large-scale ML tasks for validation of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tInadequate Justification for the α Parameter: The derivation of the α parameter lacks theoretical depth, as the paper does not substantiate the choice of logarithmic modeling. Providing empirical or theoretical evidence for using f(t)=ln(1+αt) would strengthen its validity; a comparison with alternative functions could clarify this choice.\n2.\tLimited Model Generalization: OpenCarbonEval does not convincingly show its ability to generalize across diverse ML tasks and architectures. The adaptability of the α parameter remains unclear, particularly for models outside the initial dataset. Additional validation across a wider range of model types by extending Table 1, 2 will reinforce its versatility. Detailed results for the validation of the method is required.\n3.\tLack of Explanation for Equations: The paper lacks the connection between equation (2) and equation (3), and also lacks the explanation of how the Lcomputation is used to estimate the energy consumption E. Moreover, the Clifelong needs to be elaborated in terms of how it is attained. \n4.\tComparison with results for LLMCarbon: Can the authors present the analysis of same models and hardware combinations presented in Table 4 in the LLMCarbon paper?\n5.\tJustification or Citation for Assumption: The assumption of 1-year GPU lifespan for the embodied carbon estimation lacks justification or citation from a reliable source. \n6.\tOverlooked Factors in Operational Carbon Calculation: OpenCarbonEval does not account for essential factors like Power Usage Effectiveness (PUE) in data centers, leading to potential underestimations of emissions. Including PUE in calculations would create a more realistic operational carbon estimate.\n7.\tSimplistic Treatment of Training Dynamics: OpenCarbonEval applies Little’s Law simplistically, assuming a steady state in training dynamics, which oversimplifies the training process. More practical grounding, perhaps through empirical evidence, would enhance applicability in ML contexts. LLMCarbon addresses this by using detailed hardware efficiency and optimal parallelism settings, providing a robust framework for accurately modeling training dynamics.\n8.\tEmbodied Carbon Calculation: OpenCarbonEval’s approach to embodied carbon appears oversimplified, lacking in-depth parameters that affect emissions, such as hardware-specific manufacturing and lifetime estimates. Moreover, the Clifelong needs to be elaborated in terms of how it is attained." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 1 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please comment on the following weakness:\n\n1. **Simplified yet more accurate formulation???**: The functions presented in Equations (3) through (7) lack clarity in their intended function and accuracy. While polynomial approximations may lack precision, Equation (7) is simplified even further than LLMCarbon, containing only a single parameter compared to the multi-parameter nature of polynomial approximations. Why is this single-parameter approach purported to yield higher accuracy? The authors are encouraged to offer detailed explanations or empirical validation demonstrating how and why Equation (7) leads to improved accuracy over traditional polynomial approximations.\n\n2. **Consideration of GPU count and parallelism settings**: The paper does not discuss varying GPU counts in training configurations, appearing to assume a single-GPU setup. It also does not address different training parallelism types, such as data, tensor, pipeline, or expert parallelism, all of which may affect results depending on GPU count. Without incorporating these parallelism factors, it is unclear how OpenCarbonEval achieves greater accuracy. How does this work account for different parallelism strategies, and are there empirical results confirming its accuracy across these configurations? Additionally, Figure 4 lacks context: how many GPUs are represented, why do some GPUs exhibit smaller variance, and how many GPUs are used for training in Tables 1 and 2?\n\n3. **Lack of model architecture information**: The study appears to consider only the number of parameters in ML models, without accounting for architecture specifics. While scaling laws suggest that architecture does not impact model accuracy, it significantly affects training throughput across various architectures (see Megatron paper: https://parsa.epfl.ch/course-info/cs723/papers/Megatron.pdf). The authors should provide empirical evidence to demonstrate that model architecture does not impact the carbon footprint of training.\n\n4. **Dataset limitations**: The dataset used is limited and lacks comprehensive real-world data. Among the 863 entries in the provided table (https://epochai.org/data/notable-ai-models?view=table), only 176 entries include training times, 158 provide GPU counts, and only 31 report hardware utilization, leaving most entries without training times or hardware utilization data. With such limited information, how is \\( f(x) \\) in Equation (5) trained and validated? Furthermore, 603 of the 863 entries are classified as \"likely,\" \"speculative,\" or \"no confidence.\" Does OpenCarbonEval rely on these uncertain data points for validation while claiming higher accuracy? The authors should discuss the limitations associated with the dataset quality and address the impact on the reliability of their conclusions." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The paper works on an important topic.\n2. The paper identifies the shortcoming of preivous works (Faiz et al., 2023): the polynomial approximation for the system efficiency and hardware utiliation estimation is not accurate." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The training of machine learning (ML) models significantly contributes to global carbon emissions. This paper introduces OpenCarbonEval, an advanced estimation tool designed to quantify the carbon impact of large-scale ML models based on their total training computations and hardware configurations. The tool's accuracy is validated using real-world datasets, and experimental results demonstrate that OpenCarbonEval provides more precise predictions of energy consumption and carbon emissions than previous approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Simplified yet more accurate formulation???**: The functions presented in Equations (3) through (7) lack clarity in their intended function and accuracy. While polynomial approximations may lack precision, Equation (7) is simplified even further than LLMCarbon, containing only a single parameter compared to the multi-parameter nature of polynomial approximations. Why is this single-parameter approach purported to yield higher accuracy? The authors are encouraged to offer detailed explanations or empirical validation demonstrating how and why Equation (7) leads to improved accuracy over traditional polynomial approximations.\n\n2. **Consideration of GPU count and parallelism settings**: The paper does not discuss varying GPU counts in training configurations, appearing to assume a single-GPU setup. It also does not address different training parallelism types, such as data, tensor, pipeline, or expert parallelism, all of which may affect results depending on GPU count. Without incorporating these parallelism factors, it is unclear how OpenCarbonEval achieves greater accuracy. How does this work account for different parallelism strategies, and are there empirical results confirming its accuracy across these configurations? Additionally, Figure 4 lacks context: how many GPUs are represented, why do some GPUs exhibit smaller variance, and how many GPUs are used for training in Tables 1 and 2?\n\n3. **Lack of model architecture information**: The study appears to consider only the number of parameters in ML models, without accounting for architecture specifics. While scaling laws suggest that architecture does not impact model accuracy, it significantly affects training throughput across various architectures (see Megatron paper: https://parsa.epfl.ch/course-info/cs723/papers/Megatron.pdf). The authors should provide empirical evidence to demonstrate that model architecture does not impact the carbon footprint of training.\n\n4. **Dataset limitations**: The dataset used is limited and lacks comprehensive real-world data. Among the 863 entries in the provided table (https://epochai.org/data/notable-ai-models?view=table), only 176 entries include training times, 158 provide GPU counts, and only 31 report hardware utilization, leaving most entries without training times or hardware utilization data. With such limited information, how is \\( f(x) \\) in Equation (5) trained and validated? Furthermore, 603 of the 863 entries are classified as \"likely,\" \"speculative,\" or \"no confidence.\" Does OpenCarbonEval rely on these uncertain data points for validation while claiming higher accuracy? The authors should discuss the limitations associated with the dataset quality and address the impact on the reliability of their conclusions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024opencarboneval,\ntitle={OpenCarbonEval: How much \\${CO}\\_2\\$ will your large model exhale in training process?},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zNVefjN3EP},\nnote={under review}\n}" }, "abstract": { "value": "Data, model and hardware are crucial components in the development of large scale machine learning models. The training of such models necessitates substantial computational resources, energy consumption, and raw materials, resulting in significant environmental implications. However, the environmental impact of these models has been largely overlooked due to a lack of assessment and analysis of their carbon footprint. In this paper, we present OpenCarbonEval, a carbon emission estimation framework to quantify the environmental implications of large scale machine learning models given their total training computations and hardware configurations.\nIn OpenCarbonEval, we conducted a comprehensive dynamic analysis of the interrelationships among data, models, and hardware throughout the model training process, aiming to forecast the carbon emission of large scale models more accurately. We validated our approach on real-world dataset, and experimental results demonstrate that OpenCarbonEval can predict energy costs and carbon emissions more accurately than previous methods. Furthermore, it can be seamlessly applied to various machine learning tasks without a precision decline. By quantifying the environmental impact of large-scale models, OpenCarbonEval promotes sustainable AI development and deployment, contributing to a more environmentally responsible future for the AI community." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large-scale model", "Carbon footprint", "Sustainable AI" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/21de26db970c47e03922ea15e488742bea430b6c.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/394081c87d9bafef81d82d22278e391edf52ada8.zip" }, "title": { "value": "OpenCarbonEval: How much $CO_2$ will your large model exhale in training process?" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "A few additional comments:\n - One of the problems I have is that there already exist many approaches based on the SVD. \n- Line 69/70 “Sparse Reverse of the Principle Component Analysis” —> “Principal Component Analysis”\n- You should really simplify some of your introduction. E.g. “The proposed approach maintains a high level of smoothness by iteratively finding the principal components of the matrix based on the predicted values of both the missing and the observed parts of the matrix… ”. The sentence is way too long. I think I would remove the whole paragraph. \n- I would remove at least 2 or 3 items from page 2 (lines 76 to 90) and give the algorithm. You mention your algorithm only vaguely by saying that it updates a single component. It would be a lot clearer to have a short, clear exposition of the algorithm (even informal). E.g. just give (12) - (14) with a few words of explanation. \n- line 128/129, in your exposition of ALS, it is not clear why the entries are partitioned into subsets. That would be worth and explanation. Also when is the SVD computed. You say that it is not applied at each iteration. When is it applied then?\n- line 136/137 “we refer readers ” —> “we refer readers to…”\n- You should reorganize your explanation of PCA on lines 150 to 158. Start by something like “Let M denote any data matrix. Let V denote the eigenvectors of M^T M. We then define the matrix P as P = MV. The matrix P is diagonal. (You can then give (7)) ”. You don’t need more than 2/3 lines to recall how PCA works\n- Lines 166 - 167, when you introduce P_\\Omega, the proper technical term is “mask”. \n- Lines 166-167 “the matrix that preserves the entities”, “and replaces the remaining entities” —> “entries”\n- The notation “\\Omega^\\perp” is usually used to denote the orthogonal complement. I would use $\\Omega^c$ for the complement.\n- On lines 197-198, you first say that a fair intuitive approximation for M is (11) but then you seem to indicate that SRPCA starts with the restriction to the mask X_\\Omega? Which initialization do you use? This is not clear. If you don’t use (11), this should go \n- lines 226- 229 are not clear. Do you mean that the value of the matrix on the mask does not change? This does not seem true to me as when you compute the PCA decomposition you do not necessarily maintain the values on Omega. In any case, this should be rephrased. \n- In your statement of Lemma 3.1., Why not replace P^{(k-1)}R^{(k)} by M^{(k+1)} ? Moreover, your statement does not imply convergence. The convergence would require a strict inequality (which I guess you have, yet just made a typo). Also the expression converges iteratively does not really mean anything. You can just use “converges” or in this case, you can even say “converges linearly with respect to the ||P_\\Omega()||“ semi-norm\n- line 262/263 “This is a key finding, because if the algorithm terminates for external reasons ” this is unneccesary\n- line 270 - 292, in Algorithm 1, how do you determine the number of components that you retain. E.g. how do you set the value of r on line 278?\n- The statement of Lemma 3.2. is unclear. What do you mean by “it converges at an iteration K”\n- On Figure 2, the contrast does not seem to be the same between SRPCA and ALM\n- line 300-301 “which start each iteration with teh” —> “with the”\n- Lines 327 - 334, why are some of the convergence time and reconstruction errors underlined? You should clarify this.\n- Line 312-313, the statement of Lemma 3.3. does not mean anything. What is K? From what I understand, K is defined with respect to a given tolerance epsilon_tol. This should be clarified.\n- Line 312- 313, I find Lemma 3.3. somewhat misleading. From what I understand, the Fast RPCA algorithm corresponds to a change in the stopping criterion of Algorithm 1. I.e it basically tells you you should add the criterion ||P_{\\Omega}(X - P^{(k-1)}R^{(k)})||_F^2< delta on top of the stopping criterion on the successive iterates. Why not merge this with algorithm 1? It would make sense to me to have a combined criterion of the form “if deviation between iterates < epsilon or error on Omega < delta then stop“ It sort of makes sense that if your error on Omega get very small, you won’t be able to make further progress. You can keep part of the discussion from section 3.5 if you want but to me it would make more sense to merge the Fast SRPCA and the SRPCA algorithms (possibly with a short justification from section 3.5). This would be much clearer." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The possibility to parallelize has some interest, the disadvantage being that you have to compute an eigenvalue decomposition at each step and to invert a matrix of size r by r (see (16)) which although not as computationally expensive as a full inverse can remain quite expensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an iterative algorithm for matrix completion which the authors call Sparse Reverse Princial Component Analysis (SRPCA). The algorithm is based on iteratively (1) computing a PCA decomposition and (2) updating the principal vectors through a smooth gradient descent step." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concern has to do with the fact that the algorithm remains very close to existing algorithms (e.g. SVT from Cai et al. ) based on a combination of a gradient step and a truncated SVD. There exists an extensive body of work on matrix completion and any new approach (although perhaps interesting) would require a serious analysis and comparison with the literature. Had the authors shown a clear improvement in the rate of convergence (not just numerical as shown in Table 1) or in the number of samples needed for the recovery of the matrix, I would have been more inclined to accept the paper. But given that they only show convergence, I have to say I feel a little undecided. Perhaps the improvement should be better documented. E.g. what is the rate of convergence of SVT for example?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "n/a" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* Rephrasing the presentation in terms of the singular value decomposition would help elucidate what the updates correspond to. While an attempt is made at connecting R to an eigenvector matrix, after the first iteration R is no longer an eigenvector matrix and P no longer corresponds to principal components as traditionally defined. Further, the differences between the updates proposed and those performed in ALS should be clear.\n* The authors claim that the proposed algorithm leads to “a performance boost for smooth matrices”. At a minimum, the authors should precisely define the notion of smoothness used.\n* It would be helpful to have working code that fully reproduces Figures 1-3, including all comparisons to other methods.\n* Could you make a phase transition plot indicating how performance degrades as a function of rank and proportion of missing data? See Figure 1 in https://arxiv.org/abs/0805.4471 and the associated description for the setup.\n* One of the claims in the abstract is that the proposed method results “in a significant convergence acceleration“. To demonstrate this, the authors should include convergence plots depicting the reconstruction error of the algorithm (on the y-axis) as a function of the amount of iterations or time elapsed (on the x-axis), and repeat this for all methods compared in the experiments section. See Figure 1 in https://arxiv.org/abs/1605.07051 for an example.\n\nMinor comments\n\n* Line 110: apositive -> a positive\n* Line 115: algorithm.Unfortunately -> algorithm. Unfortunately\n* Line 119: you’re relaxing the minimization problem (1), not relaxing the rank\n* Line 147-161: Variables are used before they are defined, including P, V, S_p, and n\n* Line 168: entities -> entries\n* Line 170: minimize the problem -> reduce the problem\n* Line 11 in Algorithm 1: R(k) is used before it’s defined — probably should reorder lines 11-13.\n* Line 13 in Algorithm 1: It is unclear that this matrix contains eigenvectors of any relevant matrix after the first update. It is also unclear that it will have orthonormal columns after the first update.\n* Line 300: teh -> the\n* Line 410: enignes -> engines\n* Line 510: closed for -> closed form\n* Line 519: “can influence from“ What does this mean?\n* Line 520: predicitons -> predictions" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The proposed methods (SRPCA and fastSRPCA) appear novel.\n* The numerical results shown are promising. In particular, Figure 1-3 shows that SRPCA is significantly outperforming other methods, especially in the presence of large amounts of missing data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new iterative method for solving matrix completion problems in the presence of high levels of missing data. The proposed method is similar to ALS; the algorithm alternates between finding a least-squares solution (which has the benefit of being a convex problem) and performing a heuristic update. Theoretical claims of monotonic convergence of the iterates in finitely many steps are presented. A modified, faster-converging version of the algorithm is also presented. Finally, the authors provide numerical evidence of impressive performance on three datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The key theoretical result (monotonic convergence to within epsilon tolerance in finitely many steps) is highly misleading. The termination/convergence criterion of the algorithm is previous_reconstruction_error - current_reconstruction_error < tol without absolute value bars. This means that as soon as an iterate has larger reconstruction error than the previous one, the algorithm will terminate. Trivially this means that the algorithm will monotonically decrease for as long as it runs precisely because it terminates as soon as monotonicity fails. However, this does not give any insight into whether the final iterate has, for example, good reconstruction error. \n* While the numerical results shown are promising, they are limited overall. The key motivation is performance in the presence of limited data, so a phase transition plot indicating performance as function of amount of data present would be helpful. Further, the discussion focuses on how the algorithm converges very quickly, but no convergence plots are included. Finally, important parts of the experiments are not reproducible; the code provided in the supplement that creates Figures 1-3 does not run because of a missing file. (See suggestions below for details on what kinds of numerics would be helpful.) \n* The presentation lacks clarity. The connection to ALS or alternating minimization methods is not made explicit in the text despite the fact that the algorithm can be interpreted as a variation of ALS. Specifically, the update to R(k) is solving a least-squares problem (i.e., one step of ALS). The update to P(k) is a heuristic update, but it is unclear what minimization problem this update is solving, if any. \n* The “fast” variant of the proposed method (fastSRPCA) is poorly motivated. While the authors claim that it improves convergence at the cost of slightly degraded reconstructions, there is no empirical or theoretical comparison between fastSRPCA and SRPCA. \n* Numerous other claims are unsubstantiated or false.\n * 197: Why is this a “fair and intuitive first approximation?”\n * 216: “The new update is smoother than the prior update.” This claim is never substantiated, nor is the relevant notion of smoothness defined.\n * 220: “This adds a layer of nonlinearity...” Where is the nonlinearity? All the iterative updates in SRPCA are matrix multiplications, which are linear.\n * 244: “Applying (16) is scalable for big data.” While the author points out the opportunity for parallelization, there is no analysis of the complexity of the algorithm with or without parallelization. This should be a straightforward calculation, and would lend credence to the fact that the algorithm could be scaled to large datasets." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What is the time complexity of the proposed method? Please add a time complexity analysis and provide the running time comparison in the experimental results section.\n2. As the current experiments are small-scale data, how about the performance on large-scale data such as large images or MovieLens 1M? Will the time cost of the proposed method increase significantly?\n3. The current experiment on image completion does not consider noisy situations. How about the performance of image completion with Gaussian noise?\n4. Please compare the performance with the following work if possible: Li, Chao, Wei He, Longhao Yuan, Zhun Sun, and Qibin Zhao. \"Guaranteed matrix completion under multiple linear transformations.\" In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11136-11145. 2019." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The method and theoretical analysis are technically sound. Experimental results show some advantages of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the author proposes a new matrix completion method that uses a PAC-like update regime. The convergence analysis is provided, and a fast version is developed. Experiment verifies the desired performance of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The time complexity is limited due to the utilization of SVD. The efficiency of large-scale data should be further investigated." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Q1) Why do you use the terminology \"sparse reverse of PCA\"?\n\nQ2) Why should the new estimate $M^{(k+1)}$ be smoother than $M^{(k)}$, as claimed in lines 216-217, and in which sense exactly? \n\nQ3) Line (13) is labeled \"Eigenvectors update\". It seems to refer to the eigenvectors of $(M^{(k)})^T M^{(k)}$, that is, the right singular vectors of $M^{(k)}$. But these should be orthogonal, and I don't see how the algorithm enforces this constraint. Otherwise, how is the update of $P^{(k)}$ (line 11 of Algorithm 1) supposed to work?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "n/a" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is focused on the matrix completion problem. It essentially employs a Burer-Monteiro factorization approach with alternating updates, after resetting the observed entries according to their given values at each iteration. A variant is proposed that interpolates between these observed values and the ones estimated by the algorithm, but the hyperparameter choice advocated by the authors effectively suppresses the interpolation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1) This reviewer does not see any new ideas or useful insights brought by this paper. Instead, the proposed algorithm just seems to be a straightforward combination of well-known ideas: it employs a standard low-rank factorization with factors denoted $P$ and $R$ (even though it claims in the abstract to propose a reformulation of this standard approach) and updates the factors in an alternating fashion. Worse still, it seems to claim that some obvious and widespread techniques are innovative, such as simply setting the observed entries to their given values along the iterations (see lines 201-202).\n\nW2) The paper claims multiple times to \"maintain smoothness across the reconstructed matrix\", but offers no concrete elements as to why this should be the case. In fact, based on what is promised multiple times (including in the abstract), one would expect a formalization of these ideas, and some sort of regularization or constraint being enforced for promoting smoothness. Instead, only quite vague and superficial comments are made, trying to imply that the simple, standard updates given in the paper are somehow capable of automatically promoting smoothness in the solution. In particular, in lines 216-217 it is said that the new estimate $M^{(k+1)}$ is \"smoother\" than $M^{(k)}$, but no formalization of what is meant, proof or even concrete heuristic arguments are given to support this claim.\n\n\nW3) The paper suffers from a significant lack of clarity. Many sentences are poorly written, contain unusual expressions, improper formalization, or misinterpret basic notions such as convergence. Some examples are:\n- \"converges iteratively\"\n- \"SRPCA is proved to improve in performance iteratively until convergence\" (simply meaning that the objective function is non-increasing along iterations)\n- Lemma 3.2 seems to mistake the convergence of an algorithm with its stopping. It claims to prove an upper bound on the number $K$ of iterations required for convergence, but it merely shows that the algorithm stops, as a consequence of the imposed stopping criterion.\n- The solution to the problem in (14) is denoted with the same symbol as the optimization variable.\n\nW4) The paper does not properly motivate its contents. It only vaguely mentions some standard difficulties faced by certain classes of algorithms, but does not explain in a clear way how exactly it proposes to overcome such difficulties. In particular, the alternating least-squares (ALS) approach is said to operate by randomly partitioning the observed entries into subsets (known as resampling), and then is criticized for ignoring \"the smoothness of the data set due to the random partitioning of the original matrix.\" Yet, the authors are only focusing on a particular version of this approach: there is no reason why ALS must be employed with such a random partitioning in the first place (see, e.g., arXiv:1312.0925 and arXiv:1411.8003, the latter one containing a discussion on resampling).\n\nW5) In the extension of Section 3.5, a parameter $\\alpha^* \\in [0,1]$ is introduced to interpolate the observed entries between their currently estimated and given values, instead of fixing them according to the latter. However, the authors propose a heuristic that amounts to setting $\\alpha^* = 1$ until the difference (iin Frobenius norm) between the model and the observations is small on those entries, and then switching to $\\alpha^* = 0$, at which point the algorithm is shown to stop. Therefore, the purported \"smoothing\" advantages brought by the interpolation are eliminated, and the introduced variant becomes equivalent to the previously proposed scheme. \n\nW6) There are many typos along the text, indicating that it was not properly revised prior to submission." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024srpca,\ntitle={{SRPCA}: Sparse Reverse of Principal Component Analysis for Fast Low-Rank Matrix Completion},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zNfdtV9ADQ},\nnote={under review}\n}" }, "abstract": { "value": "Supervised and unsupervised learning methods experience a decline in performance when applied to incomplete, corrupted, or noisy datasets. Matrix completion is a common task to impute the missing values in sparsely observed matrices. Given a matrix $\\mathbf{X} \\in \\mathbb{R}^{m \\times n}$, low-rank matrix completion computes a rank-$r$ approximation of $\\mathbf{X}$, where $r\\ll\\min\\\\{m,n\\\\}$, by only observing a few random entries of $\\mathbf{X}$. It is commonly applied for recommender systems, image processing, and multi-output collaborative modeling. Existing matrix completion methods suffer either from slow convergence or failure under significant missing data levels. \nThis paper proposes a novel approach, the Sparse Reverse of Principal Component Analysis (SRPCA), that reformulates matrix factorization based low-rank completion $(\\min_{\\mathbf{U},\\mathbf{V}}\\Vert\\mathcal{P}_{\\mathbf{\\Omega}}(\\mathbf{X}-\\mathbf{U}\\mathbf{V}^T)\\Vert_F^2)$\nto iteratively learn a single low-rank subspace representation by solving the convex optimization problem\n$\\min\\_{\\mathbf{V}}\\Vert\\mathcal{P}\\_{\\mathbf{\\Omega}}(\\mathbf{X}-\\mathbf{P}\\mathbf{V}^T)\\Vert_F^2$ under the principal component analysis framework, resulting in a significant convergence acceleration. SRPCA converges iteratively and is computationally tractable with a proven controllable upper bound on the number of iterations until convergence. Unlike existing matrix completion algorithms, the proposed SRPCA applies iterative pre-processing resets that maintain smoothness across the reconstructed matrix, which results in a performance boost for smooth matrices. The performance of the proposed technique is validated on case studies for image processing, multivariate time-series imputation, and collaborative filtering. SRPCA is also compared with state-of-the-art benchmarks for matrix completion." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "matrix completion", "low rank", "PCA", "collaborative filtering", "image inpainting", "time-series imputation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/428187414791443d2e36ff927d2e2b379f0623ed.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/6f7879bbf09282b0d842098269b1487efe00ccb3.zip" }, "title": { "value": "SRPCA: Sparse Reverse of Principal Component Analysis for Fast Low-Rank Matrix Completion" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I'm curious about the performance of advanced models (e.g., Whisper v3 and Llama 3) in building a cascade pipeline. Since this work aims to explore LLMs for speech processing, and LLMs are effective in handling long text and summarization tasks, how would these models compare?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is clearly structured and represents a valuable exploration of using LLMs for long speech processing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores the use of Mamba-based multimodal LLMs to process long speech segments. The authors also apply DPO to enhance alignment during the instruction fine-tuning stage. Their experiments focus primarily on the speech summarization task, showing that the model can successfully process long speech." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although the paper explores long speech processing using LLMs, the motivation is not sufficiently compelling. Notably, LLMs excel at summarization, and long speech summarization could be effectively handled by combining an ASR model with a strong LLM. While ASR may introduce some errors, LLMs are generally robust enough to manage this task. Therefore, long speech summarization may not be the most suitable task for evaluating LLMs in handling extended long speech inputs.\n- The contribution is somewhat limited, as Mamba-based multimodal LLMs and DPO have both been explored in speech instruction tuning. This work primarily combines these two methods and tests them on the speech summarization task.\n- The experiments are insufficiently comprehensive; relying only on speech summarization does not robustly support the model’s capability with long speech. Furthermore, the LLM used here is not particularly strong.\n\nOverall, this paper lacks novelty and persuasiveness and would benefit from further development." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Q-former-like mechanism should be able to compress input of any length into fixed length. Hence I'm a bit confused about the decision to compress every 0.33s of audio into 2 query vectors. This differs from previous works including arxiv:2309.13963 and arxiv:2407.02005, which considered 30~100 vectors for every 30s of audio. In this way, the contextual information further than 0.33s will not be captured by the Q-former. I guess that the reason is to avoid the high-cost quadratic cross-attention but it will be better for the author to discuss that explicitly. Also, compressing a short sequence (only dozens of vectors, as each Whisper frame takes 25ms) into merely 2 vectors is rather simple and I doubt if the complicated Q-former will really outperform a much simpler one, e.g. pooling or convolution, which may also process information within such a short context well. More ablation studies will be necessary to justify this decision, by comparing with pooling or CNN, and comparing with different context lengths.\n\nThere are many other approaches to compress speech signals into \"token-like\" embeddings to be processed by LMs, e.g. HuBERT units, speech tokens, and neural audio codec, while Q-former is somehow similar to a kind of VQ, but with continuous features. Can you elaborate on the reasons why you chose Q-former? Do you think there is any specific advantage?\n\nI am particularly concerned with the unidirectional Mamba used in Q-Mamba, and I fail to find the motivation to apply Mamba to the sequence of query vectors. Trainable query vectors should be already capable of introducing positional information. Ablation studies (e.g. by removing this Mamba layer) should be necessary to justify this choice.\n\nI also have some questions regarding the use of DPO. What is the experiment without DPO in Table 4? Using supervised fine-tuning only?\n\nIf the issue w/o DPO is that the summary will be too detailed, has the author considered any other more straightforward solution, e.g. length penalty during generation, downsampling the input sequence, or upweighting EOS during training?\n\nIt is commonly believed that instruction fine-tuning leads to better alignment, but at the cost of flexibility and adaptability to specific downstream tasks in fine-tuning, while the authors use a instruction fine-tuned version of Mamba-2.8B as the base LLM. Is there any specific reason to use Mamba-2.8B-Zephyr instead of the original Mamba-2.8B model?\n\nWhat is the LLM used in the Cascaded model? The original Mamba-2.8B or the instruction fine-tuned one? Is it further fine-tuned to summarization?\n\nCan you elaborate more on the speedup of the model compared to the cascaded one? With both of them using Whisper and Mamba (though the inputs to Mamba are different), I'm curious about the source of the extra overhead in the cascaded pipeline. Also, it can be helpful to report the average input sequence length to the final LLM model as a reference to the expected computational costs.\n\nUsing only synthesized datasets is a weak point of the empirical evaluation, particularly when the labels are also synthetic. It can be necessary to also report the results on real datasets, e.g. SLUE-SUMM, and include more examples and human evaluations.\n\nIt can be interesting to report the ASR performance of the model after either of the two Alignment stages.\n\nIt will be better to also include the original transcript in Appendix C.\n\nIn Figure 3, is Whisper frozen in the Fine-tuning Stage?\n\nMinor issues:\nL210: Figure 4.1?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The model enables efficient speech summarization by combining pretrained Mamba and Q-former.\n* Empirical results are strong." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a sub-linear complexity speech summarization model by combining Q-former and pretrained Mamba. Segmented audios are processed by Whisper and compressed by Q-former with Mamba-processed query vectors, which are then fed to Mamba LLM to generate the summarization. A 3-stage training with different tasks is carried out, i.e. short-form ASR, long-form ASR, and summarization, accompanied by DPO. Empirical studies show better results and speed compared to cascaded models and a HuBERT+LLAMA E2E model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Ablation studies are not sufficient to demonstrate the advantages of the proposed methods and to identify the impact of each component. Some design choices are yet to be well-motivated.\n* The method is more like replacing transformers in existing methods (esp. arxiv:2407.02005) with Mamba, which leads to doubt on the technical novelty.\n* Only one synthetic dataset is used.\n\nSee questions below for details." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "/" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Novel Architecture:\n\nInnovative combination of Mamba with querying-attention for speech processing.\n\nWell-motivated design choices for handling long-form speech.\n\nClear architectural improvements over existing approaches.\n\n- Strong Empirical Results:\n\nSignificant speed improvements (17x faster than cascaded baseline).\n\nCompetitive performance on standard metrics.\n\nComprehensive ablation studies validating design choices.\n\n- Technical Soundness:\n\nThorough theoretical foundation and clear mathematical formulation.\n\nWell-documented training process and implementation details.\n\nCareful experimental design with appropriate baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents SQuBa, an end-to-end speech summarization model that combines a Mamba language model with a novel querying-attention mechanism for efficient processing of long speech inputs. The key contributions are:\n\n- A query-attention Mamba projector that compresses acoustic features into compact semantic tokens.\n- Extension of Mamba-based LLM for speech summarization with a two-stage training process including bootstrapped DPO.\n- Empirical demonstration of competitive performance with significantly faster inference speeds compared to transformer-based approaches.\n\nThe model achieves this through:\n\n- Using Whisper encoder for speech features.\n- Novel Q-Mamba projector for efficient feature compression.\n- Pre-trained Mamba LLM (2.8B parameters) for generation.\n- Two-stage training: speech alignment followed by summarization fine-tuning.\n- Bootstrapped DPO for improved summary quality." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Limited Dataset:\n- Uses synthetic speech data for fine-tuning\n- Could benefit from evaluation on more diverse real-world speech datasets\n- Lack of cross-lingual evaluation\n\nArchitectural Constraints:\n- Fixed 30-second chunks due to Whisper encoder limitations\n- Query length choices could use more theoretical justification\n- Potential information loss in compression not fully analyzed\n\nEvaluation Metrics:\n- Limited human evaluation or qualitative analysis\n- No discussion of failure cases or limitations" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In Equation 1, what is h'(t) ? \n2. The \"ideal\" query length for speech transcription is likely not representative of representations necessary for speech summarization. Can the authors clarify why these ablations were done for the transcription task on Librispeech ?\n3. How does the speed of Scuba compare to that of the model by Shang et al ? \n4. The Whisper speech encoder is frozen, and it is not clear why this modeling choice was made." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper attempts to address an important challenge, i.e., summarization of longform audio through a cross-attention based temporal downsampling module. \n2. It applies the recently introduced DPO technique to speech summarization, and demonstrates improved ROUGE and METEOR scores from this." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper describes an approach to summarizing 6-minute-long audio recordings by combining the Whisper speech encoder with the Mamba LLM through a cross-attention based Mamba querying projector. The authors show that DPO improves ROUGE and METEOR metrics, and that the proposed model has a better ROUGE and METEOR score, and latency over the cascade model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I have serious concerns about the novelty of the proposed approach. From a modeling standpoint, the work is very similar to Shang et al. (2024). From a training method standpoint, the 2-stage fine-tuning approach involving speech recognition and speech summarization is well established in the field since Sharma et al. (2021), leaving only two differences: (a) having an ASR training stage over both short and long audio as opposed to just short audio, and (b) using DPO post hoc, another well established technique to improve ROUGE and METEOR numbers. All in all, it appears that there is little technical novelty in the paper. \n\n2. Validating the proposed approach on a single relatively shortform audio dataset (upto 6 minutes) comprising synthetic audio is not very convincing. Furthermore, the work is done on a custom dataset whose LLM-generated summary labels have not been validated for correctness, either through automatic or human evaluations. Since LLMs are known to hallucinate, it is hard to make a meaningful case using any numbers on this dataset. The authors should ideally consider evaluating on any real other dataset(s) with real audio. \n\nTo add more context, here is a paper [1] that used synthetic data for speech summarization but still reported a myriad of automatic and human evaluation metrics to validate that the data used was reasonable. Something similar to this might be more convincing than what is in the paper currently. \n\n3. The metrics used for speech summarization in this paper do not go far enough. It is well known that ROUGE and METEOR based evaluations for summarization are not all encompassing, and that the metrics have significant flaws. This again makes it hard to validate that the observed improvements correlate with summaries of higher quality. The authors could supplement these measures using human evaluations of coherence, consistency, factuality and relevance. \n\n4. Table 4 could be expanded to show the impact of the long audio transcription based alignment if any. \n\n5. The manner in which DPO is performed is not very convincing. The authors use the model generated responses as the non-preferred responses and the ground-truth summaries as the preferred responses. Do the authors validate that the model generated responses are in fact undesirable, and record metrics that demonstrate the same ?\n\n\n[1] J. -W. Jung, R. Sharma, W. Chen, B. Raj and S. Watanabe, \"AugSumm: Towards Generalizable Speech Summarization Using Synthetic Labels from Large Language Models,\" ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Seoul, Korea, Republic of, 2024, pp. 12071-12075, doi: 10.1109/ICASSP48485.2024.10447328." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "Refer to weaknesses" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Originality:\n1. The Mamba-based approach has not yet been utilized for speech summarization.\n\nQuality:\n1. Faster inference with better summarization performance against a cascaded and E2E baseline.\n\nClarity:\n1. The paper is mostly easy to follow.\n\nSignificance: Results could be significant to speech summarization community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an end-to-end abstractive summarization method that processes speech inputs directly. It utilizes a querying-attention Mamba projector to condense extended acoustic features into compact semantic tokens, which are subsequently fed into the Mamba Large Language Model (LLM). They further employ Direct Preference Optimization (DPO) fine-tuning to produce coherent summaries. Experiments on a TTS-synthesized speech summarization dataset demonstrate that this approach outperforms both a cascaded baseline and an end-to-end baseline." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Clarity: \n1. The biggest weakness to me is the lack of clarity on the baseline models. They use Whisper large v2 as the ASR model in the cascaded system, but it has a limit of 30 seconds. How do they use it to get ASR output? Do they feed every 30-second window of audio as input? Further, do they finetune ASR or LLM or are they used in a zero-shot manner? What are the results in both these scenarios? If LLM is finetuned, do they also use DPO finetuning? The paper should provide more details about the E2E speech summarization baseline in the main text to make the paper self-contained. \n2. It’s also unclear how the approach handles long speech sequences, which seems to be a central aspect of this work's novelty. The paper mentions chunking audio into 30-second segments, yet doesn’t address how contextual continuity is managed between chunks. Prior studies on streaming ASR (e.g., https://arxiv.org/abs/2107.09428) indicate that chunk boundary in the middle of token can result in generation inaccuracies. Clarifying whether any overlap is applied between chunks and providing additional discussion on this topic would improve the paper's depth and accessibility.\n\nSoundness: \n1. The evaluation is limited to one dataset, a TTS-generated synthetic speech summarization dataset. Including publicly available human speech datasets, such as SLUE_TED (https://huggingface.co/datasets/asapp/slue-phase-2) or AMI (https://groups.inf.ed.ac.uk/ami/corpus/), would provide a more robust assessment and ensure that the approach is tested on natural human speech data.\n2. Additional details on the synthetic dataset would be valuable, including whether it consists of single-speaker audio, and its quality (e.g., WER for the TTS output as evaluated by a pre-trained ASR model or via human relevance judgment).\n3. Further analysis is needed to pinpoint where the model outperforms a cascaded baseline. Does this improvement stem primarily from avoiding error cascading (that can potentially be addressed by improving the ASR system), or does the model also capture non-phonemic audio signals that enhance summarization quality?\n\n Significance and Originality: I feel that Mamba-based approaches have been shown to be useful for various speech processing tasks (https://www.isca-archive.org/interspeech_2024/miyazaki24_interspeech.html) and have been shown to be particularly efficient for long-form speech processing. The paper has limited novelty since it additionally verifies an expected conclusion that Mamba-based architecture is also useful for another speech processing task, namely speech summarization. Further, the lack of clarity in setup of baseline models make me question the improvements claimed in the work." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024squba,\ntitle={{SQ}uBa: Speech Mamba Language Model with Querying-Attention for Efficient Summarization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zOMa82W1HV},\nnote={under review}\n}" }, "abstract": { "value": "Abstractive Speech Summarization (SSum) becomes increasingly difficult as the input speech length grows. To address this, we present SQuBa (Speech Querying Mamba Architecture), an end-to-end model designed explicitly for efficient speech summarization. SQuBa leverages a querying-attention Mamba projector to condense extended acoustic features into compact semantic tokens, which are subsequently summarized by the Mamba Large Language Model (LLM). The architecture’s computational complexity scales linearly with input length, enabling efficient handling of longer inputs. A two-stage training framework, complemented by bootstrapped Direct Preference Optimization (DPO) fine-tuning, empowers SQuBa to generate concise and coherent summaries. Experimental results demonstrate that SQuBa delivers competitive performance while significantly improving inference speed, making it ideal for real-world applications such as podcast and meeting transcriptions." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Summarisation", "Speech", "Mamba" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/c82c28c31b6fc46fa036bffd97f0231774b30cb1.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "SQuBa: Speech Mamba Language Model with Querying-Attention for Efficient Summarization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In the limitation section is mentioned a limited due to \"first-order Taylor expansion\", but I thought influence functions rely on second order expansion, which one is it?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper unifies three unlearning objective under a single simple framework.\nThe paper directly addresses the complexity concerns the framework seems to have in two simple ways: using fast Hessian-vector product, and using mini-batches. It then provides a convincing argument for why such problem formulation would converge using SGD." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper suggest a unified and feasible way to address multiple unlearning scenarios under a single framework. It first generalizes three unlearning objectives. Then, proposes to use influence-functions to approximate the influence of datapoints requested to be unlearned. It then proposed to use PEFT to avoid full finetuning of the model, by training each adapter separately." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper is a bit hard to read. I would suggest putting most equations in the appendix and focus on high-level explanation in the body of the paper. It is unclear exactly what exactly how adapters are used in Algorithm 1. The evaluation is a bit limited, consider comparing this method performance to other unlearning approaches. Evaluating efficiency in seconds can be a nice addition but the main evaluation should compare an asymptotic running-time complexity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "see weakness. Also, how is the performance of the method on text-generation tasks? Could you evaluate your approach on some QA datasets?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The problem is important and interesting. The three unlearning tasks are practical and meaningful in real-world settings.\n\n2. The unlearning approach edits model parameters based on influence functions, which avoids finetuning or re-training the models\n\n3. The evaluations are comprehensive. \n\n4. The writing is clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduce a novel instance-wise unlearning method that utilizes influence functions to adjusting parameters for unlearning tasks. Specifically, the method can remove specific sample from the training dataset, adjust input tokens in user queries, as well as correct model response. Experimental results show that the method can perform unlearning tasks with high accuracy while ensuring efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My concern is mainly about evaluation datasets. \n\n1. Although the paper claims to address privacy concerns in LLMs, the authors primarily evaluated the method using tabular datasets for recommendation tasks and a multimodal dataset for a classification task. While these tasks are representative, they are relatively simple and straightforward. Although the authors applied prompt engineering to those data, the training data are still very similar. \n\n2. In real AI applications, LLMs are used for text generation more often, and the training data consists of documents from various sources and may contain sensitive information, such as bank accounts, addresses, and SSNs. Thus, this paper lacks evaluations in more practical and meaningful scenarios that would better demonstrate its effectiveness in addressing privacy concerns in real-world settings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1) How effectively can the proposed algorithm let the model forget the samples?\n2) How effectively can the proposed algorithm maintain the model performance on generative tasks?\n3) If a sample is also used in the pre-training, can it still be unlearned? \n4) How can we tell whether the unlearning (or failure to unlearn) is on the PEFT module but not on the pre-trained model?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) The paper shows the motivation and derivation of the proposed algorithm/framework.\n2) The proposed algorithm demonstrates that it can effectively maintain good performance on different tasks, including recommendation and relation mining tasks.\n3) The proposed algorithm is also shown to have high efficiency compared to the existing algorithms." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies machine unlearning for parameter-efficient fine-tuning (PEFT) of large language models (LLM). The authors first propose a framework called LLMEraser for three different unlearning tasks: instance removal, query modification, and response correction. The core of LLMEraser is based on the influence function, based on which the algorithm can minimize the effect of the removed or modified samples. Some experiment results are presented to show the effectiveness of the proposed algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) All the experiments show how well the models can perform after unlearning a certain amount of training samples. However, it seems the paper does not present or compare how effectively the algorithms can let the model \"forget\" those training samples, which is the original goal of the unlearning algorithm.\n2) As PEFT is fine-tuning the model, it is unclear how to distinguish the influence of a sample when it is in the fine-tuning training set or the pre-training training set. When the sample was also used in the pre-training set, it is also unclear whether it is reasonable to require the PEFT to eliminate those \n3) The experiment settings in the paper do not include the generative tasks of LLMs.\n4) Some minor typos, such as \"handling handle\" near the end of page 9." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could the authors provide a detailed analysis of estimation errors to demonstrate their impact on performance? Given the importance of precision in unlearning tasks, such an analysis would help establish whether the errors introduced by the method are indeed within an acceptable range and do not significantly detract from the efficacy of LLMEraser in high-precision unlearning scenarios.\n\n2. The LLM4Rec unlearning tasks are different from those unlearning tasks used in the baseline paper. Could the authors provide performance results on additional unlearning tasks more closely aligned with those proposed in other baselines? Additionally, how do the LLM4Rec tasks reflect the generalizability of LLMEraser across diverse unlearning scenarios? \n\n3. While LoRA was used in the experiments, it would be helpful to know the specific experimental setup, including the number of trainable parameters. Since computational complexity is directly related to the number of trainable parameters and rank values in LoRA, could the authors clarify these settings and present results across different ranks? Additionally, could they specify the sample sizes used in unlearning experiments, as sample size likely affects the overall complexity?\n\n4. While the method demonstrates computational efficiency, I am concerned about memory efficiency, particularly for broader applications. Could the authors provide an analysis of the memory cost and specify whether LLMEraser could run on a lower-memory GPU than an Nvidia A100? If the method is limited to high-memory GPUs, this could hinder its practical application. A comparison of memory usage with baseline methods is also important to clarify this point.\n\nI would be happy to engage with the authors to help improve the presentation of the method and evaluation, but my concerns are not insignificant. Clarifications would need to resolve my questions in order for my score to improve." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The paper is written clearly and is easy to follow, with clearly presented formulations.\n2. The taxonomy of unlearning tasks is well-defined, providing clarity on the types of unlearning scenarios addressed. \n3. Extensive experiments are conducted across diverse unlearning tasks with both LLMs and MLLMs, covering recommendation and multimodal relation mining applications. \n4. LLMEraser accelerates the computation of the inverse Hessian-vector-product in the influence function, enabling efficient implementations. This improvement is valuable for LLM applications, where computational efficiency is a growing concern." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces LLMEraser, a unified framework for parameter-efficient unlearning in large language models (LLMs), specifically tailored to address privacy and security concerns in domain-specific fine-tuning. The framework utilizes influence functions to perform instance-wise unlearning tasks such as instance removal, query modification, and response correction. LLMEraser allows unlearning without requiring full model retraining, and has demonstrated efficacy in preserving model performance across various unlearning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While LLMEraser reduces the need for retraining, the memory-intensive computation of inverse Hessian-vector products remains demanding. This requirement may limit scalability, particularly for very large models or environments with limited GPU memory.\n2. The method’s reliance on first-order Taylor expansion in influence functions can introduce estimation errors. A lack of detailed error analysis makes it difficult to assess the impact of these errors, especially in tasks requiring high unlearning precision.\n3. Limited generalizability is a concern, as the LLM experiments primarily focus on LLMs for recommendation (LLM4Rec). Broader evaluation across more diverse LLM unlearning tasks would strengthen the claims.\n4. Key experimental details, such as PEFT configurations and the number of trainable parameters, are not fully provided. These details are essential for evaluating the method's complexity and efficacy across different settings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024unified,\ntitle={Unified Parameter-Efficient Unlearning for {LLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zONMuIVCAT},\nnote={under review}\n}" }, "abstract": { "value": "The advent of Large Language Models (LLMs) has revolutionized natural language processing, enabling advanced understanding and reasoning capabilities across a variety of tasks. Fine-tuning these models for specific domains, particularly through Parameter-Efficient Fine-Tuning (PEFT) strategies like LoRA, has become a prevalent practice due to its efficiency. However, this raises significant privacy and security concerns, as models may inadvertently retain and disseminate sensitive or undesirable information. To address these issues, we introduce a novel instance-wise unlearning framework, LLMEraser, which systematically categorizes unlearning tasks and applies precise parameter adjustments using influence functions. Unlike traditional unlearning techniques that are often limited in scope and require extensive retraining, LLMEraser is designed to handle a broad spectrum of unlearning tasks without compromising model performance. Extensive experiments on benchmark datasets demonstrate that LLMEraser excels in efficiently managing various unlearning scenarios while maintaining the overall integrity and efficacy of the models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large Language Model Unlearning; Machine Unlearning; Influence Function" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/379d530a1f4cd8cfe45a0b082be1ab765b3b4e0f.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Unified Parameter-Efficient Unlearning for LLMs" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How do you verify the quality of converted natural language style questions?\n- What is the inter-agreement score of human annotations?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- CAQA uses KGs to generate complex QA benchmarks automatically, enabling scalability and minimizing manual annotation effort.\n- Different reasoning complexities are considered, highlighting LLMs' capabilities in handling logical relationships between facts.\n- The benchmark includes fine-grained attribution categories." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents CAQA (Complex Attributed Question Answering), a large-scale automatically generated benchmark designed to assess the attribution capabilities of QA systems, particularly Large Language Models (LLMs). CAQA leverages Knowledge Graphs (KGs) to create comprehensive attribution categories and to handle complex reasoning scenarios. The benchmark distinguishes between supportive, partially supportive, contradictory, and irrelevant evidence types and introduces reasoning complexity through different forms of evidence combination (e.g., union, intersection, concatenation)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The task setting seems very similar to NLI to me, more discussions are needed.\n- Lack of a few details about the human annotation process.\n- The distribution of the complexity is biased." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See \"Weaknesses\"." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper introduces CAQA, a large-scale benchmark for evaluating complex attributions in QA.\n2. The CAQA dataset contains various new definitions (e.g., fine-grained attribute categories and attribution complexities), and the data construction process is automatic, considerate, and comprehensive.\n3. This paper contains comprehensive experiments. In addition to model performance on CAQA, it also includes fine-grained analysis, human consistency, and out-of-distribution data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Complex Attributed Question Answering (CAQA), a large-scale benchmark designed to evaluate complex attributions in question answering (QA). CAQA is automatically generated using knowledge graphs (KGs), includes a broader range of attribution categories along with intricate attribution reasoning scenarios, and is also aligned with human annotations. Experiments with two specifically developed evaluators and nine large language model (LLM) evaluators reveal that these models struggle to identify negative attribution categories and handle complex attribution reasoning in both zero-shot and few-shot settings but mostly perform relatively well in the fine-tuning setting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper only considers GPT-3.5 and GPT-4 as closed-source LLMs, and some open-source LLMs used may be outdated (e.g., Mistral-7B has revolutionized various versions). Adding more diverse and latest models in experiments would have greater contributions and help to discover which LLMs perform best on this challenging task.\n2. There is a lack of comparisons with human performance on (a subset) of the dataset, which would better illustrate the performance gap and the challenge of the dataset.\n3. While the contribution of the paper centers on a new challenging benchmark, it would be much helpful if the authors can provide an error analysis, which will direct newcomers in future research." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Some discussion on the ambiguity (and actual errors) in the labeled categories would be useful (e.g., human annotator agreements on a sample). \n\nAlso would be good to discuss the lack of context consideration which is usually very important in real usage (e.g., the example in the paper \"Who plays Fruma Sarah in Fiddler on the Roof\" depends on which version of Fiddler on the Roof is being referenced)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The dataset is relevant for the important topic of answers with attributions from LLMs. Being able to carefully validate whether an answer actually follows from the sources is an important skill, and this dataset aims at helping with this.\n\nThe paper is well written, clearly describing the approach.\n\nThe use of the KG to create various incorrect attributions, together with using LLM to rewrite at text, seems quite effective.\n\nThe paper provides access to the full dataset for exploration which is truly helpful in assessing it.\n\nThe methods are tested on a more realistic, OOD, dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the dataset Complex Attributed Question Answering (CAQA), containing answers to questions with associated source attributions, where the attributions may or may not support the answer. The non-support attributions are divided into 3 labeled categories: Partially supported, Contradictory Irrelevant.\n\nThey evaluate how well different LLMs can classify Q+A+source into these 4 categories, finding that in many cases they struggle to do well, especially on distinguishing the non-supportive categories.\n\nThe CAQA dataset is constructed from existing KGQA datasets (GrailQA and WebQuestionsSP), making use of the associated knowledge graph to produce different types of non-supportive evidence (and using GPT-3.5 to turn KG triples into natural language sentences). The resulting dataset is quite big (137k train, 24k test), allowing for fine-tuning experiments as well. The fine-tuned models do very well in distribution, and they also do limited out-of-distribution evaluation on a subset of ALCE further annotated with these non-supportive categories, showing promising results there as well." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While breaking down the non-supportive cases into three subcategories can be helpful for understanding limitations, the boundary between them can be quite unclear. Also the prompt for the non-GPT models doesn't go into great detail (beyond some examples) on what each category means. For instance, the \"contradictory\" evidence is often for actual true facts, so they're not actually contradiction, it's just the \"wrong\" evidence.\n\nE.g., the answer \"The person who founded the United States Coast Guard also founded the United States Department of the Treasury.\" is presented as being contradicted by the source \"Alexander Hamilton is the founder of the United States Coast Guard and the Montgomery County Sheriff's Office, which is a government agency.\", but this isn't really a contradiction, it's more like missing evidence. A true contradiction should lead you to think the answer is actually false, if you trust the source.\n\nThe \"Partial support\" category also can be quite subjective, as in the case of \"The 2011 Estoril Open tournament event competition is the Men's Singles.\" being partially supported by \"The 2011 Estoril Open had a men's singles competition.\" (what's missing is apparently that \"2011 Estoril Open\" was \"tournament event competition\", but that's pretty much implied by the fact that they had a men's single compeition).\n\nBecause of this, it might also be useful to report the most important \"supported\" vs \"not supported\" scores.\n\nAnother concern is the simplicity of the dataset, with simple QA assertions attributed by short source sentences. How does good performance on CAQA transfer to more realistic settings. And can it be used to train better source attribution models as well? There is some exploration of this with the OOD ALCE dataset, but the effect (e.g., between Vicuna-13B and Vicuna-13B-finetuned) isn't as impactful as one might have hoped." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How might CAQA be adapted to handle dynamic or temporal data in QA tasks?\n2, What specific types of biases were identified or considered when using LLM-generated prompts?\n3. Are there plans to include more diverse logical operations, such as negation, in future iterations of CAQA?\n4. Could the inclusion of human-in-the-loop evaluations further enhance the quality of the generated benchmark data?\n5, Extend the benchmark's applicability by including examples from various knowledge domains (e.g., medical, legal) to test the robustness of attributions in specialized contexts.\n6. The reliance on GPT models for generating natural language representations from KGs may introduce subtle biases. Addressing how these biases are minimized or discussing potential implications would strengthen the manuscript.\n7. Discussing how CAQA could be adapted for such tasks would add value especially how to address the challenges in more diverse, such as open-domain QA systems. ." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The use of KGs for automatic construction of the benchmark is novel, making the process scalable and adaptable.\nThe research tests various models and demonstrates the needs of fine-tuning for achieving robust performance.\nThe benchmark shows high consistency with human evaluations, supporting its credibility as an effective tool for future developments in QA systems.\nThe choice to test multiple LLMs, including state-of-the-art models like GPT-4 and LLaMA variants, provides a robust analysis of performance across different model scales and settings.\nThe inclusion of a wide range of attribution types and complexity levels sets a high standard for evaluating QA systems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript aims to bridge the gap in attribution evaluators' use of knowledge graphs by providing detailed categories. This study's experiment offers multiple configurations in zero-shot, few-shot, and fine-tuned contexts, demonstrating that the fine-tuning process can significantly improve performance. This benchmark aims to address the shortcomings of existing attribution evaluators, which face challenges with intricate attribution categories and sophisticated reasoning processes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The benchmark is tailored for KG-based QA tasks, which may not reflect the challenges present in more diverse, open-domain QA systems. \nThe reliance on GPT models for generating natural language representations from KGs may introduce subtle biases.\nThe rationale behind choosing specific complexity types (e.g., concatenation, intersection) could be expanded with examples illustrating real-world implications of these complexities in QA." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024can,\ntitle={Can {LLM}s Evaluate Complex Attribution in {QA}? Automatic Benchmarking Using Knowledge Graphs},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zP8HygcAMY},\nnote={under review}\n}" }, "abstract": { "value": "The attribution of question answering (QA), which is to get evidences for supporting the generated answer, has attracted wide research attention. The current methods for automatically evaluating the attribution, typically relying on Large Language Models (LLMs), are still inadequate, particularly in recognizing subtle differences between attributions, and in measuring complex attribution reasoning. Existing benchmarks, which are primarily based on manual annotations, suffer from limited evaluation settings with incomplete and coarse attribution categories and reasoning scenarios, hindering the evaluation and advancement of attribution evaluators. To address this gap, we introduce Complex Attributed Question Answering (CAQA), a large-scale benchmark automatically generated using Knowledge Graphs (KGs), containing more comprehensive attribution categories and complex attribution reasoning scenarios. Our experiments with two specifically developed evaluators and nine LLM evaluators reveal that they struggle in identifying negative attribution categories and handling complex attribution reasoning in both zero-shot and few-shot settings, but mostly perform relatively well in the fine-tuning setting. Moreover, all evaluators perform inadequately in fine-grained attribution identification scenarios. The experiments also demonstrate that CAQA is consistent with human annotations, and is promising for selecting and developing more effective attribution evaluators in QA." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large Language Model", "Attributed Question Answering", "Knowledge Graph" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/44f117c0c2d0d66e6abedc661726e45e7a1b67f2.pdf" }, "presentation": null, "primary_area": { "value": "datasets and benchmarks" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Can LLMs Evaluate Complex Attribution in QA? Automatic Benchmarking Using Knowledge Graphs" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N.A." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. A holistic FL framework incorporating different clustered FL methods for more comprehensive FL. The research problem is important, and this work contributes to fixing the shortcomings of the existing related works.\n2. Comprehensive experimental evaluation has been conducted to illustrate the effectiveness of the proposed method.\n3. The paper is well written with a clear demonstration of the motivations and problems for solving." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript proposes a holistic federated learning framework to enhance classification performance by grouping clients into clusters. The framework comprehensively integrates the hard and soft partitional clustering, and clustering with and without automatic cluster number determination. Comphrehensive theoretical and empirical evidence have been provided to illustrate the effectiveness of the proposed method. Moreover, the paper is generally well-written and easy to follow." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This work focuses on using clustering techniques to enhance the classification accuracy of FL. The difference between this type of research and the fully unsupervised federated clustering should be discussed to avoid potential misunderstandings.\n2. The efficiency issue is listed as one of the challenges in Section 4.1. But only the final number of clusters is reported accordingly. More discussions about the time and space complexity of this work, or even corresponding evaluation results are preferable.\n3. The source code is not opened in the current version." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "What is the overall objective of FL in a learning theoretical sense, i.e. how exactly would the generalization error which is targeted be expressed?\nWhat would be naturally occurring drift/shift in such scenarios and how does this match with the shift modelled in experiments?\nWhat are results if the clustering itself is evaluated (eg having ground truth on the data distributions)?\nHow personalized are the models? And how does this scale with the required number of data as regards valid generalization?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The article addresses the important problem of efficient federated learning in the presence of data shift. \nIt provides a very detailed experimental analysis. \nIt also takes some effort to substantiate the observations with theoretical insight. \nMoreover, the authors promise to release the code open source." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The contribution formalizes a general pipeline/framework for clustered federated learning, on the one hand via a cost function, on the other hand via algorithmic choices (as it is not easy to include the objective of an optimum clustering number into costs in a meaningful way). This leads to an improved variant, where the data per client can belong to different clusters, and extensions of at the moment rare soft clustering schemes. The benefit is evaluated in both, comparative studies and ablation studies. A comparably long appendix addresses how to compute an EM scheme based on the costs, how to design data, what happens for linear parts, and some more insight." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The proposal left me a bit puzzled, as the specific contribution is somewhat unclear. On the one hand, the contribution promises a general framework/principle how to model FL with clustering. Here, the costs are rather obvious, as is the four-tier modeling, given the existing work how to model clustering; hence I am not sure what exactly is the contribution, is it the specific way of implementation, or specific guarantees which can be given? In how far is this modeling surprising/challenging and what exactly is the contribution, please specify (eg is it the better implementation? It would help if you could either provide specific benefits which arise from there which would not have been possible without this abstraction, or to provide examples where it is not obvious that the method falls under this common framework. \n\nThe improved version allows the individual assignment of data of one client. Here an according EM scheme is derived (a bit lengthy but straightforward), soft clustering is considered (which seems also straightforward given the existing work on soft clustering and its algorithms). Personally, I find the definition of a new way to measure distances w.r.t. drift most interesting, albeit very shortly presented. Here references to existing technologies how to deal with drift are missing (such as decomposition of sets of data with drift into some where the drift is homogeneous, e.g. moment trees and Kolmogorov trees). I suggest to have a closer look at the (exhaustive) literature on learning with drift in the incremental setup. \n\nI find the presentation suboptimal as the main part of the work reads almost trivial in wide parts, whereas some important insight seems to be hidden in the appendix. It would help if the main take aways of the appendix would be highlighted. Moreover (as already said before), please more clearly highlight why the holistic framework is not trivial, and benefitial with specific non-trivial results/examples." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see weaknesses. Generally, I think this is a solid work. I will adjust the score based on author response." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The problem is well motivated and interesting. The algorithm proposed is novel, with solid theoretical analysis and extensive experimental studies. The paper is well structured and written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces HCFL, a holistic framework for Clustered Federated Learning (CFL) that integrates existing methods. HCFL+ builds on this by addressing key challenges in HCFL, improving the effectiveness. Extensive experiments show the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "For clustered based FL algorithm, there is a recent work [1] to conduct clustering based on the inferred label distributions. The authors are suggested to discuss about this clustering strategy.\n\nCould the authors provide more experiments on a wider range of beta (e.g., from 0.1 to 1.0), to show the effectiveness on different levels of data heterogeneity?\n\nAlso, there are setups such as C=2,C=3 in [1]. Such settings are also suggested to be studied. If the proposed algorithm can perform well across various data heterogeneity partitions, the paper can be stronger.\n\n[1] Diao, Yiqun, Qinbin Li, and Bingsheng He. \"Exploiting Label Skews in Federated Learning with Model Concatenation.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 10. 2024." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a unified framework for clustered FL algorithms and improve the techniques within the framework." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024enhancing,\ntitle={Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zPDpdk3V8L},\nnote={under review}\n}" }, "abstract": { "value": "Federated Learning (FL) is an evolving distributed machine learning approach that safeguards client privacy by keeping data on edge devices. However, the variation in data among clients poses challenges in training models that excel across all local distributions. Recent studies suggest clustering as a solution to address client heterogeneity in FL by grouping clients with distribution shifts into distinct clusters. Nonetheless, the diverse learning frameworks used in current clustered FL methods create difficulties in integrating these methods, leveraging their advantages, and making further enhancements. \nTo this end, this paper conducts a thorough examination of existing clustered FL methods and introduces a four-tier framework, named HCFL, to encompass and extend the existing approaches. Utilizing the HCFL, we identify persistent challenges associated with current clustering methods in each tier and propose an enhanced clustering method called HCFL$^{+}$ to overcome these challenges. Through extensive numerical evaluations, we demonstrate the effectiveness of our clustering framework and the enhanced components. Our code will be publicly accessible." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Federated Learning", "Clustering" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/0a4bcc8af50fe0f24c26cc90f2c8a60d79c1656a.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/e6dad37955ac4b05dbcb64015aee5bc59667bb9c.zip" }, "title": { "value": "Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The feature average is clearly defined in two-layer networks - the linear input combinations. However, in the deeper layer, such a definition does not seem reasonable. How do we analyse features in this case? Can the authors provide one or more examples, conceptually or mathematically, about how features are averaged in deeper neural networks?\n\n2. In image tasks, reasonable and robust features are often local and isolated. However, there are other scenarios where the features are required to have global or long-term dependency. Such as time series, and natural language. In those cases, it seems like the real interpretable and robust features are some \"mixing\" of the original data. Therefore:\n a) How should we define the feature mixing in those cases?\n b) Would the theory developed in the paper be effective in those circumstances? If so, how to measure or validate? If not, should there be possible modification can improve the adaptability of the theory?\n\n3. There are researches that have out that, the neural network tends to learn low-frequency features (which are nonlocal and varied smoothly) more effectively than high-frequency features. Some connect such behaviour to the generalization capacity of over-parameterized neural networks. Do you think there is any relation between this theory and the ones developed in this paper? I want to hear about the comments of the authors. \nI list a few papers for reference:\nXu, Zhi-Qin John, et al. \"Frequency principle: Fourier analysis sheds light on deep neural networks.\" arXiv preprint arXiv:1901.06523 (2019).\nXu, Zhi-Qin John, Yaoyu Zhang, and Tao Luo. \"Overview frequency principle/spectral bias in deep learning.\" Communications on Applied Mathematics and Computation (2024): 1-38." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper gives very thorough theoretical study of the problem. The hypothesis, assumptions and derivation are reasonable and sound. The results matched with the intuition well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides a theoretical analysis of the relation between feature dependency and adversarial robustness. The authors studied a two-layer ReLU network on the simulated data and derived that the gradient descent tends to learn averaged features instead of isolated/robust ones. The results were proved thoroughly. Then, the paper provides a test on the derived theory on real-world datasets, showing that by incorporating more precise training labels, the model can learn isolated features rather than mixed ones, therefore achieving better adversarial robustness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The theory developed in this paper is phenomenological. It can be further improved if systematic and quantifying criteria can be built to design an optimization method that can help improve the robustness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "-Minor issues in notations: Big $O$ notation on line 185 is different from the others; $Theta$ notation is not defined.\n\n-Why does strong correlations indicate the existence of feature averaging? Is there a rigorous proof? How do you define the threshold of having a strong correlation?\n\n-Proposition 4.8 is used as an example of “adding more fine-grained supervision signals does not trivially lead to decoupled features and robustness”, but it does not follow the setup of your analysis, since $h=1$. How is this multi-class network obtained? Why is its parameter assumed to be the combination of $mu_s$?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "-This paper focuses on explaining the adversarial robustness of neural networks in classification problem, which of course is an important topic in ML/DL community.\n\n-This paper is generally well-written and logically organized with clear notations, assumptions, and detailed references to prior works with comparisons\n\n-This paper’s idea is intuitive: this paper characterizes the “approximate linearity” in networks under NTK assumptions as adversarial non-robustness in terms of Feature averaging." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper analyzed the training dynamics of gradient descent on a simplified two-layer ReLU networks in a binary supervised classification task, where the feature vector is sample from a gaussian mixture model with equal probability for K>2 components and the label are masked as 2 categories. The main result in section 4 showed that, under some regulatory assumptions and choices of hyper-parameters, the learned network turns to a feature averaging network, which has low generalization error but is not robust to perturbation of dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Too many strong and unrealistic assumptions are used. First, the number of neurons $M=2m$ assumed to be finite as the number of clusters of features in assumption 4.3. It is against the purpose of using the neural tangent kernel (NTK) theory to approximate infinite wide neural networks. Thus, the result in this paper is limited since “training deep neural network is a highly non-convex and over-parametrized optimization problem”. Second, as the solution to adversarial robustness, “fine-grained supervision” suggests including the original cluster labels for all data points, which is extremely unrealistic, because the response is simply assigned from the cluster labels. It is not surprising that there is an improvement in adverbial robustness as the exact latent information is provided. Feature decoupling is a simple result as the network can learn each cluster’s mean given the cluster labels. \n\n- The evidence of feature averaging is shown Figure 2 as the strong correlations/cosine value between cluster’s mean and network weights, but the scale of correlations is not consistent. In particular, the results from CIFAR-10 are not convincing, where the average cosine value can be lower than 0.1." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In the fine-grained supervision case of the multi-cluster data, I understand that the binary classification using the multi-class network results in a robust solution. But, I wonder if the learned network would still be non-robust for the multi-class classification. \n2. It is not an ask to prove the following. I am only curious to understand and know the authors thought. The fine-grained supervision setting can also be seen as learning a more expressive/complex classifier (learn k classes to do binary classification). Then, can the authors comment on the adversarial min-max training? Would similar feature decoupling, ie the weight of each neuron is aligned with one cluster feature, happen naturally if one does adversarial training of the binary class data?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The theoretical analysis of the finite time gradient descent dynamics is thorough in the feature learning regime. \n* This paper is an interesting contribution in understanding the robustness of neural networks from optimization perspective. The analysis also aided in proving a conjecture. \n* The paper is well written and the presentation is clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the implicit bias of gradient descent, which the authors call 'Feature Averaging' by theoretically analyzing two layer ReLU network with only a learnable first layer, under a specific data distribution - binary data having multiple clusters with orthogonal cluster centers. \nUnder gradient descent, weights of neurons in the hidden layer converge to an average of the cluster centers, resulting in a non-robust network. Additionally, they also prove that with more fine-grained cluster information, the network can learn a robust solution for the binary classification problem." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The assumptions on the data (orthonormal equinorm and multi-cluster) is restrictive.\n* The proof is detailed and thorough for two layer ReLU with only one learnable hidden layer. Extending to deeper networks seems challenging." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Q1: Can the fact that the network reaches an \"averaging\" solution be deduced from the KKT equations in Frei et al. (2022)?\n\nQ2: The implicit bias of GD for this problem in 2-layer neural networks is undesirable for $\\ell_2$ robustness. What about $\\ell_\\infty$ perturbations? It seems that this bias would help for $\\ell_\\infty$ perturbations, in contrast to the feature decoupling solution $NN_{FD}$. I would appreciate your thoughts on this.\n\nQ3: Relatedly, would it be possible to measure robustness in the experiments on CIFAR10/MNIST against $\\ell_\\infty$ perturbations? If there is evidence for larger $\\ell_\\infty$ robustness when training with additional supervision, then this would make the authors' claims much stronger.\n\nQ4: Figure 3: How do you measure perturbation radius? If, for example, you normalise MNIST images to be between 0 and 1, then a perturbation radius of 2.0 does not make any sense. Can you mention the ranges of the input for MNIST and CIFAR10?\n\nQ5: In line 882, you mention that \"In image classification tasks, Ilyas et al. (2019) visualized both robust and nonrobust features.\". However, this is not true to the best of my knowledge. I believe there are 2 works that have provided such visualizations [1, 2].\n\n[1]. G. Goh. A discussion of ’adversarial examples are not bugs, they are features’: Two examples of useful, non-robust features. Distill, 2019.\n[2] Tsilivis, N. and Kempe, J. (2022). What can the neural tangent kernel tell us about adversarial robustness? Advances in Neural Information Processing Systems 35." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "This is a good paper, and it was a pleasure to read and review. I would like to highlight the following contributions:\n\n- Important topic: The interplay between the implicit bias of optimization and robustness is, in my opinion, an important area of study.\n- Finite time theoretical analysis of neural networks trained in the various settings: The paper rigorously analyzes the training trajectory of the networks. I enjoyed reading and learning about interesting techniques, such as the ones outlined in Section C.2. In general, the Appendix of the paper is exceptionally polished and the results are easy to follow and verify (to the extent that this was possible during the review process - the paper is 67pages long).\n- Interesting suggestion on providing additional supervision during training for improved robustness: I found this observation interesting and the experimental results seem to suggest that these ideas may generalize beyond the toy distribution considered in the paper's analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper theoretically studies the robustness of two layer neural networks under a specific data distribution that consists of $k$ clusters in $\\mathbb{R}^d$ and the target consists of deciding on which group of clusters the input belongs to (binary classification). The authors show that these networks tend to learn an \"averaging\" solution (i.e., the first-layer weights encode the averages of the clusters), making the networks susceptible to $\\Omega(\\sqrt{d / k})$ perturbations of the input (despite the existence of networks capable of tolerating $O(\\sqrt{d})$ perturbations). Motivated by this observation, they propose modifying the training process by using each example's cluster vector as the label, and they prove that this approach indeed yields optimally robust networks. Finally, the paper presents experiments with both synthetic data and image classification benchmarks that contextualize the theoretical findings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The main conceptual observation does not appear to be new. To my understanding, Frei et al. (2022) showed that neural networks trained with gradient descent on this distribution are not robust to $\\ell_2$ perturbations of the input due to the implicit bias of optimization. In lines 153-171, the authors argue that their work differs from prior work because they do not explicitly analyze the properties of the KKT point of the max-margin problem and instead perform a finite-time analysis. Furthermore, they argue that convergence to the KKT points might be slow. However, we know from Liu & Li (2020) that these conditions are approximately satisfied as long as there is a positive margin. Thus, approximate results are conceptually consistent with the KKT analysis. Additionally, I believe that the fact that the network reaches an \"averaging\" solution can be deduced from the KKT equations in Frei et al. (2022)—see also (Q1).\n\n2) The paper argues that \"feature averaging is a principal factor contributing to non-robustness of neural networks\" (line 014). However, it focuses on one type of perturbations. While the authors provide evidence in this direction for $\\ell_2$ perturbations in a simple distribution (& experimental results in more complicated settings), it is not clear that this holds more broadly. Note that the model’s solution is exceptionally contrived, as all neurons (within each group) converge to the same solution. See also (Q2, Q3).\n\n2) While the paper is well-written, I found the discussion on prior work somewhat insufficient. I would appreciate a bit more clarity on what is novel about the first part of the contributions (i.e., the proof that networks reach feature-averaging solutions). Furthermore, in line 110, the authors mention the work of Min & Vidal (2024) and their conjecture for the first time. It would be more appropriate to introduce this work and its conjecture earlier in the introduction, making it easier for readers to appreciate the contribution.\n\n4) A list of non-significant typos/grammatical errors I spotted:\n- line 091: features-> feature\n- line 107: exists-> exist\n- line 167: analyze -> analysis\n- line 261: neural -> neuron\n- line 893: echos -> echoes\n- line 1098: We -> we\n- line 1100: margin -> margins\n- lines 1307-1310 require rephrasing (they do not make sense)," }, "withdrawal_confirmation": null }, { "TLDR": { "value": "In this paper, we provide a theoretical explanation for non-robustness in trained neural networks." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024feature,\ntitle={Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zPHra4V5Mc},\nnote={under review}\n}" }, "abstract": { "value": "In this work, we investigate a particular implicit bias in the gradient descent training process, which we term “Feature Averaging”, and argue that it is one of the principal factors contributing to non-robustness of deep neural networks. Despite the existence of multiple discriminative features capable of classifying data, neural networks trained by gradient descent exhibit a tendency to learn the average (or certain combination) of these features, rather than distinguishing and leveraging each feature individually. In particular, we provide a detailed theoretical analysis of the training dynamics of gradient descent in a two-layer ReLU network for a binary classification task, where the data distribution consists of multiple clusters with orthogonal cluster center vectors. We rigorously prove that gradient descent converges to the regime of feature averaging, wherein the weights associated with each hidden-layer neuron represent an average of the cluster centers (each center corresponding to a distinct feature). It leads the network classifier to be non-robust due to an attack that aligns with the negative direction of the averaged features. Furthermore, we prove that, with the provision of more granular supervised information, a two-layer multi-class neural network is capable of learning individual features, which is able to induce a binary classifier with the optimal robustness under our setting. Besides, we also conduct extensive experiments using synthetic datasets, MNIST and CIFAR-10 to substantiate the phenomenon of feature averaging and its role in adversarial robustness of neural networks. We hope the theoretical and empirical insights can provide a deeper understanding of the impact of the gradient descent training on feature learning process, which in turn influences the robustness of the network, and how more detailed supervision may enhance model robustness." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "deep learning theory", "feature learning", "adversarial robustness", "implicit bias" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/b90fb22bf29ecec13a7835f702e8bc726ebe4c88.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In the proposed method, when an incorrect solution is generated, the model immediately moves on to the next iteration. Could the authors explore the effect of incorporating feedback on the incorrect code before proceeding to the next iteration? How would this impact the experimental results?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The method effectively incorporates self-correction and self-refinement techniques into the RLHF framework, leading to significant sample reduction.\n- It demonstrates enhanced utilization of multi-step feedback, contributing to improved performance and efficiency in iterative code generation tasks.\n- The approach addresses a critical limitation of current LLMs by successfully leveraging execution feedback, providing a practical and innovative solution for complex code synthesis problems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel approach aimed at enhancing the performance of Large Language Models (LLMs) in multi-turn code generation tasks. The primary challenge addressed is the difficulty LLMs face in leveraging execution feedback for iterative code improvement, a crucial aspect of achieving reliable outcomes in multi-step code generation. The authors propose an end-to-end reinforcement learning method designed to teach LLMs to effectively use execution feedback. The paper makes three key contributions: (1) the development of a reinforcement learning framework that grounds LLMs in execution feedback, (2) demonstrated improvements in performance on code synthesis tasks, and (3) practical benefits including significant sample reduction and enhanced utilization of multi-step feedback." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The method's novelty is limited, as it mainly adapts self-correction and self-refinement techniques to the RLHF framework. Similar approaches have already been explored, such as Anthropic's work on iterative code correction[1] and DeepMind's adaptation of self-correction in RLHF for improved training and inference efficiency[2]. Additionally, the core idea of using compiler feedback as a reward signal for RLHF lacks sufficient comparative analysis with contemporary methods that integrate RLHF with compiler feedback, such as RLTF, Execution-based Code Generation using Deep Reinforcement Learning, CodeRL, StepCoder, and $\\mathcal{B}$-Coder.\n2. The experiments are limited to the CodeContest dataset, with a relatively small number of validation and test samples. The paper should validate its effectiveness on more datasets, such as APPS, which is commonly used for training in RLHF+compiler feedback research.\n3. The paper claims that the multi-turn iterative approach of SFT performs significantly worse than RL (as mentioned in the appendix). However, self-improvement methods like critic+SFT have proven effective in code tasks. Providing the SFT and RL code as open-source could help validate this experimental conclusion.\n\n[1] Improving Code Generation by Training with Natural Language Feedback\n\n[2] Training Language Models to Self-Correct via Reinforcement Learning" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How robust is RLEF to feedback of varying quality beyond random feedback? For instance, how does the model perform if unit tests are incomplete or only cover trivial cases? How does the performance of RLEF change with different types or amounts of execution feedback? Is there an optimal amount or type of feedback for maximizing performance improvements?\n\n2. The paper sets a fixed turn limit in the multi-turn setup. How does this choice affect model performance, and could RLEF benefit from a dynamic turn limit based on feedback content or problem complexity? How sensitive is the model to the number of feedback turns, particularly with respect to diminishing returns after a certain number of turns?\n\n3. The paper mentions using a hybrid approach for value estimation (response-level) and policy optimization (token-level). Can you elaborate on why this approach works better than optimizing both at either the turn or token level?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The experimental results show general improvement when applying RLEF with Llama 3 series models on the CodeContests dataset, demonstrating the effectiveness of the proposed method.\n\n2. The authors demonstrate that Llama 3.1 models trained on CodeContests with RLEF can generalize to other datasets like HumanEval+ and MBPP+, especially the 70B model. This generalization capability makes the proposed method more appealing and suggests broader potential applications.\n\n3. The behavioral analysis post-RLEF training provides valuable insights into how the model learns to handle feedback differently. The detailed error analysis (e.g., fewer repetitions, more targeted edits, reduced reliance on random sampling) gives empirical weight to claims about the improved robustness of RLEF-trained models in multi-turn setup." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces RLEF, a reinforcement learning framework that enhances code synthesis for large language models (LLMs) by utilizing execution feedback during training. RLEF allows models to iteratively refine code solutions based on execution feedback from test cases, significantly improving solve rates on competitive programming tasks and reducing sample requirements. The authors benchmark RLEF with Llama 3 models on CodeContests and demonstrate state-of-the-art performance, with generalizability to HumanEval+ and MBPP+ benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The scientific novelty of RLEF is somewhat limited. Although it extends previous work by introducing a multi-turn setup, the core concept of using unit test feedback as a reward signal has already been proposed in the literature, such as in RLTF [1], which uses unit test feedback within an online RL framework for code generation. The primary advancement in RLEF lies in iterating on this approach by incorporating execution feedback as input text across turns, making this more of an empirical extension than a conceptual breakthrough. \nTo clarify the novel contributions, the authors could explicitly outline how RLEF builds upon and differs from RLTF and related works. Including a citation to RLTF would also better contextualize RLEF within the existing literature, helping to position its contributions more clearly.\n\n2. While the experiments on CodeContests show the effectiveness of RLEF, high-quality unit tests like those provided in CodeContests are hard to obtain in practice or from other data sources. The authors didn't study how the quality of the unit tests (beyond just random feedback) would affect the effectiveness of RLEF. For example, if the public unit tests only cover simple input cases, the execution feedback might struggle to find bugs, imperfections, or inefficiencies in the code, thus providing less useful feedback. This limitation might hinder the application range of this method. \nIt would strengthen the paper to include an analysis on RLEF's sensitivity to unit test quality. The authors could consider testing the method with varying qualities of unit tests, or provide a discussion on approaches to generate high-quality tests in real-world environments where comprehensive unit tests are scarce. Specific experiments that address this limitation would add valuable depth to the study.\n\n3. The paper’s presentation, particularly in the Method section, needs improvement. (a) The organization is overly condensed, and crucial methodological details are packed into dense paragraphs, making it challenging for readers unfamiliar with PPO or RL for code generation. Key adjustments in the RL setup are mentioned briefly without adequate justification, making reproduction difficult and potentially deterring readers from fully engaging with the methodology. Expanding on crucial areas, particularly the PPO implementation details and any task-specific adjustments, could improve clarity. Breaking down dense sections into distinct, digestible subsections would also enhance readability. (b) Additionally, the primary results section compares models across different n@k metrics with varying n and k values, which may not be intuitive for readers unfamiliar with this metric. A more consistent comparison framework or additional explanation of these metrics would improve clarity.\n\n\n[1] Liu, J., Zhu, Y., Xiao, K., FU, Q., Han, X., Wei, Y., & Ye, D. RLTF: Reinforcement Learning from Unit Test Feedback. Transactions on Machine Learning Research. (2023)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See above weakness" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This method is very intuitive, and modeling reinforcement learning as an interactive task is reasonable.\n- The experimental results are good in training language models for coding." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors introduce a new approach for LLMs coding abilities that utilizes execution feedback from generated code through reinforcement learning to enhance performance in code synthesis tasks. The core insight is that LLMs can effectively follow user instructions while leveraging execution feedback to identify and correct errors in the generated code. The authors conceptualize the code generation process as an interactive task, proposing an environment where LLM-based agents generate code and receive real-time execution feedback. Their reinforcement learning-based method, termed RLEF, optimizes the performance of LLMs through this interactive framework. The results demonstrate that RLEF improves the correctness of the generated code compared to existing methodologies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Considering previous methods for training LLM coding and reasoning capabilities with reinforcement learning [1][2], the innovation is limited, as they also used some specially designed reward functions.\n- There is no ablation on the reward function, which is important for the paper.\n- There is a lack of experiments on more models to verify the generality of the method.\n- It lacks some related work on training LLMs for reasoning and coding with reinforcement learning [1][2].\n\n[1] Coderl: Mastering code generation through pretrained models and deep reinforcement learning.\n\n[2] Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written and clearly presented. I appreciate the clarity and organization in the presentation of the results.\n2. The proposed method appears to be effective, substantially enhancing code generation performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel method to enhance LLM code synthesis through the use of execution feedback. In contrast to existing works that use unit test feedback as binary indicators or fractions of unit test pass rates to improve code generation, the feedback in this work is provided as language descriptions, including error messages and unit test results. Additionally, the proposed method is iterative, with the model trained to self-correct its previous responses based on received execution feedback. Experimental results indicate that this approach significantly improves LLM code synthesis capabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited Practical Application:**\n While effective, the proposed method seems highly constrained in its practical application. The reliance on unit tests for feedback is a significant limitation, as generating accurate unit tests for arbitrary user prompts is often as challenging as solving the problems themselves. This confines the method to specific OJ style problems, where unit tests are readily available, and cannot be trivially extended to more general user scenarios where unit tests are not available.\n\n2. **Advantage Computation:**\n The way how the authors compute the advantages seems to be weird. The authors claim the action space is defined on the token level, but the advantages for every token in the same turn is the same. In that sense, how does the critic be updated? Can the authors intuitively / empirically explain why they do this?\n\n3. **Effectiveness of Multi-turn Feedback:**\n The empirical success of multi-turn feedback is intriguing. It is unclear whether the iterative nature of the algorithm or the unit tests' guidance contributes more to this success. An ablation study could clarify this by training RLEF iteratively while providing intermediate/final feedback with only numerical values, avoiding specific details on failed unit tests and expected outputs. For example, assigning smaller reward signals at the end of each turn and a larger signal at the end of the episode based on whether the generated code is correct or the fraction of unit tests passed could yield insights.\n\n#### Overall Assessment:\nWhile the method shows promising empirical results, I am inclined to reject this submission due to its limited applicability and some missing crucial details and ablation results. However, I am open to revising my score if the authors can provide evidence demonstrating the method's applicability to real-world user scenarios and offer more detailed explanations about the algorithm's design and efficacy." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We apply reinforcement learning to ground LLMs in execution feedback for effective multi-turn code generation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024rlef,\ntitle={{RLEF}: Grounding Code {LLM}s in Execution Feedback with Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zPPy79qKWe},\nnote={under review}\n}" }, "abstract": { "value": "Large language models (LLMs) deployed as agents solve user-specified tasks over multiple steps while keeping the required manual engagement to a minimum. Crucially, such LLMs need to ground their generations in any feedback obtained to reliably achieve desired outcomes. We propose an end-to-end reinforcement learning method for teaching models to leverage execution feedback in the realm of code synthesis, where state-of-the-art LLMs struggle to improve code iteratively compared to independent sampling. We benchmark on competitive programming tasks, where we achieve new start-of-the art results with both small (8B parameters) and large (70B) models while reducing the amount of samples required by an order of magnitude. Our analysis of inference-time behavior demonstrates that our method produces LLMs that effectively leverage automatic feedback over multiple steps." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large language models", "automatic code generation", "reinforcement learning", "LLM agents" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/84d5bdcceff45e8b4abef9ae9bd2ff30b62feccc.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1)\tIn addition to the logistic regression model used, is the proposed method suitable for more complex models? In Appendix E, a surrogate model is employed to compute the influence, but could the ResNet itself be directly used for influence calculation? \n2)\tSince the accuracy of the pseudo-labels depends on the quality of the validation set, how does the performance of the proposed method vary with different validation set sizes?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Solid experiments are conducted on both tabular and image datasets, integrating recent data selection methods from AL and SSL. The proposed method demonstrates promising empirical results without the need for human-annotated labels. In addition, the method is validated on a LLM fine-tuning task, which further underscores its potential for application in different domains." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel pseudo-labeling approach for unlabeled data using influence functions. The proposed method estimates the influence of each possible label on the validation loss for a given unlabeled data point and assigns the label with the most significant improvement in loss as its pseudo-label. Subsequently, a subset of the unlabeled data with the highest improvement in validation loss is selected to update the model. Extensive experiments are conducted to validate the effectiveness of the method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the method is framed in part within the active learning context, its approach—assigning pseudo-labels to unlabeled data via a self-training mechanism—seems more aligned with semi-supervised learning. The introduction could benefit from adjustments to reflect this alignment more accurately. Another concern is the limited technical contributions. It uses the influence function to score and pseudo-labeling the unlabeled data. Although this is an interesting application, it may not represent a substantial methodological advance. Furthermore, in the experiments, a query budget of 10 examples is set, yet only the first 10 rounds of performance are reported. Providing results for additional rounds or scenarios with a larger query budget would offer a more comprehensive evaluation of the method’s long-term effectiveness. In Appendix D, results from querying 1% of the data reveal that the proposed method underperforms relative to other baseline methods. This needs further exploration and explanation. My last concern is that the paper’s approach relies on setting aside 20% of the data for validation, which may be impractical for certain active learning settings, where labeled data is typically scarce. Performances on different sizes of validation set should also be explored." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "My main concern is the weakness 1. I do not understand the value of their new proposed task. I think this setting is similar to the unsupervised learning with the pseudo labels. In my understanding, I think the human annotation is helpful during training. This is the reason why we study active learning. If we fully abandon human intervention, this is totally another area which is unsupervised learning." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The motivation and paper writing are clear. \n2. The experiment is sufficient \n3. The method is fully automatic without human annotation" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes salutary labeling, which automatically assigns the most beneficial labels to the most informative samples without human annotation. Specifically, they utilize the influence function, a tool for estimating sample influence, to select newly added samples and assign their salutary labels by choosing the category that maximizes their positive influence." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. They do not discuss the difference with the unsupervised learning methods, such as \n\n[1] Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID\n[2] Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification\n\nIf the human intervention is removed from active learning, it will be transformed to unsupervised learning that assigns the pseudo labels to the samples. Could you discuss the difference ?\n\n2. How to tackle with the situation that the selected labels are wrong? Could you discuss potential error correction mechanisms, or to analyze the impact of incorrect labels on model performance?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- The proposed approach appears closely related to training with noisy labeled data methods mentioned in the Weakness section. An analysis and comparison with this literature in the related work would be beneficial.\n- A detailed analysis of the assigned salutary labels would be interesting. It would be helpful to know if labels different from the ground truth sometimes enhance performance or if performance improvement primarily comes from refining noisy labels." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The *Salutary Labeling for Active Learning* framework was new to me. If we can effectively annotate data automatically without human annotator, it could demonstrate remarkable potential for machine learning as a whole.\n\n- Overall, the writing is well-structured and easy to follow, making it straightforward to understand the main concepts.\n\n- The proposed method consistently improved performance across nine datasets. This study also conducted detailed ablation studies to analyze the effectiveness of *salutary labels* through the influence function." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study defines a task called salutary labeling, which involves selecting a subset of data from an unlabeled data pool predicted to be the most beneficial for training, similar to traditional active learning, and then using this selected data for model training. The key difference in the proposed salutary labeling approach is that, instead of using labels from human annotators, it assigns pseudo labels expected to improve performance on the validation set. Specifically, it measures the influence of each data point and assigns a salutary label that maximizes validation set performance. The data with the highest influence from the assigned salutary labels is then used for training. This method was tested on nine datasets and demonstrated improved results compared to traditional active learning approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The motivation behind combining salutary labeling with active learning is not fully clear to me.\n - The core motivation of active learning is to label only a small number of informative data points to reduce annotation costs. If automatic labeling without human annotation costs is feasible, applying salutary labels to all available data without the selection process in active learning should suffice.\n - In Algorithm 1, it seems that salutary labels are generated for the entire unlabeled pool during selection. Figure 5 also suggests that model performance improves as the number of salutary labels increases, so the necessity of sampling is unclear.\n - **Additional explanation on why the active learning framework remains effective despite not requiring human labor for labeling would help clarify my understanding.**\n\n- I am not fully convinced about how the salutary labeling task is novel compared to existing label-efficient tasks.\n - To me, the salutary labeling task appears to combine elements from various existing label-efficient tasks. If the goal is to automatically label training data, this could be viewed as a pseudo-labeling process in semi-supervised or self-supervised learning. On the other hand, methods for correcting label noise align more closely with the learning with label noise task [a, b, c].\n - In Appendix A, a key difference from other label-efficient learning tasks is described as a focus on active learning (line 836). However, as mentioned above, without a clear reason for combining salutary labels with the active learning framework, I remain uncertain about the task’s distinction from others.\n - **A clearer explanation on why salutary labels need to be combined with active learning and how this approach fundamentally differs from other label-efficient learning methods would be helpful.**\n\n- There seem to be potential fairness concerns in the comparative experiments with existing AL baselines.\n - The proposed experimental setup uses 20% of the dataset as a validation set and utilizes this validation set’s annotations for influence function calculations and labeling. If labeled data is essential for the proposed method, it would be fairer to allocate part of the training budget for this purpose.\n - Specifically, in Table 1, the budget sample size is set as low as 10, which makes it impractical to reserve 20% of the dataset as a validation set while only training on about 10 samples. This setup might give the impression that validation set labels are indirectly being used for salutary labeling of the unlabeled data.\n - As shown in Figure 5 of Appendix D, one potential reason for the reduced gain of the proposed method as the budget size increases may be the dilution of the validation set’s supervisory effect.\n - **The following additional experiments might strengthen the justification for salutary labeling:**\n - Using a smaller validation set (e.g., 1% to 5% instead of 20%).\n - Including the validation set size in the training budget for all methods (e.g., allowing AL baselines to train on the validation set or at least part of it).\n\n[a] Xiao, Ruixuan, et al. \"Promix: Combating label noise via maximizing clean sample utility.\" arXiv preprint arXiv:2207.10276 (2022).\n\n[b] Chen, Wenkai, Chuang Zhu, and Mengting Li. \"Sample prior guided robust model learning to suppress noisy labels.\" Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2023.\n\n[c] Liu, Sheng, et al. \"Robust training under label noise by over-parameterization.\" International Conference on Machine Learning. PMLR, 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1.In the last paragraph of Section 4, the authors mentioned that the time complexity of salutary labeling is O(nd). However, the proposed salutary labeling algorithm need to calculate the influence estimation of every data point in each iteration of active learning. How much will this slow down the entire training process? Can the authors provide the running time comparison results of each method in Table 1?\n\n2.In the experiment, the author set active rounds R = 10 and query budget b = 10. When b and R are larger, is it impossible to prove that the proposed salutary labeling is effective?\n\n3.The first paragraph of Section 2 is too long and a little bit difficult to read. It should be adjusted appropriately.\n\n4.The legend in Figure 1 obscures part of the polyline and may need to be further refined." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.The writing is good and the supplementary materials are relatively sufficient.\n\n2.The authors propose a simple-sounding but effective active learning method which eliminates the need for human annotation. Judging from the comparative experimental results provided by the authors, this idea is effective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose salutary labeling for active learning which is human annotation-free. They adapt the influence function to calculate the sample influence by assessing the impact of each sample across all possible labels and assigning the label that yields the greatest positive influence. The authors conducted experiments on different datasets to verify the effectiveness of the simple idea." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.In the last paragraph of Section 4, the authors mentioned that the time complexity of salutary labeling is O(nd). However, the proposed salutary labeling algorithm need to calculate the influence estimation of every data point in each iteration of active learning. How much will this slow down the entire training process? Can the authors provide the running time comparison results of each method in Table 1?\n\n2.In the experiment, the author set active rounds R = 10 and query budget b = 10. When b and R are larger, is it impossible to prove that the proposed salutary labeling is effective?\n\n3.The first paragraph of Section 2 is too long and a little bit difficult to read. It should be adjusted appropriately.\n\n4.The legend in Figure 1 obscures part of the polyline and may need to be further refined." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024salutary,\ntitle={Salutary Labeling with Zero Human Annotation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zPRQ7wtwhb},\nnote={under review}\n}" }, "abstract": { "value": "Active learning strategically selects informative unlabeled data points and queries their ground truth labels for model updates. The prevailing assumption in the active learning paradigm is that the acquisition of ground truth labels optimally enhances model performance. However, this assumption may not always hold or maximize learning capacity. Moreover, ground truth annotations incur significant costs due to the need for intensive human labor. In contrast to traditional active learning, this paper proposes salutary labeling, which automatically assigns the most beneficial labels to the most informative samples without human annotation. Specifically, we utilize the influence function, a tool for estimating sample influence, to select newly added samples and assign their salutary labels by choosing the category that maximizes their positive influence. This process eliminates the need for human annotation. Extensive experiments conducted on nine benchmark datasets demonstrate the superior performance of our salutary labeling approach over traditional active learning strategies. Additionally, we provide several in-depth explorations and practical applications including large language model fine-tuning." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Active learning", "influence function" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/be6fa69740004965645dc66e1c95ec688eb801af.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Salutary Labeling with Zero Human Annotation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "077 what is the role of this threshold $S(\\theta)$ is it on the sharpness? If so we don’t pick it but algorithm trajectory can enforce it, so the sentence “beyond which training is thought to destabilize” is misleading.\n\nIs the theoretical claim about diagonal linear network true for any matrix dimension and depth?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Authors study a new phenomenon about reorientation of eigenvectors of the Hessian during the instability phase of training NNs so that the trajectory can go to flatter regions, which seems new and interesting; even though intuitively it is not clear to me why this is necessary for going to flatter regions. Their claim oppose the claim in Damian et al (2023) that the gradient is well-aligned with the gradient of the sharpness during the instability phase." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors study a new phenomenon in training NNs which they call \"Progressive flattening\" during the instability regime when inverse of learning rate is larger than sharpness. In this regime, they claim, via empirical investigations and a theoretical result about a toy model for diagonal linear network (DLN), that loss goes to flatter regions by changing the eigenvector directions of the Hessian, while in the stable regime it seems the eigenvector directions are reinforced. Experiments show this regularization (starting with large learning rates) and gonig through these instability phases, also has a better effect on generalization They also challenge sharpness as a measure of generalization.\nMoreover, their results seem to contradict that of Damian et al 2023 about the behavior of the instability regime." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major concerns: \nThe relation between “growth of parameters” that the authors discuss with the progressive flattening and reorienting the eigenvectors, phenomenon is unclear. \n\nThe claims seems to oppose the sharpness largest eigenvector alignment with the gradient in the instability regime, so I think the experiments seem insufficient to back their claim, especially about reorienting the eigenvectors \n\nAuthors refer to reorienting the eigenvectors of Hessian as the major effect that algorithms goes back to stability, I can imagine reorienting the eigenvectors of the sharpness can have such effect, but the former is not clear to in why it should help. Authors seems not to explain this in their theorem for the toy diagonal model that. Additionally authors interchangeably also talk about the eigenvector of the sharpness Hessian instead of the Hessian of the loss, which is confusing which one is exactly the target of their claims.\n\nOther issues:\n033: is dependent —> depends\n054: in generalization —> in the generalization \n075: grow without bound —> grow unboundedly\nLine 087 definition of $\\alpha$ has an issue, there are two dots used after the two gradients\n092 last sentence not clear\n123: what is a “two-parameter” network?\n133: “sharpness hessian eigenvectors” -> you mean loss hessian eigenvectors?\n148: very unclear: what does “sharpest parameter” mean?\n157: “despite originating from same loss function” doesn’t make sense, are you referring to different coordinates of gradient being different? Also you are referring to $z$ as loss function, but isn’t that the link function in your notation?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* In Figure 5, the authors mark the 'stability threshold' using dotted vertical line. Is this stability threshold wrt initialization? \n* Can the authors contrast their work with existing literature referred above?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The analysis of Hessian eigenvectors rotation during instabilities is new, best to my knowledge.\n* The paper is clearly written and is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This research paper investigates the relationship between gradient descent instabilities and generalization performance in neural network training. Using toy model and experiments on small scales, they demonstrate that these instabilities cause rotations in the eigenvectors of the loss Hessian. Through various experiments on benchmark datasets like fMNIST and CIFAR10, the authors provide evidence supporting their claim that larger learning rates can significantly enhance generalization by promoting this implicit regularization effect." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The authors show that instabilities move training towards flatter regions of the loss landscape. However, This has been shown extensively show in prior works such [1, 4, 5, 6, 7].\n\n* The authors claim that the effect of large learning rates pertains late in training, which they dub it as progressive flattening. First, the phrase can be misunderstood as progressive decrease in sharpness, which has been studied in [4, 5]. I would suggest to use some other phrase for this phenomena. Second, I think the result directly follows from learning rate decay experiments from [9]. If the learning rate is decayed later in training, it would require more time to reach the new EoS threshold. If all models are trained for fixed training time, the sharpness will take more time to increase in the case where learning rate is decayed later in training. I would suggest the authors to train the models long enough such that all models reach the new EoS threshold and then compare the performance.\n\n* The authors claim that their empirical results reveal a clear phase transition where generalization benefits only emerge for\nlearning rates beyond the stability threshold. However, looking at Figure 5(b) the transition is not sharp, rather its continuous. I would not recommend using the phrase transition. Also, similar results are already shown in prior works [1, 4, 5, 6].\n\n* The authors claim that they identify that sharpness is not a great predictor of performance. However, this is already known in prior literature, such as [8]. In addition, authors claim that the rotation can be a better predictor than sharpness. But the experimental results are not conclusive. \nMoreover, my intuition is that such predictors can be misleading for very flat initializations (Figure 3. of [4]) which do not require instabilities to train at high learning rates. Therefore, I would predict there won't be any significant change in eigenvector rotation.\n\n* The authors claim that they identify that the effects of large learning rate are observed without progressive sharpening. Various prior works have already found that instabilities occur without progressive sharpening [1, 3, 4, 5, 6] and this result is not new. I would like to request the authors to highlight their new claims wrt these works. \n\n* The authors claim that they prove that for any network with depth > 1, instabilities cause rotation along the principle components of Hessian. This is only demonstrated theoretically for deep diagonal networks. This claim should be modified accordingly. \n\n* Similar models to the two parameter model studied in Section 3.1 is extensively analyzed in prior studies, such as [1, 2, 3]. \nIn particular, when z is quadratic (which is mainly considered in this work) and the global minima is at zero at origin $\\theta_1 \\theta_2 = 0$ has been analyzed in [1] to demonstrate the catapult effect. As discussed in [3], this model does not exhibit progressive sharpening (also can be seen in equations in [1]), this model does not exhibit progressive sharpening and therefore, this model does not capture real-world behavior well. \n\n[1] The large learning rate phase of deep learning: the catapult mechanism\nA Lewkowycz, Y Bahri, E Dyer, J Sohl-Dickstein, G Gur-Ari\narXiv preprint arXiv:2003.02218\n\n[2] Understanding Edge-of-Stability Training Dynamics with a Minimalist Example \nXingyu Zhu, Zixuan Wang, Xiang Wang, Mo Zhou, Rong Ge\nICLR 2023\n\n[3] Universal Sharpness Dynamics in Neural Network Training: Fixed Point Analysis, Edge of Stability, and Route to Chaos\nDayal Singh Kalra, Tianyu He, Maissam Barkeshli\nICLR 2024 BGPT Workshop\n\n[4] Why Warmup the Learning Rate? Underlying Mechanisms and Improvements\nDS Kalra, M Barkeshli\narXiv preprint arXiv:2406.09405\n\n[5] Catapults in SGD: spikes in the training loss and their impact on generalization through feature learning\nL Zhu, C Liu, A Radhakrishnan, M Belkin\narXiv preprint arXiv:2306.04815\n\n[6] Quadratic models for understanding neural network dynamics\nL Zhu, C Liu, A Radhakrishnan, M Belkin\narXiv preprint arXiv:2205.11787\n\n[7] A loss curvature perspective on training instabilities of deep learning models\nGilmer et al.\nICLR 2022\n\n[8] On the maximum hessian eigenvalue and generalization\nS Kaur, J Cohen, ZC Lipton\nhttps://arxiv.org/abs/2206.10654\n\n[9] Gradient descent on neural networks typically occurs at the edge of stability\nJ Cohen, S Kaur, Y Li, JZ Kolter, A Talwalkar\nInternational Conference on Learning Representations" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- The authors claim that instabilities along the top Hessian eigenvector cause the changes in $S$. Is there any evidence for this? It seems that the main changes in sharpness in Figures 3,12 come from the overall curve changing, not from the sharpness differing at different parts of the curve.\n- What is the precise definition for progressive flattening?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The loss landscape visualizations in Figure 3 and Figure 12 are a nice way to visualize the sharpening/flattening process" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the dynamics of gradient descent in the edge of stability regime. It begins by analyzing gradient descent on a two parameter model $L(\\theta) = z(\\theta_1 \\theta_2)$. It then shows that in realistic settings, the eigenvectors of the Hessian rotate during the EOS dynamics, and conjectures that these rotations are responsible for the edge of stability dynamics. They also propose a \"progressive flattening\" mechanism which argues that training with large learning rates for longer will lead to flatter solutions, and studies the connection between learning rates, sharpness, eigenvector alignment, and generalization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While Figure 2 supports the claim that the eigenvectors of the Hessian rotate at EOS, this is a correlational observation and doesn't support the claim of the paper that this rotation is relevant for stabilizing training, giving rise to the EOS dynamics, or exploring flatter regions of the loss landscape.\n- The paper does not precisely define \"progressive flattening\" – see my question below. In addition, the use of cross entropy acts as a strong confounder in these experiments, as the sharpness will converge to $0$ at the end of training. Therefore, a simple explanation for Figures 4,6 is that using a larger learning rate for longer will decrease the loss more, leading to a smaller overall sharpness. To establish progressive flattening as a distinct phenomenon, it would be valuable to repeat these experiments with MSE loss and train all models to near-zero training loss.\n- The paper mischaracterizes the progressive sharpening factor in Damian et al. 2023. Assumption 1 in this paper assumes $\\alpha > 0$ along their \"constrained trajectory,\" not the actual gradient descent trajectory. It is expected to be negative along the gradient descent trajectory since the sharpness decreases when self-stabilizing. Their experiments (Appendix E, top left) show $\\alpha > 0$ in all of the settings studied in Cohen et al. 2020.\n- This paper would benefit from a proper literature review. The only theoretical analyses of EOS cited are Damian et al. 2023 and Arora et al. 2022. There are many relevant missing papers, including papers that study EOS on diagonal linear networks [1,2,3], and [4] which observed a phenomenon possibly related to progressive flattening.\n\n[1] Chen and Bruna 2023: Beyond the edge of stability via two-step gradient updates\n\n[2] Even et al. 2024: (S)GD over Diagonal Linear Networks: Implicit bias, Large Stepsizes and Edge of Stability\n\n[3] Zhu et al. 2023: Understanding Edge-of-Stability Training Dynamics with a Minimalist Example\n\n[4] Kreisler et al. 2023: Gradient Descent Monotonically Decreases the Sharpness of Gradient Flow Solutions in Scalar Networks and Beyond\n\n\nMinor Points:\n- line 154: $theta$s -> $\\theta$s" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- I find the comparison to Damian et al. (2023) in the Related Work and Appendix D to be inaccurate. That paper proves that, during the unstable regime, 1) the parameters will oscillate and grow in the top eigenvector direction, and 2) via the third-order Taylor expansion, such oscillations will lead to an implicit regularization effect which decreases sharpness and drives the model to stability. This self-stabilization effect occurs whenever training is unstable, and does not require progressive sharpening. The progressive sharpening assumption only characterizes the fact that the sharpness increases during gradient descent for all learning rates, and hence training will eventually becomes unstable. To me, it seems that the \"progressive flattening\" mechanism presented in this work can be explained by the \"self-stabilization\" mechanism in Damian et al., (2023).\n\n- Line 244 \"Even after the instability is resolved, the similarity among individual eigenvectors fall while the subspace comparison remain largely similar\": This doesn't seem to be true in Figure 2, as after the instability the similarity is close to 1.\n\n- Much of the derivation in Section 3.1 is hard to follow. What are the quantities $\\gamma_\\Theta, \\gamma_{|\\theta_2^2 - \\theta_1^2|}$?Why is the stability limit $\\eta_{eos}$ defined to be twice the value of the asymptote?\n\n- In Figure 7, what are these quantities in the table? How is $\\rho$ defined? (i.e with respect to what is similarity calculated?)\n\nMinor comments/typos:\n- The descent lemma (line 68) does not necessarily require convexity.\n- Line 254: \"Figure E\" -> Figure 12." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- It is an important question to understand the dynamics of gradient descent in the large learning rate, or \"edge of stability\" regime.\n- The proposal of eigenvector rotation as a mechanism for reducing instability has not, to the best of my knowledge, appeared before in the literature.\n- The theoretical derivations in Section 3 appear to be sound." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the \"edge of stability\" phenomenon, where neural networks can be trained with a larger learning rate than prescribed via the descent lemma. The central claim of this paper is that during instability, the eigenvectors of the Hessian rotate, thus moving the model parameters to a flatter region of the loss landscape -- the authors call this behavior \"progressive flattening.\" The paper provides theoretical support via a diagonal linear network derivation, and shows empirically that the top eigenvectors do rotate during periods of instability, and that large learning rates lead to improved generalization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- My main concern with this paper is that it does not adequately support the claim that eigenvector rotation is the *cause* of instabilities being resolved. The paper provides evidence that during instability, the top Hessian eigenvectors rotate, and moreover afterwards the training becomes stable. However, there is no justification that such eigenvector rotation causes the return to stability. Justification for such return to stability has already been given in prior works such as the self-stabilization mechanism of Damian et al., (2023) or the catapult effect of Lewkowycz et al. (2020), and to me it appears that eigenvector rotation is itself caused by instability (yet goes away when other mechanisms cause a return to stability).\n- The definition of \"progressive flattening\" is not clear. My interpretation is that progressive flattening is the phenomenon that higher order derivatives of the sharpness are reduced during instability. However, this effect is not quantified nor easy to observe in the experiments in Section 3.2. \n- The justification in Section 3.3 is also insufficient. One alternate hypothesis to Figure 4 that I find more plausible is the following. First sharpness is constrained to $2/\\eta$ when training with learning rate $\\eta$. Later when $\\eta$ is decreased progressive sharpnening occurs, yet the degree of progressive sharpening is proportional to the loss gradient, which is smaller later in training. Therefore reducing $\\eta$ later leads to a lower final sharpness.\n- Many of the results in Section 4 on the generalization benefit of large learning rates are known in the literature. This benefit is because training with a learning rate of $\\eta$ implicitly constrains the sharpness to be at most $2/\\eta$; lower sharpness is believed to correlate with better generalization." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "For Figure 8:\n - Is 5 samples enough to estimate a rank correlation? Could you report the uncertainty in the rank correlation plot?\n - Why does the sharpness often have a high rank correlation with generalization performance, but a varying sign? (compared to rho which has a more consistent but smaller absolute correlation on average)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is well-written. In particular, the theoretical parts (sections 2 and 3) are clear and didactic. Similarly, while I'm not familiar enough with the literature to assess its exhaustivity, I found the related work well-written and easy to follow, giving context to the authors' work.\n\nI did not carefully check the computations but the theoretical parts seem rigorous.\n\nSection 4 provides rigorous experiments. We particularly appreciate that the authors try to empirically understand which specific factor of instabilities impacts generalization the most, and that they provide quite realistic experiments (section 4.4)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tries to understand the instabilities during neural network training when using large learning rates and their impact on generalization.\n\nThe paper theorizes that unstable, high learning rates improve generalization by rotating Hessian eigenvectors, helping the model explore broader regions in the loss landscape. Using a Diagonal Linear Network (DLN) as a toy model, the authors show that unstable parameter growth leads to these beneficial rotations. They confirm this effect experimentally with an MLP on fMNIST, where instabilities cause rotations of the sharpest Hessian eigenvectors.\n\nThen, the authors directly study generalization. They empirically study the impact of large learning rates in different settings (MLP/VGG, CIFAR/MNIST), showing that increasing the learning rate way beyond the theoretical stability limit improves generalization. The authors extend their analysis to decreasing learning rate schedules and try to untangle the impact of different variables correlated with instabilities (learning rate, eigenvector rotation, sharpness), finding that all these variables impact generalization performance, though the eigenvector rotation seems to play a crucial role (compared to sharpness which is often studied in the literature)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Though it allows for interesting derivations, the DLN model is quite limited, and the paper will be stronger if it studies a more realistic model.\n\nI'm wondering how robust the conclusions in Figure 8 are (see questions).\n\n(Minor) Though I understand that it would require more compute that academic researchers might not have, experiments on harder tasks with more modern models would enrich the paper.\n\nSome figures are quite hard to read. In general, I would recommend increasing the font size, especially for Figures 5 and 6.\n\nFor Figure 2: it can seem that the $S(\\theta)$ plot is done in the lr reduction setting \n\nFor Figure 7, I find the 3d plot hard to read.\n\nFor Figure 8, it took me some time to understand what was represented at the top.\n\nFor Figure 9, it might be interesting to report the result for lr=6.4 to show that 3.2 is the last stable value.\n\nThe code is not currently available, though it seems that it contains interesting contributions (for instance a new Jax implementation of Hessian-Vector product).\n\n\n### Typos\n\n- Arora et al. (2022) demonstrated that gradient descent can lead to alignment between gradient the sharpest eigenvector of the Hessian \n- where exactly derivations are possible" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This work identifies the implicit regularisation effects of instabilities in gradient descent, attainable through the use of large learning rates." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024can,\ntitle={Can Stability be Detrimental? Better Generalization through Gradient Descent Instabilities},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=zPaTnGjgpa},\nnote={under review}\n}" }, "abstract": { "value": "Traditional analyses of gradient descent optimization show that, when the largest eigenvalue of the loss Hessian - often referred to as the sharpness - is below a critical learning-rate threshold, then training is ‘stable’ and training loss decreases monotonically. Recent studies, however, have suggested that the majority of modern deep neural networks achieve good performance despite operating outside this stable regime. In this work, we demonstrate that such instabilities, induced by large learning rates, move model parameters toward flatter regions of the loss landscape. Our crucial insight lies in noting that, during these instabilities, the orientation of the Hessian eigenvectors rotate. This, we conjecture, allows the model to explore regions of the loss landscape that display more desirable geometrical properties for generalization, such as flatness. These rotations are a consequence of network depth, and we prove that for any network with depth $> 1$, unstable growth in parameters cause rotations in the principal components of the Hessian, which promote exploration of the parameter space away from unstable directions. Our empirical studies reveal an implicit regularization effect in gradient descent with large learning rates operating beyond the stability threshold. We find these lead to excellent generalization performance on modern benchmark datasets." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Gradient Descent", "Generalization", "Optimization", "Stability" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/12415da43e1311d6a1358172ecac64373ec80a7a.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Can Stability be Detrimental? Better Generalization through Gradient Descent Instabilities" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]