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": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "Yes, Discrimination / bias / fairness concerns" ] }, "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 you please explain how quality control is conducted in the construction of data across different evaluation dimensions? For example, how is a human verification process conducted, and what is the proportion of data that gets filtered?\n\n2. Given the criticisms of CLIP-based methods as 'bag-of-words' and the lack of interpretability and reproducibility when directly asking VLMs to output scores, evaluating and comparing alternative evaluation methods is crucial. It's important to include popular alignment approaches that avoid these weaknesses, such as decomposition methods from [T2i-compbench](https://proceedings.neurips.cc/paper_files/paper/2023/file/f8ad010cdd9143dbb0e9308c093aff24-Paper-Datasets_and_Benchmarks.pdf) and [Davidsonian Scene Graph](https://arxiv.org/pdf/2310.18235), as well as answer probability methods like [VQAScore](https://arxiv.org/pdf/2404.01291)." }, "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 logically coherent. The proposed benchmark is comprehensive, covering aspects such as alignment, safety, image quality, and bias." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MJ-BENCH, a novel benchmark that includes a comprehensive preference dataset to evaluate multimodal judges in providing feedback for image generation models across four key perspectives: alignment, safety, image quality, and bias. Detailed experimental analyses were conducted on CLIP-based scoring models, open-source VLMs, and close-source VLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I would consider increasing my score if these concerns are addressed or resolved. \n\n**1. Comprehensiveness and fairness of the evaluation**: The paper evaluates models like CLIP, LLaVA and GPT-4, but lacks some popular alternative alignment evaluation models such as the Decompose method ([Davidsonian Scene Graph](https://arxiv.org/pdf/2310.18235)) and the Answer Probability Method ([VQAScore](https://arxiv.org/pdf/2404.01291)). Additionally, the paper claims that the alignment dataset was collected from Pick-a-pic, HPDv2, and ImageRewardDB, so evaluating PickScore-v1, HPS-v2.1, and ImageReward in the experiments is unfair because these models have already been trained on similar data or dataset formats." }, "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": 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": "Please see in Weaknesses" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The writing and presentation of the entire paper are good.\n* A comprehensive benchmark is provided, which can further advance research in the related field.\n* It further promotes research on RLAIF (Reinforcement Learning from AI Feedbacks) and provides solid evidence for its effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "* The paper introduces MJ-BENCH, a benchmark to improve feedback for text-to-image models on alignment, safety, image quality, and bias. \n* Some interesting points are found: e.g., Closed-source VLMs like GPT-4o outperform others overall, while smaller models excel in alignment and quality, and VLMs perform better on safety and bias. Human evaluations confirm MJ-BENCH’s effectiveness for model fine-tuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Please provide a comprehensive discussion that covers the related work. For example, some studies have also made efforts in aligning text-to-image models with feedback from MLLMs, such as VisionPrefe r[1], which also discusses the varying effectiveness of different MLLMs as annotators for alignment data in text-to-image tasks.\n\n* It would be beneficial to discuss further the situations in which human judges and MLLM judges disagree, as this could provide valuable insights for future work.\n\n* Besides, to better demonstrate the effectiveness of MJ-BENCH, the authors are recommended to present some visualization cases of MJ-BENCH to help to offer a clearer and more comprehensive understanding of how well the dataset.\n\n**Reference**\n[1] Wu X, Huang S, Wei F. Multimodal Large Language Model is a Human-Aligned Annotator for Text-to-Image Generation[J]. 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": 3 }, "primary_area": null, "questions": { "value": "- I found that very limited details about how to use GPT-4o/4v/Gemini/Claude as the reward model for DPO/DDPO is provided in the paper. Do you use online or offline RL? If online, how to make use of these APIs and obtain the reward for DPO/DDPO fast? What if the API calls fail? If offline, what is the dataset used to finetune the model with DPO/DDPO?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Evaluating and benchmarking the capability of LMMs is an important research direction of the community nowadays. This paper proposes a comparably comprehensive benchmark set and provides the comparison of a large amount of LMMs, which is valuable to the community.\n- I like the design of the benchmark questions based on (win, lose, prompt) triplets, which, in my opinion, can make the human annotation easier compared with direct quality rating." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a benchmark for evaluating LMMs about the judgement capability for T2I generation. The proposed benchmark set contains a few thousands (win, lose, prompt) triplets covering T2I evaluation aspects: alignment, safety, quality, and bias. To show that the proposed benchmark set can provide a fair platform to compare LMMs, this paper uses LMMs as the reward model and use DPO/DDPO to finetune SD1.5. The results show that the model finetuned with a better LMM (evaluated with the proposed benchmark) is preferred by human." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- In terms of the presentation, I am unsure whether claiming evaluating \"reward models\" is a better idea than claiming evaluating \"LMMs\". Close-sourced LMMs such as GPT-4o/4v/Claude are not widely adopted T2I reward models due to their high cost and low throughput. In addition, Table 10 summarizes the number of evaluation questions for each category of the benchmark, which is of the pivotal importance and should be put in the main paper.\n- In Table 10, how the number of evaluation questions for each category is determined? I found that most categories, scenarios, and subsets have random numbers of evaluation questions. If so, does this evaluation benchmark introduce bias by itself? For example, the questions for object (250) is more than 4 times of counting questions (55). Since all LMMs are evaluated by the averaged metrics, does the proposed benchmark biased towards LMMs that are better at object questions?\n- Table 2 and 3 show the results of SD1.5 finetuned with DPO and DDPO, which, according to my understanding, is the \"evaluation\" to the proposed benchmark. However, I found that the LMM that work best as the reward (Table 2) is not aligned with the LMM having the highest evaluation scores based on the benchmark (Table 1). Does it mean the proposed benchmark is not good to evaluate T2I reward models? In addition, in Table 3, GPT-4o and GPT-4v achieve the best results with DPO and DDPO, respectively. I think this result suggest that the evaluation of reward model should be performed for certain RLHF methods. I suggest the authors to provide more discussion about 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": { "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": "Please refer 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": "1. The paper tackles a crucial issue in evaluating text-to-image models. \n2. The paper includes detailed ablation studies with some insightful findings. For instance, the analysis of the consistency of the preference of the judges w.r.t different image modes. \n3. The proposed dataset offers many formats (e.g., ranking and voting), which can enable a wider variety of preference modeling." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper organizes a new preference dataset called MJ-Bench, which captures four different aspects of text-to-image generation: text-image alignment, safety, image quality, and bias. \n\nThe authors propose two methods to obtain feedback from multi-modal judges: single-input and multi-input judge. The paper also conducts detailed analysis, such as the consistency of the preference of the judges w.r.t different image modes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper's novelty seems limited, as the whole pipeline has been proposed in the previous method. Besides, The selection of these four aspects lacks in-depth analysis.\n2. The paper employs human evaluators for experimental evaluation multiple times, yet it fails to report the number of human evaluators involved. Since human evaluators may introduce bias, it is recommended to report this metric. If the number is small, it is advised to increase the scale of human evaluators.\n3. The scale of the dataset was not compared with other existing datasets. As a result, the application scope of the dataset may be limited.\n4. Since this work is a benchmark study, the quality of the samples within the benchmark was not evaluated. It is recommended to supplement the study with an experiment to assess the quality of the samples in the dataset.\n5. The dataset only involved feedback from six judges to train a model (as indicated on page 7). It is suggested to supplement the study with an experiment where feedback is directly constructed from the benchmark data to train a model, in order to observe the outcomes." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A novel benchmark using a comprehensive preference dataset to evaluate multimodal judges across multiple key perspectives" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024mjbench,\ntitle={{MJ}-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation?},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vxutwN3xQN},\nnote={under review}\n}" }, "abstract": { "value": "While text-to-image models like DALLE-3 and Stable Diffusion are rapidly proliferating, they often encounter challenges such as hallucination, bias, and the production of unsafe, low-quality output. To effectively address these issues, it is crucial to align these models with desired behaviors based on feedback from a multimodal judge. Despite their significance, current multimodal judges frequently undergo inadequate evaluation of their capabilities and limitations, potentially leading to misalignment and unsafe fine-tuning outcomes. To address this issue, we introduce MJ-BENCH, a novel benchmark which incorporates a comprehensive preference dataset to evaluate multimodal judges in providing feedback for image generation models across four key perspectives: alignment, safety, image quality, and bias. Specifically, we evaluate a large variety of multimodal judges including smaller-sized CLIP-based scoring models, open-source VLMs (e.g. LLaVA family), and close-source VLMs (e.g. GPT-4o, Claude 3) on each decomposed subcategory of our preference dataset. Experiments reveal that close-source VLMs generally provide better feedback, with GPT-4o outperforming other judges in average. Compared with open-source VLMs, smaller-sized scoring models can provide better feedback regarding text-image alignment and image quality, while VLMs provide more accurate feedback regarding safety and generation bias due to their stronger reasoning capabilities. Further studies in feedback scale reveal that VLM judges can generally provide more accurate and stable feedback in natural language (Likert-scale) than numerical scales. Notably, human evaluations on end-to-end fine-tuned models using separate feedback from these multimodal judges provide similar conclusions, further confirming the effectiveness of MJ-Bench." }, "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 Reward Models", "Foundation Models Alignment", "Reinforcement Learning from Human Feedback" ] }, "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/52020be06ecbfd54fd78d6d18ddca0681b7e9d26.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": "MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-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": 1 }, "primary_area": null, "questions": { "value": "In the second-level bullet points before Proposition 2, there are various claims about convegence, why are they true?\n\nHow is Proposition 3 \"the formal and clean version of Theorem 2\", e.g. it contains nothing like the last part of Theorem 2? And in what sense is Theorem 2 informal and dirty?\n\nThe remarks after Proposition 3 seem to contradict the last part of Theorem 2, i.e. \"we can prove that no matter the step size and the initialization we have linear convergence\" is at odds with \"we have convergence but it could be logarithmically slow\"?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This is a simple yet challenging setting, chosen well for this in-depth theoretical analysis.\n\nThe results are non-trivial and interesting, and proofs are provided in the appendix.\n\nThe plots from the experiments are detailed and complement the theoretical results well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper builds on recent studies of gradient descent on a two-layer linear network with scalar inputs and outputs. The main results are on the location and speed of convergence of gradient descent. By a careful analysis of the training dynamics which includes the edge of stability phenomenon, the main results elucidate the interplays among the learning rate, the implicit regularization, and the speed of convergence. The theoretical results are supplemented by plots provided by numerical experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The grammar and spelling are poor throughout the paper, and especially in Sections 5-7, sometimes causing near unreadability.\n\nIn Sections 5-7, the presentation is confusing. It is not even clear what the purpose of those sections is, I suppose they are meant to show some components of the proofs of Theorems 1 and 2, since they consist of only Definitions, Lemmas and Propositions, however that is not explained. If that is true, then it is not clear how and to what extent Sections 5-7 provide outlines of the proofs of Theorems 1 and 2.\n\nThere are no suggestions for future work.\n\nThere are no details about the experiments, and no code as supplementary materials.\n\nMore minor issues:\n- In the last paragraph on page 2, I think the multiplier of Q(t) in parentheses should be in absolute value.\n- In equation (8), I am not sure where the $\\sqrt{3}$ above $\\approx$ comes from, and also what the meaning of writing a condition like this above $\\approx$ is.\n- In Theorem 2, after \"If\", what is $\\epsilon$, is it meant to be $\\epsilon(0)$?\n- It seems Section 6 should have been Section 5.2.\n- \"Fig 3\" versus \"Figure 2\", please be consistent.\n- The citation of Wang et al. at the end of Section 6 should not be in parentheses.\n- Proposition 3 contains $\\eta < \\eta$.\n- Only one item in the References has a clickable link." }, "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": "The model treated in this paper is quite simple, while the same model has been treated in the existing literature. Is there any implication for more complex models, such as matrix factorization with multi-dimensional input?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The analysis of the EoS phenomenon is one of the significant topics in the deep learning theory literature. The convergence results exhibited in Theorem 1 and 2 provide informative insights into this problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper analyzes the effect of taking a large learning rate in training two-layer linear neural networks with single input and single output, formalized by $\\mathbf{a}^\\top\\mathbf{b}$. First, the authors evaluate the imbalance between $\\mathbf{a}$ and $\\mathbf{b}$ and ensure that by taking a larger learning rate, GD converges to a more imbalanced solution.\nThe authors also analyze the convergence rate and show that a larger learning rate leads to a slower convergence speed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the topic treated in this paper is interesting, the latter section, specifically from section 5, needs to be completed. For example, I could not find the proof of Propositions 2 and 3 in the Appendix. Moreover, some results are demonstrated without detailed explanation. Regarding these drawbacks, I vote 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": 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": 3 }, "primary_area": null, "questions": { "value": "- Do the authors have any idea how to extend some of the exposed results to non-scalar inputs? For example, instead of $\\vec x\\in \\mathbb{R}$, at least $\\vec x\\in \\mathbb{R^2}$ ?\n- Section 5 is less than a quarter page long. Ether there is a problem/ something missing, or it should be merged together with Section 6 and better contextualized.\nMinor:\n- I think displaying $\\lambda$ in bold can be misleading, since in the rest of the paper the bold is used for vectors.\n- Typo line 430: $\\eta<\\eta$." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- All results presented are precise and analytical, presented with mathematical rigor and well written. The model is extensively analyzed and all possible aspects are discussed in detail.\n- I think the most valuable takeaway from this paper is the proof that, even in a very simple model, gradient flow dynamics are inherently different from gradient descent. In particular, the paper shows that GD regularizes better in this setting, but more generally can be used as a proof that the common use of GF as a theoretical tool for understanding GD is not always well founded." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper aims to study the GD dynamics of two-layer linear networks with scalar input and output. The authors present a detailed analysis of the model they introduce, being able to identify analytically all the relevant aspects of the setting. In particular, they show the ability of GD to converge for unexpectedly large learning rates (edge of stability), and a full characterization of the implicit regularization that GD brings. Finally, they merge these two aspects and show a trade-off between speed and regularization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I think the claim of studying flat linear networks is an overstatement. The setting is too simple to claim that it is a faithful model for networks. It seems to me that related work analyzing similar settings is either not as simplified as this, or is motivated by something else like a matrix factorization problem (or both).\n- Alongside the previous point, I am not sure that the results presented are relevant enough to pass the bar of the 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": 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": "See above\n\nand l. 068\n\n> More importantly, the dynamics when optimizing (2) with gradient descent are qualitatively similar to the dynamics of training more complex models\n\nCould the authors elaborate on this point?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The topic is relevant and the approach of studying simplified models is worthwhile. \nThe paper includes clear and well-designed figures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper analyzes the GD dynamic of a 2-layer linear network with scalar input and scalar output. The authors study the rate of convergence of GD and provide some properties of the reached solution by using conserved quantities during a gradient flow dynamic." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "General impression: The paper appears to have been written hastily, lacking structure, with numerous typographical errors, missing hypotheses, and unsatisfactory proofs that are difficult to follow. Major revisions requiring another round of review are necessary in my opinion.\n\nRegarding the results: The contributions seem overall quite weak and insufficient for acceptance at ICLR.\n\nSome comments/suggestions:\n\n1. The authors state \n> Despite its simplicity, the objective (2), which has also been studied by prior work (Lewkowycz et al., 2020; Wang et al., 2022; Chen & Bruna, 2023; Ahn et al., 2024), is a useful object of study because...\n\nBy quickly checking some of these references, it appears they do not restrict their analysis to the case $x_i, y_i \\in \\mathbb{R}$ only.\n\n2. The authors claim\n> In addition to showing how fast gradient descent converges to some global minimizer, we can also describe which of the many possible solutions, a ⊤b = Φ, gradient descent will converge to (l. 083) \n\nThe corresponding result in section 3 does not specify which solution is reached, but only describes certain properties of that minimizer (which could be satisfied by multiple solutions).\n\n3. Furthermore, the analysis for Theorem 1 requires $Q(0) \\neq 0$ at initialization. This should be stated in the hypotheses of Theorem 1 and in the paragraph following equation (7). Additionally, $Q(0) = 0$ is also a common hypothesis, this should be discussed.\n\n4. The paragraph following equation (8) motivates Takeaway 1 but falls short of rigorously proving it.\n\n5. Sections 4, 5, and 6 severely lack structure (Section 5, subsection 5.1??).\n\n6. Appendix C is unclear:\n- First, the conserved quantity is preserved for gradient flow (not necessarily gradient descent). Therefore, the conclusion\n> These two lemmas imply that the norm of the solution found by gradient descent will always be smaller than λ∞ and since ε∞ = 0 we can compute it using the formula above\n\n(note: two → one lemma?) about a solution found by *gradient descent* requires proof.\n- There appears to be confusion between $\\lambda$ before Lemma 7 (continuous) and after (discrete); the link is not examined.\n- The computations for the \"proof\" of Lemma 7 are incorrect. While the result may be true, a proper proof is needed.\n- not clear at what point is defined the maximal sharpness (the formulation of Lemma 7 suggests it is at initialization).\n- For Lemma 8, $\\eta$ needs to be less than 2/maximal sharpness; this should likely be also the case in Lemma 7 (the first inequality of Lemma 7 appears to lack any attempt of proof)?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024gradient,\ntitle={Gradient Descent Converges Linearly to Flatter Minima than Gradient Flow in Shallow Linear Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vxvgZ0kTFv},\nnote={under review}\n}" }, "abstract": { "value": "We study the gradient descent (GD) dynamics of a depth-2 linear neural network with a single input and output. We show that GD converges at an explicit linear rate to a global minimum of the training loss, even with a large stepsize--about $2/\\textrm{sharpness}$. It still converges for even larger stepsizes, but may do so very slowly. We also characterize the solution to which GD converges, which has lower norm and sharpness than the gradient flow solution.\nOur analysis reveals a trade off between the speed of convergence and the magnitude of implicit regularization.\nThis sheds light on the benefits of training at the ``Edge of Stability'', which induces additional regularization by delaying convergence and may have implications for training more complex 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": [ "Gradient Descent", "Implicit Regularization", "Shallow Networks", "Linear Networks" ] }, "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/31e58cf7d7b6653d71c36696273550dcf4b1067e.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": "Gradient Descent Converges Linearly to Flatter Minima than Gradient Flow in Shallow Linear 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": 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": "> The number of decoder layers.\n\nIt seems decoder only have one layer of CA and 6 layers of SA?\n\n> Speed of inference.\n\nHow fast can the QFree-Det be for some sparse scenario, i.e., only a instance in a image.\n\n> Performances in line 476\n\nIs the same experiments (47.1) as that of line 248?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "> 1. Motivation:\n\nThe paper introduces the QFree-Det model, addressing the variable number of instances in different images, which is a pertinent issue in object detection. The proposal for a dynamic query selection approach is innovative and could significantly enhance the flexibility and accuracy of detection models.\n\n> 2. Performance Improvement:\n\nThe paper demonstrates some performance improvements on the WiderPerson dataset, particularly in detecting small objects, which is a challenging aspect of computer vision." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel query-free detector, QFree-Det, aimed at addressing the limitation of a fixed number of object queries in Transformer-based detectors such as DETR and DINO. The authors propose an Adaptive Free Query Selection (AFQS) algorithm and a sequential matching method, which significantly enhance the model's capability to dynamically detect a variable number of objects across different input images. Additionally, a new loss function, PoCoo, is designed to improve detection capabilities. The experimental results demonstrate that QFree-Det achieves remarkable performance on multiple datasets and various backbone networks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "> Comparison with DETA and DDQ-DETR:\n\nThe discussions in Section 3.1.3 regarding the role of Self-Attention (SA) bear similarities to the disscussions taken in \"NMS Strikes Back\" (DETA) [1] and \"Dense Distinct Query for End-to-End Object Detection\" (DDQ-DETR) [2]. Both papers explore the effectiveness of traditional non-maximum suppression (NMS) and the role of SA integrated with transformer-based detectors, which is a relevant area of comparison for QFree-Det.\n\n[1] https://arxiv.org/pdf/2212.06137\n\n[2] Dense Distinct Query for End-to-End Object Detection (CVPR2023)\n\n> Comparative Experiments in Table 3:\n\nWhile the paper provides comparisons with models like DINO, it lacks a comprehensive set of comparative experiments with other state-of-the-art Transformer detectors, such as Stable DINO and CO-DINO. Including these comparisons would provide a more holistic view of QFree-Det's performance relative to the current landscape of object detection models.\n\n> Performance Comparison with Fixed Detectors:\n\nThe paper compares the performance of QFree-Det (AP 50.5) with DINO (AP 49.0). However, it is noted that improved versions of DINO, utilizing focal loss, have achieved an AP of 50.0, as demonstrated in the MMDETECTION repository. This raises the question of whether other state-of-the-art fixed detectors, when combined with loss functions like PoCoo or IA-BCE, could potentially outperform QFree-Det. Further experimentation and comparison with these models would be beneficial to assert the superiority or uniqueness of QFree-Det's approach." }, "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": "-How does QFree-Det perform on other challenging datasets beyond COCO and WiderPerson? Are there specific scenarios where QFree-Det might underperform?\n\n-Can you provide a more detailed analysis of the trade-offs between performance and computational cost in QFree-Det?\n\n-How does the AFQS algorithm scale with increasing image complexity and object density?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "-The AFQS algorithm and LDD framework provide a novel solution to the challenges of fixed query limitations and detection ambiguities that are common in traditional transformer-based detectors like DETR and DINO.\n\n-The model achieves state-of-the-art results on the COCO2017 and WiderPerson datasets, particularly excelling in small object detection. This demonstrates the effectiveness of the proposed method over existing leading techniques.\n\n-By enabling dynamic query selection, QFree-Det effectively adapts to varying scene complexities, providing a robust framework for diverse detection scenarios.\n\n-The introduction of the PoCoo loss function significantly enhances the detection capabilities for small objects, addressing a common challenge in object detection." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes QFree-Det, a novel query-free detector based on transformers, aiming to address the limitations of fixed query numbers and detection ambiguity in existing transformer-based detectors. The main contributions include the Adaptive Free Query Selection (AFQS) algorithm and the Location-Deduplication Decoder (LDD) framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-The authors propose AFQS algorithms and LDD frameworks, but lack in-depth theoretical analysis. There is no detailed explanation of why removing location queries (PQ) does not significantly affect model performance, nor is there a theoretical proof of how LDD can effectively solve the detection ambiguity problem. \n\n-Insufficient comparison with state-of-the-art methods, In Table 3, the authors mainly compare with older methods such as DINO, but lack detailed comparisons with more recently published methods (e.g. DAC-DETR, Co-DETR, etc.). This makes it difficult to assess the position of QFree-Det in the current research frontier. \n\n\n-The author mentions that AFQS can reduce redundant queries, but does not provide a detailed computational complexity analysis or efficiency comparison with other methods. Lack of specific inference time and memory usage data.\n\n-Although the improvement of the PoCoo loss function for small object detection is shown in Table 7, there is no in-depth analysis of why this loss function can be particularly effective in improving small object detection 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": 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 the weakness part" }, "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 analyze the ambiguity problem overlooked in current one-to-many detectors, and propose a new sequential matching method to tackle this problem, which is not explored before.\n2. The paper is well-written and the experiments are comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose QFree-Det, a query-free detector that select dynamic queries from encoders and decouples one-to-one and one-to-many matching in sequential way. It achieves remarkable improvements compared to the DINO baseline." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The dynamic query selection mechanism is similar to the distinct query selection process in DQS[1]. And the sequential matching process is similar to the hybrid layers in H-DETR[2] and MS-DETR[3]. It is better to add citation and discussion with 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": 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": "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": "1. The idea of producing a flexible number of queries to predict a flexible number of detections, is interesting.\n2. The paper is easy to follow, and the presentation is good.\n3. The performance of the proposed QFree-Det compared to SOTAs looks good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Transformer-based detectors like DETR and DINO have a fixed capacity for detecting objects, leading to missed detections and false positives. To overcome these limitations, QFree-Det is proposed as a dynamic, query-free detector that adapts to varying object counts in input images. It features an Adaptive Free Query Selection (AFQS) algorithm for flexible query selection and a Location-Deduplication Decoder (LDD) that separates one-to-one and one-to-many detection processes to reduce ambiguities. Additionally, a unified PoCoo loss enhances small object detection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The contribution of AFQS is not strong enough. The main idea of the proposed Adaptive Free Query Selection (AFQS) algorithm: The fixed content query (CQ) is replaced by the proposed flexible self-adaptive decoder query (SADQ), where SADQ is obtained by sorting the classification scores of all N encoder tokens and selecting the scores above a certain threshold as M SADQ. In short, the AFQS algorithm is a simple strategy to reduce N queries into M queries by sorting the classification scores. I don't think it is enough to be a main contribution.\n2. The novelty of sequential matching is incremental, where sequential matching combines Box Locating Part (One-to-many matching), and Deduplication Part (One-to-one matching). One-to-many matching and One-to-one matching are the existing methods.\n3. The framework of Box Locating Part (One-to-many matching) and Deduplication Part (One-to-one matching), looks complex. Deduplication Part (One-to-one matching) can be replaced by the parameter-free NMS. The Deduplication Part (One-to-one matching) looks more complex than NMS.\n4. PoCoo loss is useful for small object detection. However, the connection between PoCoo loss and the motivation of designing a query-free detector, is very weak.\n5. Dynamic query selection and the dual-step decoding process may require more computational resources, potentially impacting processing speed in real-time detection applications.\n6. According to Table 5, free-query with AFQS only is worse than fixed-query." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024qfreedet,\ntitle={{QF}ree-Det: Query-Free Detector with Transformer and Sequential Matching},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vyF5aim4US},\nnote={under review}\n}" }, "abstract": { "value": "Transformer-based detectors, such as DETR and DINO, often struggle with a specific limitation: they can detect only a fixed number of objects based on the predefined number of queries set. This limitation leads to missed detections when the scene exceeds the model’s capacity and increases false positives when the scene contains fewer objects. In addition, existing approaches often combine one-to-one and one-to-many matching label assignment methods in the decoder for accelerating the model training and convergence. However, this operation can introduce a new detecting ambiguity issue, which is often overlooked by those methods. To address these challenges, we propose QFree-Det, a novel query-free detector capable of dynamically detecting a variable number of objects across different input images. In particular, we present an Adaptive Free Query Selection (AFQS) algorithm to dynamically select queries from the encoder tokens, which resolves the issue of fixed capacity. Then, we propose a sequential matching method that decouples the one-to-one and one-to-many processes into separating sequential steps, effectively addressing the issue of detecting ambiguity. To achieve the sequential matching, we design a new Location-Deduplication Decoder (LDD) by rethinking the role of cross-attention (CA) and self-attention (SA) within the decoder. LDD first regresses the location of multiple boxes with CA in a one-to-many manner and then performs object classification to recognize and eliminate duplicate boxes with SA in a one-to-one manner. Finally, to improve the detection ability on small objects, we design a unified PoCoo loss that leverages prior knowledge of box size to encourage the model to pay more attention to small objects. Extensive experiments on COCO2017 and WiderPerson datasets demonstrate the effectiveness of our QFreeDet. For instance, QFree-Det achieves consistent and remarkable improvements over DINO across five different backbones. Notably, QFree-Det obtains a new state-of-the-art of 54.4% AP and 38.8% APs on val2017 of COCO with the backbone of VMamba-T under 1× training schedule (12 epochs), higher than DINO-VMamba-T by +0.9% AP and +2.2% APs. The source codes will be released upon acceptance." }, "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": [ "free-object prediction", "query-free", "detecting ambiguity", "location-deduplication decoder" ] }, "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/72c8c03c82dc62f1b0d6067dd7bf6945dd8748c4.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": "QFree-Det: Query-Free Detector with Transformer and Sequential Matching" }, "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": "- I find the two loss functions redundant. In fact, according to the Ablation Study, it seems that the contrastive loss is of little value. Section 3.5 details the two cost functions, but IMO it is not clear what kind of information is expected to be captured by the Contrastive Loss that cannot be obtained from the ECB loss. Could you please elaborate on this?\n\n- Related to this first question, ECB Loss and Contrastive Loss have different scales, but in the total loss function, no regularisation is applied on these scales. Have you tried to apply some kind of regularisation on the Contrastive Loss, which has wider ranges (especially at the beginning of the pre-training) to mitigate this difference in scales?\n\n- I see you are training every method for 50 epochs. What is the rationale behind it? I am expecting when removing the Contrastive Loss, the method will converge faster and maybe a better weights configuration can be achieved with less amount of epochs and benefit from early stopping according to loss values from a evaluation dataset split." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Multimodal Learning is a powerful approach. Studies like this one about how to incorporate the information from reports in order to enhance the learning of the model are needed.\n\n- The core of the paper, incorporating this supervised objective based on the clinical entities appearing in the original report makes sense. In addition, they demonstrate that it improves the overall performance of the method compared to other methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a new multimodal method for optimizing an encoder to process ECG signals, enhancing the information in the reports associated with such tracing. They claim the limitations of simply aligning the report representations with the ECG tracing representations and incorporating a new loss function that assesses the ability to infer which relevant clinical entities appear in the original report. \n\nFurthermore, they suggest that when processing multi-lead ECG signals with a transformer architecture, lead embedding should be added to increase the learning potential of the model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- My biggest concern is that IMO the paper is not properly contextualized according to the existing literature. Methods like MedKLIP, KAD, or MAVL are briefly mentioned in line 106 and no further explanation is provided about what makes K-MERL stand out compared with them. I am finding the proposed framework mirrors KAD's framework without adding any novelty to the Multi-Modal framework. \n\n- Authors claim that they design a novel \"lead-specific tokenization\". (Line 63) I do not see any differences between the way they embed lead information compared with other studies such as ST-MEM paper (Also used as a benchmark), which also includes this kind of lead embedding in its framework.\n\n- Even the amount of baselines used during the evaluation is significant, most of them (all except one) are just trained on single-modality data. In addition, Most of those are not ECG-specific methods but image-processing ones. I am missing some relevant baselines such as PCLR [1], MAEFE [2], or DEAPS [3]. (See references).\n\n[1] Nathaniel Diamant, Erik Reinertsen, Steven Song, Aaron D. Aguirre, Collin M. Stultz, and Puneet\nBatra. Patient contrastive learning: A performant, expressive, and practical approach to electro-\ncardiogram modeling. PLOS Computational Biology, 18(2):1–16, 02 2022. doi: 10.1371/journal.\npcbi.1009862. URL https://doi.org/10.1371/journal.pcbi.1009862.\n\n[2] Huaicheng Zhang, Wenhan Liu, Jiguang Shi, Sheng Chang, Hao Wang, Jin He, and Qijun Huang.\nMaefe: Masked autoencoders family of electrocardiogram for self-supervised pretraining and\ntransfer learning. IEEE Transactions on Instrumentation and Measurement, 72:1–15, 2023. doi:\n10.1109/TIM.2022.3228267.\n\n[3] Adrian Atienza, Jakob Bardram, and Sadasivan Puthusserypady. Contrastive learning is not optimal\nfor quasiperiodic time series. In Kate Larson (ed.), Proceedings of the Thirty-Third International\nJoint Conference on Artificial Intelligence, IJCAI-24, pp. 3661–3668. International Joint Con-\nferences on Artificial Intelligence Organization, 8 2024. doi: 10.24963/ijcai.2024/405. URL\nhttps://doi.org/10.24963/ijcai.2024/405. Main Track." }, "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": "1. **Ambiguities of text encoder**:\n - It is unclear whether the text encoder is fixed or trained during pre-training. The role of the text encoder needs more clarity.\n2. **Ambiguities of lead-specific tokenization**:\n - The lead-specific tokenization process is not well explained. Questions arise on how experiments like those in Table 2(b) were conducted, especially regarding the application of lead-specific spatial positional embedding without lead-specific tokenization.(The experiment which earns 68.47 for 1 Lead and 74.23 for 12 Leads)\n3. **Ambiguities of seen classes and unseen classes**: \n - The zero-shot evaluation lacks clarity on how the 35 fixed classes were determined across different datasets.\n4. **Differnece between other baseline methods**: \n - How does the \"lead-specific processing\" proposed in this paper differ from the approach used in ST-MEM, which you cited?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The strength of this paper lies in its innovative integration of lead-specific processing and cardiac-related entities from LLMs to enhance ECG-text alignment, demonstrating superior performance across various downstream tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper aims to address the limitations of MERL by introducing lead-specific processing and leveraging cardiac-related entities extracted from large language models (LLMs) to improve alignment between ECGs and text reports. It evaluates the approach on various downstream datasets, demonstrating superior performance in linear probing and zero-shot classification compared to other baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Contribution Clarity and Experimental Validation**:\n - The paper highlights improvements over MERL, specifically in alignment between ECGs and free-text reports and a method considering spatio-temporal aspect of ECGs. However, there are no experiments demonstrating superior alignment in K-MERL compared to MERL, nor is there an ablation study showing that using LAMA for cardiac entity extraction results in less noise.\n - The spatio-temporal aspect is not convincingly validated. While partial lead input experiments are presented, they are insufficient. Comparisons with other spatio-temporal methods are missing.\n\n2. **Presentation Issues**:\n - The token size \\( p \\) is not mentioned in the tokenization section.\n - Figure 7 lacks clear labeling between parts (a) and (b). \n - Figures are poorly aligned with the text, placed too closely together, and not self-contained, particularly Table 2.\n - The inclusion of baseline performance in Figure 7 for ablation results is unnecessary and could have been presented more effectively in a table.\n - The notation for loss calculation is problematic, as \\( L \\) is the mini-batch size and \\( N \\) the total dataset size, which seems inappropriate.\n - If I have understood correctly, the \bformula in Section 3.2 under **Lead-specific Spatial Positional Embedding**, [$lead_{1}$ + **W**$e_{i}^{l}$[$p_{1}$], ..., $lead_{1}$ + **W**$e_{i}^{l}$[$p_{M}$], ..., $lead_{12}$ + **W**$e_{i}^{l}$[$p_{1}$], ... , $lead_{12}$ + **W**$e_{i}^{l}$[$p_{M}$]] should be updated to: [$lead_{1}$ + **W**$e_{i}^{1}$[$p_{1}$], ..., $lead_{1}$ + **W**$e_{i}^{1}$[$p_{M}$], ..., $lead_{12}$ + **W**$e_{i}^{12}$[$p_{1}$], ... , $lead_{12}$ + **W**$e_{i}^{12}$[$p_{M}$]]\n\n\n3. **Prompt statement**:\n - The prompt statement selection lacks justification, unlike the detailed description provided in MERL.\n\n4. **Missing Performance Metrics**:\n - The paper does not show results for full fine-tuning or partial fine-tuning.\n - Training and inference times compared to MERL are not reported, which is essential for understanding the practical implications of the proposed 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": 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": "Besides the questions/comments in the above weakness section. The authors should improve the presentaiton of this paper. For example, in the motivation of this paper, the authors first mentioned suboptimal alignment and then leading modalities, while for the method part, the authors first introduce the method to tackle the second challenging and then the first challenge. Moreover, for section 3.1, the first and second paragraphs seems redudant." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed framework is overall reasonable. \n2. The idea of tackling lead missing (or arbituray leads) seems clinically relevant and important.\n3. The proposed framework achieves good performance on the benchmark datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Knowledge-enhanced Multimodal ECG Representation Learning (K-MERL), a new framework that improves ECG diagnostic accuracy by addressing limitations in current ECG multimodal learning models. K-MERL utilizes large language models to extract structured knowledge from free-text ECG reports, enhancing the alignment of ECG signals with textual information. Furthermore, it introduces a lead-aware ECG encoder with dynamic lead masking, enabling the model to handle arbitrary ECG lead combinations, a critical feature for diverse clinical settings. Extensive evaluations on six external datasets demonstrate K-MERL’s superior performance, achieving state-of-the-art results in both zero-shot classification and linear probing, especially in scenarios with partial lead inputs, underscoring its adaptability and robustness in real-world applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The rationale and advantages of the \"lead-aware ECG encoder\" should be further elaborated. The basic idea is to treat signals from different leads as separate tokens and to apply different position embeddings to these leads. While this approach can help learn and distinguish information from various leads, it may also complicate network training due to the increased introduction of raw but noisy signals as input. Therefore, it is essential to further clarify and validate the rationale and advantages of this approach. \n\n2. One of the main contributions is the claim regarding the ability to handle missing leads in deployment. Although the proposed lead random masking strategy is reasonable, its effectiveness and advantages should be further validated. The baseline compared (MERL with zero mean filling) is suboptimal, and it is expected that this baseline would demonstrate inferior performance. It is challenging to assert that the proposed framework achieves its goals based solely on the limited experimental results presented. A better comparison might be MERL with random masking during training.\n\n3. The proposed framework utilizes a fixed classifier for classification. Therefore, it remains unclear whether the claim of \"zero-shot classification\" is accurate. Moreover, the comparison with the previous state-of-the-art (SOTA) MERL may also be questionable, as MERL adopts a CLIP-based approach for final classification.\n\n4. The entire framework appears to be an ad-hoc combination of several existing components, making the advancements in this field unclear. The authors should further clarify their technical contributions or insights, in addition to presenting the positive experimental 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": "It would be beneficial for the authors to clearly address the concerns raised above regarding the limited originality compared to MERL, as well as the limitations of considering ECG and text reports as truly multimodal ECG representations." }, "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-structured, offering clear explanations of the model components and experimental setup. Each section logically guides readers through the model's rationale and implementation with clarity. The descriptions of key contributions—such as the lead-aware encoder and entity extraction process—are both detailed and accessible.\n\nThe research is bolstered by comprehensive experimental results across six external ECG datasets. The authors conduct thorough evaluations, including zero-shot classification, linear probing, and ablation studies. These assessments effectively demonstrate the robustness of the proposed framework and its individual components." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses limitations in existing multimodal electrocardiogram (ECG) representation learning, particularly issues in aligning ECG signals with free-text reports due to the unstructured nature of medical language. The authors propose a new framework, Knowledge-enhanced Multimodal ECG Representation Learning (K-MERL), which utilizes large language models (LLMs) to extract structured cardiac knowledge from free-text reports to improve the ECG learning process.\n\nKey contributions include:\n\n1. Structured Knowledge Extraction: K-MERL converts free-text ECG reports into structured cardiac entities using LLMs, enhancing the quality of multimodal ECG learning.\n2. Lead-aware ECG Encoding: A lead-specific encoder captures spatial-temporal characteristics of individual leads with a dynamic lead masking strategy, enabling flexibility in handling arbitrary lead combinations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The primary difference between K-MERL and the previous MERL (Liu et al., 2024) approach is the extraction of entities from cardiac pre-text; however, this alone does not merit a high score in terms of originality and novelty.\n\nAdditionally, the experiments used to validate the hypotheses largely replicate those conducted in the MERL (Liu et al., 2024) paper, lacking significant variation or new insights.\nFor instance, since text reports are harder to obtain compared to raw ECG data, it would be meaningful to conduct experiments to determine how many ECG-text pairs are necessary for this approach to be effective. Additionally, if utilizing diagnoses based on ICD codes yields better results than extracting cardiac entities from the ECG text reports, it would reduce the significance of using text reports. Instead, it would suggest a novel approach that combines ECG data with diagnostic history as a new modality.\n\nFurthermore, while the authors claim that K-MERL enables learning of multimodal ECG representations, the text reports in the MIMIC-ECG dataset are essentially rule-based text diagnoses initially generated by the ECG equipment provider and subsequently reviewed by medical professionals. Thus, these ECG text reports are more a structured representation of ECG features than an entirely new modality (e.g., blood tests, vital signs, or medical images) distinct from the ECG modality itself." }, "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": "Main questions:\n- Line 84: Citing 5 works from the same group of authors Liu et al., 2023a,c,d,e,f for \"biomedical applications\" is excessive and could be interpreted as inflating citations. Other relevant work could include [1,2,3] and many more. Similarly, line 40, citation of Liu et al., 2023b seems to be unrelated to annotation effort and therefore does not support this claim. Can you justify why this group of authors is most relevant to support these claims in that excess compared to other citations?\n- Section 3.3:\n - You mention to mask {9,10,11} leads. Only later it becomes clear that they are only partially masked. Please clarify this.\n - Why do you mask up to 11 leads? In real applications, you will encounter only a small set of variations in ECG lead recordings. This includes 1 lead ECG with specific leads measured and 5 and 6 lead ECG with only specific leads measured. Thus having all random permutations of masked leads is not a realistic setup that you will encounter. \n - Do you use all 12 leads? If so, why? This is redundant since only 8 leads in a 12 lead ECG are mathematically independent.\n - Line 247: In Figure 2 it is unclear whether masking occurs over a complete lead or only on a segment of a lead. Please clarify\n - Line 248: Why is it problematic if \"some leads have more tokens than others\"? But above you say you mask up to 11 leads meaning that one lead will not be masked at all. This is not clear.\n - Line 251: Why do you choose the masking ratio as 0.25, what is the rationale? Does this mean that 25% of each lead or of all tokens are masked? Please justify and clarify.\n- Line 273: Why do you use an LLM to check for the true availability of extracted entities? This can again introduce errors. Instead, you can simply check if the text includes these entities with a string compare operation. What is the performance difference between a string match and your LLM approach or why do you opt for the LLMS approach?\n- Eq. 2/3: Can you justify why you need to algin the ECG to both, the report and the entities? Since the entities are originated in the report, there is no additional new information.\n- Table 1: The performance gains over MERL (ResNet) are small. This needs discussion on why this is the case and why K-MERL should be preferred. Can you provide evidence why in linear probing your method is not improving significantly and what the difference is to your other experiments where you show larger performance improvements?\n\nMinor points:\n- Figure 1: unclear notation especially $\\mathcal{F}_{xx}$ is not defined. There are two losses (contrastive and BCE) and in the figure it is unclear what each of them does. Please update the figure to be understand from the information given until its first mentioning.\n- Figure 2,4,5,6,7 and Table 1: why is the caption font smaller?\n- Line 145: remove full stop before \"adaptable\" \n- Line 154: Starting a section with \"To this end\" makes it unclear what this refers to.\n- Line 197: All your experiment ECGs have a length of 10sec, how large do you choose M then? How do you deal with different sampling frequencies?\n- Line 268: remove the abbreviation KG. Only used once.\n- Line 278: \"from [the] whole dataset\".\n- Line 297: did you check that all 277 entities are meaningful and medically correct entities?\n- Line 307: \"Afterward[s]\" \n- Line 308: The projections $\\mathcal{P}_{e}$ and $\\mathcal{P}_{t}$ are not in Figure 1. Please correct.\n- Line 323: Is this $\\mathcal{F}_t$ the same as for the contrastive loss? So two losses are operating on this function and one on all others? \n- Line 360: If you don't have 4.2.2, then it doesn't make sense to have 4.2.1\n- Figure 4: I first tought the caption refers to the figure below it, which is actualy fig 5c. Please modify to make this clearer.\n- Figure 4: Avg AUC as a separate bar is misleading since it looks like its own test set. Please change.\n- Line 414: writing incomplete. \"to full showcase the our method's\" \n- Figure 5c: What do you mean with \"free with PE\". Your method does only barely outperform MERL with PE in this setting which needs discussion.\n- Figure 6: Not all lead combinations are actually clinically useful. Thus this comparison is ill-posed.\n- Line 533: \"lead[ ]&[ ]segment\". Best to remove the & at all.\n\n---\nReferences:\n\n[1] Stahlschmidt, S.R., Ulfenborg, B. and Synnergren, J., 2022. Multimodal deep learning for biomedical data fusion: a review. Briefings in Bioinformatics, 23(2), p.bbab569\n\n[2] Acosta, J.N., Falcone, G.J., Rajpurkar, P. and Topol, E.J., 2022. Multimodal biomedical AI. Nature Medicine, 28(9), pp.1773-1784\n\n[3] Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H. and Gao, J., 2024. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Aligning the ECG features with the ECG reports is a meaningful problem and the authors manage to improve this alignment by extracting cardiac related entites from the reports. \n- A rigorous comparison a large set of method is performed\n- Analysis with ablations to all relevant model components is presented" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present a new method called K-MERL to align ECG recording with ECG reports to enhance evaluation of ECGs. The method is agnostic to the number of leads used and trained in self-supervised way. Comparison with other methods are favorably to the presented method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It is unclear why lead masking is clinically useful in the way it is performed in the paper since not all combinations of leads are measured.\n- The method is an improvement over MERL but it the results over MERL (especially in Table 1) needs discussion since they seems incremental.\n- It is unclear why alignment with the entities helps the model since this information is practically already in the ECG report. This seems to be the main novelty 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": 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 quite curious whether the authors anticipated the issue that the signal segment required for knowledge alignment could potentially be masked, thus introducing noise. How did they handle this issue?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The main strength of this study is that it achieves SOTA performance in the ECG classification task, particularly by achieving good performance with relatively small amounts of data. Additionally, it seems that this structured report-based approach using large models is applicable to any medical imaging/signal classification tasks, demonstrating good generalizability. Moreover, the experiments in this paper are very solid." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a cross-modal foundation model training method that incorporates external knowledge. It generates representations of signal data and text reports through encoders, and then aligns them using contrastive learning. To enhance the model’s robustness, the authors introduce a dynamic signal masking scheme. Subsequently, the model uses LLM to structure the text reports, allowing it to participate in multi-label prediction tasks during pretraining. This strategy significantly expands the supervisory labels for the model, leading to a more effective foundation model that improves performance in downstream applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper has two main shortcomings. The first is that the handling of ECG-entity alignment appears somewhat rough. The masking of the ECG signals is done randomly, which means that it is possible for the part of the signal corresponding to an entity was coincidentally masked. This could negatively affect the performance of knowledge alignment. I believe this is an obvious potential issue, but it seems that the authors have not addressed it in any way. The second issue is that the formatting of the paper is somewhat disorganized, and it appears that in an effort to compress the paper to within 10 pages, the authors removed a lot of content. This makes the technical section seem incomplete, and I had to rely on the figures to guess some implementation details. If this paper is accepted, I hope the authors can restore this content and place some of the experiments in the supplementary materials." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose K-MERL, the first framework to integrate structured knowledge from free-text reports for ECG multimodal learning. It achieves superior performance and supports arbitrary lead inputs, surpassing the limitations of fixed 12-lead setups." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024knowledgeenhanced,\ntitle={Knowledge-enhanced Multimodal {ECG} Representation Learning with Arbitrary-Lead Inputs},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vyFSyfiOIu},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in multimodal representation learning for electrocardiogram (ECG) have moved onto learning representations by aligning ECG signals with their paired free-text reports. \nHowever, current methods often result in suboptimal alignment of ECG signals with their corresponding text reports, thereby limiting diagnostic accuracy. This is primarily due to the complexity and unstructured nature of medical language, which makes it challenging to effectively align ECG signals with the corresponding text reports. \nAdditionally, these methods are unable to handle arbitrary combinations of ECG leads as inputs, which poses a challenge since 12-lead ECGs may not always be available in under-resourced clinical environments.\n\nIn this work, we propose the **K**nowledge-enhanced **M**ultimodal **E**CG **R**epresentation **L**earning (**K-MERL**) framework to address these challenges. \nK-MERL leverages large language models (LLMs) to extract structured knowledge from free-text reports, enhancing the effectiveness of ECG multimodal learning. \nFurthermore, we design a lead-aware ECG encoder to capture lead-specific spatial-temporal characteristics of 12-lead ECGs, with dynamic lead masking. This novel encoder allows our framework to handle arbitrary lead inputs, rather than being limited to a fixed set of full 12 leads, which existing methods necessitate.\n\nWe evaluate K-MERL on six external ECG datasets and demonstrate its superior capability. \nK-MERL not only outperforms all existing methods in zero-shot classification and linear probing tasks using 12 leads, but also achieves state-of-the-art (SOTA) results in partial-lead settings, with an average improvement of **16%** in AUC score on zero-shot classification compared to previous SOTA multimodal methods[^1].\n\n[^1]: All data and code will be released upon acceptance." }, "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": [ "Electrocardiogram", "healthcare", "physiological signals" ] }, "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/40932d0ca6ea3578ff9aabf76c8534982961b993.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": "Knowledge-enhanced Multimodal ECG Representation Learning with Arbitrary-Lead 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": 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": "None." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The phenomenon of instruction following without LLM post-training is observed in practice. It would be interesting to find the root cause or theoretical foundation of such phenomenon." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on exploration of approaches to enable LLM’s instruction following capabilities without instruction tuning. It discusses three different approaches including training solely on responses, instruction-response training on narrow-domain data, and changing a model’s distribution to yield instruction following. Though studying an interesting direction not widely adopted by main-stream approaches, the paper still has multiple drawbacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. All three approaches explored in this paper perform significantly worse than their comparable instruction-tuned versions in terms of win rates against the instruction-tuned models. It is unclear about the motivation of this work from the practical aspect.\n\n2. Besides missing practical motivation, the paper lacks theoretical analysis or parameter-level insight on why response-only tuning, narrow-domain instruction tuning, or distribution perturbation of certain tokens helps on instruction following as claimed in the paper. Some related work discussing how instructing tuning modifies LLMs could be referred to for this line of study.\n\n[1] Wu et al., “From Language Modeling to Instruction Following: Understanding the Behavior Shift in LLMs after Instruction Tuning”, NAACL, 2024.\n\n3. It is still unclear whether the instruction following capability has already been partially obtained through pretraining. Indeed, the experiments with OLMo, in which no instruction-tuning data was intentionally included in its pretraining, show similar conclusions as the experiments with Llama-2. However, existing pretraining corpus could include instruction tuning components from contents like human dialogues from novels, question-answer pairs from websites, etc. It requires careful investigation on the roles of pretraining, instruction tuning, as well as preference learning to reach persuasive conclusions about LLM’s instruction following capability." }, "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. More response examples should be provided for all the settings including response tuning, single-task tuning, and rule-based tuning." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The phenomenon proposed by this paper, Instruction Following without Instruction Tuning, is novel and interesting. \n2. This paper is well-written, easy to follow, and explicitly illustrates the points that might lead to misunderstanding, e.g. the definition of instruction-following behavior in line 183; the use of templates in line 178; and the potential rephrasing in line 280." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes that the instruction-following ability can be obtained without explicitly using instruction. It presents findings that models can follow instructions when trained only on response data, without explicit pairing with instructions. Additionally, single-task finetuning—training on narrow data domains like poetry or coding—also elicits broad instruction-following behavior across different tasks, such as recipe generation. The study also introduces a rule-based approach with simple modifications, showing that small changes to a language model’s distribution can yield instruction-following capabilities. This work challenges traditional views on instruction tuning, suggesting that implicit methods can lead to general instruction-following, raising questions about the necessity of explicit instruction tuning in developing adaptable, instruction-capable language models​" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper tries to reveal a transformative finding in the area of Instruction Tuning, while the experiments conducted in this paper are too limited to support the finding: \n(1) Only 2 LLMs are experimented, and all of them are slightly out-dated. Although as mentioned in line 101, the authors use these 2 LLMs because it is common to include instruction-tuning data during pretraining in modern days, only presenting results on 2 LLMs is still so limited. It gives the impression that your finding is only useful for slightly earlier LLMs. \n(2) For the response-tuning setting, only the LIMA dataset is utilized, which is also really limited. How can we know whether the instruction-following ability you mentioned can only be derived from the LIMA dataset? The LIMA dataset has some unique characteristics, e.g. all of them are long and deal with complex instructions. What if standard instruction data is utilized? What if the datasets used in single-task fine-tuning settings are used in response tuning? \n(3) Only the AlpacaEval metric is used to judge if the model has the instruction-following ability, which is not enough and is potentially biased. \n\n2. Related to the previous point, I think this phenomenon (response tuning to get instruction-following ability) might probably only work for some types of tasks. For example, if you only finetune the response on your Chess task, I don’t think the LLM can obtain a reasonable instruction-following ability. Thus more detailed experiments and discussion should be provided. \n\n3. The definition of instruction-following behavior in line 183 is not convincing to me and your experiments of rule-based response adapter actually deepened my suspect: If doing Slowly upweight EOS, Uniform token changes, and Encourage word diversity leads to better win rates, how can you ensure the gain on win rates is caused by real instruction-following ability or the drawbacks of this evaluation metric, which will give higher scores as long as the output is long and diverse. I think more controlled experiments and discussions are required for this part." }, "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": 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": "There are no specific questions. Kindly refer to the weakness. Additionally, ICLR does not have a best paper nomination. However, I consider this work worthy." }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "This is undeniably an exemplary piece of work. The authors courageously put forward a new paradigm for instruction tuning and meticulously present their arguments in a step-by-step manner.\n\nFrom a conceptual perspective, the conclusions of this paper have significantly transformed my understanding of instruction following. The authors demonstrate that instruction following can largely be accomplished by directly optimizing the response, and that these responses need not overly emphasize diversity.\n\nFrom an experimental standpoint, I highly appreciate the extensive experiments carried out by the authors. Naturally, I will add some of my personal viewpoints in the \"weaknesses\" section.\n\nRegarding the writing, I must apologize that I did not find the abstract particularly fluent to read. However, the instruction section is exceptionally well-written. The step-by-step, question-and-answer format is particularly captivating." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper conducts an exploration of instruction tuning of large language models (LLMs) in three progressively deeper layers. Firstly, the authors present what appears to be a radical conclusion—that fine-tuning solely on the answers while disregarding the instructions can significantly enhance the model's ability to follow instructions. Following this surprising conclusion, the authors further emphasize that the effectiveness of this \"response tuning\" method has no necessary correlation with the diversity of responses. In certain tasks, even employing extremely narrow, single-task responses for fine-tuning can improve the model's instruction-following performance. Finally, the authors provide their ultimate conclusion: this implicit instruction tuning method merely requires a slight adjustment in the model’s parameter distribution. The authors developed a rule-based language model and found that minimal alterations to the model’s output distribution can achieve remarkable improvements in instruction-following capabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have several minor suggestions regarding this paper. However, these suggestions by no means diminish its outstanding merits.\n\nFirstly, as previously mentioned, the experimental section requires further consideration. At this year's International Conference on Machine Learning (ICML), Allen Zeyuan Zhu presented his work titled \"The Pythics of Language Models.\" In this work, he maintained extremely strict control over variables from the pretraining phase, ensuring that there was no overlap or interference between the pretraining data and the data used for fine-tuning. Given that the conclusions of this paper are similarly groundbreaking, I suggest that, if resources permit, the authors could consider controlling variables from the pretraining phase. This would alleviate concerns about potential interference from the pretraining data of the LLaMA model affecting the conclusions. Naturally, pretraining costs are extremely high, and I do not consider this a shortcoming on the part of the authors. Nevertheless, conducting experiments using a wider range of base models might render the conclusions more robust.\n\nAdditionally, in the section on single-task tuning, although the conclusions are quite remarkable, I believe that the discussion of two exceptional datasets (one of which pertains to chess) could be enhanced.\n\nFinally, I found the abstract to be less engaging than the introduction." }, "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": "NA" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "+ The object of study lies at the central place of post-training technique. I find the selected problem important and fundamental. \n+ The article provides new insights into the post-training process, along with literature like LIMA, that post-training process could be more lightweight than we assume currently. The caveat in the conclusion is also important and worth highlighting that both adding AND removing behaviors picked up in pretraining is difficult." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper ran a comprehensive ablation study on instruction tuning, testing out several dimensions of the original instruction tuning design --- (1) tune conditionally on prompt? (2) tune over all tasks? (3) even tuning? The results show that current instruction tuning is kind of an overkill, thus suggesting finding simpler and more principled post-training techniques." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ Experiments are well designed and insightful. Even though, I wonder whether the inclusion of other instruction following benchmarks could add to the solidness of the results, e.g. MT Bench, or Arena hard.\n+ Worth noting that none of the ablated methods out-performs original instruction tuning. Therefore, this paper is not contributing new techniques to the field, as intended. Maybe introducing another dimension of measurement could be useful, e.g., language distribution drift to pretrained model, or alignment tax as measured by drop of performance on capability benchmarks." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Many adaptations that are deficient compared to instruction tuning also yield instruction-following language models" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024instruction,\ntitle={Instruction Following without Instruction Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vyHFTsOUWu},\nnote={under review}\n}" }, "abstract": { "value": "Instruction tuning commonly means finetuning a language model on instruction-\nresponse pairs. We discover two forms of adaptation (tuning) that are deficient\ncompared to instruction tuning, yet still yield instruction following; we call this\nimplicit instruction tuning. We first find that instruction-response pairs are not necessary: training solely on responses, without any corresponding instructions, yields\ninstruction following. This suggests pretrained models have an instruction-response\nmapping which is revealed by teaching the model the desired distribution of re-\nsponses. However, we then find it’s not necessary to teach the desired distribution\nof responses: instruction-response training on narrow-domain data like poetry still\nleads to broad instruction-following behavior like recipe generation. In particular,\nwhen instructions are very different from those in the narrow finetuning domain,\nmodels’ responses do not adhere to the style of the finetuning domain. To begin\nto explain implicit instruction tuning, we hypothesize that very simple changes to\na language model’s distribution yield instruction following. We support this by\nhand-writing a rule-based language model which yields instruction following in a\nproduct-of-experts with a pretrained model. The rules are to slowly increase the\nprobability of ending the sequence, penalize repetition, and uniformly change 15\nwords’ probabilities. In summary, adaptations made without being designed to\nyield instruction following can do so implicitly." }, "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": [ "instruction tuning", "instruction following", "ablation", "rule-based" ] }, "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/05b1e750b8c9b117a5362d77ebe387bf74d268f6.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": "Instruction Following without Instruction Tuning" }, "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": 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": "The generation DB-SYNTH consists of four steps intended to capture\nsynthetic task examples from workflows in published works. In\nparticular, a task dataset is constructed in the third step by\ngenerating synthetic data using various sampling strategies. The\nauthors do not explain how these steps, and particularly the data\ngeneration step, provide good representations of the task examples from\nworkflows in published work (notice that there is a missing reference\nto a section in the data generation paragraph). This justification\nshould be part of this work, as if the synthetic tasks are not good\nrepresentatives of the real examples, then we could be asking the LLMs\nto solve tasks that may be too simple or too complicated, which could\nreduce the quality of the tasks in the benchmark." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "S1) The authors introduce a simple yet expressive notion of\ndata-driven hypothesis. Discovering and validating such hypotheses\nfrom a dataset is a challenging problem for LLMs.\n\nS2) The authors develop a comprehensive benchmark to test the\ncapabilities of LLMs in discovering data-driven hypotheses.\n\nS3) The authors use the benchmark to test some popular LLM-based\nreasoning frameworks, drawing useful conclusions about the\ncapabilities of these systems in discovering data-driven hypotheses." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors consider the question of how capable\nstate-of-the-art LLMs are at automated data-driven discovery. More\nprecisely, the authors present a benchmark for this task, designed not\nonly to assess LLMs’ capability in discovery tasks but also to provide\ninformation for improving these capabilities.\n\nThe benchmark presented in this paper consists of 264 manually\ncollected tasks, along with 903 tasks that were synthetically\ngenerated. This benchmark is used to evaluate several LLM-based\nreasoning frameworks, providing useful information on the current\ncapabilities of such systems in automated data-driven discovery." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1) As part of the benchmark, the authors developed some synthetic\ntests. These tests are supposed to capture synthetic task examples\nconstructed from workflows in published works. However, the authors do\nnot clearly explain in what sense these synthetic tests properly\nrepresent these workflows." }, "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 can it be ensured that the data alterations in the proposed datasets (page 6) do not lead to contradictions with the discovery goal, resulting in only one answer—the target hypothesis?\n\nMinor remarks:\n - Pg 3 : discopagevery --> discovery\n - Pg 6: Sec ?? --> update the number of the section" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The role of LLMs in the scientific method is still unknown (if one exists), and evaluating their capacity to accelerate the process of knowledge discovery is a topic of significant interest. This work represents an incremental advancement in this domain, providing a formal definition of data-driven hypotheses and expanding the search space for these hypotheses. The paper also proposes evaluating proposed hypotheses against a gold standard using semantic similarity measures. In my opinion, this work is of genuine interest to the community.\n\n- The proposition is well described, and the code is available on GitHub. The authors have done impressive work collecting datasets, defining objectives for these datasets, and constructing gold standard hypotheses.\n\n- The results are clearly presented and discussed, demonstrating the limitations of current LLMs. This methodology can be utilized in the future to observe potential progress in the field of LLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The advances in data science and AI seem to be evolving the process of scientific discovery (the scientific method) by (drastically) reducing the time needed to formulate hypotheses, conduct experiments, and evaluate results. This paper proposes a benchmark to evaluate the potential role of LLMs in the automatic formulation of hypotheses, starting from a dataset and a natural language question about that dataset.\n\nTypically, for a classification or regression task, a human will identify certain attributes of interest (A_1, …, A_n), a target attribute Y, and attempt to build a relationship between these elements (A_1, …, A_n → Y). This is the current process for generating new knowledge. In this work, the authors aim to determine whether using LLMs allows for the production of valid knowledge in a more general setting: providing a dataset, posing a general question in natural language to the system using an LLM, and obtaining a response in the form of a scientific discovery. The main question the authors seek to answer is how well this approach performs using state-of-the-art LLMs.\n\nThere are some recent tools that evaluate the performance of LLMs in testing well-defined hypotheses (clearly formulated questions on a dataset). These methods utilize a constrained search space to look for hypotheses. DiscoveryBench expands the search space by finding hypotheses in a broader context, closer to a human approach. As expected, the performance of LLMs decreases in this setting. Indeed, the paper shows that the best results achieve 25% performance (which I interpret as 25% of responses being valid knowledge, even if it is a bit more complex to interpret)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The process of finding a data-driven hypothesis can be time/energy-consuming. This aspect is not discussed in the paper, and I understand that the page limit might not allow space for such discussions. Although this is not the main focus of the paper, it could be useful to explore whether there is a relationship between the size of the search space and the performance of the results.\n\n- The proposed formalism for defining data-driven hypotheses, discovery goals, and task difficulty is somewhat limited. At the same time, I understand that human expertise is essential in this process, and complete automation is difficult, if not impossible, to achieve." }, "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": "There are concerns regarding data safety/privacy/proprietary data -- these issues arise with any benchmark that is released to the public as this one. Can the authors please quickly comment on these issues, e.g., if it is safe to release their benchmark and what steps have they taken to ensure that the benchmark complies with data protection schemes." }, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety", "Yes, Responsible research practice (e.g., human subjects, data release)" ] }, "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 read my comments/questions from the above field. In addition: \n\n- One suggestion that I have for the authors is to start with the example given in lines 247 and show how G, h, \\psi, c, u, r, and \\mathcal{V}_D look like for this example. \n- Please clarify whether a hypothesis is represented as a sentence, a semantic tree, or a ternary tuple, and to explain how these different representations relate to each other if multiple are used.\n- Please define the space of valid verification procedures V_D and clarify whether finding V_D itself is part of the task or if it's given. \n- Another suggestion that I have for the authors is to look into automated theorem proving, (e.g., “Handbook of Automated Reasoning”, Volume 1, 2001), where the notions of hypotheses, theorem proving, proof trees, are formally defined. I would suggest using this notation to formally define the data-driven discovery task. I believe that this work will benefit a lot from a more rigorous formulation, e.g., to my understanding, the elements in \\mathcal{V}_D map to proof trees." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors attempt to formalize the problem of data-driven discovery, which, in my opinion, is a crucial step towards solving this task. In addition, a benchmark is introduced -- unfortunately, I did not have the chance to look at it in more detail, but the authors claim that they will publicly release it." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work deals with the problem of data-driven discovery using large language models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am confused with the formulation of the problem of data-driven discovery as in Section 3. In line 049, the authors give an example of a data-driven discovery task: “How did urban land use affect the invasion of introduced plants in Catalonia?” The answer to this task is the ways urban land used affected the invasion of introduced plants in Catalonia (if this is the case). However, in Section 3, line 152, you define a hypothesis as a tuple of three elements and in line 141, you say that a hypothesis maps to supported or unsupported. So, if a hypothesis maps to true or false (supported or unsupported), then the example task that you have in line 049 cannot map to your formulation. Please clarify what is the case and, please specify concretely how the task in line 049 translates to your formulation. \n\nI have a few other comments/questions regarding the formulation in Section 3:\n- In line 139, the authors say that a hypothesis is a sentence/semantic tree, and then in line 152 the authors define a hypothesis as a ternary tuple. Please fix this inconsistency, as it is confusing to the readers. \n- In line 141, the authors talk about a verification procedure \\mathcal{V}_D mapping a hypothesis to supported and unsupported. Is my understanding correct, that the goal of the data-driven task is both to find whether a hypothesis is supported and unsupported and to return the exact \\mathcal{V}_D? If my understanding is correct, then how do you define the space of valid verification procedures \\mathcal{V}_D? I would kindly ask the authors to clarify this part, as it is missing. \n- In line 152, the authors write h := \\psi(c,u,r). That formulation essentially means that h (the hypothesis) is defined as its answer, \\psi(c,u,r). I believe what the authors want to say is that \\mathcal{V}_D(h) = \\psi(c,u,r), i.e., that the answer to h is defined as \\psi(c,u,r). Please fix the notation accordingly or clarify what you meant. \n- Continuing with my previous comment, I would kindly ask the authors clarify what is the relationship between \\mathcal{V}_D and \\psi. \n- In line 161, the authors give another definition of the data-discovery task: that of given a goal G, to find a hypothesis h. From Section 3, I understood that the objective is: given a hypothesis h = (c,u,r) to derive \\psi (or \\mathcal{V}_D). I would kindly ask the authors to define concretely the problem they are dealing with, addressing all my comments from above, as now it is difficult for the reader to understand what is happening. \n- In line 293, the authors talk about hypothesis semantic trees, giving an example in Figure 1. However, in line 152, the hypothesis is a ternary tuple. \n- In line 240, the authors talk about the implementation workflow, however, they give no definition/specification of them in Section 3. \n\nThe above issues, make it difficult for the reader to understand this work and the proposed benchmark. I am willing, however, to increase my score if the authors address my comments/questions." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present DiscoveryBench, the first comprehensive benchmark that formalizes the multi-step process of data-driven discovery designed to systematically assess current LLM capabilities in discovery tasks." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024discoverybench,\ntitle={DiscoveryBench: Towards Data-Driven Discovery with Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vyflgpwfJW},\nnote={under review}\n}" }, "abstract": { "value": "Can the rapid advances in code generation, function calling, and data analysis using large language models (LLMs) help automate the search and verification of hypotheses purely from a set of provided datasets? To evaluate this question, we present DiscoveryBench, the first comprehensive benchmark that formalizes the multi-step process of data-driven discovery. The benchmark is designed to systematically assess current model capabilities in discovery tasks and provide a useful resource for improving them. Our benchmark contains 264 tasks collected across 6 diverse domains, such as sociology and engineering, by manually deriving discovery workflows from published papers to approximate the real-world challenges faced by researchers, where each task is defined by a dataset, its metadata, and a discovery goal in natural language. We additionally provide 903 synthetic tasks to conduct controlled evaluations on data-driven workflows that are not covered in the manually collected split. Furthermore, our structured formalism of data-driven discovery enables a facet-based evaluation that provides useful insights into different failure modes. We evaluate several popular LLM-based reasoning frameworks using both open and closed LLMs as baselines on DiscoveryBench and find that even the best system scores only 25%. Our benchmark, thus, illustrates the challenges in autonomous data-driven discovery and serves as a valuable resource for the community to make progress." }, "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": [ "data-driven discovery", "data analysis", "large language models", "hypothesis generation", "hypothesis verification" ] }, "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/3ff33a4564855e71c5cf33aee5a6afba51333a29.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": { "value": "/attachment/97e5d79d727f833c094a3c95e64d0437b4c5ced2.zip" }, "title": { "value": "DiscoveryBench: Towards Data-Driven Discovery with 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": 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) What is the effect of C++ vs Python?\n2) Please elaborate on the effects of variable size allocations on the Memgraph generation\n3) How does it compare to other existing work-conserving, priority-based algorithms?" }, "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 and the examples are helpful\n- The problem addressed in this work is interesting and relevant\n- Using LLaMA and LoRA in the experiments makes this work applicable" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents TURNIP, a GPU Runtime that incorporates CPU-RAM offload into the execution graph of AI models under memory constraints. Specifically, the authors create a MemGraph and GPU memory mapping by simulating task graphs. Comparing with PyTorch-based systems, they show faster runtimes of LLMs primarily for short sequence lengths." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The idea of \"nondeterministic\" is not new. The described scheduler is simply a work-conserving dynamic scheduler.\n- The simulation of the task graph to create the MemGraph is omitting any possible knowledge of execution times. Therefore, the resulting memgraph is only dependent on the topological sort and thus likely suboptimal.\n- The authors only present memory allocation sizes of 1 unit each and claim that \"In the 'real life' case where tensors are variably-sized, the algorithm does not change appreciably\". Intuitively, the problem should get significantly harder with variable size allocations. The authors should elaborate on this.\n- Section 7 could benefit from a more formal analysis.\n- In Section 8: \"Note that these are PyTorch-based systems, whereas TURNIP is not.\" Does that mean the authors are running their system in C++ while the others are running in Python? Does this explain the initial performance benefit for low sequence lengths? They should run a test of a model that has no memory problems. In that test all algorithms should be the same.\n- Section 8, Nondeterministic vs deterministic: In the deterministic case precedence constraints are added based on a topological order which has many solutions. Therefore, it is unsurprisingly weaker. The authors should compare to any other scheduling algorithm, such as priority-based algorithms (e.g. priority-based list scheduling).\n\nMinor issues:\n- l147 takes is -> takes as\n- Pacement of figures in Sec 4 is not ideal. Try placing them where they are discussed.\n- Fig. 7: Why is there a memory dependency from 4 to 3. It should at least be dashed.\n- Figure 8 -> Algorithm 1, Figure 9 -> Algorithm 2\n- Figure 8: Assignment and comparison of v_allocHzn.loc in the first \"if\" clause is difficult to understand" }, "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": "How is the simulation obtained to generate the memgraph?\n\nHow does the system ensure that operations are executed in the best order at runtime?\n\nHas the system been open-sourced?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Provided detailed explanations of how to generate memgraph\n \nImplemented a working system" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a runtime system for executing AI applications via GPU-to-CPU offloading. The system dynamically determines whether to allocate AI computations to GPU or CPU memory. It begins by transforming a task graph into a memgraph and implementing the memgraph engine through simulations. This approach enables efficient execution of AI computations through best order of actions (offloading or reloading)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper lacks detailed discussions on how Turnip differs from existing systems mentioned in the related work section.\n\nIt is unclear how the simulation for memgraph is generated.\n\nThe simulation-based memgraph implementation raises questions on how the system guarantees optimal action sequencing during runtime.\n\nHas the system been opensourced?\n\nThere are language issues here and there:\nPage 3: Turnip takes [as] input ...\nThree GPUs, each [having] five ...\nWith seven tensors [in] total ..." }, "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": "The reviewer is happy to raise the rating if some/all the experiment issues stated are addressed.\n\n1. Do you already have any plan to address your paper's known limitations? Since you build a customized framework that doesn't involve Python or PyTorch, For the second one, I believe a separate experiment, when no CPU offloading is involved, comparing your framework with Pytorch on top of the existing experiment parameters can somewhat contribute to the understanding of the performance differences.\n2. Do you have any comments on the experiments that I think are missing? (In the weakness section, Others, 1-4)\n3. Do you have any explanation for the 5th point in the weakness section under the others sub-section?\n4. Just out of curiosity, would it be possible to apply TURNIP to generic graphics or data processing? I feel like it's highly possible.\n5. Last question that doesn't contribute to the paper's rating: Will you open-source the code once the paper gets accepted in the future?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Originality: TURNIP accepts an abstracted version of a generic GPU computation. Other systems are more specifically targeted to certain categories of models, optimization algorithms, or specific tasks such as training or inference. None of the previous work considers the effect of the non-determinism of offload and reload operations on system performance, nor does it focus on the system runtime.\n2. Quality: The quality of the design is demonstrated by experiments under various hardware conditions (A100 and P100), using LLAMA 7B & 65B.\n3. Clarity: The details of the idea are clearly demonstrated with graphs and explanations in section 4-7.\n4. Significance: TURNIP provides a generic solution to democratize AI computing, especially LLMs, which extends the functionality of older GPUs and enables large-scale AI model operations in constrained environments, making it valuable to both the AI and systems communities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents TURNIP, a GPU runtime that supports offloading data to CPU RAM to remove GPU memory restrictions in AI computations. TURNIP addresses the challenges of nondeterministic execution introduced by the slow data transfer rate between CPU RAM and GPU RAM, which harms computational efficiency. TURNIP’s main contribution is its use of a dependency graph (MEMGRAPH) that allows for overlapping execution of GPU kernel operations and memory transfers. The system dynamically chooses the best order of execution based on real-time events, thus minimizing idle times caused by memory transfer delays. TURNIP outperforms existing systems like PyTorch-based ones in constrained GPU environments, contributing to efficient memory management in AI workloads." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "2 limitations that the author stated in the conclusion section of the paper, which potentially harms the significance of contribution (1) and soundness (2):\n1. The biggest limitation of TURNIP is that the input computation (in the form of a TASKGRAPH) must be static and known completely beforehand so that the MEMGRAPH can be constructed. This is not of consequence during model training, but can be an issue during any recursive, generative AI computation. This includes LLM inference, where the next token must repeatedly be generated and the KV -cache increases in size. There are some naive solutions to this (such as pre-compiling a MEMGRAPH for a specific number of token generations in the case of an LLM) but more work will be necessary to create a satisfactory solution to recursive\ncomputation.\n2. Another limitation is that while the experiments showed that TURNIP has certain performance advantages, it is impossible to be sure where those advantages come from. Our ablation shows that a fixed execution order slows TURNIP, suggesting that non-determinism is important. But unlike ZeRO and FlexGen, TURNIP was implemented from the ground up and does not rely on\nPython or PyTorch—that fact alone could account for some of the performance differences.\n\nOthers:\n1. The experiments conducted in this paper use the LLAMA 1st generation model, while the newest one up to date is already LLAMA 3.2. The paper will benefit from using the SOTA LLMs to demonstrate the effectiveness of TURNIP.\n2. As nowadays, the sequence length has grown rapidly (e.g., LLAMA 3.1 models have already supported 128K context window), the paper only did experiments on max 16K, so experiments on longer sequence length will be appreciated to demonstrate the effectiveness of TURNIP.\n3. more diverse hardware, such as V100 and H100, can be used in the experiments if possible. Nowadays, people rarely use P100, while some still stick with V100. Meanwhile, H100 represents cutting-edge technology.\n4. We already have the 405B model in LLAMA 3.1, so seeing how it performs with TURNIP can make the results more convincing.\n5. In Figure 10: Time for LLaMA first token (prefill) inference, A100 server, TURNIP seems to show some scalability issues (e.g., 7B inference batch size 8, TURNIP starts to demonstrate the worst run time when the sequence length is 16K, compared to FlexGen and ZeRO Inference). Although the author mentioned that FlexGen uses very low precision arithmetic to save RAM and speed computing, this doesn't explain the issue here.\n\nA minor issue:\n1. Figures 8 & 9 are actually not figures but details of algorithms. You should call them Algorithms 1 & 2." }, "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": "What is the definition of the MemGraph referenced in Figure 5? Does it represent solely a data dependency graph with memory dependencies, or does it include additional information like control flow dependency?\n\nIn Figure 6, if the tensor locations on GPU devices are predetermined at the compilation stage, will different tensor sizes lead to GPU memory fragmentation? \n\nHow does TURNIP compare to TensorRT or other GPU-only solutions for inference? This information will help us understand its overhead. \n\nZeRO-Offload and ZeRO-Infinity adopt different strategies for allocating tensors between GPU devices and CPU memory. To provide a more comprehensive comparison, could you include GPU-only and ZeRO-Offload solutions in Figure 13?\n\nLastly, why was ZeRO-Infinity not configured with an NVMe as secondary storage, as shown in Figure 13? Doing so could potentially extend the input sequence length or support larger model sizes." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "An interesting approach for incorporating dependency analysis during the compilation phase to optimize offloading and reloading operations. The evaluation was conducted using a state-of-the-art LLM." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Turnip, a software framework to support CPU memory offloading for training and inferencing deep neural networks. The key idea is to introduce a MemGraph, generated at the compilation phase, which is then used to maximize computation and memory transfer overlaps. Turnip was evaluated on 7B and 65B LLaMA models and compared against ZeRO-Inference. The experimental results demonstrate that Turnip reduces first-token latency during inference and accelerates LoRA-based model fine-tuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "A straightforward approach where the positions of tensors reloaded to GPUs are predefined at specific slots through a compilation process based on the data flow graph. This is fine for simple model architectures with a static execution flow but can struggle to deal with dynamic networks like MoEs where tensors/variables are activated based on the current input data?\n\nThe evaluation can be improved. For example, I would like to see a scalability evaluation by testing Turnip on various model sizes, as well as models with f16 during inference. \n\n\nThis may speed up simple model architecture, but is difficult to work with or achieve better performance on complex/dynamic model structures. Second, the experiment design could introduce more workloads such as 13B, 175B, 400B models and more scenarios such as half precision full parameter training and even more solutions such as ZeRO offload, seen 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": 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": "While this paper applies a nondeterministic MEMGRAPH approach to AI computing, similar techniques have been extensively explored in compiler research for general-purpose workloads. What are the unique challenges of applying such memory management strategies to AI-specific workloads?\n\nThe non-deterministic, event-driven scheduling works best for tasks with loose dependencies. However, heavily dependent tasks are common in AI workloads (e.g., transformer models with layer-by-layer computations). When dependencies are heavy, it may leave GPUs idle because tasks are queued. How does the proposed approach work under such a case? \n\nThe paper does not seem to optimize GPU assignments or balance workloads effectively across GPUs, especially when some GPUs are busy while others are idle. This could lead to bottlenecks and inefficient use of resources in multi-GPU settings.\n\nWhat is the time complexity of the MEMGRAPH construction process, particularly for very large TASKGRAPHS? How much overhead in this process?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "++ The paper targets important problem of addressing GPU utilization" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents TURNIP, a GPU runtime system optimized for AI computations under constrained GPU memory by utilizing CPU RAM offloading. TURNIP addresses the challenge of nondeterministic execution by introducing a dependency graph -- MEMGRAPH, which allows for flexible operation sequencing. The approach dynamically decides the best operation order at runtime, minimizing GPU stalls caused by memory transfers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-- The proposed approach may perform poorly when handling complex, dependency-heavy workloads\n\n-- It is not clear what unique challenge in nondeterministic graph-based scheduling for AI computing" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We describe TURNIP, a system for running AI computations using CPU offload using a dependency graph with nondeterministic execution" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024turnip,\ntitle={{TURNIP}: A {\\textquotedblleft}Nondeterministic{\\textquotedblright} {GPU} Runtime with {CPU} {RAM} Offload},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vyzPMQ5weJ},\nnote={under review}\n}" }, "abstract": { "value": "An obvious way to alleviate memory difficulties in GPU-based AI computing is via CPU offload, where data are moved between GPU and CPU RAM, so inexpensive CPU RAM is used to increase the amount of storage available. While CPU offload is an obvious idea, it can greatly slow down a computation, due to the relatively slow transfer rate between CPU RAM and GPU RAM. Thus, any system for CPU offload needs to ensure that when such a transfer needs to happen, no computation is blocked waiting for the transfer to finish. One of the key challenges when using CPU offload is that memory transfers introduce nondeterminacy into the system: it is not possible to know before runtime when the transfers will finish, and hence what is the best order of operations to run to ensure there is no blocking. In this paper, we describe TURNIP, which is a system for running AI computations using CPU offload. The key innovation in TURNIP is the compilation of the AI computation into a dependency graph that gives the TURNIP runtime freedom to run operations such as GPU kernel calls in many different orders; at runtime, TURNIP chooses the best order in response to real-time events." }, "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": [ "CPU Offload", "Memory Management", "Nondeterministic Execution", "Machine Learning System" ] }, "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/7c9821e0bfe42e7a4a3a42e3992d9c741e560843.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": null, "title": { "value": "TURNIP: A “Nondeterministic” GPU Runtime with CPU RAM Offload" }, "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. Could the authors elaborate more on how the long-term transition data is collected? The phrase from the appendix still confusing to me, specifically, how is the reward of the state helps measure the long-term transition." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The jumpy prediction technique within the long-term imagination framework is innovative as it departs from the fixed interval approach prevalent in previous work, offering increased flexibility in jumpy prediction\n2. The paper is well-organized and clearly written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel hierarchical model-based reinforcement learning (MBRL) framework named Long Short-Term Imagination (LS-Imagine), designed to address the challenge of short-sightedness in open-world decision-making with high-dimensional visual inputs, such as MineDojo. In LS-Imagine, imagination horizon can be extended with a specialized world model. Specifically, a affordance maps are predicted to guide the jumpy switch. Two world models are trained to capture transitions at different temporal resolutions, short-term state transition and long-term state transition. Agent learning is conducted during imagination, a common strategy utilized in most background planning method. The authors provided experimental results on Harvest tasks from the MineDojo benchmark, showed superior results compared with baseline methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method employs a hierarchical structure, yet the baseline comparisons are made with flat learning methods. Including comparisons with hierarchical MBRL methods like Director[1] could greatly strengthen the paper.\n2. Equation 9 appears to have an inconsistency in the time indexing; should the bootstrapping term $R^\\lambda_{t+1}$ be $R^\\lambda_{t+\\hat{\\Delta}_{t+1}+1}$ ? \n3. The use of  $\\lambda$ -return in evaluating the policy might introduce bias since it should be evaluated with on-policy data,, but the predicted jumpy state, $\\hat{z}_{t+1}$, might not aligned with the learning policy.\n4. The paper focuses on Harvest tasks. Including results from other complex, long-horizon tasks, such as the Tech Tree task group from MineDojo, would better demonstrate the framework’s effectiveness.\n\n[1]: Hafner, Danijar, et al. \"Deep hierarchical planning from pixels.\" Advances in Neural Information Processing Systems 35 (2022): 26091-26104." }, "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 term \"Gaussian matrix\" in line 215 is confusing. Based on the context later, I think you are referring to Gaussian kernel. If that is correct, what is the $\\sigma$ you use in the paper? Is the method sensitive to different value of $\\sigma$?\n- In Section 3.2.2, you mentioned the affordance map generator is pretrained on the random exploration data. How do you make sure that the task relevant observations are present in the random data to train a meaningful generator? For example, if the task is to collect a diamond, and there certainly won't be any diamond seen by the random agent. Will the method still help to solve the task?\n- I am confused by the model analysis part with the parallel pathway. Things explained in Section 4.3 sound the same with Section 3.4. Could you elaborate what is the difference between the sequential pathway and the parallel pathway?\n- Some relevant citations are missing: [1] studies the hierarchical world model which uses a similar strategy of doing the long-term imagination on the higher-level states; [2] studies a similar problem that short imagination horizon may not enough to cover behaviour of interest. I would suggest the author to discuss these works.\n- There are no error bars for the middle and right plots in Figure 7. Are these results only based on one seed?\n- What is the computation cost of this method?\n- Maybe I missed some parts, but how does the policy learn the behaviour that the long-term imagination skip? For example, in the \"Harvest log in plains\" task, the go toward the tree is always skipped by the long-term imagination if the model learns well. Since the skip transition doesn't have gradient attached to the action, I wonder how does the policy learn to walk toward the tree.\n\n[1] Gumbsch *et al.*, Learning Hierarchical World Models with Adaptive Temporal Abstractions from Discrete Latent Dynamics, ICLR 2024 \n\n[2] Hamed *et al.*, Dr. Strategy: Model-Based Generalist Agents with Strategic Dreaming, ICML 2024" }, "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 mostly well written. \n- The method proposed is novel and the results are promising comparing to the baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies Model-based RL in Open-World environment, specifically Minecraft. The authors propose a long short-term imagination method to improve the imagination process of the dreamer algorithm. The main idea is to use an affordance map to identify the distance object of interest and train a long-term transition model to skip the approaching steps so that the imagination can be more task-relevant. Results from 5 tasks in Minecraft showcase the improvement of the method over Dreamerv3 and other two baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although the high-level idea is straight-forward, the implementation is overcomplicated. \n- The method feels very ad-hoc to the Minecraft tasks studied in this paper. It doesn't come into my mind about any other relevant tasks other than Minecraft where the proposed method can be applied." }, "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": "- Why was the word 'jumpy' chosen? Could a more precise word be used instead?\n- The authors say on L42 that MBRL approaches optimize short-term experiences of \"typically 50 time steps\". Where did 50 come from? My understanding is that MBRL methods such as DreamerV3 commonly use an even shorter horizon of $\\sim 16$ timesteps.\n- Could the limitations be expanded upon to cover the general application of LS-Imagine beyond Minecraft?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- **Significance**\n - Long-horizon world modeling and reinforcement learning in open-world environments are important problems.\n - The proposed approach is insightful and successfully addresses these problems.\n- **Originality**\n - The proposed approach involves the combination of multiple novel and inisightful components.\n- **Quality**\n - Overall the quality of the paper is relatively high, with the method reasonably clearly explained and analyzed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an exploration approach, LS-Imagine, to improve model-based reinforcement learning from pixels on open-world environments (with large state spaces). \n\nThe paper aims to address the short-hoirzon limitation of model-based approaches relying on repeated autoregressive single timestep prediction. This is achieved by first generating affordance maps of observations by zooming into different patches and determining if there are objects associated with the current task in those patches using a vision-language reward model (MineCLIP in this work) distilled into an affordance map predictor. This affordance map is used to generate an intrinsic reward associated with the observation. The kurtosis of the affordance map is then used to determine if the object associated with reward is near or far away (binary, based on a threshold). If the object is deemed to be near, a standard 'short-term' world model using autoregressive single timestep prediction is used to generate trajectories for an actor and critic to learn from. Alternatively, if the object deemed to be far away, a 'long-term' world model using autoregressive multi-timestep interval prediction is used (although ignoring actions). This enables the world model to generate a trajectory over a larger number of environment timesteps in the same number of autoregressive prediction steps, therefore potentially enabling the generated trajectory to get closer to the object of interest. This long-term trajectory is used to train the critic but not the actor. The combination of these models is termed a Long Short-Term World Model, giving the approach its name: LS-Imagine.\n\nLS-Imagine is applied to 5 MineDojo tasks and outperforms the baselines in terms of success rate and per-episode steps after training. Ablations demonstrate the benefit of both affordance-based intrinsic reward and the use of long-term imagination.\n\nIn summary, the paper attempts to tackle the important problem of long horizon world modelling and provides an interesting and novel approach to address this problem in 3D environments. While it is a relatively complex method and somewhat limited to Minecraft, I believe the ideas and insights warrant acceptance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Clarity**\n - Some aspects of the paper are not particularly clear. The main one is the use of the word 'jumpy' throughout the paper. The meaning of this word is assumed, but is not defined in the paper or standard usage as far as I'm aware, and is relatively unscientific, so I feel it is not the right word to use. 'Multi-step' state transitions seems more appropriate. If the authors were attempting to highlight that the number of steps can vary, then 'variable-step' transitions would be better. At the very least, 'jumpy' should be properly defined at the beginning of the paper.\n - Similarly affordance maps may not be familiar to all readers and the exact meaning of this term can vary. For example, a short clarification early in the paper such as ''...affordance maps, that elucidate which parts of an observation may be associated with reward, ...\" would be helpful.\n - Some other unclear aspects/minor mistakes include:\n - L326: \"employs *an* actor-critic algorithm to learn behavior *from* the latent state sequences...\"\n - L351: Grammar is slightly wrong and confusing, should be: \"Notably, since long-term imagination does *not* involve actions, we do not optimize the the actor when long-term imagnation is adopted.\" \n - L354: Also worth highlighting the difference with the DreamerV3 loss. \"The loss of the actor is therefore equivalent to DreamerV3, with an additional factor of $(1-\\hat j_t)$ to ignore updates from long-term imagination:\"\n - L361: \"on the chellenging...\"\n - L500: Doesn't make sense. Maybe \"Our work is also related to affordance maps for robotics.\" is sufficient?\n - \"*Learning to Move with Affordance Maps*\" [1] should likely also be compared and cited here.\n\n- **Limitations**\n\n - This approach has important limitations that are not mentioned. In particular, the approach is limited to embodied agents navigating a 3D environment in which there are objects associated for which reward is obtained by approaching them. Therefore the approach assumes, for example:\n - Observations are of a 3D environment\n - Actions are navigation actions of an embodied agent\n - Rewards are assoiated with identifying or moving towards objects\n - A reward model is available to identify high reward regions of observations\n - The experiments are limited to Minecraft for which these assumptions hold. This approach would likely not work as well even in Crafter [2] for example, which provides a 2D 'open-world' analogue of Minecraft, since objects do not become larger as you move towards them.\n - The approach also relies on both the long-term and short-term models being used, given only the short-term model is able to update the actor. While the thresholding of $P_{jump}$ can partially be used to address this, this is not particularly robust, and still requires some close-up objects in initial exploration for the standard one-step world model to be used, so the approach may not work as well in very sparse environments.\n - There is still significant value of the approach despite these limitations, and the paper is reasonably scoped, but they should be included in the limitations at the end of the paper, which are currently overly brief and narrow.\n\n **References:**\n\n [1] \"*Learning to Move with Affordance Maps*\", Qi et al., 2020\n\n [2] \"*Benchmarking the Spectrum of Agent Capabilities*\", Hafner, 2021" }, "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. Jumping transitions\n\n (a) I don't get the reason and rule for this dynamics threshold on lines 248 to 250. Could you clarify this?\n\n (b) The jumpy transition in the world models is a novel contribution, and I think it's worth more explanation for intuitive understanding. Like how many times would the jumping transition be triggered during the training, and what is the predicted average $\\Delta'_t$ in the imagination (how many short-term steps are saved)?\n\n2. I suggest adding a numerical result (possibly in the appendix) for the results in Figures 4 and 5.\n\n3. For the parallel imagination in section 4.3. Is the post-jumping state connected with the bootstrapping $\\lambda$-return of the pre-jumping states? How is the conclusion on lines 464 to 466 be made?\n\n4. For the model-based RL review on lines 487-489, I suggest adding another two world model papers for completeness: (1) Zhang, Weipu, et al. \"STORM: Efficient stochastic transformer based world models for reinforcement learning.\" Advances in Neural Information Processing Systems 36 (2024). and (2) Alonso, Eloi, et al. \"Diffusion for World Modeling: Visual Details Matter in Atari.\" arXiv preprint arXiv:2405.12399 (2024). (NeurIPS24 Spotlight)" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. It introduces a method for generating target states in MineDojo (or possibly in other 3D-RPG games).\n2. It demonstrates the feasibility of training a world model with jumping transitions and optimizing policies over such transitions.\n3. The illustration is clear and the writing organization is good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a long short-term world model to address the limited imagination horizon of traditional world models.\nThe authors leverage MineCLIP to detect task goals and generate intrinsic rewards, enabling the model to jump directly to these goal states when feasible.\nFor policy optimization over these jump-based transitions, they train modules to predict future value, interval timesteps, and cumulative rewards.\n\nThe paper is well-organized, and its key contributions are:\n(1) It introduces a method for generating target states in MineDojo (or possibly in other 3D-RPG games).\n(2) It demonstrates the feasibility of training a world model with jumping transitions and optimizing policies over such transitions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Since MineCLIP is an important tool for this work, I suggest the author include a brief introduction of what MineCLIP can do in section 3.2/3.2.1 or an appropriate position. This would help readers who are not familiar with research on MineDojo to understand this paper.\n\n2. In the abstract, \"We argue that the primary obstacle in open-world decision-making is improving the efficiency of off-policy exploration across an extensive state space.\" This seems not closely connected to the main contribution of this paper. I suggest paraphrasing it to highlight \"across a long horizon\" or something that is more related to the topic.\n\n3. Though the method sounds promising for solving MineDojo tasks, it may not be a general method for all kinds of open-world games. Such as in 2D games or fixed camera control tasks.\n\n (a) Before seeing the target for the first time in the training process, there won't be a reasonable goal-directed reward or jumping option, the exploration still requires extensive enumerates.\n\n (b) The crop and resize operation (or assumption) is only useful for 3D visual navigation tasks.\n\n (c) When the target is occluded, the world model still needs to perform step-by-step imagination.\n\nIf these are true, I suggest the authors include a sentence or so in the limitation section to clarify it." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024openworld,\ntitle={Open-World Reinforcement Learning over Long Short-Term Imagination},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vzItLaEoDa},\nnote={under review}\n}" }, "abstract": { "value": "Training visual reinforcement learning agents in a high-dimensional open world presents significant challenges. While various model-based methods have improved sample efficiency by learning interactive world models, these agents tend to be “short-sighted”, as they are typically trained on short snippets of imagined experiences. We argue that the primary obstacle in open-world decision-making is improving the efficiency of off-policy exploration across an extensive state space. In this paper, we present LS-Imagine, which extends the imagination horizon within a limited number of state transition steps, enabling the agent to explore behaviors that potentially lead to promising long-term feedback. The foundation of our approach is to build a $\\textit{long short-term world model}$. To achieve this, we simulate goal-conditioned jumpy state transitions and compute corresponding affordance maps by zooming in on specific areas within single images. This facilitates the integration of direct long-term values into behavior learning. Our method demonstrates significant improvements over state-of-the-art techniques in MineDojo." }, "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": [ "World models", "reinforcement learning", "visual control" ] }, "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/27f9c01eb634372364261ff18305771579cb344d.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/1ffa69aa8ca220b00e8418dfa2bf7a0b80ef63e5.zip" }, "title": { "value": "Open-World Reinforcement Learning over Long Short-Term Imagination" }, "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": "- Figure 2 shows comparisons of models based on accuracy and inference time. Does that mean accuracy and inference time are dependent on each other? If not, these graphs create confusion. Currently, from the graphs it seems that if the inference time is longer, the accuracy could be higher, and vice versa. If this is not the case, I suggest keeping inference time separate from accuracy.\n\n- Line 669 of Algorithm 1 states, \"Update teacher model parameters via backpropagation\". Typically, knowledge distillation keeps the teacher fixed once pre-trained, so if this algorithm intends otherwise, the authors should provide a justification.\n\n- In Figure 6, can the authors keep the range of accuracy axis the same for all the graphs? It would be easier to visualize the differences. \n\n- It would be interesting to see how an adaptive knowledge distillation approach performs compared to the proposed fixed distillation approach. For instance, the model could prioritize contrastive learning for highly connected hypergraphs to capture complex relationships and rely more on soft labels for simpler ones to reduce model complexity even further." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The dual knowledge distillation approach (soft labels and high-order structural knowledge) effectively transfers complex relationships, addressing limitations found in existing methods.\n\n- Comparative accuracy with minimal inference time.\n\n- The paper is well structured and well written.\n\n- The paper includes a thorough experimental setup, covering multiple datasets and evaluation metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed DistillHGNN, a novel framework designed to improve the inference speed of Hypergraph Neural Networks (HGNNs) without sacrificing accuracy. DistillHGNN utilizes a teacher-student knowledge distillation approach, where the teacher model comprises an HGNN and a Multi-Layer Perceptron (MLP), while the student model, TinyGCN, is a lightweight Graph Convolutional Network (GCN) paired with an MLP. The framework incorporates a dual knowledge transfer mechanism, passing both soft labels and structural knowledge from the teacher to the student. Trained on labeled data and teacher-provided soft labels with contrastive learning to retain high-order relationships, DistillHGNN achieves performance comparable to traditional HGNNs with significantly lower resource requirements." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The authors claim the proposed approach is memory efficient everywhere in the paper. In line 416, they mention, \"As illustrated in Figure 2, DistillHGNN maintains competitive inference times and memory efficiency without sacrificing accuracy\"; however, Figure 2 shows accuracy vs. inference time. Moreover, no other evidence shows that the proposed approach is memory efficient.\n\n- Among the previous works in knowledge distillation, the authors only compared with LightHGNN. Since the authors compared with GNN and GCN, the authors are suggested to compare with GNN and GCN-based distillation approaches as well. \n\n\nMinor:\n\n- In line 152, the authors mention \"Framework 1\". Since there is no Framework 1 in the paper, I suppose it should be \"Figure 1\".\n\n- The caption for Figure 1 is too big. The authors should keep it concise and incorporate the details into the actual text of Section 3. Also, the authors didn't mention what happens with the soft labels after they are generated from the teacher model in the caption." }, "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. The proposed method converts a hypergraph into a homogeneous graph, potentially leading to the loss of high-order structural information. How do you address this information loss, and can you incorporate mechanisms similar to LightHGNN’s high-order soft-target constraints to better preserve high-order knowledge? Additionally, could you provide visualizations to enhance the interpretability of your method?\n2. Since DistillHGNN depends on homogeneous graphs during inference, how does it handle inductive learning scenarios where new samples are added? Specifically, what strategies do you propose to minimize the need for graph reconstruction and complex preprocessing steps in such cases?\n3. The experimental evaluation primarily focuses on datasets from academic papers and movie databases. Do you have plans to evaluate DistillHGNN on larger-scale or synthetic datasets to demonstrate its scalability and effectiveness across a broader range of applications?\n4. Could you provide more technical details and visualizations regarding how contrastive learning and soft labels facilitate the transfer of high-order knowledge from the teacher model to the student model? This would help in better understanding the mechanisms behind knowledge preservation and transfer." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed method combines knowledge distillation with contrastive learning in an innovative way, allowing for a more comprehensive transfer of both soft labels and high-order structural knowledge from the complex HGNN to the lightweight TinyGCN. This dual transfer mechanism enhances the student model's ability to capture intricate dependencies while maintaining computational efficiency.\n2. The experimental evaluation is thorough and spans multiple real-world datasets, including CC-Citeseer, CC-Cora, IMDB-AW, and various DBLP datasets. The results consistently demonstrate that DistillHGNN achieves a favorable balance between accuracy and inference speed, often outperforming existing methods such as LightHGNN in both metrics.\n3. The paper is well-structured and provides a clear and detailed explanation of the methodology, including the architecture of both the teacher and student models, the training process, and the loss functions used. This clarity facilitates understanding and potential replication of the proposed framework." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces DistillHGNN, a novel knowledge distillation framework designed to enhance the inference speed and memory efficiency of Hypergraph Neural Networks (HGNNs) while maintaining high accuracy. DistillHGNN employs a teacher-student model where the teacher consists of an HGNN and a Multi-Layer Perceptron (MLP) that generates soft labels. The student model uses a lightweight Graph Convolutional Network (TinyGCN) paired with an MLP, optimized for faster online predictions. Additionally, contrastive learning is utilized to transfer high-order and structural knowledge from the HGNN to the TinyGCN. Experimental results on various real-world datasets demonstrate that DistillHGNN significantly reduces inference time and achieves accuracy comparable to or better than state-of-the-art methods like LightHGNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The method involves converting a hypergraph to a homogeneous graph, which can lead to the loss of high-order structural information inherent in hypergraphs. While contrastive learning helps mitigate this issue, the paper does not sufficiently address how the conversion process preserves the complex relationships that hypergraphs naturally capture.\n2. DistillHGNN relies on homogeneous graphs during the inference phase, which poses a significant limitation for inductive learning scenarios. When new samples are introduced, the need to reconstruct the graph and perform complex preprocessing steps can result in increased computational and time costs, limiting the model's applicability in dynamic or real-time environments.\n3. The evaluation is primarily conducted on datasets related to academic papers and movie databases. There is a lack of assessment on larger-scale or synthetic datasets, which raises concerns about the method’s scalability and generalizability to other domains or more complex real-world applications.\n4. The selection of baseline methods, while covering traditional GNNs and some knowledge distillation approaches, does not include more recent or advanced hypergraph neural network models. This omission may limit the understanding of DistillHGNN's relative performance against the latest state-of-the-art techniques." }, "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": "See Weaknesses part" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The dual use of soft labels and structural knowledge distillation through contrastive learning enhances the capability of the student model, addressing limitations of previous distillation methods.\n2. DistillHGNN achieves significant reductions in inference time (up to 80%) compared to traditional HGNNs, making it suitable for real-time applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes **DistillHGNN**, a knowledge distillation framework designed to accelerate the inference of hypergraph neural networks (HGNNs) while maintaining their high accuracy. DistillHGNN leverages a teacher-student structure where the teacher model, an HGNN, captures complex relationships within hypergraphs and generates soft labels. The student model, a simplified TinyGCN, learns from these soft labels and employs contrastive learning to capture high-order structural knowledge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **(Minor)** The writing could be improved for clarity and simplicity. For instance, the caption in Figure 1 (lines 175–188) could be condensed to improve readability, ideally to half its current length. If extensive detail is necessary, consider moving some information to the main text.\n2. **(Minor)** The paper could use `\\citet{}` for citations when referring to the authors directly. For example, in lines 280–281, instead of \"Multilayer Perceptron (MLP) by \\cite{author2020},\" using `\\citet{}` could enhance readability.\n3. When mentioning the use of six popular datasets, it would be helpful to list the features and characteristics of each dataset to provide more context for the experiments. For instance, some datasets may have dense connections while others are sparse; some may benefit more from heterogeneous graphs compared to homogeneous ones. Without a detailed explanation of the datasets' biases or inductive properties, understanding the generalization of the proposed method becomes challenging.\n4. **Scalability**: The scalability of the proposed method is unclear. According to Table 3, all datasets used in the experiments are relatively small, with node counts between 3k and 6k, similar in scale to the Cora dataset. It would be beneficial to provide insights on how the method scales to larger datasets, such as ogbn-arxiv or ogbn-products, if they have hypergraph structures." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Hypergraph Neural Networks (HGNNs) using knowledge distillation to transfer high-order information to a lightweight GCN, achieving faster inference with minimal resource use while maintaining high accuracy." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024distillhgnn,\ntitle={Distill{HGNN}: A Knowledge Distillation Approach for High-Speed Hypergraph Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=vzrs42hgb0},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we propose a novel framework to significantly enhance the inference speed and memory efficiency of Hypergraph Neural Networks (HGNNs) while preserving their high accuracy. Our approach utilizes an advanced teacher-student knowledge distillation strategy. The teacher model, consisting of an HGNN and a Multi-Layer Perceptron (MLP), not only produces soft labels but also transfers structural and high-order information to a lightweight Graph Convolutional Network (GCN) known as TinyGCN. This dual transfer mechanism enables the student model to effectively capture complex dependencies while benefiting from the faster inference and lower computational cost of the lightweight GCN. The student model is trained using both labeled data and soft labels provided by the teacher, with contrastive learning further ensuring that the student retains high-order relationships. This makes the proposed method efficient and suitable for real-time applications, achieving performance comparable to traditional HGNNs but with significantly reduced resource requirements." }, "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": [ "Knowledge Distillation", "Hypergraph Neural Networks", "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": null, "pdf": { "value": "/pdf/bcc066b06bd546080c00ed4d7a83b1a49092a5fd.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong 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/978400769ba3c7381e480ed8760d95bafc055c81.zip" }, "title": { "value": "DistillHGNN: A Knowledge Distillation Approach for High-Speed Hypergraph 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": 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 the perturbation generation strategy be optimized or adapted to incorporate other noise categories (e.g., adversarial perturbations)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The idea of optimizing to reach a flat landscape in the full parameter space while maintaining the advantages of parameter efficiency is innovative and well-justified.\n2. The methodology is generally clearly articulated, and lots of experiments show considerable improvements over existing LoRA-based methods, which validates the efficacy of Flat-LoRA proposed.\n3. The explanation of the Bayesian expectation loss objective function and the perturbation generation strategy is very thorough and definitely contributes to the transparency of the proposed method.\n4. This work is equipped with practical implications, as fine-tuning LLMs efficiently is increasingly important in current ML field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Flat-LoRA, a novel extension to the Low-Rank Adaptation (LoRA) framework, designed to optimize model fine-tuning via discovering solutions within a flatter loss landscape in the full parameter space. Different from traditional LoRA, which may result in sharp solutions that impact generalization negatively, Flat-LoRA incorporates random weight perturbations and a Bayesian expectation loss objective to maintain efficiency. The approach seeks to combine parameter efficiency with enhanced generalization capabilities across both NLP and CV tasks, demonstrating improvements over conventional LoRA in various experimental setups." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Althougth the paper emphasizes the computational overhead and the minimal memory, the perturbation generation and its integration into the mixed-precision training could be simplified or clarified furthermore.\n2. The method concentrates mostly on linear layers in the transformer-based models. Despite the fact that the authors acknowledge this as a limitation, extending such approach to other parameter categories would make the method more versatile.\n3.: Although the comparisons with approaches such as SAM are very insightful, deeper analysis with more recent variations of sharpness-aware algorithms could strengthen the study and contribution." }, "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": 4 }, "primary_area": null, "questions": { "value": "What do you think is the essential difference or advantage between adding perturbations to weights and adding perturbations to data samples? Is adding perturbations to data samples simpler?" }, "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 well done, with motivations and insights clearly explained in an intuitive manner, accompanied by reasonable mathematical assumptions to introduce the design of full-rank noise perturbations.\n2. The design for storing random seeds for memory efficiency and integrating into mixed-precision training is very clever, saving additional overhead." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Based on the consideration of local optimal values, this paper proposes that the flat optimal value learned by LoRA is not necessarily flat at full rank. Corresponding mathematical explanations are provided. Following this, using the idea of SAM, a method is designed to add full-rank noise perturbations to search for a global optimal solution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. On line 125, the sentence \"the LoRA matrices. better initialization strategy for LoRA parameters.\" should probably use a comma instead of a period.\n2. The work on adding full-rank noise has been done in Noisetune ([https://arxiv.org/abs/2202.12024](https://arxiv.org/abs/2202.12024)) and LORASC ([https://arxiv.org/abs/2407.01491](https://arxiv.org/abs/2407.01491)), especially in LoRASC, where the exact same approach is used, adding a full rank noise to each LoRA optimization process. \n3. Overall, the core technical implementation of this paper is to add random perturbations at each step of LoRA training. While the approach is elegant, it’s not particularly novel, and there doesn’t seem to be very convincing experimental results. The proposed Flat-LoRA series offers limited improvement across various LoRA variants and tasks, with most gains being around a few tenths of a percentage in accuracy (except for significant improvements in gsm8k and Human-Eval). It would be helpful to include tasks like apaca and other SFT tasks. I’m particularly interested in the practical significance of this work—flat local optima should theoretically bring stronger generalization, such as supporting OOD (e.g., Alpaca, instruct-eval, etc.) and robustness (e.g., image-R, image-C) evaluations, proving that a flatter local optimum could enhance out-of-domain generalization. This would be valuable since even sharp in-domain optima can perform well, and this might be why performance improvements are modest." }, "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": 3 }, "primary_area": null, "questions": { "value": "no" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Advantages:\n\nThe paper is written clearly\nThe core idea seems reasonable" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Flat-LORA, which adds noise to the training process: W + AB + ε, where ε is stored using a seed. The authors attempt to achieve SAM-like effects through this method.\n\nAdvantages:\n1. The paper is written clearly\n2. The core idea seems reasonable\n\nDisadvantages:\n1. The paper lacks mathematical rigor in several places, particularly in equations 8 and 9\n2. Insufficient experimental validation:\n - Should test on larger models (e.g., LLAMA 13B or 70B)\n - Should evaluate on SuperGLUE\n3. Lacks necessary ablation studies, particularly regarding σ^2\n\nTechnical Issues:\n1. Equations 8 and 9 have fundamental problems:\n - var(X) should be a covariance matrix\n - var(W'_{i,:}X) should be a scalar\n - These dimensions are inconsistent and cannot be equated\n2. In Equation 7, n should be sqrt(n), as large n values would result in negligibly small epsilon values added to the weight matrix\n\nAdditional Concerns:\n1. No memory usage results are reported\n2. Table 3 lacks full-tuning baseline results\n3. this paper should report results on more diverse datasets\n4. More comprehensive ablation studies on σ^2 are needed" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Disadvantages:\n\nThe paper lacks mathematical rigor in several places, particularly in equations 8 and 9\nInsufficient experimental validation:\nShould test on larger models (e.g., LLAMA 13B or 70B)\nShould evaluate on SuperGLUE\nLacks necessary ablation studies, particularly regarding σ^2\nTechnical Issues:\n\nEquations 8 and 9 have fundamental problems:\nvar(X) should be a covariance matrix\nvar(W'_{i,:}X) should be a scalar\nThese dimensions are inconsistent and cannot be equated\nIn Equation 7, n should be sqrt(n), as large n values would result in negligibly small epsilon values added to the weight matrix\nAdditional Concerns:\n\nNo memory usage results are reported\nTable 3 lacks full-tuning baseline results\nthis paper should report results on more diverse datasets\nMore comprehensive ablation studies on σ^2 are needed" }, "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. How does it perform on the SuperGLUE benchmark, SQuAD, XSum, and CNN/Dailymail?\n2. How does it perform on Stable Diffusion?\n3. How does it compare between AdaLoRA and DyLoRA?\n4. How is the performance on Llama3 with the alpaca dataset?" }, "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 introduces Flat-LoRA, a novel method that improves upon traditional Low-Rank Adaptation (LoRA) techniques by targeting a flat region of the loss landscape. This aims to enhance generalization performance by avoiding sharp minima that can degrade model performance on unseen data.\n\n2. Flat-LoRA addresses the computational and memory cost issues associated with fine-tuning large-scale models. By optimizing only a low-rank matrix and using random weight perturbations, it achieves parameter-efficient fine-tuning without additional memory costs during inference.\n\n3. The paper not only proposes a new technique but also details the underlying mathematical foundations. It discusses the optimization objective, perturbation generation, and integrates a Bayesian expected loss objective to maintain training efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper discusses how to improve the generalization of LoRA, with good writing, innovative ideas, and credible experimental performance. However, I think the article overlooks one point: LoRA only fine-tunes a very small number of parameters, which is unlikely to overfit on a specific task and can inherently maintain good generalizability. If I am wrong, please convince me." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is no significant improvement, the limited enhancement on the GLUE benchmark does not prove the lack of generalizability of methods like LoRA. I need to know if your method is effective on more datasets. More datasets and models should be compared. For example: the SuperGLUE benchmark, SQuAD, XSum, CNN/Dailymail, and some LoRA training on Stable Diffusion.\n\n2. There is a lack of extensive comparisons, such as with methods like DyLoRA[1] and AdaLoRA[2]. \n\n3. More relevant articles, such as AWP[3], should be cited, which is an effective method to improve generalization.\n\n[1] Valipour M, Rezagholizadeh M, Kobyzev I, et al. Dylora: Parameter efficient tuning of pre-trained models using dynamic search-free low-rank adaptation[J]. arXiv preprint arXiv:2210.07558, 2022.\n[2] Zhang Q, Chen M, Bukharin A, et al. AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning[J]. arXiv preprint arXiv:2303.10512, 2023.\n[3] Wu D, Xia S T, Wang Y. Adversarial weight perturbation helps robust generalization[J]. Advances in neural information processing systems, 2020, 33: 2958-2969." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose Flat-LoRA that aims to optimize the sharpness of the loss landscape for low-rank adaptation using efficient random weight perturbation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024flatlora,\ntitle={Flat-Lo{RA}: Low-Rank Adaption over a Flat Loss Landscape},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w0389y0W9D},\nnote={under review}\n}" }, "abstract": { "value": "Fine-tuning large-scale pre-trained models is prohibitively expensive in terms of computational and memory costs. Low-Rank Adaptation (LoRA), a popular Parameter-Efficient Fine-Tuning (PEFT) method, provides an efficient way to fine-tune models by optimizing only a low-rank matrix. Despite recent progress made in improving LoRA's performance, the connection between the LoRA optimization space and the original full parameter space is often overlooked. A solution that appears flat in the LoRA space may exist sharp directions in the full parameter space, potentially harming generalization performance. In this paper, we propose Flat-LoRA, an efficient approach that seeks a low-rank adaptation located in a flat region of the full parameter space. Instead of relying on the well-established sharpness-aware minimization approach, which can incur significant computational and memory burdens, we utilize random weight perturbation with a Bayesian expectation loss objective to maintain training efficiency and design a refined perturbation generation strategy for improved performance. Experiments on natural language processing and image classification tasks with various architectures 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": [ "low-rank adaption", "flat minima", "efficient training" ] }, "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/f19f43c15168c72850f51d314f71f5e1f9936979.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/91857800a3fc7b58247939c4747b970d1da57add.zip" }, "title": { "value": "Flat-LoRA: Low-Rank Adaption over a Flat Loss Landscape" }, "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": 3 }, "primary_area": null, "questions": { "value": "1. In line 250, \"during the training phase\", is this refer to training of reward models? Is it correct that we need $r_{score}$ for policy training?\n2. What does Value Model in Table 8 mean?\n3. In table 1, what are the differences between preference accuracy and intensity accuracy." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper structure is well-organized and the main concepts are well-articulated. The methodology section is clear and sound with sufficient summary of the previous methods at start. The differences between the proposed method with baselines are explicitly demonstrated through the figures. In the experiments section, the outline at the front is followed by corresponding verification of the experiments results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new multi-objective LLM alignment method which trains a single multi-head classification reward model for policy optimization and manipulates diverse preferences through prompting during inference. The method MOSLIM achieves both training and inference efficiency compared to two baselines MORLHF and RSoup. MOSLIM also enables granularity control over preference intensity level." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Experiment designs may need improvements.\n- Although the authors mention the latest work RiC and CDPO, it seems that there are not direct comparison results shown in the manuscript. If feasible, it would be more persuasive if the authors could show the superiority of MOSLIM over these two methods through the comparisons in the performance or training cost. \n- For Figure 5, I am not sure whether it is a fair comparison with two baselines since these are not trained with certain granularity of preference intensities.\n2. There are lots of typos throughout the whole manuscript. Please check the grammatical errors carefully. Here are some examples:\n- LIne 293, it should be \"three SFT models\"\n- Line 295 296, \" is scaling law still consist\" and \"weather\".\n- Table 2 it should be \"RSoup\"\n3. Some important details about experiments are missing. \n- It could be useful if some details about the evaluation process in Table 2 are added. Do different data types mean different forms of training dataset used in policy optimization? Are you taking the granularity level (i.e. the number of preference intensity classes) into account during evaluation? \n- For Table 8, providing time breakdown of reward model and policy model training would make the comparisons more pronounced.\n\n4. Others. \n- In equation (10), since the number of classes K could be different for different preference dimension. it is not appropriate to use single K." }, "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": "- Minor: Reward Soups must be abbreviated as RSoup, not Rsopu.\n- How does MOSLIM handle cases where preference intensities are not explicitly defined or vary significantly in real-world scenarios?\n- Could the authors provide more detail on the decision process for choosing specific intensity levels within the reward mapping function?\n- Unlike baseline approaches, the authors have emphasized the advantage of removing SFT in MOSLIM, but there have been no experiments verifying such advantages." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- By employing a single reward model for diverse preferences, MOSLIM significantly reduces computational overhead. This enables off-the-shelf models without requiring fine-tuning for each new preference.\n- The framework shows potential scalability, given its ability to handle multiple preference objectives without complex adjustments during training." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MOSLIM, a multi-objective alignment framework for Large Language Models (LLMs) that optimizes diverse human preferences through a single reward and policy model. Unlike traditional methods, MOSLIM uses a multi-head reward model for classification-based preference alignment and prompt-driven policy optimization, reducing the need for extensive retraining on specific preferences. The authors validate their approach across multi-objective benchmarks, demonstrating that MOSLIM is more efficient in GPU usage and improves alignment flexibility compared to existing methods like MORLHF and Rewarded Soups." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While MOSLIM addresses multi-objective alignment, the core approach builds on established methods, particularly prompt-driven alignment and multi-head reward models. The contributions are incremental rather than ground-breaking, as the framework primarily refines and consolidates existing techniques.\n- The writing is hard to understand, and the word usage is inconsistent in the paper (e.g., Both RSoup and RSopu are used.)\n- Limited comparison with a few baselines (Rsoup and MORLHF). Please add more baselines dealing with multi-objective RLHF (RiC [1], RLPHF [2]).\n- The paper compares MOSLIM primarily against a few specific methods, but additional comparisons with other state-of-the-art frameworks in multi-objective alignment could provide a more comprehensive evaluation.\n\n---\n**References**\\\n[1] Rui Yang et al., Rewards-in-context: Multi-objective alignment of foundation models with dynamic preference adjustment, ICML 2024.\\\n[2] Joel Jang et al., Personalized Soups: Personalized Large Language Model Alignment via Post-Hoc Parameter Merging, https://arxiv.org/pdf/2310.11564" }, "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": "Please refer to the Weaknesses part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- MOSLIM outperforms existing methods such as MORLHF and Rewarded Soups, while achieving controllable alignment across different preference dimensions and intensities.\n- The model’s effectiveness is thoroughly validated across several benchmarks (MT-Bench, HaluEval 2.0, Hackaprompt). The study explores various model scales and compares different algorithms (PPO, RLOO, Online-DPO), demonstrating MOSLIM’s robustness across configurations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents MOSLIM, a framework designed to align LLMs with multiple, diverse user preferences through a single reward and policy model. MOSLIM uses a single multi-head reward model and policy model to align with various preference objectives and intensities at training. At inference, it adapts flexibly to different preference intensity combinations using a prompt-driven mechanism. MOSLIM enables efficient, scalable preference alignment, allowing integration with existing off-the-shelf models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Contribution and Novelty: The approach of using a multi-head RM for preference alignment has been introduced in prior work [1,2], which may also support multi-objective preference classification. Additionally, the fast inference strategy utilized by MOSLIM shares similarities with previous efforts to dynamically adjust preferences, such as Rewards-in-Context [3]. While MOSLIM mentions that some methods use SFT loss primarily to enhance core abilities (lines 55-62), this claim may overlook the use of rejection sampling for alignment in works like Llama 2 [4] and Gao et al. (2023) [5].\n- Explanation for Outperformance of Baselines: The paper does not sufficiently explain why MOSLIM outperforms MORLHF and Rewarded Soups, in terms of the policy performance. It would be beneficial to include both empirical comparisons and theoretical explanations of the different elements of MOSLIM with MORLHF and Rewarded Soups, for example, the performance of RM (here only the model size is compared). 
\n- Representation: Figure 4 may be unnecessary, as it duplicates information already presented in Table 1. \n\n[1] Li, Lihe, et al. \"Continual Multi-Objective Reinforcement Learning via Reward Model Rehearsal.\"\n[2] Yang, Adam X., et al. \"Bayesian reward models for LLM alignment.\" arXiv preprint arXiv:2402.13210 (2024).\n[3] Yang, Rui, et al. \"Rewards-in-context: Multi-objective alignment of foundation models with dynamic preference adjustment.\" arXiv preprint arXiv:2402.10207 (2024).\n[4] Touvron, Hugo, et al. \"Llama 2: Open foundation and fine-tuned chat models.\" arXiv preprint arXiv:2307.09288 (2023).\n[5] Gao, Leo, John Schulman, and Jacob Hilton. \"Scaling laws for reward model overoptimization.\" International Conference on Machine Learning. PMLR, 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": 1 }, "primary_area": null, "questions": { "value": "* \"For computational convenience, we aggregate the losses\": could you explain why is it more convenient?\n* when you say: \"we categorize a question-answer (Q,A) sequence into preferences such as helpfulness, harmlessness, or honesty\", do you do it automatically or with human in the loop?\n* Could you discuss the connections between your RM classification approach and the paper \"Stop Regressing: Training Value Functions via Classification for Scalable Deep RL\"?\n* How do you explain that Rewarded soups perform better than MORLHF? and actually, which interpolating coefficients did you choose for those methods?\n\nSee weaknesses for more questions." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Multi-objective RLHF is an area of interest for the ICLR community.\n* Using an architecture with multiple heads is a practical approach to combine multiple RMs into a single one.\n* Using a classification approach for reward estimation is an interesting direction, notably regarding the recent findings in \"Stop Regressing: Training Value Functions via Classification for Scalable Deep RL\"?." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents MOSLIM, a novel approach to multi-objective alignment of LLMs. The main contribution is to fit the multiple rewards into a single architecture with multiple heads. Then by prompt conditioning, the policy can optimize a linear combinaison of those rewards. At inference time, prompting the network correctly can lead to diverse behaviours, improving over alternatives MORLHF baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper is very unclear. Some sections are overly complicated to explain simple things such as softmax or cross-entropy. This complexity makes things way more complex that required. For example, rather thansaying \"a scalar reward derived from a mapping function that converts classification results from reward model into reward scores\" you could just say \"combine the multiple rewards by weighted sum\".\n* Weak contributions. \"The prompt-driven mechanism enables flexible adaptation to varying preference intensity combinations\" has previously been used. The choice of the reward architecture is not novel and not sufficiently ablated. The reward mapping is actually a weight sum.\n* The experiments lack ablations. What the pros and cons of such architecture; do you have transfer across tasks, or in contrast do you lose some?\n* Nit. the sum in Equation (11) contains k+1 elements.\n* Nit. Multiple typos in Figures: \"RSopu\" in Fig1, \"Multiple Lable\" in Fig2, \"helpfulness or helpful\" in Fig 7..." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024moslimalign,\ntitle={{MOSLIM}:Align with diverse preferences in prompts through reward classification},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w0MAu8vjwj},\nnote={under review}\n}" }, "abstract": { "value": "The multi-objective alignment of Large Language Models (LLMs) is essential for ensuring foundational models conform to diverse human preferences. Current research in this field typically involves either multiple policies or multiple reward models customized for various preferences, or the need to train a preference-specific supervised fine-tuning (SFT) model. In this work, we introduce a novel multi-objective alignment method, MOSLIM, which utilizes a single reward model and policy model to address diverse objectives. MOSLIM provides a flexible way to control these objectives through prompting and does not require preference training during SFT phase, allowing thousands of off-the-shelf models to be directly utilized within this training framework.\nMOSLIM leverages a multi-head reward model that classifies question-answer pairs instead of scoring them and then optimize policy model with a scalar reward derived from a mapping function that converts classification results from reward model into reward scores. We demonstrate the efficacy of our proposed method across several multi-objective benchmarks and conduct ablation studies on various reward model sizes and policy optimization methods. The MOSLIM method outperforms current multi-objective approaches in most results while requiring significantly fewer GPU computing resources compared with existing policy optimization 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": [ "Large Language Models", "Multi-objective alignment", "Reward modeling" ] }, "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/a8cdd1e8c684da4aa963b92a95f4101fd11c0e76.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": "MOSLIM:Align with diverse preferences in prompts through reward classification" }, "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": "Wizard-70b was used as the auxiliary model. This is a strange setup when attacking weaker models such as Vicuna and LLaMA-2-7b. Why would you need to attack a weaker model when you already have a stronger model that is capable of manipulating the weaker model?\nClarify the process for training the human judge to classify harm. Is it just a single human? Is the judge independent of the authors with a single or double-blind setup?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The topic of this paper is important, jailbreaking attacks demonstrate LLMs are not robust and easily give harmful content outside of a single-turn question.\n\nThe work is relatively easy to understand and the authors provide a good amount of ablation studies to justify the design of their methods. It takes inspiration from prior work to leverage the context, but does not use any optimization against a specific model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Contextual Interaction Attack (CIA) against LLMs, which involves prepending a series of prompt, response pairs that gradually lead to a harmful question. The prompts in these conversations are generated via few-shot prompting the Wizard-70b LLM, and response generated by the target model. The authors show results competitive with other attacks on 3 open-source and 3 closed-source LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Novelty:\nThe main results on 50 prompts do not show significantly more success than prior attacks. The open source models evaluated are not exciting, they’re somewhat outdated and not claimed to be robust to attacks. This work would benefit from attacks on more recent models such as Llama 3 or 3.1, or attacking a larger model/models with refusal training.\nThe method is fairly simple, it boils down to prompting another LLM to mimic how a human jailbreaker would attack an LLM. It would be great to elaborate more on the novelty of this approach relative to recent work on multi-turn jailbreaking.\n\nSoundness:\nA human judge was used on AdvBench without much detail given, this affects the soundness of the paper. Please elaborate more. Have you considered testing your method with a controlled judge, such as HarmBench [1], which has a trained static classifier?\nThe transferability experiments (Table 2) feel contrived. (1) Attacks such as PAIR don’t assume a threat model to transfer the instance-level outputs of the attack, it is not a fair comparison. However, the attack itself only needs black box access, it doesn’t need to transfer the instance-level output. (2) The closed-source models tested have specific defenses implemented against popular attacks such as GCG, it is not a fair comparison to measure transferability. The authors should report the results in the original attack paper whenever possible, which shows a much stronger transferability attack against GPT-3.5 and GPT-4 models.\n\nDefenses:\nPerplexity based defenses tested are specifically designed against GCG-style attacks, it is expected that they are not actually detecting harm. This work would benefit from showing strong results on general defenses such as DERTA [2] or Llama Guard [3], or more robustly trained models such as CYGNET.\n\nPresentation:\nPage 1 line 45: “jailbreak prompts generated by automated attacks\noften do not perform well when transferred to other LLMs, which is a major bottleneck”. This contradicts the citation on GCG immediately prior, which is transferable to other LLMs. The reason why GCG performs worse in this paper when transferred is likely due to model providers providing specific defenses against it. This does automated attacks transfer poorly.\nRelated work section only cites 3 papers involving jailbreaking LLMs, the rest are on background information such as in-context learning and autoregressive LMs that should be assumed as basic knowledge among this community. It would be good to further contextualize the novelty of this attack relative to prior work.\nMinor: Figure 1 has a typo “Repharse” -> “Rephrase”.\n\n[1] Mazeika, M., Phan, L., Yin, X., Zou, A., Wang, Z., Mu, N., ... & Hendrycks, D. (2024). Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. arXiv preprint arXiv:2402.04249. \n[2] Yuan, Y., Jiao, W., Wang, W., Huang, J. T., Xu, J., Liang, T., ... & Tu, Z. (2024). Refuse whenever you feel unsafe: Improving safety in llms via decoupled refusal training. arXiv preprint arXiv:2407.09121. \n[3] Inan, H., Upasani, K., Chi, J., Rungta, R., Iyer, K., Mao, Y., ... & Khabsa, M. (2023). Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674." }, "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. Explanation: Please give an illustrative example to support your demonstration - ” compared to harmful information provided by the user (exogenous input), LLMs tend to perceive their own generated content (endogenous output) as safer.” \n\n2. Experiment: Please add Crescendo[a] as a multi-turn attack baseline for comparison in experiment section\n\n[a] Russinovich, Mark, Ahmed Salem, and Ronen Eldan. \"Great, now write an article about that: The crescendo multi-turn llm jailbreak attack.\" arXiv preprint arXiv:2404.01833 (2024)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Interesting topic and insightful research, especially the mechanism analysis\n2. Well-written and visual logic\n3. Abundant and convincing experiment" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel multi-turn attack - Contextual Interaction Attack, using the model interactions (here is the benign preliminary questions to interact with the LLM) to elicit harmful responses. Specifically, this multi-turn attack leverages the autoregressive nature of LLMs to guide the model construct a context that is semantically aligned with the attack target, based on the question-responses pair." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of some detailed evidence to support the demonstration\n2. Lack of comparison with the prevalent multi-turn attack Crescendo" }, "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": 3 }, "primary_area": null, "questions": { "value": "1. What is the human judge setup? \n2. Why the attack success rate of Contextual Interaction Attack against Claude-2 is much lower than other models? Are there any distinctive behaviors of Claude-2 compared with other models?\n\nI would like to raise my score if the authors could address my concerns and questions." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ The presentation of the methodology part is clear and easy to follow. \n+ The experimental results are comprehensive, which include both advanced closed-source and open-source LLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a multi-turn jailbreak, which crafts a series of harmless adversarial queries to guide the victim model to output harmful content. Experimental results demonstrate the effectiveness of their attack method against different large language models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Lack of Novelty**: The core idea of this paper closely resembles that of Crescendo [1], a multi-turn jailbreak method that starts with an innocuous query about a harmful task and gradually guides the model to generate harmful content through benign incremental steps. Given this similarity, the concept presented in this paper does not appear to introduce significant novelty over Crescendo.\n\n Furthermore, the implementation in this paper also mirrors Crescendo’s approach. Crescendo employs in-context learning with a series of carefully crafted examples—a strategy similarly adopted in this paper. For instance, Crescendo’s attack sequence begins with “introduce the history of [harmful target]” (Figure 1), while this paper uses a similar prompt, “Write a paper on the history of [harmful target],” as the starting point of its attack (Figure 1).\n\n It would be valuable for the authors to elaborate on how their approach diverges from or improves upon Crescendo's methodology. Are there any distinctive methodological or experimental results that could underscore the novelty of this work?\n\n- **Insufficient Experiments**: The experiments in this paper are limited to a subset of the AdvBench benchmark, while other benchmarks, such as HarmBench [2] and SorryBench [3], offer more balanced and less noisy alternatives. Including results from HarmBench or SorryBench would enhance the robustness and comprehensiveness of the experimental evaluation, and the authors should consider incorporating these benchmarks in future work or as additional experiments.\n\n- **Lack of Comparisons with Other Multi-Turn Attacks**: Besides Crescendo, several other multi-turn attack methods have recently emerged, including [4, 5, 6]. This paper would benefit from a concise comparison with these approaches, outlining key similarities or differences, and including a discussion of these methods in the related work section to provide a more comprehensive review of the landscape in multi-turn attacks.\n\n1. *Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack.*\n2. *HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal.*\n3. *SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal Behaviors.*\n4. *Chain of Attack: A Semantic-Driven Contextual Multi-Turn Attacker for LLMs.*\n5. *Speak Out of Turn: Safety Vulnerability of Large Language Models in Multi-Turn Dialogue.*\n6. *Imposter.ai: Adversarial Attacks with Hidden Intentions Towards Aligned Large Language 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": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "Not Needed" }, "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": "Questions and Suggestions\n1. Clarification on Sequential Tailoring: Section 3.2 claims that the model’s responses to preliminary questions are used to tailor subsequent prompts for attack purposes. However, further details on how these responses influence the attack's progression would add clarity to the method and demonstrate the practical application of this approach.\n \n2. Details on Evaluation and Human Judgement: The number of examples used to compute jailbreak percentages is not specified, making it difficult to assess the reliability of the reported success rates. Additionally, incorporating an automated LLM-based judge to compare human evaluation scores could enhance the objectivity of the results and add depth to the validation process.\n \n3. Further Validation Against Advanced Detectors: The current evaluation does not include finetuned adversarial prompt detectors, which could provide a more comprehensive test of the method’s robustness. Testing against finetuned models, such as [6] and [7], specifically designed to detect malicious prompts, would offer valuable insights into the effectiveness of the proposed attack.\n\nReferences:\n1. [Multi-turn and contextual jailbreak attacks research](https://arxiv.org/abs/2408.15221)\n2. [CFA: Contextual attacks on LLMs](https://arxiv.org/abs/2408.04686)\n3. [RedQueen: Multi-turn attacks](https://arxiv.org/pdf/2409.17458)\n4. [MART: Adversarial prompt sequences](https://arxiv.org/pdf/2311.07689)\n5. [PAIR: In-context attack prompt generation](https://arxiv.org/abs/2310.08419)\n6. [Prompt-Guard-86M](https://huggingface.co/meta-llama/Prompt-Guard-86M)\n7. [Llama-Guard-3-8B](https://huggingface.co/meta-llama/Llama-Guard-3-8B)" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper is well-structured and thoroughly validates the proposed method on multiple LLMs, providing a comprehensive assessment of its effectiveness.\n- The authors conduct an extensive empirical analysis, highlighting the jailbreak potential across different LLMs and presenting their findings in a detailed and systematic manner." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a multi-turn approach to interact with large language models (LLMs) and elicit a jailbreak condition by using an auxiliary LLM to automatically generate a sequence of benign prompts. The authors validate their proposed method on both closed and open-source LLMs, reporting jailbreak success rates of ≥ 90% on two open-source LLMs and ≥ 80% on two closed-source LLMs. Additionally, the paper explores cases where the proposed method does not achieve high jailbreak percentages on one closed-source and one open-source LLM. The approach aims to automate the creation of the attack prompt sequence, contributing to the automation of jailbreak attacks on LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The concept of multi-turn attacks on LLMs is already well-established in the literature (e.g., [1, 2, 3, 4]), with [2] particularly exploring the use of context to craft jailbreak sequences. Authors should clarify how this attack is different from the contextual attacks in [2]\n- The method for generating benign prompts closely resembles that in [5], where in-context learning capabilities of the LLM are utilized to generate new attack prompts based on given examples. Authors should clearly state the key difference in their approach for generating prompts compared to [5]\n- Section 4.1, which discusses evaluation metrics, mentions that the judge function relies entirely on human evaluation. However, there is insufficient clarity regarding the number of examples used to compute the jailbreak percentage and how human evaluation scores correlate with an automated LLM judge score. Authors should clearly specify the following points in regard to the paper:\n1. The exact number of examples used to compute jailbreak percentages\n2. Details on how many human evaluators were involved\n3. Any analysis comparing human evaluation scores to automated LLM judge scores, if available\n\nReferences:\n1. [Multi-turn and contextual jailbreak attacks research](https://arxiv.org/abs/2408.15221)\n2. [CFA: Contextual attacks on LLMs](https://arxiv.org/abs/2408.04686)\n3. [RedQueen: Multi-turn attacks](https://arxiv.org/pdf/2409.17458)\n4. [MART: Adversarial prompt sequences](https://arxiv.org/pdf/2311.07689)\n5. [PAIR: In-context attack prompt generation](https://arxiv.org/abs/2310.08419)\n6. [Prompt-Guard-86M](https://huggingface.co/meta-llama/Prompt-Guard-86M)\n7. [Llama-Guard-3-8B](https://huggingface.co/meta-llama/Llama-Guard-3-8B)" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Jailbreak LLM through multi-turn dialogues" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024leveraging,\ntitle={Leveraging the Context through Multi-Round Interactions for Jailbreaking Attacks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w0b7fCX2nN},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) are susceptible to Jailbreaking attacks, which\naim to extract harmful information by subtly modifying the attack query. As de-\nfense mechanisms evolve, directly obtaining harmful information becomes increas-\ningly challenging for Jailbreaking attacks. In this work, inspired from Chomsky’s\ntransformational-generative grammar theory and human practices of indirect con-\ntext to elicit harmful information, we focus on a new attack form, called Contextual\nInteraction Attack. We contend that the prior context—the information preced-\ning the attack query—plays a pivotal role in enabling strong Jailbreaking attacks.\nSpecifically, we propose first multi-turn approach that leverages benign preliminary\nquestions to interact with the LLM. Due to the autoregressive nature of LLMs,\nwhich use previous conversation rounds as context during generation, we guide the\nmodel’s question-responses pair to construct a context that is semantically aligned\nwith the attack query to execute the attack. We conduct experiments on seven\ndifferent LLMs and demonstrate the efficacy of this attack, which is black-box, and\ncan also transfer across LLMs. We believe this can lead to further developments\nand understanding of the security in LLMs" }, "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": [ "LLM jailbreak", "trustworthy ML", "safety 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/17d2a86bc69a27f2a8c7908530859fcb0ce23c2c.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": "Leveraging the Context through Multi-Round Interactions for Jailbreaking Attacks" }, "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": "1. The paper evaluates the performance of GPT-4 and Llama but does not extend the analysis to other state-of-the-art models such as GPT-4o, Claude, or a broader selection of open-source alternatives. Including these models could provide a more comprehensive assessment of current capabilities in automatic R&D.\n2. The study's focus on financial domain data raises concerns about the generalizability of the benchmark. Without empirical evaluation across diverse domains, it remains unclear whether the models would exhibit comparable performance in other research areas, which limits the applicability of the findings." }, "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 propose an interesting task. The concept of automating the R&D process to minimize human intervention is innovative and highly valuable.\n2. The proposed RD2Bench appears well-conceived, focusing not only on a model’s understanding ability but also on the complex interaction between multiple abilities, such as data extraction, method selection, and implementation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a benchmark named RD2Bench for evaluating data-centric automatic Research & Development (R&D) using Large Language Models (LLMs). The benchmark aims to assess and improve the efficiency of the R&D process by evaluating a model’s ability to understand, extract, and implement methods described in raw research materials. The focus is on leveraging LLMs to perform data-centric R&D in a real-world context, with an emphasis on minimizing manual labor for researchers. The authors also present an evaluation of current state-of-the-art LLMs, like GPT-4, in different stages of data-centric R&D." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper presents an interesting and novel task for large language models, contributing valuable insights to the field. However, despite its strengths, there are still some weaknesses that prevent me from assigning a higher score and lead me to believe that the work, in its current form, is not yet sufficiently developed or complete.\n1. While the goal of RD2Bench is to evaluate models across a broad spectrum of R&D tasks, the current focus on only financial reports and stock trading data is a significant limitation. The models' performance on financial data may not be indicative of how well they would perform in fields with different data characteristics or domain-specific challenges.\n2. The paper mentions that RD2Bench includes human-annotated ground-truth results, but it provides insufficient detail about the annotation guidelines and mechanisms used. Given that human annotation quality is vital for evaluating automated systems, a more comprehensive explanation of how annotation challenges were overcome would be valuable.\n3. The performance metrics reported in Tables 1, 2, and 4 show values that are frequently close to 0.9 or even nearly 1.0, which raises concerns about the benchmark's effectiveness in distinguishing the capabilities of different models. Such high scores across various models suggest that the benchmark may lack the complexity or sensitivity needed to reveal meaningful performance differences, potentially limiting its utility for assessing model strengths and weaknesses comprehensively." }, "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. How do the authors envision RD2Bench being adapted or extended for domain-specific applications, such as finance or biotech, where specialized knowledge is often critical? Are there plans to incorporate domain-specific adaptations in future work?\n\n2. Could the authors elaborate on the complexity of the R&D tasks included in RD2Bench? Introducing more advanced, iterative R&D tasks could better reflect real-world challenges. Would the authors consider integrating more complex task workflows?\n\n3. What are the most common sources of errors encountered in the benchmark tasks, and how could they inform improvements for language models? An in-depth error analysis might reveal valuable insights into model limitations and areas for enhancement.\n\n4. The paper mentions quality control for human annotations but provides limited details. Could the authors clarify the mechanisms used to ensure annotation consistency, such as inter-annotator agreement metrics or other quality checks?\n\n5. Given the computational demands of RD2Bench, have the authors considered optimizations to reduce resource requirements? An analysis of potential optimizations would be beneficial for research environments with constrained computational resources." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. Innovative and Practical Benchmark: RD2Bench introduces a unique data-centric benchmark specifically designed for automating R&D tasks, moving beyond traditional evaluations by assessing multiple model capabilities in real-world applications.\n\n2. Focus on Synergistic Model Abilities: Unlike other benchmarks that assess isolated abilities, RD2Bench emphasizes the interaction and synergy between language understanding and technical implementation, providing a comprehensive evaluation framework for large language models.\n\n3. Detailed and Insightful Experimental Findings: The paper provides experimental results that reveal specific strengths and weaknesses of state-of-the-art models like GPT-4, offering actionable insights for future research on improving automated R&D capabilities.\n\n4. Robust Metric Design: The benchmark incorporates multiple detailed metrics, such as running success rate, format success rate, correlation, and accuracy, allowing nuanced assessments of model performance across diverse R&D tasks.\n\n5. Structured Literature Review: The paper contextualizes its contributions within existing research on language models and automated R&D, effectively highlighting gaps and motivating the need for a new benchmark like RD2Bench.\n\n6. Practical Applications for Productivity Enhancement: RD2Bench has clear practical implications, as it is designed to improve the efficiency and productivity of scientific research processes, making it highly relevant for both academic and industrial applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces RD2Bench, a benchmark designed to evaluate and advance data-driven automated R&D processes. RD2Bench assesses large language models on their performance in automated R&D tasks, focusing on two core abilities: (1) language understanding to accurately extract implementable methods from research materials, and (2) technical implementation skills to develop reliable and trustworthy solutions. Unlike existing benchmarks, RD2Bench evaluates not just individual capabilities but also the synergistic effects of multiple abilities in real-world R&D tasks. The contributions of the paper include providing a structured framework to assess model performance in automated R&D workflows and establishing metrics to measure task success, accuracy, and stability. Experimental results reveal the potential of advanced models like GPT-4 in these tasks, while also identifying areas needing further improvement to achieve full automation in scientific and data-driven fields." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of Analysis on Model Error Sources: While performance metrics like accuracy and correlation are well-covered, the paper could strengthen its impact by analyzing common sources of model errors, such as misunderstanding prompts or misinterpreting data. A deeper error analysis could highlight specific improvement areas in LLMs for R&D.\n\n2. Limited Detail on Human Annotation Quality Control: Although human annotation plays a significant role in RD2Bench, the paper could provide more information on quality control mechanisms for annotators beyond training and double-checking. Expanding on inter-annotator agreement or including validation checks would reinforce the reliability of the benchmark.\n\n3.Insufficient Discussion on Computational Efficiency: RD2Bench implies the need for repeated trials and extensive processing, especially for large models. An analysis of the benchmark’s computational requirements and possible optimizations could make it more practical for widespread research use, especially in settings with limited resources." }, "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.\tIs there a more detailed explanation of the task difficulty levels?\n2.\tRegarding method extraction in Section 3.3: Figures 1 and 2 only show the formula extraction results. Are the model extraction results also in a key-value format? What exactly do the evaluation metrics measure? Is it the matching degree between expected keys and extracted keys?\n3.\tIn the Implementation Step, are the method extraction ground truths used as inputs, or are the actual extracted results (even if potentially incorrect) used as inputs?" }, "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 the first to formalize the task of automating research and development (R&D), which is an important and impactful endeavor.\n2. The motivation of the paper is well-grounded and clearly articulated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a benchmark called \"RD2Bench\" designed to evaluate the process of Data-Centric Automatic R&D (D-CARD). RD2Bench focuses on the interaction and synergy between the abilities of large language models (LLMs) in areas such as language understanding, code implementation, and data selection during the R&D process. The benchmark simulates real-world data-centric R&D scenarios to study the performance and challenges of these models in handling automated R&D tasks. Experimental results indicate that while current state-of-the-art LLMs perform well on some simple R&D tasks, more complex tasks still pose significant challenges." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The dataset size and scope are quite limited, containing only 27 formulas and 6 models. The formulas are solely from the financial domain, and the models are only graph neural networks, representing a very small part of \"Data-Centric Research and Development.\n2. The authors claim RD2Bench evaluates the interaction and synergy of various model capabilities, but the results only assess \"information extraction\" and \"method implementation\" separately. Existing benchmarks, like [1], already combine these capabilities, and RD2Bench's unique contribution in this regard is not clearly demonstrated.\n3. The evaluation metrics, task descriptions, and several experimental details remain unclear, particularly with the absence of Appendix A.\n\n[1] Mialon, G., Fourrier, C., Swift, C., Wolf, T., LeCun, Y., & Scialom, T. (2023). Gaia: a benchmark for general ai assistants. arXiv preprint arXiv:2311.12983." }, "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 above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Innovative approach to automating data-centric R&D with a comprehensive benchmark.\n2. Thorough methodology encompassing data collection, human annotation, and robust evaluation metrics.\n3. Insightful experimental results demonstrating the capabilities and limitations of current models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces RD2Bench, a benchmark designed to enhance real-world data-centric automatic R&D (D-CARD) by evaluating the synergistic effects of various model capabilities. The authors aim to improve research efficiency and foster the automation of data-centric R&D processes. The study highlights the strengths of state-of-the-art models like GPT-4 while identifying areas for further improvement." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Data is limited to the financial domain; a broader variety of datasets would enhance the benchmark's applicability.\n2. The paper would benefit from providing related implementation codes and datasets to facilitate reproducibility.\n3. There is a lack of comparison with model combinations, which could provide insights into potential performance improvements.\n4. Insufficient discussion on the limitations and potential biases in the data collection process.\n5. Need for more clarity on how the benchmark could be adapted for various domains beyond the current focus." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024rdbench,\ntitle={{RD}2Bench: Toward Data-Centric Automatic R\\&D},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w0es2hinsd},\nnote={under review}\n}" }, "abstract": { "value": "The progress of humanity is driven by those successful discoveries accompanied by countless failed experiments. Researchers often seek the potential research directions by reading and then verifying them through experiments. The process imposes a significant burden on researchers. In the past decade, the data-driven black-box deep learning method has demonstrated its effectiveness in a wide range of real-world scenarios, which exacerbates the experimental burden of researchers and thus renders the potential successful discoveries veiled. Therefore, automating such a research and development (R&D) process is an urgent need. In this paper, we serve as the first effort to formalize the goal by proposing a Real-world Data-centric automatic R&D Benchmark, namely RD2Bench. RD2Bench benchmarks all the operations in data-centric automatic R&D (D-CARD) as a whole to navigate future work toward our goal directly. We focus on evaluating the interaction and synergistic effects of various model capabilities and aiding in selecting well-performing trustworthy models.\n Although RD2Bench is very challenging to the state-of-the-art (SOTA) large language model (LLM) named GPT-4, indicating ample research opportunities and more research efforts, LLMs possess promising potential to bring more significant development to D-CARD: They are able to implement some simple methods without adopting any additional techniques. We appeal to future work to take developing techniques for tackling automatic R&D into consideration, thus bringing the opportunities of the potential revolutionary upgrade to human productivity." }, "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": [ "Real-world Data-centric automatic R&D Benchmark", "data-centric automatic R&D", "trustworthy 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/4c1c0d6fcc72e13a6d9e5dc5990b43db9fab7d03.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": "RD2Bench: Toward Data-Centric Automatic R&D" }, "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": "Please see the 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 decomposes the inference process and provides a detailed analysis of the reasons behind the Detection-Generalization Paradox.\n\n2. This paper validate the phenomenon of the Detection-Generalization Paradox from the perspectives of landscape and sharpness.\n\n3. Experimental results demonstrate that DR-SAM simultaneously enhances the performances of OOD-D and OOD-G ability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the concept of the Detection-Generalization Paradox and analyzes the detailed reasons why existing OOD-D and OOD-G methods lead to this phenomenon. It proposes DR-SAM to simultaneously enhance the model's detection and generalization capabilities for OOD data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method DR-SAM lacks innovation. It appears to combine OE and SAM as optimization objectives, with an additional data augmentation to calculate perturbation factor $\\epsilon$.\n\n2. The analysis of the method is not detailed enough. In Algorithm 1, should $f_{\\theta+\\epsilon}$ in lines 3 be $f_{\\theta}$?\n\n3. In Algorithm 1, does using data augmentation to calculate the perturbation factor $\\epsilon$ in line 4 affect the model's performance on $D_{ID}^{test}$ compared to vanilla SAM?\n\n4. Is the capability for OOD-G derived from data augmentation or SAM? The authors should include relevant ablation experiments to clarify this.\n\n5. Data augmentation seems to be the most significant innovation in DR-SAM, and the authors should include experiments to demonstrate the impact of having or not having data augmentation." }, "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": 4 }, "primary_area": null, "questions": { "value": "Some questions to the authors are:\n\n1. In line 082, Analysis on logit space: For better OOD_D method enlarges the gap of prediction confidence between D_ID and D_CS. Is this a typo? Shouldn't it be D_ID and D_SS instead of D_ID and D_CS?\n\n2. Fig. 2 is not clear. Atleast not very explaining. Why the FPR is in the range of -ve and why the OOD-G accuracy of DR-SAM around 1.7? A delta term is used for both FPR and ACC. What is this delta? Is it the difference? It needs to be clarified both in the caption and the actual. It is difficult for a normal reader to apprehend whats going in this figure.\n\n3. In Fig. 5 the sharpness value is larger and in 6 and 7 the sharpness value is smaller. Is this the preferred characteristics of the curves for DR_SAM? Shouldn't the sharpness value in the Fig.5d remain almost steady across the value of rho? Because this characteristic contradicts with the statement made in the line 255-256.\n\n4. Why is the ID and OOD accuracy of DR-SAM less than that of vanilla SAM? As per the result, the gain can only be seen in the AUROC which is the metric for detecting semantic shifts. What about for the covariate shift part? Shouldn't the OOD accuracy be at least on par or better than the reference and vanilla SAM as per the claim of breaking the detection-generalization paradox? Please clarify.\n\n5. Regarding the experimental results, why the results are not compared with recent approaches that has been studied for both detection and generalization? Also, why the comparison has not been made for the Imagenet-200 benchmarks in terms of OOD-G methods in Table 3?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper delves into important aspect of the non-trivial problem in existing machine learning models. One of the most important strength of this paper is the manuscript is structured in a very proper way. The introduction section is very clear, the motivation is also clear and it is very clear what the authors are trying to do. The way in which the authors conducted the in-depth analysis of the behavior of models in the representation, logits, and loss space to show the actual detection-generalization paradox is worthy of admiration. The paper also stands out well in terms of contribution, where they have presented a new methodology \"Distribution Robust Sharpness Aware Minimization\" which is fairly intuitive and proving effective in maintaining the detection-generalization balance in the classification systems. The authors have provided fairly good amount of literature review and conducted extensive experiments on the available benchmark datasets, and also compared with the existing references in the field. Considering the reproducibility, the authors have provided full algorithm, and released the source codes as well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides research on one of the most important topic in machine learning, which is dealing with out of distribution data. Specifically, the paper present a detection-generalization paradox that exists in current machine learning systems. The authors analyze this paradox by\nproviding in-depth analysis of the models trained under various paradigms, focusing on representation, logits, and loss across\ndifferent data shifts. The authors propose a new idea for breaking this paradox and support their findings with extensive experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are some weaknesses associated with the paper. The paper has severe typos, and a thorough proof-read is required. For instance, covariate is written as covariant in many places. The provided source code is very hard to reproduce as it does not have a readme file, and to replicate the exact experiment is difficult. There are irrelevant and lot of details in the appendices of the related works section, which is not necessary at all. The experiments and results are promising, but it could have been done better by comparing with OOD-G methods too because there is a lack of proof indicating DR-SAM can beat existing OOD-G methods. The results are majorly focused on semantic shifts based methods only." }, "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": "See above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "## Strength\n\n- Interesting topic. The investigated problem is realistic, practical, and important. Combining these two tasks is necessary and critical.\n- Clear writing and good organization. The logic of the most part in this paper is smooth which makes it easy to follow. I enjoy the clear writing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on the relationship between two widely studied problems, i.e., OOD generalization and OOD detection. The authors conducted an empirical study and found that these two tasks conflicted with many previous OOD detection methods. To address this issue, a novel optimization framework named DR-SAM is proposed to balance OOD generalization and detection performance. The proposed method obtains the optimal model parameters by pursuing not only a clear decision boundary between clean ID data and semantic OOD data but also simulated covariant-shifted data and semantic OOD data. And thus better overall performance can be expected. Experiments on commonly used benchmarks can support the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Weakness\n\nMajor concerns:\n\n- Vague contributions. The authors claim that the main contribution of this paper is to identify the detection-generalization paradox. However, as far as I could tell, the trade-off/conflict relationship has been pointed out by several previous works [1] [2] [3]. Thus such a claim should be either toned down or a clear explanation about the difference from previous work should be provided.\n- Unclear motivation. In lines 266-269, the authors claim that the ideal model should yield low sharpness on both ID and covariate-shifted data. They claim that this cannot be adopted OOD-G method. The logic here is hard for me to follow. Why solely using OOD-G method can not ensure low sharpness for ID and covariate-shifted data? I guess this is a typo. Is the covariate-shifted data here should be replaced with semantic-shifted data?\n- Lock of essential comparison. Although the experiments in Section 5 encompass a few representative works in OOD detection and generalization, some of the most related works are missed. Several recent methods also jointly consider OOD detection and generalization [1] [3] [4]. Thus, the comparison in this current version is biased and incomplete. Besides, I also feel that some SOTA OOD detection methods are also missed. For example, POEM [5], NPOS[6], and NTOM[7]. As far as I can tell, POEM substantially surpasses OE, MCD, and MixOE in terms of OOD detection on CIFAR benchmarks. SCONE, WOODS, and DUL which jointly pursue OOD detection and generalization can achieve much better overall performance compared to the baselines in Table 1 2 and 3. The reviewer suggests comparing the proposed method with these methods.\n- Experimental settings. The authors claim that they deploy brightness as data augmentation (Appendix C). I have concerns about whether using brightness augmentation during training can result in information leakage from the test covariate-shifted distribution. Since all the other corruptions in CIFAR10/100-C or ImageNet-C may also alter the brightness of images. As far as I could tell, in standard OOD generalization settings, the test-time covariate-shifted distribution should be kept unknown during training. Besides, the authors seem to tune the augmentation and make such a choice as they said in lines 466-468. Thus, the dependence on manually selected augmentation is a notable limitation that makes the OOD generalization problem more like domain adaption or even a trivial problem.\n\nMinor concerns:\n\n- Similar to other training-required OOD detection methods, the proposed method also needs to access semantic-shifted data during training. This limitation widely exists in many previous OOD detection methods, but still worth noting here.\n\n- In Figure 3(d), why there is no blue area (ID data) in the figure?\n- I am unsure whether the formulation of OOD generalization in Eq. 3 is correct. $D_{CS}$ is mentioned by the text above Eq.3. However, there is no such notation in the following equation. The formulation should be revised carefully and a proper citation should be provided here.\n- The authors post a visualization of the loss landscape in Figure 10. However, such a visualization contains limited information. The reviewer suggests comparing with SAM, OE, and the original ERM.\n\nOverall, the quality of this paper in its current version does not meet the expectations of ICLR. However, I may adjust my score according to opinions from other reviewers if a strong argument is provided.\n\n[1] Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. ICML'23\n\n[2] Unified out-of-distribution detection: A model-specific perspective. CVPR'23\n\n[3] The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection. NIPS'24\n\n[4] Training ood detectors in their natural habitats. ICML'22\n\n[5] Poem: Out-of-distribution detection with posterior sampling. ICML'22\n\n[6] Non-parametric outlier synthesis. ICLR'23\n\n[7] Atom: Robustifying out-of-distribution detection using outlier mining. ECML PKDD'21" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024breaking,\ntitle={Breaking the Detection-Generalization Paradox on Out-Of-Distribution Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w0jk3L3IjV},\nnote={under review}\n}" }, "abstract": { "value": "This work studies the trade-off between out-of-distribution (OOD) detection and generalization. We identify the Detection-Generalization Paradox in OOD data, where optimizing one objective can degrade the other. We investigate this paradox by analyzing the behaviors of models trained under different paradigms, focusing on representation, logits, and loss across in-distribution, covariate-shift, and semantic-shift data. Based on our findings, we propose Distribution-Robust Sharpness-Aware Minimization (DR-SAM), an optimization framework that balances OOD detection and generalization. Extensive experiments demonstrate the method's effectiveness, offering a clear, empirically validated approach for improving detection and generalizationability in different benchmarks." }, "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": [ "Trustworthy Machine Learning; Out of distribution data" ] }, "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/bf80cbfd1188a08b31ada628430558c0f2b53d87.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": "Breaking the Detection-Generalization Paradox on Out-Of-Distribution Data" }, "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": "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. The paper is well-organized and easy to follow. Figures are clear to read, such as Figure 2. \n2. The story is complete: they propose hypothesis and then use experimental results to verify them in Sec 3.3 with clear ablation studies. \n3. The problem setup is clear. They also provide enough details to reproduce the work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work first presents the conflict distribution issue in preference datasets, where preferred and less-preferred images generated from the same project exhibit significant overlap. For this issue, they introduce the Captioning and Perturbation methods: generate a caption based on the image and the prompt, create three levels of perturbation from the prompt. They also explore the irrelevant prompt issue in previous DPO methods and propose Dual Caption Preference Optimization (DCPO) to improve diffusion model alignment. Lastly, they show promising results compared to the existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My biggest concern is about the generalization of the approach method in the development of diffusion models. For example, in Figure 2, it is easy to distinguish the preferred and less-preferred image as the latter one even does not align with the original prompt. What if the model's development is already beyond the alignment stage? The current positive/negative samples are only about alignment, what about more advanced difference if both have enough alignment? \n\n2. Line 188-189, could you explain more details on how to get the preferred and less-preferred images? Human annotation? \n\n3. It would be beneficial to highlight the difference between medium and strong permutation. Do we have a way to quantify the difference between them? Are they controllable generated? Why do we need medium permutations? Would weak/strong be enough? \n\n4. In terms of GPT-4o evaluation, does it matter for showing the images together or showing them separately? And how about the order of showing them to GOT-4o if showing separately?" }, "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": "Despite the experiments suggests the proposed approach is better, it is unclear to me why this would be the case, any proofs or intuitions will help reader better understand it.\n\nSeveral papers appear multiple times in the References section, please dedupe." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is well written with right amount of details in both main text and appendix. The proposed method is clear, and relativly straightforward to implement. \n\nOn a popular open source diffusion model ( SD 2.1), several experiments are done to ablate the design details of the proposed approach. The used set of metrics are comprehensive, including both single side evaluation such as HPSv2, as well as side by side evaluation such as the one using GPT4-o as judge." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose an improved method for aligning text to image diffusion models using human labeled preference datasets. Instead of using the original caption that is used to generate the preferred and less preferred image pair, the method propose to generate new captions from the generated images or original captions so as to increase/decrease the text image alignment for the generated images, which make the the distribution difference between preferred and less preferred data larger than using original captions.\n\nExperiments using diffusion DPO method are conducted in various versions of caption image combinations, which show adding perturbed captions for less preferred image helps finetuned model get better performance on automatic metrics, including itemized metrics such as HPSv2, as well as side by side metrics using GPT-4o as judge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The motivation behind the proposed approach is not clear to me.\nFor the conflict distribution challenge, when the distribution overlap becomes larger, the dataset is proposing a harder problem for the model to optimize, but it isn't necessary an issue as long as the two distributions are not identical. When the diffusion models's quality gets better, the two distribution will inevitably become more and more similar, as both preferred and less preferred images from an optimized model will be closer to real human preference. So it's more of the nature of the task itself, unless the task is defined differently.\n\nFrom the description of L175-L180, the irrelevant problem is hardly a problem either. It is an inherently part of the objective in Eqn (1), where one way of minimizing Eqn(1) is to decrease $\\log(p_{\\theta}(x_{0:T}^l|z^l)$, which makes the model less likely to generate the less preferred image. So to me this is a desired behavior instead of a problem. \n\nBy changing the captions, authors changed a prefer/less prefer pair into two separate samples. In this sense it is no longer the original DPO problem, yet there is no clear connection between the original DPO formulation and the new problem e.g. is the new one an upper-bound of the original so minimizing the new problem potentially minimize the original one? or why solving the new problem will necessarily give better results than original DPO?\n\nThe change of captions made the problem closer to the KTO problem referenced in the paper, where text-image data are labeled by like and dislike binary labels. Please describe the connection and difference between the modified problem represented by the new data and the KTO problem formulation above.\n\nIt is great to conduct extensive experiments on SD 2.1, but the paper will be stronger if there are experiment results on other diffusion models, even if the experiments are not as complete as on SD 2.1." }, "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": 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 address my concerns above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The dual caption framework is reasonable. DCPO introduces a dual-caption system that effectively addresses the problem of overlapping distributions in existing datasets.\n2. This paper achieves better performance. Demonstrated improvements across multiple metrics (e.g., Pickscore, CLIPscore) and benchmarks (e.g., GenEval) show that DCPO enhances image quality and relevance significantly.\n3. The experimental results are analyzed in detail. The paper includes extensive quantitative and qualitative analysis, supporting the effectiveness of DCPO with various baselines and ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a preference optimization technique called Dual Caption Preference Optimization (DCPO). This method aims at improving text-to-image diffusion models. DCPO tackles issues inherent in current preference datasets, namely conflict distribution and irrelevant prompts, by introducing separate captions for preferred and less preferred images. This dual-caption approach is implemented through three methods: captioning, perturbation, and a hybrid method, all aimed at enhancing the clarity of distinctions between preferred and non-preferred images. Experimental results demonstrate that DCPO outperforms existing models across several benchmarks and metrics, including Stable Diffusion 2.1 and Diffusion-DPO." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method depends on the caption quality. The quality of generated captions significantly affects performance, and challenges remain in creating effective captions for less preferred images without straying out-of-distribution.\n2. While DCPO demonstrates quantitative improvements across several metrics, the qualitative results (e.g., Figure 1) indicate that the visual distinctions between images generated by DCPO and baseline methods are not significant. This subtle difference may limit the perceived impact of DCPO in practical applications.\n3. The DCPO has limited generalizability compared to real-world large-scale datasets. Although leveraging preferred and non-preferred images is a novel approach for enhancing diffusion models, high-quality, large-scale datasets from real-world settings often provide stronger improvements in model performance. This reliance on real-world data diminishes the relative advantage of DCPO, potentially limiting the distinctiveness of its contributions in scenarios where comprehensive datasets are available.\n4. The LAION-2B and MSCOCO datasets are widely regarded benchmarks for image generation tasks, yet they are not discussed or evaluated within this study. The absence of experiments or comparisons involving LAION-2B raises questions about DCPO’s general applicability." }, "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": 3 }, "primary_area": null, "questions": { "value": "The issues of conflict distribution and irrelevant prompts seem like two aspects of the same problem—both involve a single prompt (C) corresponding to two different images, which can lead to unstable optimization. Therefore, I think they could be consolidated into a single issue." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "As a reviewer from a broader field, I am not very familiar with the specific domain of this paper. Therefore, I am reviewing this paper from a generalist’s perspective. The strengths of this paper are:\n\n1. It provides sufficient theoretical support for the motivation, which aligns well with the characteristics of ICLR papers.\n2. The issues raised seem quite reasonable.\n3. Extensive quantitative and qualitative experiments support the arguments presented." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Dual Caption Preference Optimization (DCPO) to enhance text-to-image diffusion models by aligning them with human preferences. Traditional methods face issues like overlapping distributions and irrelevant prompts. DCPO addresses these using two distinct captions for each image, mitigating conflicts in preference data. The authors introduce the Pick-Double Caption dataset to support this approach. They apply three strategies—captioning, perturbation, and hybrid methods—to generate unique captions. Experiments show DCPO improves image quality and prompt relevance. DCPO outperforms prior models on multiple metrics, validating its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "However, I still have a few concerns:\n\n1. The issues of conflict distribution and irrelevant prompts seem like two aspects of the same problem—both involve a single prompt (C) corresponding to two different images, which can lead to unstable optimization. Therefore, I think they could be consolidated into a single issue.\n2. When comparing generated images, the improvements achieved by the proposed method could be highlighted more clearly; otherwise, it’s often not immediately obvious, as in Figure 1.\n3. In fact, the explanations of conflict distribution and irrelevant prompts in the abstract and introduction are quite obscure and difficult to understand. I had to reread these sections several times, only gaining clarity after reading the methods section. This part may need reorganization." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024dual,\ntitle={Dual Caption Preference Optimization for Diffusion Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w0lhe9prqH},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in human preference optimization, originally developed for Large Language Models (LLMs), have shown significant potential in improving text-to-image diffusion models. These methods aim to learn the distribution of preferred samples while distinguishing them from less preferred ones. However, existing preference datasets often exhibit overlap between these distributions, leading to a conflict distribution. Additionally, we identified a performance issue in previous optimization methods, where using the same prompt for preferred and less preferred images, known as the irrelevant prompt issue, restricts model performance. To address these challenges, we propose Dual Caption Preference Optimization (DCPO), a novel approach that utilizes two distinct captions to mitigate irrelevant prompts. To tackle conflict distribution, we introduce the Pick-Double Caption dataset, a modified version of Pick-a-Pic v2 with separate captions for preferred and less preferred images. We further propose three different strategies for generating distinct captions: captioning, perturbation, and hybrid methods. Our experiments show that DCPO significantly improves image quality and relevance to prompts, outperforming Stable Diffusion (SD) 2.1, SFT-Chosen, Diffusion-DPO and MaPO across multiple metrics, including Pickscore, HPSv2.1, GenEval, CLIPscore, and ImageReward, fine-tuned on SD 2.1 as the backbone." }, "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": [ "Preference Optimization", "Diffusion Models", "Alignment" ] }, "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/97c5def3ed53f91241db16fdf0e9c7236a9a7b71.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/374cb01dc3465628d70200d0c6c8424ea24beff3.zip" }, "title": { "value": "Dual Caption Preference Optimization for Diffusion 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": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No ethics is foreseen in this work." }, "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": "Would be nice to understand the difference of eq. 16 and the standard VIB." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "If the paper were clearer, the paper presents many contributions and analyses of the proposed terms. \n\nTwo experiments (image and text classification) compare the \"vanilla\" and VIB with the new loss. \n\nHistorical overview of IB." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper claims to provide the \"theoretical optimal approach to data modeling\", and to extend the framework, by deriving the a variational bound to resolve some problems of the previous framework. \n\nMy understanding is that eq. 16 (derived by eq.3) is the contribution of this work. The paper provides derivations to justify eq. 16. \n\nIn the experimental session, the authors evaluate the performance in image and text classification of the new loss and the robustness to adversarial attack. \n\nThe authors claim that the new loss outperforms the previous loss." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper is largely unclear. \n\nIt starts with the history of the IB, more than presenting the contribution of the work.\n\nThe presentation is not clear on the steps of the new loss. \n\nIn the new loss, there seems to be a new contribution on the predictor, but the predictor (or classifier) is already included in the loss in the VIB. \n\nThe impression is that the new loss introduces a new regularization term, but its justification is not clear. \n\nThe abstract is unclear, what are the two points of the \"dual role\"? What is the \"theoretically optimal approach to data modeling\"?" }, "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. Could you please provide more experiments on PGD and AutoAttack?\n2. Could you please visualize the latent representations of SIB and compare them with VIB?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "SIB performs better than VIB regarding classification accuracy and adversarial robustness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper revisits the variational information bottleneck and extends it to a supervised variational information bottleneck. The experiment on ImageNet and text classification shows that SIB achieves better results than VIB." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, the novelty is insufficient, and the motivation is unclear. Some related works are missing.\n\n### Novelty:\n1. Using variational lower or upper bonds to approximate mutual information is not novel.\n2. SIB's application focuses on traditional image classification and adversarial attacks. It doesn't include other applications like time series or more challenging scenarios like out-of-distribution or few-shot learning. \n3. Compared to VIB, SIB adds $H( \\hat{Y} \\mid Z)$, involves new hyperparameters and new terms to approximate, and cannot guarantee to reach an accurate value for $H( \\hat{Y} \\mid Z)$.\n\n### Motivation:\n1. The title of the paper uses \"revisiting,\" which refers to why the VIB is revisited and what the problem of VIB is. These two parts are not clear in the paper. Figure 1 cannot demonstrate well since an overfitted decoder can also exist in SIB if the $\\lambda$ is not set appropriately. \n\n### Missing Reference:\n\n[1] Kolchinsky, Artemy, Brendan D. Tracey, and David H. Wolpert. \"Nonlinear information bottleneck.\" Entropy 21.12 (2019): 1181.\n\n[2] A. Zaidi and I. E. Aguerri, “Distributed deep variational information bottleneck,” in Proc. IEEE 21st Int. Workshop Signal Process. Adv. Wirel. Commun., 2020, pp. 1–5\n\n[3] S. Sinha, H. Bharadhwaj, A. Goyal, H. Larochelle, A. Garg, and F. Shkurti, “DIBS: Diversity inducing information bottleneck in model ensembles,” in Proc. AAAI Conf. Artif. Intell., 2021, pp. 9666–9674\n\n[4] S. Mai, Y. Zeng, and H. Hu, “Multimodal information bottleneck: Learning minimal sufficient uni modal and multimodal representations,” IEEETrans. Multimedia, vol. 25, pp. 4121–4134, 2022\n\n[5] K. W. Ma, J. P. Lewis, and W. B. Kleijn, “The HSIC bottleneck: Deep learning without back-propagation,” in Proc. AAAI Conf. Artif. Intell., 2020, pp. 5085–5092.\n\nThe paper should compare the above methods as well and also put them into a related work 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": 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 think the theoretical contribution is solid and valuable to share with the community, but I think the empirical treatment is weak. I would be very happy to increase my rating if the experiments were improved, even if they did not show that SVIB is reliably better than VIB or CEB in all of the settings considered. Whatever the outcome for SVIB on more careful preliminary experiments would be a valuable scientific contribution." }, "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 and easy to read.\n- Constrained maximization of H(\\hat{Y}|Z) will clearly achieve the goal of preventing the classifier from overfitting to the representation.\n- The theoretical approach is plausibly useful. A careful set of experiments could demonstrate its value beyond using VIB or CEB." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "- The paper picks up an issue identified in the Deep Variational Information Bottleneck paper, where the classifier can overfit to the learned representation, Z, of a VIB model, and proposes a new framework for supervised learning with IB, which they call Supervised Information Bottleneck (SIB), and a corresponding variational approach, SVIB.\n- The core theoretical contribution is to add a constraint to the IB and VIB objectives that minimizes an upper bound on I(\\hat{Y},Z), which is equivalent to maximizing a lower bound on H(\\hat{Y}|Z). The paper shows that this new constraint is tractable in the SVIB setting.\n- The paper provides experiments comparing SVIB to VIB and “vanilla” Maximum Likelihood models (trained with cross entropy) on ImageNet and natural language sentiment analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- In general, the experiments are of the correct form (comparisons between different IB approaches and Maximum Likelihood on clean and adversarial test sets), but they are unconvincing at supporting the main claim that SVIB substantially improves on other proposed tractable IB approaches, as pointed out in more detail below.\n- One shortcoming of all of the experiments is that the VIB models are not given the same amount of hyperparameter tuning as the SVIB models – it appears that in all cases, the SVIB models get three times as many runs with different hyperparameters to find a setting that outperforms the VIB models.\n- VIB on classification tasks often benefits from having a mixture distribution for r(z), whether learned or just distributed across part of the domain of Z, rather than having a single isotropic Gaussian distribution for r(z). It’s likely that your selected values of \\beta would perform better in that setting, as it becomes easier for the model to learn to assign classes to different mixture elements as it sees fit, which makes the model more powerful (more powerful models can tolerate higher compression/higher values of \\beta). This would likely benefit SVIB as well, so that it more reliably outperforms the Maximum Likelihood baseline on the test set.\n- The paper is missing an important citation: CEB Improves Model Robustness, Entropy 2020. Overlooking this reference is a major shortcoming of the paper, since it studies the same question on one of the same datasets using the same Information Bottleneck framework, and it achieves substantially better results on that dataset than reported in this paper (its VIB results are also stronger than your VIB and SVIB results).\n- The ImageNet table highlights SVIB results in settings where the VIB results appear to strongly overlap – it seems a stretch to claim that SVIB is doing better than VIB with a result of 53.4%+/-1.8% compared to 53.5%+/-0.2% for FGS with \\epsilon=0.1, for example (and similarly but to a lesser extent for FGS with \\epsilon=0.5).\n- For all experiments, hyperparameter selection for VIB is questionable, as test set performance on the clean data appears to still be improving substantially at the smallest value of \\beta. As \\beta goes to 0, its performance should match the vanilla model on the clean data, but you stop exploring \\beta when the test set performance is substantially worse than the vanilla model, indicating that probably neither the VIB nor the SVIB models are very close to optimally configured.\n- The Conditional Entropy Bottleneck paper showed that CEB reliably outperforms VIB on both clean and adversarial examples on a variety of image datasets. The CEB Improves Model Robustness paper further explores that in detail on ImageNet. Since implementing CEB can be made parameter-equivalent to implementing VIB (and consequently SVIB), it seems like an important point of comparison. \n- In Figure 1, right-hand side, the H(\\hat{Y}) circle is drawn in a way that does not respect the Markov chain constraint Y-X-Z-\\hat{Y}. It is not possible to have H(\\hat{Y}) overlap H(Y) in any area where H(Z) does not also overlap H(Y). Compare this to the Venn diagrams in the Conditional Entropy Bottleneck paper you cite, where similarly the Markov chain Z-X-Y prevents H(Z) from overlapping H(Y) anywhere that H(X) does not also overlap H(Y).\n- Line 360: repeated word: “is uninformative about about Y”." }, "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": "Here are some detailed comments and questions:\n\nUse \\log in latex and format the integral d.\n\nThe writing may be a little verbose. Examples: lines 260-264 restates things. (7) follows trivially from (4). The sentences preceding both (4) and (7) are also similar. lines 217-254 is repeat well known material.\n\nLine 198 why carry p(x,y) around if everything is conditional on those later? I suspect dropping that saves some hassle later.\n\nLines 271 to 296 would be better expanded after moving lines 217-254 to an appendix. You could be explicit about the use of the chain factorization etc (although maybe the previous point about line 198 can avoid needing to deal with this?).\n\nLine 249 who’s -> whose\n\nLine 334 why is Z left as an r.v.? Please explain how to handle this with sampling. I feel like this is just the entropy of the y given the sampled z, so it should be written explicitly as such?\n\nLine 452 Val column bolded wrongly, the vanilla method should be shown as the winner not the proposed method?\n\nTable 1: it seems as though VIB best performance is at the boundary of your sweep, so we can’t tell if SVIB beats VIB?\n\nTable 2: as previous comment.\n\nA plot instead of tables 5 and 6 would be easier to absorb.\n\nTables 1 and 2: can you not fix lambda and show improvement generally? Varying this in-sample looks like overfitting to the untrained eye (but I think this is an illusion and the results are good). It just seems like sub optimal presentation given tables 5 and 6 show good robust performance over lambda. A plot >> tables of numbers.\n\nSection 4: showing robustness is nice, and the methodology seems very good i.e. adversarial approaches.\n\nLine 478 private -> special" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper boils down to a simple to implement and intuitive loss function." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper extends the variational information bottleneck by adding an entropy regularizer to the model that predicts the target y given the latent z. This is motivated by adding and variational bounding a second info bottleneck." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There is a lot of justification that is somewhat verbose and subjective. The derivation is long and elaborate for what boils down to an extra regularizer with an extra tuning parameter." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A new variational adaptation of the IB for supervised DNN optimization" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024revisiting,\ntitle={Revisiting the Variational Information Bottleneck},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w10KdRwcMk},\nnote={under review}\n}" }, "abstract": { "value": "The Information Bottleneck (IB) framework offers a theoretically optimal approach to data modeling, though it is often intractable. Recent efforts have optimized supervised deep neural networks (DNNs) using a variational upper bound on the IB objective, leading to enhanced robustness against adversarial attacks. In these studies, supervision assumes a dual role: sometimes as a random variable with the empirical distribution of the data, and at other times as a random variable distributed according to the classification performed. This work proposes an extension to the framework, and consequently to the derivation of the bound, that resolves this duality. Applying the resulting bound as an objective for supervised DNNs induces substantial empirical improvements." }, "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": [ "information bottleneck", "information theory", "representation learning", "adversarial attacks", "regularization", "supervised 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/780a7b77086b47d3497331fc0b85e68b5361c969.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/0534395a96dd4fc2f39fae56eade10cdc9f3aa3d.zip" }, "title": { "value": "Revisiting the Variational Information Bottleneck" }, "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": "Minor Issues:\n1. The text in Figure 4 is difficult to read, and the radar chart lacks specific performance values, making it challenging to interpret the results accurately." }, "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 defines a novel PDL to address OOW requests that agents may encounter during workflow execution. A comprehensive and unified framework like this facilitates further research in this field.\n2. The authors also introduce a new evaluation framework specifically designed to assess workflow agents' performance in OOW scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces FlowAgent, a framework for integrating workflows into LLMs that balances procedural compliance with flexibility, addressing the limitations of existing prompt-based and rule-based methods. To achieve this, the authors propose a Procedure Description Language (PDL) that combines the natural language flexibility and programming-like precision needed for diverse workflows." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The introduction of PDL lacks thorough analysis. After observing the impacts of OOW queries, it is unclear what specific considerations led to the development of PDL. Additionally, the completeness of PDL requires further examination to demonstrate its capability to handle more complex, real-world workflows effectively.\n\n2. The evaluation of experiments is incomplete; the authors only assess GPT-4 and Qwen2-72B, with a brief note that “weaker models could not handle more complex workflow tasks.” However, there is no detailed analysis on what specific issues smaller models faced. Further exploration is needed to show how smaller models perform on simpler workflow tasks to provide a clearer picture of model scalability across task complexity.\n\n3. Several key details are missing, such as the hyperparameters used during LLM inference, the prompts employed during data collection, dataset construction details, and relevant examples from the datasets. Including these would improve the reproducibility and clarity of the experiments." }, "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. For pass rate in turn-level evaluation, do you use an LLM to check whether the output is correct by turning it into a binary classification problem? \n2. Why do you use three types of OOW categorization? Changing their previous answers could be another scenario of OOW.\n3. I believe that the Star dataset also contains OOW scenarios; why did you add more such dialogues, and how do you generate such queries?\n4. Can you provide some stats or descriptions for your in-house dataset?\n5. It is unclear how you converted existing flows into natural language, code, and flowcharts. Who made these conversions? Xiao et al. 2024 uses GPT to convert the text into NL, code, and flowcharts, whereas you write your own PDL. I don't think the comparison is fair here.\n6. How do you construct reference sessions from tasks for turn-level evaluation using GPT-4o? Please provide more information.\n7. Can you please provide an example conversation with the simulated users? There are works that demonstrate that user simulation is non-trivial in a conversational setting. [1]\n8. Do you have any experiments with users or simulated users to suggest that PDL can handle WIKIHOW like examples?\n9. In the WIKIHOW example, what would happen if the user directly asks how to find a website's publication date using code? Since the write PDL says `if publication_date is None:` then use Google search or other tools. In this, the publication date will not be None, right? This makes me believe that you must define several \"flows\" for the PDL to work in all scenarios.\n10. Table 1 typo, for # Turn in Star turn-level row.\n11. Who add the OOW nodes to the DAG?\n12. What metric did you use to decide the flexibility and compliance score in Figure 1 (c)? What is the scale for the plot?\n\n[1] Zhou, X., Su, Z., Eisape, T., Kim, H., & Sap, M. (2024). Is this the real life? is this just fantasy? the misleading success of simulating social interactions with llms. arXiv preprint arXiv:2403.05020." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The work tries to add flexibility and compliance to the conversational agents, which is a real-world challenge.\n- They experimented with three augmented datasets and showed that FlowAgent with PDL designed by humans works the best compared to NL, Code, and flowcharts created by GPT-o. \n- The writing is clear and easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper creates a framework (FlowAgent) for task-oriented agents that can offer flexibility and compliance with LLMs. They propose a new language, PDL, that creates a Directed Acyclic Graph with out-of-workflow (OOW) slots, making the agent flexible and using pre- and post-controllers for compliance. They perform an extensive evaluation on augmented datasets for showing how PDL handles flexibility and compliance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper has several missing details, especially regarding the experimental setup. I don't think there is enough evidence to suggest that PDL works for the WikiHow case study.\n- The authors simulate users. However, as several studies have suggested, simulated users still do not capture real-world cases. The authors should have done a real-user study.\n- With PDL, as the number of tools scale, it might be difficult for the developer to define all the pre_conditions. Especially since you need to define multiple flows and the logic is non-trivial to write to cover all the cases. See question 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": 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": "NA" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper was easy to read and addresses an interesting problem in the literature that is very relevant to industry applications. The solution also seems simple to implement making it easy to adopt in use cases. The experimental section shows the benefits of the authors' proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose FlowAgent, an agent built using LLMs and incorporates workflow rules to balance between flexibility and compliance. The agent leverages both pre-decision and post-decision controllers to adjust the agent's behavior. Experimental results on existing and custom datasets show FlowAgent's superiority compared to other approaches from the literature." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper contains a few gaps in the both the presentation of the work and the experimental section. \n\nFirst, it is unclear how the controllers were implemented, i.e., there isn't sufficient information in the write up for one to reimplement or even have an idea of what approach was adopted. It seems that the controllers rely on deterministic syntax and logic checks but that is pure speculation on my part as the paper only has a few lines describing both pre and post decision controllers, namely focusing on their purpose as opposed to their implementation. It would be helpful if authors could either include a pseudocode of their controllers or a diagram describing their operations. \n\nSecond, the experimental section does not discuss the computational cost and additional overhead of FlowAgent compared to other approaches. Since FlowAgent operates in a conversational setting, a level of responsiveness is expected but we have no way of knowing from the paper how much overhead the added controllers are causing from both responsiveness and computational cost perspectives. It would be helpful if the authors included runtime comparisons or latency measurements between FlowAgent and the baseline approaches in their experimental results. Additionally, an analysis of how the controllers impact the overall response time in a conversational setting and the accuracy of the end to end system would be very helpful \n\nWhile figure 2 serves as the main architecture diagram describing the approach, it is very abstract. The paper could benefit from another figure showing how the controllers were working since those seem to be a key contributor. Again, I assume they rely on some sort of graph algorithm based on the DAG in figure 3 but more information is needed. Including a flowchart showing the decision-making process of the pre-decision and post-decision controllers, or a diagram illustrating how they interact with the DAG structure. This would help clarify the relationship between the controllers and the workflow representation." }, "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": "Please see my questions/comments in the above field. In particular: \n\n- Please give a concrete example from a domain of your interest (e.g., booking a hotel room) showing how prompt-based and rule-based methods from paragraph in 042--052. differ on their responses to a user query and stressing the limitation of the prior art and the contributions of this work. Please also give exact references-- there is no reference in this paragraph. \n- Please create a table or analyse the key differences between PDL and previous established workflow languages (see for example Serge Abiteboul, Pierre Bourhis, Victor Vianu. Comparing workflow specification languages: A matter of views. ACM Transactions on Database Systems, 2012, 37 (10)), particularly focusing on aspects relevant to LLM control.\n- Please provide a dedicated subsection defining out-of-workflow queries, along with a few concrete examples demonstrating how these queries differ from in-workflow queries and how PDL handles them.\n- Please experimentally compare your framework against one recent constrained-decoding technique using regular expressions, e.g., “Validating Large Language Models with Relm” by Michael Kuchnik et al. The experiment should demonstrate how the authors' approach differs from or improves upon previous constrained-decoding methods in the context of workflow control for LLMs. \n- Please provide a specific toy example showing how PDL handles a new out-of-workflow demand, and include an experimental comparison demonstrating the flexibility of PDL versus rule-based approaches in adapting to new demands." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed topic is interesting and, certainly, useful. However, several issues in the presentation make it difficult to understand the contributions of this work and the limitations of the state-of-the-art. For example, the authors talk about out-of-workflow queries, however a clear definition/example of such queries is missing. \n\nAs I state at the end of comments/questions to the authors (please see below), I am willing to increase my score if the authors address my comments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper deals with an important topic: that of controlling the responses of an LLM based on a workflow. The authors introduce a new workflow specification language for controlling the LLMs’ responses, called PDL." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "*Typos*\n\nSentence in line 014-015 difficult to understand. Please revise. \n\n*Presentation issues*\n\nIn the paragraph starting in line 042, the authors discuss the pros and cons of prompt-based and rule-based methods. Two comments: \n- References to prompt-based and rule-based methods are missing, making difficult for the readers to understand how the aforementioned approaches work.\n- A concrete example demonstrating the issues describe in the paragraph in lines 042--052 is missing (i.e., how the responses of a state-of-the-art prompt-based method and of a state-of-the-art rule-based method a state-of-the-art differ on a specific user question based for a given LLM). I recommend the authors adding such an example as that would help the readers understand concretely the limitations of previous work and appreciate the contributions of the proposed work. \n\n*Related work*\n\n- A more elaborated discussion on previous work is missing. The only discussion I found is in Section 2.2. \n- Please add a discussion and conduct experiments against work on constrained-decoding using regular expressions, which I find it very relevant to this work:\n(1) “Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning” by Sabio Geng et al. \n(2) “Validating Large Language Models with Relm” by Michael Kuchnik et al. \n\n*Technical questions*\n\n- How PDL differs from other workflow specification languages, not necessarily designed to support LLMs – some references for the authors to consider: \n(1) Serge Abiteboul, Pierre Bourhis, Victor Vianu. Comparing workflow specification languages: A matter of views. ACM Transactions on Database Systems, 2012, 37 (10).\n(2) W.M.P. van der Aalst and A.H.M. ter Hofstede .YAWL: yet another workflow language. Information Systems. Volume 30, Issue 4, June 2005, Pages 245-275.\n- Following up from the previous question: why cannot we use/extend existing workflow specification languages to constrain the responses of LLMs? \n- In line 050 the author state: “for an existing rule-based workflow, if we want it to support a new demand outside the original procedure, such as helping a user check the weather (the yellow diamond in the figure), significant modifications to the original workflow are required, which becomes impractical as more out-of-workflow demands are required.” How exactly PDL overcomes this crux? This should be demonstrated both via a (toy) example and via experiments. \n- The authors say that their framework can support out-of-workflow queries, however a clear definition/example of such queries is missing, making it difficult to assess the importance of the proposed work. \n\n*Overall assessment*\n\nA clear positioning of the contributions against the state-of-the-art is missing, making it difficult to understand the novelty of this research. I am willing to increase my score if the authors address my comments/questions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024flowagent,\ntitle={FlowAgent: a New Paradigm for Workflow Agent},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w1MEIGDepc},\nnote={under review}\n}" }, "abstract": { "value": "Combining workflows with large language models (LLMs) allows LLMs to follow specific procedures, thereby extending their application to more real-world scenarios. However, incorporating workflows often compromises the flexibility of LLMs. For example in the case of Task-Oriented Dialogue (TOD), workflow atomize the function of LLM while programmatically imposing restrictions on execution path making the dialogue obstructed and less flexible when facing out-of-workflow (OOW) queries. Prompt-based methods offer soft control but sometimes fail to ensure procedure compliance. This paper introduces a new agent paradigm to address this challenge. Specifically, we first propose a novel Procedure Description Language (PDL) that integrates the flexibility of natural language and the precision of code for workflow expression. Additionally, we present a comprehensive framework that enables LLM to handle OOW queries while keeping execution safe with a series of controllers for behavioral regulation. This includes pre-decision and post-decision methods, where the dependency relationships between workflow nodes are modeled as a Directed Acyclic Graph (DAG) to validate node transitions. Beyond the primary objective of compliance considered in previous work, we introduce a new approach to evaluate the agent's flexibility in OOW situations. Experiments on three datasets demonstrate that FlowAgent not only adheres well to workflows but also responds better to OOW queries, showcasing its flexibility. Furthermore, exploration on WikiHow data confirms that the PDL effectively represents broader formats of workflow, inspiring further research on workflow-based QA tasks." }, "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": [ "workflow", "LLM-based agent", "task-oriented dialog" ] }, "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/7220dce5d52d13299a743f3149941e841963d9db.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": "FlowAgent: a New Paradigm for Workflow Agent" }, "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. It would be valuable for the authors to discuss how the proposed method compares to existing works that transfer 3D Gaussian representations directly into mesh or other representations better suited to capturing surface characteristics. Many state-of-the-art approaches (such as SuGaR) can transform Gaussian splats into mesh or similarly structured representations, which often enhances geometric detail and surface fidelity. The manuscript would benefit from clarifying the unique advantages or improvements this hybrid representation offers compared to these transformation-based approaches. Specifically, it would be helpful to see a comparison in terms of accuracy, computational efficiency, or other aspects that highlight this method’s distinct contributions.\n2. In Figure 1, why is it difficult to ascertain the correctness or improvement in the Gaussian representation within the highlighted red box regions, as the ellipsoid visualizations appear similar between the baseline and the proposed method? Additionally, what accounts for the significant black hole in the GSSLAM representation in the second example?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The manuscript is clear and easy to understand.\n2. The results demonstrate the effectiveness of the hybrid 3D representation with minimal additional time overhead." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce a novel SLAM approach that leverages a hybrid 3D representation by integrating an implicit, multi-resolution TSDF hash grid with explicit Gaussian primitives. This combination enriches the traditional Gaussian-based SLAM pipeline by enhancing mapping performance, which, in turn, improves pose accuracy. Experimental results demonstrate the method's effectiveness, highlighting its potential in advancing SLAM performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors lack a deep understanding of basic concepts such as Gaussian representations, mapping, and localization, which results in superficial and imprecise descriptions throughout the manuscript, particularly in the Introduction. This lack of rigor affects the overall clarity and logical flow of the paper.\n\n1. The authors consistently use the term “Gaussian ellipsoid” throughout the manuscript, which lacks rigor and accuracy. A Gaussian function represents a probabilistic distribution, characterizing the influence or effectiveness of a primitive within its surrounding region through a soft representation anchored at the primitive center. Using the term \"ellipsoid\" to describe a Gaussian primitive is misleading and stems from a simplification of Gaussian visualizations. I recommend revising this terminology to more precisely reflect the Gaussian function's role in soft spatial representation, rather than suggesting a fixed, ellipsoidal form.\n2. In lines 37–39, the authors suggest that Gaussian Splatting (GS) is suboptimal for representing 3D geometric structures, an assertion that is intuitively understandable. However, the manuscript would benefit from a clearer explanation of how this limitation affects pose estimation. Specifically, the impact of GS on pose accuracy seems more indirect compared to its effect on mapping. I recommend that the authors conduct an in-depth analysis, supported by experiments, to examine how the hybrid map representation directly enhances mapping accuracy and, by extension, indirectly improves localization performance. A comparison of mapping and localization outcomes with and without the hybrid representation would offer valuable insights.\n3. The authors currently assess mapping performance using depth error alone, which provides only a limited perspective on mapping improvements. Given the hybrid 3D representation proposed in this work, depth error alone does not sufficiently capture the contribution of the TSDF representation to mapping accuracy. A more comprehensive evaluation metric, such as mesh accuracy or 3D reconstruction quality, would more effectively demonstrate the mapping enhancements introduced by the TSDF component. Including these metrics would provide a clearer validation of the hybrid approach’s impact on spatial fidelity and mapping robustness. \n4. In lines 46–47, using submaps is one approach to address the large number of Gaussian splats, but not the only one; the statement should be revised for accuracy.\n5. The ablation study should include mapping performance and present additional 3D visualizations to better illustrate the effects of the proposed method.\n6. The manuscript lacks the visualization for the proposed TSDF representation." }, "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": "Why is memory usage lower than Co-SLAM? This method requires both 3D Gaussian and hash-grid encoding, which would suggest a higher memory requirement." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Hybrid map representation combining implicit and explicit methods.\n- Competitive benchmark performance in tracking, mapping, memory consumption, and runtime analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an RGB-D SLAM approach that uses a hybrid representation combining implicit TSDF and explicit 3D Gaussians. The method leverages SDF as a smoothed geometry estimator to initialize Gaussians. It demonstrates competitive results on benchmark datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Limited technical novelty. The method is a straightforward integration of existing components; MonoGS (3DGS system), Co-SLAM (hashgrid), and ESLAM (TSDF supervision).\n- Integrating grid and point representations in a simple way can weaken the unique strengths of each approach. Point-based representations, for example, allow flexible, on-demand resource allocation without the need for predefined scene boundaries or resolutions. However, introducing a neural field can reduce this flexibility.\n- Despite these limitations in novelty, the method offers only a marginal insight—namely, that smoothed geometry provides better initialization of Gaussian locations. The method is an engineering case study rather than a technical contribution to ML community." }, "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": "According to Figure 2 in SAD-GS [1], directly using render depth loss may fail to provide correct surface reconstruction, I wonder if the author can do some ablation study to determine whether the two depth losses in your formula (9) provide positive optimization effects.\n\n[1] Kung, Pou-Chun, et al. \"SAD-GS: Shape-aligned Depth-supervised Gaussian Splatting.\" CVPRW. 2024." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper conducts extensive experiments and shows that their method achieves state-of-the-art tracking and mapping accuracy with high efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper combines the implicit and explicit representation by introducing Truncated Signed Distance Function (TSDF) hash grid to constrain the distribution of Gaussian ellipsoids. This methodology enhances both the quality of the reconstruction and the accuracy of the tracking process." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) I appreciate the authors' integration of TSDF with a 3DGS-based SLAM framework. However, this approach feels somewhat incremental, given that TSDF has already been employed in NeRF-based SLAM. It would be valuable to include an analysis comparing the advantages of this method against modified 3DGS techniques [1].\n\n2) A notable limitation of 3DGS-based SLAM methods is the efficient detection of loop closures, which remains unaddressed in this paper. While the paper emphasizes improvements in tracking, it does not provide visual results demonstrating loop closure. Although I do not expect this method to surpass established techniques such as photo-SLAM or other traditional SLAM-based modules, omitting loop closure limits the scope of innovation and reduces the paper's overall impact.\n\n3) Regarding Figures 4 and 6, it would be beneficial to include emphasized or highlighted areas, as shown in Figure 1. The current figures do not demonstrate advantages in rendering quality or geometric accuracy over other 3DGS-based methods.\n\n[1] Lu, Tao, et al. \"Scaffold-gs: Structured 3d Gaussians for view-adaptive rendering.\" CVPR. 2024." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a dense RGB-D SLAM system that leverages an implicit TSDF hash grid to constrain the distribution of Gaussian ellipsoids." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024geometrically,\ntitle={Geometrically Constrained Gaussian Splatting {SLAM}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w1Pwcx5hPp},\nnote={under review}\n}" }, "abstract": { "value": "3D Gaussian Splatting (3DGS) has emerged as a promising technique in SLAM due to its rapid and high-quality rendering capabilities. However, its reliance on discrete Gaussian ellipsoid primitives limits its effectiveness in capturing essential geometric features crucial for accurate pose estimation. To overcome this limitation, we propose a novel dense RGB-D SLAM system that integrates an implicit Truncated Signed Distance Function (TSDF) hash grid to constrain the distribution of Gaussian ellipsoids. This innovative approach enables precise estimation of the scene's geometric structure by smoothing the discrete Gaussian ellipsoids and anchoring them to the scene's surface. Acting as a low-pass filter, the implicit TSDF hash grid mitigates the inductive biases inherent in traditional 3DGS methods while preserving rendering quality. Our geometrically constrained map also significantly enhances generalization capabilities for depth estimation in novel views. Extensive experiments on the Replica, ScanNet, and TUM datasets demonstrate that our system achieves state-of-the-art tracking and mapping accuracy at speeds up to 30 times faster than existing 3DGS-based systems." }, "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": [ "3DGS", "SLAM", "Robotics" ] }, "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/d60f4924c852ff7bbbe645b9e299090552418bc1.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": null, "title": { "value": "Geometrically Constrained Gaussian Splatting SLAM" }, "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": "Please refer to the Weaknesses section." }, "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-structured, making the methodology and findings easy to understand.\n- The paper provides thorough experimental evaluations, including ablation studies on key components of the I-Max framework." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the I-Max framework, designed to address resolution extrapolation in text-to-image models using Rectified Flow Transformers. This paper incorporates a Projected Flow for fidelity and an inference toolkit to enhance model generalization at extrapolated resolutions. Experiments on the Lumina-Next-2K and Flux.1-dev models demonstrate that I-Max effectively improves the stability and detail of extrapolated high-resolution images, showing its potential for practical applications where tuning-free resolution extrapolation is needed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The evaluation relies heavily on a single metric (GPT-4 preference) to assess the quality of generated images, limiting the objectivity of the results. This reliance may affect the demonstration of the proposed method’s effectiveness.\nIf evaluating high-resolution images with widely used metrics (e.g., FID) is challenging, as noted in the manuscript, a toy experiment using a pretrained model on a lower-resolution dataset, such as CIFAR or ImageNet, could offer a feasible benchmark and enable standardized metric comparisons.\n- In line 254, the paper mentions the use of a low-pass filter for projection but does not specify the type. Additionally, exploring the impact of different low-pass filters could offer insights into how they affect stability and quality during resolution extrapolation." }, "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": "Why is the proposed method not shown on only Lumina-Next instead of the self-trained Lumina-Next-2K? Does the method require native resolution also to be high? Can the proposed I-Max work for low native resolutions?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Attempting to extrapolate resolution of trained models is interesting and practically useful given the issues of data quality and fine tuning costs. This paper addresses the task of resolution extrapolation for trained Rectified Flow Transformers for the first time.\n\nRelevant prior works have been discussed appropriately." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the task of tuning-free resolution extrapolation for text-to-image Rectified Flow Transformers (RFTs) from which one can obtain samples at a much higher resolution than the resolution at which the model was originally trained. While directly training high-resolution generative models is practically difficult, this paper aims to adapt trained RFTs to generate images of high resolutions (such as 4096X4096) without the need for fine tuning.\n\nThe proposed scheme named I-Max involves low-resolution guidance named projected flow. Here, the low-resolution\nspace is treated as a low-dimensional projection of the high-resolution space, and thereby the low-resolution\nflow can be regarded as the projection of the ideal high-resolution flow. Considering the linear\ninterpolation characteristic of rectified flow, I-Max incorporates guidance in the projected space at each timestep.\nAdditionally I-max incorporates inference techniques tailored for RFT to enhance the model’s ability to generalize to extrapolated resolution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The guidance mechanism in eq 7 does not exactly correspond to the Classifier Free Guidance. It is indeed some sort of a guidance function. Could the authors explain the relationship between their guidance mechanism and Classifier Free Guidance?\n\n Additionally, it is not clear how the first term in the RHS of eqns 6 and 7 (v_{theta} at the extra resolution) is obtained. Is the same model trained at the native resolution used for this?\n\n\nThe steps followed to generate the high resolution image could have been summarized in the form of an algorithm.\n\n\n\nEvaluation is based only on GPT-4o. More qualitative examples wherein one can see improvements as shown in Fig 2 could have been shown in the supplementary material. Since GPT-4o is not necessarily trained for image quality assessment, other measures should have been used for comparison. Time aware scaled ROPE (Fig 7) also has good performance according to this measure.\n\nTypo 'for butter efficiency' line 419.\n\n\nSome of the ideas incorporated are based on existing works. Specifically, the inference techniques in section 2.3 are based on prior works. \nCould the authors clarify their novel contributions in the techniques used in section 2.3?" }, "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": 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 like the authors to clarify the following points:\n1) Provide visual examples of failures at very high resolution (as pointed out in sec 3.3 for Figure 6);\n2) Why is the projected flow implemented via the classifier-free guidance? Is this the only way?\n3) Could you show how you would explain the transition to eq. (5) with more technical details?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I-Max integrates a number of simple but important components to make a rectified flow model generalise to higher resolutions at inference. In particular, the projected flow strategy makes sense as a method to ensure more stability. As far as I know this is original and the specific implementation in the style of a classifier-free guidance seems original too. \nThe results achieved in the experimental section show also that the proposed projection with the other inference techniques are quite effective in the resolution extrapolation task." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a method to extrapolate the resolution of the generated images at inference time. The authors focus on rectified flow transformers. The key ideas are a projected flow strategy that is designed to ensure more stability at inference, and a number of implementation techniques to enhance the quality of the extrapolation, such as NTK-aware scaled RoPE, SNR resolution adjustment, attention re-scaling, and text duplication." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The presentation is at times not optimal.\nFor example, the split in the introduction into How to guide and How to infer does not seem very clear to me. At lines 93-95 the explanations do not seem to match the names of the two perspectives.\nOverall, the use of the English language could be better. I would suggest to have the paper revised but a native English speaker to correct typos.\nCould you check the following?\nLine 220: Eq. 2 illustrates the equivariance of the flow wrt the projection rather than its invariance.\n\nThe other concern is regarding the method (see also the Questions below). It would be useful to the reader to better explain the technical choices by providing the motivation/rationale behind each of 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": 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": "As shown in Weaknesses" }, "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-structured.\n- The proposed method achieves excellent visual results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes the I-Max, designed to maximize the resolution potential of Text-to-Image Rectified Flow Transformers (RFTs). I-Max includes a novel Projected Flow strategy and an advanced inference toolkit, enhancing generative stability, improving image detail, and correcting artifacts during resolution extrapolation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Lack of Quantitative Comparisons:**\n The paper lacks any quantitative comparisons, making it difficult to demonstrate the superiority of the proposed method. Metrics such as FID (Fréchet Inception Distance) and IS (Inception Score) could be used to provide concrete quantitative comparisons.\n\n2. **Need for User Study:**\n A user study is necessary to validate the visual effectiveness of the method. This study should focus on aspects such as detail preservation, artifact reduction, and overall image quality of the generated images, which would further enhance the quality of the paper.\n\n3. **Comparison of Model Parameters and Runtime:**\n The paper should include comparisons of model parameters and runtime to provide a comprehensive picture of the method’s efficiency. Reporting the generation time at different resolutions, such as 1K and 2K, is crucial for understanding the practical applicability and efficiency of the proposed framework." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024imax,\ntitle={I-Max: Maximize the Resolution Potential of Pre-trained Rectified Flow Transformers with Projected Flow},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w2BELPYbU0},\nnote={under review}\n}" }, "abstract": { "value": "Rectified Flow Transformers (RFTs) offer superior training and inference efficiency, making them likely the most viable direction for scaling up diffusion models. However, progress in generation resolution has been relatively slow due to data quality and training costs. Tuning-free resolution extrapolation presents an alternative, but current methods often reduce generative stability, limiting practical application. In this paper, we review existing resolution extrapolation methods and introduce the I-Max framework to maximize the resolution potential of Text-to-Image RFTs. I-Max features: (i) a novel Projected Flow strategy for stable extrapolation and (ii) an advanced inference toolkit for generalizing model knowledge to higher resolutions. Experiments with Lumina-Next-2K and Flux.1-dev demonstrate I-Max's ability to enhance stability in resolution extrapolation and show that it can bring image detail emergence and artifact correction, confirming the practical value of tuning-free resolution extrapolation." }, "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 Model", "Generative Model", "Image Generation", "High-resolution" ] }, "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/3592b94351ee4ceb19900a8fe892b89f1eaab075.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": "I-Max: Maximize the Resolution Potential of Pre-trained Rectified Flow Transformers with Projected Flow" }, "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": 1 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "need to be checked for LLM-generated content." }, "flag_for_ethics_review": { "value": [ "Yes, Other reasons (please specify below)" ] }, "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": "-" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "-" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "I believe this paper either has a high chance of being generated by a language model, or was rushed into submission. In both cases, I recommend a rejection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**I stopped reviewing this paper, as there were multiple issues. The objective and the problem setting are not clear, assumptions are not clearly stated, related work are missing, and writing is very poor.**\n\nThe definition of a \"system\" in the first place is not clear. The authors casually drop \"properties\" of this said system/multi-system, which vaguely correspond to some concepts, such as Nash equilibria, but not clear at all if those are assumptions/setting the authors consider. \n\nThe paper is almost impossible to read, with abuse of notation used absolutely for no reason. It is not really clear what problem the paper is dealing with and it is very poorly formulated throughout.\n\nIt is not clear how the COVID example fits into the Constraints 1 & 2. Why Eq. (4) needs to hold for the COVID case for instance?\n\nRelated work section mentions mostly datasets and equilibrium examples from different domains, but I could not spot any ML related work that consider a similar problem. \n\nWhy do you define ${\\cal M} {\\cal S}$ to be the sum of the target variables in systems, whereas it was first defined as the set of systems? The notation should be improved across the board, but this is just one example.\n\nWhy does Eq. (1) hold *in general*? Is that a setting you consider?\n\nYou cannot refer to Constraint 1 when defining Constraint 1 itself?\n\nLine 158 - Please refer to where did you describe your ESE method as claimed.\n\nWhat is the takeaway from Figure 1 exactly?\n\nThe connection to the Nash equilibrium is never formally introduced or motivated. It is not clear where Lemma 1 comes from. There are no proper citations as well." }, "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. Convergence Mechanism: Since Nash equilibrium and the zero-sum constraint do not inherently ensure convergence to equilibrium, does the model include any dynamic process to guide the system toward equilibrium from any initial state? Would the authors consider adding time-based dynamics or stability conditions to ensure convergence?\n2. Exclusion of Oscillatory and Chaotic Behavior: In traditional multi-compartment models, stability conditions can be defined to ensure convergence to equilibrium. However, the equilibrium condition in this paper does not seem to exclude stable oscillatory or chaotic behavior. Could the authors clarify if such behaviors are possible in their model? If so, how might these be addressed to ensure the system reaches a steady-state equilibrium?\n3. Role of Cointegration: Cointegration is typically used to identify statistical relationships between non-stationary variables, assuming they share a long-term equilibrium path. Is cointegration in this model intended merely as a statistical tool for parameter estimation, or is it expected to actively drive the system toward equilibrium? Could the authors clarify how cointegration contributes to achieving or maintaining equilibrium in the system?\n4. Distinction from Multi-Compartment Models: The ESE model shows strong similarities to multi-compartment models, which also use conservation principles to maintain balance. Beyond the conceptual framing of Nash equilibrium and payoff functions, is there a fundamental structural difference that distinguishes ESE from these existing models? Specifically, does ESE offer any new insights or capabilities that go beyond what is possible with traditional multi-compartment approaches?\n5. Computational Complexity: The paper claims linear computational complexity but does not provide a formal proof. Could the authors offer a theoretical complexity analysis to substantiate this claim? Additionally, how does the model handle scenarios with extensive interactions or feedback loops between subsystems, which might increase computational demands?\n6. Applicability to Non-Stationary or Oscillatory Systems: If the system exhibits non-stationary or oscillatory behavior, how reliable is the equilibrium index (EI) as a measure of stability? Would the model need adjustments to handle such cases, or does it implicitly assume stationarity and convergence to a single steady state?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The primary strength of this paper lies in its effort to apply a novel conceptual framework for modeling interconnected systems. It draws on ideas from Nash equilibrium, zero-sum constraints, and cointegration. This approach offers an interesting perspective by reinterpreting multi-system predictions through game-theoretic concepts, particularly utilizing Nash equilibrium principles and payoff functions to illustrate the mutual influences among subsystems. While these concepts are mainly interpretative, they provide a fresh way to think about multi-system interactions, which could have significant implications for interdisciplinary applications across fields such as economics, epidemiology, and regional forecasting. The paper demonstrates the practical feasibility of the proposed model, ESE, by applying it to a real-world COVID-19 dataset. This highlights the model's potential to capture complex interdependencies and offers insights into multi-region transmission dynamics. The presentation is generally clear, with structured explanations of equilibrium concepts and the role of each component, such as the equilibrium index. Furthermore, if the model’s empirical performance in terms of accuracy and computational efficiency is further substantiated, it suggests that ESE could serve as a competitive alternative for high-dimensional applications where subsystems are highly interconnected." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a new model called Equilibrium State Evaluation (ESE) for predicting the collective behavior of interconnected multi-system structures. ESE is designed to manage systems with multiple interdependent subsystems, each affecting and being affected by the others. The goal of the model is to determine an equilibrium state for the entire multi-system by applying concepts from Nash equilibrium along with a zero-sum constraint.\nIn ESE, each subsystem is given a payoff function that assesses its \"benefit\" based on the states of all the subsystems. The equilibrium state is defined as the configuration in which each subsystem achieves its optimal state (maximum payoff), given the conditions of the others, consistent with Nash equilibrium principles. The zero-sum constraint enforces a balance across subsystems, ensuring that any gain in one subsystem is countered by a loss in another, thereby maintaining overall stability.\n\nTo estimate the long-term equilibrium trends of the system, the model utilizes cointegration, an econometric technique that identifies shared equilibrium paths among interdependent variables. Additionally, an equilibrium index (EI) is introduced to quantify how closely the system's current state aligns with equilibrium, providing a measure of system stability.\n\nThe paper validates the ESE model using multi-regional COVID-19 transmission data, demonstrating that it achieves competitive prediction accuracy while maintaining linear computational complexity. The authors contend that ESE's equilibrium-based approach enables scalable and efficient prediction of complex multi-system dynamics, making it potentially valuable in high-dimensional environments where subsystems are closely interconnected." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "A significant weakness of the paper lies in its insufficient support for the claim that the system will reach a true equilibrium state. While the model employs Nash equilibrium and zero-sum constraints to define static equilibrium conditions, these concepts alone do not provide a mechanism to ensure that the system will naturally progress toward equilibrium over time. Without a dynamic framework or time-dependent interactions—such as differential equations or explicit stability conditions—there is no mathematical basis for assuming that the system will move from an arbitrary initial state toward equilibrium. Furthermore, although the paper incorporates cointegration in its training process, this is a statistical technique that assumes the existence of a long-term equilibrium relationship rather than actively guiding the system toward it. This distinction weakens the model’s theoretical foundation, as it does not establish how equilibrium is achieved dynamically. Another critical limitation is that the equilibrium definition based on Nash equilibrium and payoff functions does not exclude the possibility of stable oscillatory behavior. This means the system could theoretically settle into a stable oscillation rather than converging to a true steady state. Additionally, the ESE model shows strong mathematical similarities to traditional multi-compartment models, which use conservation principles to maintain balance across subsystems. While the introduction of game-theoretic concepts like Nash equilibrium and payoff functions offers a fresh interpretive layer, it does not constitute a substantial mathematical advancement over established multi-compartment models. To improve the model’s robustness, the authors would need to provide a formal analysis of convergence conditions, explicitly address the exclusion of oscillatory states, and better differentiate their approach from conventional multi-compartment modeling frameworks." }, "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": "The authors could provide a more detailed algorithm description and better justification of experimental and methodological design" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "There aren't any significant strengths in the paper" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new prediction approach called Equilibrium State Evaluation for handling multi-system prediction problems where multiple interacting systems must be predicted simultaneously. Where the authors treat each system as independent time series and consider multiple-interacting systems as components of a \"super-system\". The method maintains an Equilibrium Index (EI) that measures the distance between the current state and the equilibrium state." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper is confusing to follow. The introduction is poorly written. The mathematical notation is very dense without sufficient explanation. \n- Complex method workflow is not fully explained\n- It is not apparent to me why proportions must follow zero-sum rule. Doesn't explain how to handles growing-shrinking total system values.\n- It is unclear why is specific normalization is chosen λᵢ,ⱼ = (αᵢ,ⱼ - mean(atⱼ)) / (max(atⱼ) - min(atⱼ)), why is this normalization beneficial as compared to other ways to normalize? \n- Algorithm 1 is vaguely described, and testing and training process aren't well defined. It doesn't map succinctly with the figure 2 of the paper." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A new method that can predict for multiple systems in just one run." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024integrated,\ntitle={Integrated Multi-system Prediction via Equilibrium State Evaluation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w2C7gJqaai},\nnote={under review}\n}" }, "abstract": { "value": "This study presents a new paradigm of prediction, Equilibrium State Evaluation (ESE), which excels in multi-system prediction where systems interact with each other and every system needs its own prediction. Unlike mainstream prediction approaches, ESE views each system as an integral part under one structure and predicts all systems simultaneously in one go. It evaluates these systems' equilibrium state by analyzing the dynamics of their attributes in a holistic manner, instead of treating each system as an individual time series. The effectiveness of ESE is verified in synthetic and real world scenarios, in particular COVID-19 transmission, where each geographic region can be viewed as a system. So cases spreading across regions against the medical competency and demographic traits of these regions can be considered as an equilibrium problem rather than a time series problem. Extensive analysis and experiments show that ESE is linear in complexity and can be 10+ times faster than SOTA methods, yet achieving comparable or better prediction accuracy. More importantly, ESE can be integrated with these prediction methods to achieve both high accuracy and high speed, making it a powerful prediction mechanism, especially for scenarios that involve multiple systems. When the dimensionality of the multi-system increases, e.g. more systems joining, the advantages of ESE would become even more apparent." }, "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": [ "Multi-system", "Equilibrium", "Prediction" ] }, "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/74a739790d9441f398e64899642b7ddccda314e2.pdf" }, "presentation": null, "primary_area": { "value": "learning on time series and dynamical systems" }, "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/4364b94f15b5060a6fd4d6c1262cb7a3545e330b.zip" }, "title": { "value": "Integrated Multi-system Prediction via Equilibrium State 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": 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": 3 }, "primary_area": null, "questions": { "value": "All my questions have been included in the weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper is well-structured and easy to follow.\n- The proposed method is well-motivated and novel. \n- The details of the experiments are revealed, and the code is released, which will ease the reproducibility of this paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed method aims to tackle overlooking the missing data within the guidance for diffusion models. Specifically, they first detect the missing data by constructing local and global models and checking the continuity and stability. Based on the detected missing data, the authors then train diffusion models with uncertainty-aware guidance. Extensive results are provided to support the effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some typos exist. For example, the first character in line 72 should be capitalized.\n- This paper is not well-motivated. Why is Thompson sampling the best choice to derive the guide signal for a diffusion-based recommendation? Besides, no related experiments can verify the best of the Thompson sampling strategy compared to other sampling strategies.\n- I noticed that the scale of datasets used in the experiments is relatively small, with interaction counts of fewer than one million. I recommend that the authors conduct experiments on larger datasets to further demonstrate the effectiveness of the proposed method, or the method's efficiency may be questioned." }, "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 refer to the Weaknesses." }, "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 generally well-written.\nThe problem of missing data is crucial for developing sequential recommenders." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes DreamMiss, a novel approach to handle missing data in sequential recommendation systems using denoising diffusion models (DDMs). The key innovation is a dual-side Thompson sampling (DTS) strategy that simulates missing data stochastically while preserving user preference patterns. The approach uses two probability models: 1) A local model that captures continuity between adjacent items A global model that evaluates sequence stability. 2) DreamMiss is implemented using denoising diffusion implicit models (DDIM) for faster sampling and achieves consistency regularization through uncertainty-aware guidance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. Conceptual Clarity and Methodology Concerns.\nThe paper lacks clear theoretical justification for how generating additional missing data helps address the original missing data problem.\nThe approach of simulating missing data from already incomplete sequences raises questions about potential error propagation.\nThe methodology appears counterintuitive compared to traditional approaches that aim to recover or compensate for missing data.\nThe paper would benefit from a more rigorous theoretical analysis of why this approach is superior to data completion methods.\n\nW2. Limited Validation of Dual-side Thompson Sampling (DTS).\nThe paper does not sufficiently justify why DTS is specifically effective for diffusion-based sequential recommenders.\nThere is inadequate theoretical analysis or empirical validation of the reliability of the continuity and stability metrics.\nThe generalizability of DTS to other recommendation architectures needs more thorough investigation.\nThe robustness of the probability models used in DTS requires more comprehensive validation.\n\nW3. Incomplete Baseline Comparisons.\nNotable omissions of important state-of-the-art baselines, particularly:\n[1] Lin, Y.; Wang, C.; Chen, Z.; Ren, Z.; Xin, X.; Yan, Q.; de Rijke, M.; Cheng, X.; and Ren, P. A Self-Correcting Sequential Recommender. In TheWebConf 2023.\n[2] Zhang, C.; Han, Q.; Chen, R.; Zhao, X.; Tang, P.; and Song, H. SSDRec: Self-Augmented Sequence Denoising\nfor Sequential Recommendation. In ICDE 2024.\nThese omissions make it difficult to fully assess the comparative advantages of the proposed method.\nThe evaluation would be more convincing with a more comprehensive comparison against recent approaches.\n\nW4. Scalability Limitations.\nThe computational complexity of the proposed method may limit its practical applications.\nInsufficient discussion of performance on large-scale recommendation systems.\nLimited analysis of computational resource requirements for real-world deployment.\nNeed for more detailed discussion of potential optimization strategies for larger datasets.\n\nW5. Dataset Limitations.\nThe evaluation relies on relatively small-scale datasets.\nQuestions about generalizability to larger, more complex real-world recommendation scenarios.\nNeed for validation on more diverse and larger-scale datasets.\nLimited demonstration of effectiveness across different domains and data distributions." }, "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": 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": "Sorry, I am not familiar with this research area so I can't give a valuable score. Please overlook my score." }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "Sorry, I am not familiar with this research area so I can't give a valuable score. Please overlook my score." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Sorry, I am not familiar with this research area so I can't give a valuable score. Please overlook my score." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Sorry, I am not familiar with this research area so I can't give a valuable score. Please overlook my score." }, "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. The line 7 of Algorithm 1 seems a bit different than other DDPM models. There's no square root above the $(1 - \\alpha_{\\tau_s})$. Is it a typo?\n2. Can the authors explain why accelerated sampling works? It seems to reduce thousands of iterations to less than 100 rounds, which is a huge improvement.\n3. Based on the performance of Dreasmiss, can we say the DreamRec has a overfitting problem? With much more rounds of iteration DreamRec has inferior performance." }, "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 are sound.\n2. The paper is overall well written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a dual-side Thompson sampling (DTS) strategy to simulate the stochastical mechanism of missing data, and integrates it into denoising diffusion models for sequential recommendation. Extensive experimental results show that DreamMiss significantly outperforms baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The missing data issue is not well discussed. Thompson sampling can not tackle all the missing data issues. How the proposed model can address which kind of missing data should be discussed in the paper.\n2. The rationale of accelerated sampling is not well explained.\n2. The hypothesis of stability scores is not reasonable. In recommendation model training, data are often shuffled to remove the dependency among samples. However, the stability scores seems the opposite way; it used the batch information for sampling. It contradicts with the traditional way." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a dual-side Thompson sampling strategy to create uncertainty-aware guidance for DDMs in sequential recommendation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024uncertaintyaware,\ntitle={Uncertainty-aware Guided Diffusion for Missing Data in Sequential Recommendation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w2HL7yuWE2},\nnote={under review}\n}" }, "abstract": { "value": "Denoising diffusion models (DDMs) have shown significant potential in generating oracle items that best match user preference with guidance from user historical interaction sequences. However, the quality of guidance is often compromised by the unpredictable missing data in the observed sequence, leading to suboptimal item generation. To tackle this challenge, we propose a novel uncertainty-aware guided diffusion model (DreamMiss) to alleviate the influence of missing data. The core of DreamMiss is the utilization of a dual-side Thompson sampling (DTS) strategy, which simulates the stochastical mechanism of missing data without disrupting preference evolution. Specifically, we first define dual-side probability models to capture user preference evolution, taking into account both local item continuity and global sequence stability. We then strategically remove items based on these two models with DTS, creating uncertainty-aware guidance for DDMs to generate oracle items. This can achieve DDMs’ consistency regularization, enabling them to resile against uncertain missing data. Additionally, to accelerate sampling in the reverse process, DreamMiss is implemented under the framework of denoising diffusion implicit models (DDIM). Extensive experimental results show that DreamMiss significantly outperforms baselines in sequential recommendation." }, "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 Models", "Recommender Systems", "Missing Data" ] }, "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/513988bc657b77e310c06eac224339d258bc3a9c.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": "Uncertainty-aware Guided Diffusion for Missing Data in Sequential Recommendation" }, "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": "Please refer to the weakness." }, "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 iseasy to follow and well-written.\n(2) The proposed approach of using HOI to filter exocentric videos for egocentric learning is compelling, meanwhile the exo-to-ego rephraser and ego narrator prove effective. The paper provides extensive experiments across diverse datasets and benchmarks, offering strong support for its methodology." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents EMBED (Egocentric Models Built with Exocentric Data), a novel approach for adapting exocentric video-language data to improve egocentric video representation learning. Although exocentric data is rich and varied, it lacks the close-up hand-object interactions and narratives central to egocentric perspectives. EMBED bridges this gap by identifying specific video clips that emphasize hand-object interactions and pairing them with action-focused language narrations. Experiments show EMBED's effectiveness, achieving state-of-the-art performance with a 4.7% improvement on Epic-Kitchens-100 multi-instance retrieval and a 6.2% increase on EGTEA classification in zero-shot settings. Furthermore, EMBED enhances egocentric model performance on exocentric tasks and demonstrates strong generalization across diverse exocentric datasets, showcasing its potential to unify egocentric and exocentric video learning and capitalize on the unique strengths of exocentric data for egocentric applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The technical contribution of this work appears limited, as it primarily proposes a data filtering process for exocentric videos, utilizing pretrained LLMs to refine HowTo100 captions into an egocentric format and to generate additional egocentric narrations—an approach already widely explored in previous research. These methods seem more suited for the Ego4D workshop and competition, raising questions about their suitability for an ICLR submission.\n\nAnother key concern is the reliance on egocentric retrieval and classification benchmarks, which have been heavily used in recent years. It’s widely acknowledged that improvements on these benchmarks may not necessarily indicate a truly understanding of egocentric content or applicability in real-world scenarios. If the authors could provide additional experiments demonstrating consistent improvements on more challenging tasks, such as grounding or manipulation tasks, I would be inclined to reconsider my 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": 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 HT100M consists a large part of instruction video during data collection. So it is natural that this work can select suitable clip. If possible this work also works in other datasets like K600, Moments In Time etc.?\n\n2. The HTM-AA dataset is 247K videos, how about the samples number of HTM-EMBED in Tab.4 ?\n\n3. Even though the EMBED data selection need to run only once for multiple experiments. Could you add the time comparison for LaViLa-B with LaViLa-B+EMBED." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The motivation is clear: there remains a challenge in accessing high-quality ego-centric video data.\n\n2. LaViLa-B demonstrates a clear improvement over both EgoVLP and EgoVLP V2 on the EK-100 MIR and EgoMCQ tasks.\n\n3. The approach is straightforward and well-presented, making it easy to understand and practical for real-world applications.\n\n4. Unlike previous ego-centric models, this paper proposes evaluating the model in a human-object interaction (HOI) setting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Large-scale web-based datasets are commonly available; however, high-quality, ego-centric video data remains challenging to obtain. While datasets like EGO4D have been introduced, their scale still falls short in meeting the demands of scaling laws in Multi-modal Large Language Models (MLLMs). \n\nIn this work, they propose a method to extract ego-centric clips from web videos, primarily leveraging the HowTo100M dataset, which contains a substantial collection of instructional videos. This approach serves as an efficient way to retrieve valuable ego-centric clips from existing large-scale video data, expanding the resource pool for ego-centric research. Beyond common tasks explored in related ego-centric works, the author also conduct a series of experiments on HOI set and show inspiring improvement." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The primary concern is the comparison and discussion of the advantages over retrieval-augmented methods, such as the Neural Data Server series. The critical aspect of EMBED lies in its need to access original data, denoted as X^{exo} in this work.\n\n2. EMBED requires multiple offline models and must iterate through all candidates in a large corpus, such as HT100M in this work, making it time-intensive. Additionally, LANGUAGE NARRATION GENERATION relies on off-the-shelf LLMs, like LLAMA2. These steps involve substantial engineering efforts in data filtering, rather than being directly learned or optimized as part of the core model training.\n\n3. From Tab4 we observe original noisy HT100M also boosts the performance a lot. This shows the improvement bring by EMBED is limited on UCF." }, "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": "See weaknesses above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.This method has achieved a significant performance improvement across several tasks.\n2.The authors conduct thorough ablation experiments.\n3.The authors propose a framework that utilizes a large amount of third-person video data to assist in understanding egocentric video data, which is significant for advancing first-person video understanding." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents EMBED, a framework for enhancing egocentric video representation learning using exocentric video-language data. The method involves: 1.Video Clip Curation: Selecting video clips from exocentric datasets that emphasize hand-object interactions and refining them spatially to focus on these regions. 3.Language Narration Generation: Using an exo-to-ego rephraser to transform existing narrations and an ego narrator to generate new ones in egocentric style. The experimental results indicate that this method achieves performance improvements across egocentric tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It is unclear whether the authors intend to release their models and datasets.\n2. There have already been some works on refining annotations using LLMs and leveraging exocentric videos to assist with egocentric videos, making the entire framework not novel.\n3. How is the quality of annotations on exocentric videos using an HOI detector and how do the authors handle erroneous data?" }, "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": "- **About Selecting Visually Aligned Sentences:** Why is the selection conducted by classifying the text alone? Wouldn't the mismatch between video and sentences be determined by both the video and the text? It seems that the classification focuses on distinguishing HOI from non-HOI sentences. Please correct me if I’m wrong.\n- **Limited Improvement with SSv2:** Despite SSv2 being more similar to egocentric visual content, the improvement is even smaller than with the noisier K700 dataset. I'm curious about the underlying reasons for this.\n- **Obvious Typo in Abstract, Line 20-21**: \"\"..we construct datasets thar are...\"\", thar-->that." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- **Comprehensive Experimental Evaluation of EMBED**: I appreciate the authors' efforts in conducting diverse and thorough evaluations. They primarily validate their exocentric video-text pairs from the HTM-AA dataset across a wide range of commonly used egocentric benchmarks, as well as simpler exocentric benchmarks like HMDB and UCF101. They also attempt to demonstrate the effectiveness of EMBED using other third-person datasets like Kinetics-700, though this yields only minor improvements.\n- **Promising Experimental Results**: EMBED improves LaViLa-B and LaViLa-L when continue to pretrain on Ego4D and HTM, offering a promising approach for egocentric video understanding by using filtered exocentric data (particularly HTM-AA) that highlights hand-object interactions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes EMBED, a data generation framework that leverages exocentric datasets for egocentric video-language pretraining. Unlike previous works that primarily learn from paired ego-exo data or unpaired full-scale exocentric videos (e.g., Ego-Exo, EgoExoNCE), EMBED focuses on learning HOI-related representations from zoomed-in, HOI-specific exocentric videos and egocentric-style text descriptions. The authors conduct comprehensive experiments across various egocentric and some simple exocentric benchmarks that demonstrates promising results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Limited Technical Contribution:** The tools (HOI Detector, Narrator), data (Howto100M), and architecture (LaViLa) are mostly well-known in the egocentric domain. From prior works like Ego-Exo[CVPR2021], EgoVideo [CVPR 2024 EgoVis challenge champions] and EgoExoNCE [CVPR 2024], I've seen that incorporating exocentric data can offer some benefits. The approach of extracting HOI regions for training is somewhat similar to GroupRandomSizedCrop (if the crop just happens in HOI region and zoomed-in), a useful augmentation in the egocentric domain. Additionally, rephrasing exocentric text into egocentric text is an obvious and naive way to address the text domain gap. Overall, I don’t find much novelty here.\n- **Concerns About Using the Exocentric Dataset HTM-AA:** While I appreciate the extensive experiments, the improvements seem to stem primarily from using the clean version of the HowTo100M dataset (HTM-AA). HTM-AA offers cleaner data and includes a significant amount of kitchen activities, which overlap with scenarios in Epic-Kitchens and EGTEA, likely contributing to the larger improvements in these downstream tasks. Table 5 further shows that EMBED gains little and incurs high costs when pretraining with additional noisy datasets like K700, strenghthening my concerns about the overall effectiveness of EMBED.\n- **Concerns About Initialization with LaViLa:** I have concerns of the approach to initialize with LaViLa (mentioned in Lines 356-357) rather than pretraining from scratch when fine-tuning all parameters. By initializing with LaViLa, the overall computational cost—combining the training of LaViLa on Ego4D and fine-tuning on Ego4D+HTM—becomes twice that of LaViLa alone, which raises concerns about a potentially unfair comparison in Table 1." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present a framework that constructs video-language data from exocentric sources for egocentric video representation learning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024unlocking,\ntitle={Unlocking Exocentric Video-Language Data for Egocentric Video Representation Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w2HYVwXhMh},\nnote={under review}\n}" }, "abstract": { "value": "We present EMBED (Egocentric Models Built with Exocentric Data), a framework designed to mine video-language data from exocentric sources for egocentric video representation learning. Large-scale exocentric data covers diverse activities with significant potential for egocentric learning, but inherent disparities between egocentric and exocentric data pose challenges in utilizing one view for the other seamlessly. In this study, we propose leveraging hand-object interactions and language narratives as cues to incorporate exocentric data into egocentric training. Specifically, we focus on identifying specific video clips that emphasize hand-object interactions and pairing them with action-focused language narrations. By applying our framework to exocentric datasets such as HowTo100M, we construct datasets thar are effective for egocentric video-language pretraining. Our extensive evaluations reveal that EMBED achieves state-of-the-art performance across various egocentric downstream tasks, including a 4.7\\% absolute improvement in multi-instance retrieval on the Epic-Kitchens-100 benchmark and a 6.2\\% improvement in classification on the EGTEA benchmark in zero-shot settings. Furthermore, EMBED enables egocentric video-language models to perform competitively in exocentric tasks. Finally, we showcase EMBED's application across various exocentric datasets, exhibiting strong generalization capabilities when applied to different exocentric 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": [ "video-language pretraining", "egocentric video" ] }, "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/ce70e3bc5b04a04377df6436c74a92bc319982bb.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": "Unlocking Exocentric Video-Language Data for Egocentric Video 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": 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": 3 }, "primary_area": null, "questions": { "value": "* Is the issue that models struggle with distinguishing background and foreground only present in scenarios with relatively limited data, or is this a problem for larger datasets? \n* Can this method be regarded as distilling the knowledge from SAM and diffusion models, which have been trained on large-scale datasets?\n* What is the computational complexity of the method when processing images, and is it feasible to apply it to larger datasets?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The method of dividing images into CDP and CIP regions sounds reasonable and the experiment results show its effectiveness.\n* The paper demonstrates the effectiveness of the method on different networks and datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a method to decouple an image into its class-dependent part (CDP) and class-independent part (CIP), and subsequently processes these two components individually for data augmentation. Specifically, the method first uses SAM to segment out CDP and CIP. Then, it appropriately augments the CDP part using a diffusion model. After that, it randomly combines the CDP with CIP from different images to increase image diversity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper only conducted experiments on datasets with relatively limited data. It did not illustrate the effectiveness on larger and more general datasets. Can this method work on larger datasets like ImageNet, similar to how mixup or cutmix do? \n* In the experiments comparing with RandAugment, only the results of “DE-DA” and “DE-DA + RandAugment” are provided. I believe that further results for “RandAugment only” should be included. Because it would be better to demonstrate that the “DE-DA + RandAugment” method outperforms “RandAugment only”." }, "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 uncertain about the appropriateness of using PSNR to evaluate diversity in Figure 5b. PSNR measures pixel-level deviation, meaning that a simple Mixup operation could also result in low PSNR, but it would be difficult to argue that Mixup inherently promotes high diversity. I would appreciate further clarification from the authors on this point." }, "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 a simple yet effective method that balances fidelity and diversity in synthetic data. The core idea is to decouple CIPs and CDPs using an off-the-shelf segmentor, augmenting them separately and then combining them to create new samples. This approach presents a methodological innovation compared to prior generative data augmentation techniques. \n2. De-DA utilizes layer-based composition to generate synthetic samples, making the synthesis process highly efficient. This is beneficial in resource-constrained settings,\n3. In Table 1 and Figure 2, the authors provide a comprehensive overview of various data augmentation methods from the perspectives of diversity and fidelity, which offers valuable insights for readers aiming to understand advancements in this area." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a data augmentation framework called Decoupled Data Augmentation (De-DA) to tackle the fidelity-diversity dilemma. This approach involves a decoupling strategy that separates images into class-dependent parts (CDPs) and class-independent parts (CIPs) using SAM. The method then applies text inversion and SDEdit to the CDP, and subsequently combines it with randomly selected CIPs to generate new images. This process aims to maintain semantic consistency while enhancing diversity" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.Although the Online Randomized Combination method is efficient, it may lead to semantically unnatural compositions. For instance, birds in the synthetic samples do not always appear naturally perched on branches (see the last row of birds in Figure 6, where proper positioning on branches is rare). While semantic naturalness may not always be critical for classification, this could reduce the generalizability of synthetic data.\n\n2.The authors should include more visualizations of De-DA results, especially showcasing foreground variations in CIP and the inpainting results for CDP. Additionally, it would be valuable to discuss some lower-quality samples and include a discussion on the limitations of the current approach.\n\n3.Some experimental settings are overly brief. For instance, details regarding the multi-label classification implementation are missing, and the authors provide insufficient explanation on how samples are constructed in this scenario." }, "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": "De-DA generates augmented images using SAM and diffusion models. How does it compare with other baselines in terms of efficiency? \n\nThe proposed method is evaluated only on fine-grained classification tasks. Some similar mixing methods are evaluated on object detection and instance segmentation tasks. For example, Copy-Paste augmentation [2] is a low-cost augmentation method that copies and pastes a random object into another image. How does De-DA compare with this baseline in instance segmentation tasks?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper is well-written and is easy to understand\n\n- The experimental results are promising" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes De-DA, which is a framework for sample-mixing and generative data augmentation. Specifically, it decouples the frontal object and background of an image using SAM. During training, De-DA applies generative diffusion models to transform the object and paste it to an extracted background from another image. The method helps to reduce the background noise when transforming the object and shows strong experimental results on fine-grained classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Limited Novelty. The approach of using background changes to create augmented data has been studied before. For example, InSP[1] swaps the saliency part of two images from the same class and is tested on CUB, Stanford Car, and FGVC-Aircraft datasets. Copy-paste augmentation [2] is a low-cost augmentation method that copies and pastes a random object into another image, for instance segmentation. Applying textual inversion and SDEdit to transform objects was suggested in DA-Fusion. The SDEdit and SAM models are also off-the-shelf methods proposed in previous works. The proposed idea of transforming extracted objects and pasting them into the backgrounds from other images is somewhat incremental to existing works. \n\nDe-DA proposes to extract class-dependent parts of an image using SAM. If the SAM model is not pre-trained on the target domain, it may fail to capture the class-dependent part, such as medical images. In addition, the class-dependent parts of an image are not limited to the background. For example, in gesture classification, the gesture is also a class-dependent part; in scene classification, the whole image should be considered. It seems that the proposed method cannot capture non-background class-dependent features and is limited to fine-grained object classification tasks.\n\nLimited evaluation of the tested datasets and tasks. It appears that all three tested datasets are fine-grained classification datasets. The backgrounds in fine-grained datasets are relatively easy to transfer as the objects are mostly the same. However, this may not be the case if the object classes are semantically different. For example, swapping the background of a marine animal and a desert animal may not be appropriate. Therefore, I suggest the authors test their method on large-scale datasets with more diverse objects, e.g., the ImageNet dataset. \n\n[1] Zhang, Lianbo, Shaoli Huang, and Wei Liu. \"Intra-class part swapping for fine-grained image classification.\" Proceedings of the IEEE/CVF winter conference on applications of computer vision. 2021.\n\n[2] Ghiasi, Golnaz, et al. \"Simple copy-paste is a strong data augmentation method for instance segmentation.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021." }, "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. Long-tailed image classification is a natural data-scarce scenario to test the effectiveness of data augmentation methods. I am curious about the paper's performance on long-tailed datasets like classical Places-LT& ImageNet-LT, and CUB-LT & Flower-LT used in Diff-Mix.\n\n2. Without SAM, could other instance segmentation methods achieve similar foreground-background separation? How significant is SAM's impact? I believe this is crucial for demonstrating the robustness of the proposed method. If it relies heavily on SAM, its effectiveness could be significantly diminished, as previous methods did not use additional tools like SAM for foreground-background separation. I am curious to see if the method is still effective under poor separation conditions.\n\nI believe the current version does not meet the standards for acceptance. However, I acknowledge the authors' motivation and their effective use of pre-trained models to address the data augmentation problem. I am willing to increase my score if the authors can adequately address these weaknesses and questions during the rebuttal." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The motivation is sound. The paper proposes separating foreground and background before targeted augmentation, aiming to maintain fidelity while increasing diversity. This approach is reasonable and effective, contrasting with previous methods that targeted the entire image.\n2. The paper effectively leverages existing technologies like SAM and LayerDiffuse to propose a new data augmentation framework, yielding valid experimental results.\n3. Extensive experimental results and open-source code suggest the method is reproducible." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a decoupled data augmentation framework for vision classification tasks, which separately augments the class-dependent part (CDP) and the class-independent part (CIP). Specifically, by employing SAM to initially separate CDP and CIP, the paper then designs a transparency img2img pipeline to augment CDP, inspired by SDEdit and LayerDiffuse, and finally generates synthetic samples by combining the generated CDPs with randomly sampled CIPs. Experiments on various datasets demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Major:**\n\n1. Limited technical contribution. The paper utilizes existing techniques such as SAM, LayerDiffuse, and text inversion. While these technologies are well-utilized to produce seemingly credible results, there is a lack of novel technical contributions and design. Additionally, performance improvements gained by using SAM are unsurprising and come with increased computational costs. Would the method still work without SAM? Are there alternatives to SAM?\n2. The writing needs further polish. There are confusing descriptions, particularly in the Method section's Conservative Generation of CDP. For instance, in Eq. 3, the variable $ c $ seems to refer to a text prompt, but its definition is missing. Moreover, as a critical part of augmentation, there is no experimental discussion on it. Also, if Line 233’s $\\lfloor T_{s}\\rfloor$ indicates a timestep, what does $\\lfloor S_{T_{s}}\\rfloor$ in Eq. 1 mean? $ S $ is also undefined—perhaps it means total steps? Numerous other issues are noted in the Minor weaknesses.\n3. I recommend adding a Background section to provide a brief review of SDEdit and the diffusion model used, along with clear symbol definitions.\n4. Experimental results are confusing. For example, the source of the experimental results is unclear—are they reproduced by the authors or cited from another paper? In Table 2, the authors label “training from scratch,” yet the vanilla results for DiffuseMix on ResNet-50@448 are 65.50, 80.29, and 85.52, compared to this paper’s 72.54, 71.53, and 91.32, which are significantly higher than those reported in DiffuseMix’s Table 14. The authors need to clarify the experimental setup and result sources. Additionally, in Table 5, results for Vanilla, CutMix, DA-Fusion, and Diff-Mix are cited from Diff-Mix, but the source for the remaining results is missing.\n\n**Minor:**\n\n1. Inconsistent use of \\citet{} and \\citep{}. Please verify correct usage in Line 058, Line 061, Line 068, and Line 346.\n2. Incorrect citation. The citation for DiffuseMix in Line 315 seems wrong.\n3. Incorrect Y-axis label. Why is the metric for multi-label classification in Figure 5(a) PSNR?\n4. Line 731 seems incomplete. \"For inpainting the missing part of For the training ...\" What does this mean?????" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nchen2024decoupled,\ntitle={Decoupled Data Augmentation for Improving Image Classification},\nauthor={Ruoxin Chen and Zhe Wang and Ke-Yue Zhang and Shuang Wu and Jiamu Sun and Shouli Wang and Taiping Yao and Shouhong Ding},\nyear={2024},\nurl={https://openreview.net/forum?id=w2qzdlvPMK}\n}" }, "abstract": { "value": "Recent advancements in image mixing and generative data augmentation have shown promise in enhancing image classification. However, these techniques face the challenge of balancing semantic fidelity with diversity. Specifically, image mixing involves interpolating two images to create a new one, but this pixel-level interpolation can compromise fidelity. Generative augmentation uses text-to-image generative models to synthesize or modify images, often limiting diversity to avoid generating out-of-distribution data that potentially affects accuracy. We propose that this fidelity-diversity dilemma partially stems from the whole-image paradigm of existing methods. Since an image comprises the class-dependent part (CDP) and the class-independent part (CIP), where each part has fundamentally different impacts on the image's fidelity, treating different parts uniformly can therefore be misleading. To address this fidelity-diversity dilemma, we introduce Decoupled Data Augmentation (De-DA), which resolves the dilemma by separating images into CDPs and CIPs and handling them adaptively. To maintain fidelity, we use generative models to modify real CDPs under controlled conditions, preserving semantic consistency. To enhance diversity, we replace the image's CIP with inter-class variants, creating diverse CDP-CIP combinations. Additionally, we implement an online randomized combination strategy during training to generate numerous distinct CDP-CIP combinations cost-effectively. Comprehensive empirical evaluations validate 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": { "value": [ "~Ruoxin_Chen1", "~Zhe_Wang45", "~Ke-Yue_Zhang1", "~Shuang_Wu7", "~Jiamu_Sun1", "~Shouli_Wang2", "~Taiping_Yao2", "~Shouhong_Ding3" ] }, "authors": { "value": [ "Ruoxin Chen", "Zhe Wang", "Ke-Yue Zhang", "Shuang Wu", "Jiamu Sun", "Shouli Wang", "Taiping Yao", "Shouhong Ding" ] }, "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": [ "Data augmentation", "Diffusion", "Image classification" ] }, "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": "chen|decoupled_data_augmentation_for_improving_image_classification" }, "pdf": { "value": "/pdf/cab8c7cb99d4857f621438247ab52c532b9d10a4.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/0ca9d1357a6054dcb831ed09ee3887a618313d87.zip" }, "title": { "value": "Decoupled Data Augmentation for Improving Image Classification" }, "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": "-The main premise, could it be applied to more problems? Are there related works that are directly related? This is a nice twist in a famous problem and I am curious if this has been studied for more traditional clustering problems like k-means or other graph partitioning problems." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+cute problem for correlation clustering where multiple instances are present. This is a nice twist in a famous problem and I am curious if this has been studied for more traditional clustering problems like k-means or other graph partitioning problems.\n\n+overall, the statements are clean for approximation and interesting.\n\n+well-motivated problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper extends the literature on the fundamental problem of Correlation Clustering. The plot twist here is that there are many correlation clustering instances that we have to solve on the same set of n vertices. The paper proceeds by formalizing the problem using the notion of multilayer-disagreements vector and then the authors give approximation algorithms for this. The goal is to find a common clustering\nof V that is consistent with as much as possible all layers. The main algorithm attains an Llogn-approximation where L is the number of layers. Moreover, they study the problem with probability constraints, where on each layer, there are ‘+’ and ‘−’ edge labels, with nonnegative weights in [0, 1] whose sum is equal to 1, hence the name probability constraints.\n\nNotice that the multilayer-disagreements vector the authors introduce has dimension equal to the number of layers L and every element of represents the disagreements of the clustering on the corresponding layer. The objective used is ell-p norm minimization on the said vector. For the case of probability constraints the authors give an (\\alpha+2)-approximation where we can use as a black box existing algorithms to get \\alpha approximation for the standard correlation clustering problem. In some cases, they slightly improve upon this generic (\\alpha+2)-approximation result." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-one major concern I have is that there is limited novelty. Introducing a new problem is always interesting however in terms of techniques the paper heavily relies on prior works. The L layers in the input are handled in a relatively straighforward way and the analysis is a bit incremental, given the large bode of works for correlation clustering. I like the paper, but this is an important concern that I have." }, "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": "Line 135: \"2.5 approximation, respectively\" -- means what? this is unclear. \nLine 138: Just to understand better, if the - weights satisfy triangle inequality, are the - label weights themselves positive values, or are the negative values? This becomes clearer later, but was unclear at this point. \n\nLine 231: \"Note however that for Problem 1 of the unweighted case\" -- what does this mean?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Well written paper\nResults are interesting from a theoretical perspective\nPaper could spark nice follow-up work as it leaves many interesting challenges open\nIt is rare to find a theory paper run experiments of the kind this paper does, so much credit to the authors :)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In classical correlation clustering, we have a complete graph, where every edge is either labeled + or -. And there are non-negative weights on edges. The goal is to cluster the nodes into parts, so that the total disagreement is minimized. That is, total weight of + edges going between parts, and total weight of - edges going inside parts. \nGeneral problem admit O(log n) approximation using region growing approach, and special case when weights are all 1, admits O(1) approximation ratio, which has been improved over many prior work.\n\nThis paper studies a new generalization. Imagine we have L such graphs with weights and labels, and wish to find a \"single\" clustering which works well for all L graphs. How to aggregate the scores? let (D_1, D_2, .., D_L) denote the L scores for a given clustering. Then we can consdier the l_p norm of this vector to be the quantitiy we are trying to optimize.\n\nPaper also studies one more \"probability\" version, where each edge has two weights w+ and w-, which add up to 1. \n\nFor the general problem, they show O(L log n) approximation factor.\nFor the probability version (where the weights add up to 1 for each layer), they show two results: one alpha+2 approximation and one 4-approximation, where alpha is the single-layer approximation factor." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Not sure how suitable the paper is for ICLR audience, as it is more of an SODA/ALENEX type paper in my humble opinion. \n(Not taking anything away from the technical merits!)\n\nIn Section 5.1, Authors could do much more justice in explaining how they use Problem 2 to solve the general problem. In particular, what metric they use, what are x1, .., x_L and what is F? Are these the different solutions we get from the convex program? and metric space is the space of all solutions? Adding this details would make it more readable and interesting. Also stress that Problem 3 is challenging only because the metric space could be huge. \n\n6.1 Are there no real-world datasets without any semi-synthetic aspect to sampling the weights? Especially sampling negative weights from the positive weights.\n\nWhy is pick the best not run for the larger datasets?\n\nAre there any other baselines one can think of? Perhaps some combination of adding the weights and pick a best? Perhaps using some approximation for l_p norm and then inferring a sampling strategy based on that? Like Multiplicative Weights method to weight the different layer instances?" }, "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 additional information can be gained from multilayer correlation clustering compared to simply utilizing aggregated weight functions?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The main contribution is to propose a polynomial time algorithm to output a solution of $O(L \\log(n))$ accuracy for the generalized correlation clustering problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a generalization of correlation clustering problem, termed as multilayer correlation clustering. In addition, polynomial time $O(L \\log(n) )$-accurate approximation algorithms are proposed to solve the generalized problem. The main idea is to relax the original problem to a convex problem." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1, The proposed multilayer correlation clustering problem lacks motivations. The aggregation of information from multiple weight functions $w_{l}^{+}$, $w_{l}^{-}, l=1,2, \\dots, L$ can be done through more convenient and efficient ways. For instance, one can aggregate information by aggregating the weight functions by considering $w^{+} = \\max_l w_{l}^{+}$, $w^{-} = \\max_l w_{l}^{-}$ or $w^{+} = \\sum_l w_{l}^{+}$, $w^{-} = \\sum_l w_{l}^{-}$ or $w^{+} = (\\sum_l (w_{l}^{+})^p)^{1/p}$, $w^{-} = (\\sum_l (w_{l}^{-})^p)^{1/p}$. The benefits of solving the multilayer correlation clustering problem are not be well-established in the paper.\n\n2, In section 6.1, since the case $p=\\infty$ is considered, it's fairer to compare with aggregated functions $w^{+} = \\max_l w_{l}^{+}$, $w^{-} = \\max_l w_{l}^{-}$.\n\n3, Two baseline optimization methods for solving problem 1 are compared in the simulations. However, a method for comparing information gain and clustering accuracy for problem 1 is currently lacking." }, "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": "- Only arXiv version is cited for several papers — are they published in a conference? Please check.\n\n- The baselines (Pick-a-Best and Aggregate) are relatively trivial algorithms, but in the experiments, their results seem to also approach the optimal solution of the LP. Does this suggest that there may be some issues with the experimental setup?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The definition of \"multilayer correlation clustering\" looks natural, and the motivation is clear.\n\n- Under the new model, the proposed algorithm achieves a good approximation ratio. In the case of probability constraints, the algorithm can achieve a constant approximation ratio.\n\n- The paper is easy to read. The explanation of the convex programming problem and the algorithm is clear.\n\n- The experimental results are good, obtaining near-optimal solutions in various real-world datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies multilayer correlation clustering, which is a generalization of correlation clustering. Each layer represents a distinct correlation clustering instance on the same set of vertices. The goal is to find a consistent clustering for all layers while minimizing the p-norm of the multilayer-disagreements vector. \n\nThe main result is an $O(L\\log n)$-approximation algorithm for multilayer correlation clustering and improved algorithms for the special case of the problem with a probability constraint. The authors provide theoretical proofs and experimental evaluations to demonstrate the effectiveness of the algorithms." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of theoretical justification for the effectiveness of the results: There is no comparative analysis of the algorithm with related work (e.g., MCCC, Bonchi et al. (2015)), nor is there a lower bound provided.\n\n- The approach requires to solve CV or LP which are heavy for larger datasets. (In the experiments, only $p = \\infty$ was tested; I suspect that if other values of $p$ are used, the running time will be longer due to the need to solve CV.)" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We establish Multilayer Correlation Clustering, a novel generalization of Correlation Clustering to the multilayer setting, and design several approximation algorithms." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024multilayer,\ntitle={Multilayer Correlation Clustering},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w2uIJiHTIA},\nnote={under review}\n}" }, "abstract": { "value": "We establish Multilayer Correlation Clustering, a novel generalization of Correlation Clustering to the multilayer setting. In this model, we are given a series of inputs of Correlation Clustering (called layers) over the common set $V$ of $n$ elements. The goal is to find a clustering of $V$ that minimizes the $\\ell_p$-norm ($p\\geq 1$) of the disagreements vector, which is defined as the vector (with dimension equal to the number of layers), each element of which represents the disagreements of the clustering on the corresponding layer. For this generalization, we first design an $O(L\\log n)$-approximation algorithm, where $L$ is the number of layers. We then study an important special case of our problem, namely the problem with the so-called probability constraint. For this case, we first give an $(\\alpha+2)$-approximation algorithm, where $\\alpha$ is any possible approximation ratio for the single-layer counterpart. Furthermore, we design a $4$-approximation algorithm, which improves the above approximation ratio of $\\alpha+2=4.5$ for the general probability-constraint case. Computational experiments using real-world datasets support our theoretical findings and demonstrate the practical effectiveness of our proposed algorithms." }, "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": [ "Clustering", "Correlation Clustering", "Multilayer Networks", "Approximation 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/6bf7e32a203fb4f87c42e4851e0571c79af3eace.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": "Multilayer Correlation Clustering" }, "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": "Please address the issues outlined in the weaknesses section. The effectiveness of the proposed method is significantly limited by the low quality of the experiments, particularly the empirical results. It is essential to include extensive comparisons with related works to more convincingly demonstrate the method’s effectiveness, beyond merely relying on the proposed frequency metric." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The research motivation of the paper, specifically the comparison of decision patterns and frequencies between RL and humans, is compelling.\n- The paper establishes a strong connection to biological fundamentals, providing relevant examples and insights throughout.\n- The main idea—designing an RL algorithm inspired by biological principles, where each component operates at different frequencies—is novel, and the storyline leading up to the experiment section is smooth and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new model-based RL method that enables a lower frequency for action decision-making, slower than the frequency of sensing and execution. To accomplish this, the proposed method, Sequence RL (SRL), predicts a sequence of actions at each decision step, allowing for a lower decision frequency while employing a high-frequency dynamic model learner and action-value estimator. This setup mimics biological behavior, where the brain’s computation frequency is lower than that of sensing and actuating processes. Additionally, the paper introduces a new metric, FAS, to assess an RL method’s robustness and adaptability across different frequency settings, demonstrating that SRL achieves a high FAS score." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Lack of related works\n- This paper primarily focuses on its connection to biological insights and motivation but overlooks relevant efforts in the RL literature addressing frame-skipping, action repetition, long-horizon exploration, and action correlations. Several studies, such as [1,2,3,4], have explored similar topics from different perspectives. Although these works are not biologically motivated, their contributions are highly relevant to this paper and should not be ignored.\n\n- The way of using GRUs in sequence action generation is very similar to the work in [4], please consider adding discussion.\n\n## Technical Issues\n- From line 252 to line 259, it is unclear that why using max entropy RL techniques, such as SAC, can address the issue of additive noise and automatically lower entropy for deeper actions arther from the observation. \n\n- Inproper declaration from line 292 to 296. There are several works that focus on credit assignment for action sequences, such as [2, 5].\n\n- Does the current work fully address the issues listet in line 93? Namely the sparse rewards, jerky control, high compute cost, and catastrophic failure due to missing inputs. If not, it is better to remove this misleading content. \n\n## Poor Experiment Quality (Critical Issue for this Paper)\n- The empirical results only include SAC as a baseline method, which was introduced in 2018. This is clearly insufficient. All the methods mentioned above have demonstrated advantages over SAC by employing different techniques for modeling action sequences or predicting temporally correlated actions. These approaches should theoretically achieve good FAS scores as well. Please consider including some of these methods to enrich the experiments and enhance the technical rigor of the paper.\n\n- Even with SAC as the only baseline, the proposed method, SRL, only outperforms SAC in 6 out of 10 tasks. Despite its advantage in the proposed FAS metric, SRL's performance over SAC is marginal and unconvincing.\n\n- In 5 out of 10 tasks (InvPendulum, Hopper, InvDPendulum, Reacher, and Swimmer), the Online Planning method outperforms SRL, contradicting the description in the caption of the \"second Table 3\" on page 9.\n\n- Many important results are placed in the appendix, while the main paper is disproportionately occupied by the introduction to biological concepts.\n\n## Excessive Discussion of Biological Concepts\n- Section 6 should be either removed entirely or moved to the appendix, as it detracts from the paper’s main focus on learning representation rather than biology.\n\n## Minor issues\n- It is better to use different math symbols to distinguish the macro action and per-step action, especially in equation 1. Both sides used $a_{t'}$\n- There are two \"Table 3\" captions in the paper: one on page 8 and another on page 9.\n- The layout of figures and text in the paper is of low quality and requires adjustments and proofreading. For example, the title of Section 6 is oddly placed.\n\n## References\n[1] Raffin, Antonin, Jens Kober, and Freek Stulp. \"Smooth exploration for robotic reinforcement learning.\" Conference on robot learning. PMLR, 2022.\n\n[2] Li, Ge, et al. \"Open the Black Box: Step-based Policy Updates for Temporally-Correlated Episodic Reinforcement Learning.\" ICLR 24.\n\n[3] Chiappa, Alberto Silvio, et al. \"Latent exploration for reinforcement learning.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[4] Zhang, Haichao, Wei Xu, and Haonan Yu. \"Generative planning for temporally coordinated exploration in reinforcement learning.\" ICLR 22.\n\n[5] Ni, Tianwei, et al. \"When do transformers shine in rl? decoupling memory from credit assignment.\" Advances in Neural Information Processing Systems 36 (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": 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": "1. Like SAC, SRL is using batched experience replay. Can you elaborate on how this interferes with biological plausibility of the method?\n2. What is the $\\alpha$ that shows up in eq. 3? What does the $\\alpha \\log \\pi$ term represent?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The introduced SRL method is explained well and is easy to understand while also showing significant improvement in the introduced FAS score. Experimental evaluation is good and the usefulness of the FAS score is adequately demonstrated. The proposed SRL framework is well motivated using recent neuroscientific discoveries." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This model introduces Sequence Reinforcement Learning (SRL), a framework inspired by biological learning agents that allows prediction of action sequences of variable length. The method builds on the Soft-Actor-Critic algorithm, comprising a policy and a state-action-value function. Additionally, a model of the environment is trained and used for updating the policy. The policy includes a GRU RNN allowing the computation of arbitrary length action sequences starting from a single observation. While the model and the critic are trained on truly experienced trajectories only, training of the policy is done by predicting intermediate states using the model and the critic for assigning state-action-values. The method is evaluated on several continuous control tasks. The authors introduce a metric they call Frequence-Averaged Score (FAS) which is defined as the area under the curve of the reward vs decision frequency plot, and find that the SRL framework achieves significantly higher FAS-scores compared to the SAC baseline. They show that this metric is useful in predicting the average reward achieved on a randomized timestep version of the environment, arguing that a policy trained with SRL is better equipped for bridging the sim-to-real gap. The paper ends with a comprehensive comparison to structures found in the mammalian brain." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Demonstrating the improvement in sim-to-real transfer would add to the quality of the paper.\nLikewise, including methods that use action repetition and macro-actions could be an interesting addition.\n\nMinor mistakes:\n- Line 104: demonstrate**s**\n- Line 376: performance of the policy **in** when the frequency is not constant\n- Line 535: we introduce**s** the Frequency-Averaged-Score (FAS) metric\n\nFinally, I can't say that I fully agree with the statement \"simple tasks like walking can be performed without input states if learned properly\". Biological agents also use a range of inputs to maintain gait." }, "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": "- What variations of SRL are included in “all the policies” in line 374? If e.g., SRL-2 is included how would it provide open-loop actions for 16 steps if 16 is sampled from the uniform distribution in line 376?\n- The setting on Figure 2 when considering the left-most action sequence length appears similar to an MPC setting for the SRL-J agents, where performance over model-rollouts is learned but only the first action is executed before predicting a new sequence. Could you elaborate on the performance delta of SRL-8/16 to SAC? Is this due to poor model fit?\n- Could you elaborate on what you mean by the following sentence: “The SAC algorithm addresses this by maximizing the entropy of each action in addition to the expected return, allowing our algorithm to automatically lower entropy for deeper actions farther from the observation.” How does this lower the entropy of “deeper” actions?\n- Is there a good argument of why Gym-v2 versions of the tasks were chosen?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper investigates an interesting problem, considering learning of temporally consistent open-loop action sequences that require fewer queries to the policy during deployment. While the approach is reliant on the quality of the underlying learned model, this enables learned MPC-like control at one end (deterministic 1-step) while also accommodating stochastic sampling periods by using open-loop actions until the next query is possible. The authors evaluate the method across several environments, and compare across 5 seeds against the strong SAC baseline. The paper is generally well-written and clearly motivated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Sequence Reinforcement Learning (SRL) as a method to lower effective control frequency by predicting multi-step open-loop action sequences. The action sequences are optimized on model rollouts at the given low-level frequency, while the policy only needs to be queried at a lower rate during deployment. The policy can then be deployed over the full open-loop horizon, or re-queried at intermediate points, where total reward obtained across action sequence lengths favors longer SRL variations over SAC." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The selection of baselines and or environments should be expanded, as the paper makes a quantitative argument of performance. Other model-free agents (e.g. D4PG, etc.) as well as model-based agents would help to better put the results into perspective. It could furthermore be interesting to run these experiments on the DeepMind Control suite as well, as Gym and DMC tasks have interesting differences in their underlying dynamics and resulting agent performance.\n- It would be very informative to see how SAC performs when learning actions at the same rates as SRL-J, e.g. by applying action repeats of J steps. If “SAC-J” would significantly underperform SRL-J, it would provide a stronger argument for adding complexity via model learning.\n- Potential inspiration for additional real-world-inspired environment modifications can be found in [1]\n- I partially agree with the statement in line 334 that SRL-J is at a disadvantage. However, the agent still uses access to J=1 information via model and critic learning and so the learning curves do provide an accurate comparison across time steps.\n- Instead of a learned model, it would be interesting to see an ablation with the ground-truth dynamics to determine how well SRL-J works w/o this “handicap” (ideal model)\n- The sentence starting in line 160 highlights that SRL has a model complexity of zero after training. It should be made more clear that this actually applies to many MBRL algorithms (e.g. references in lines 138-143).\n- Line 204: consider providing a brief summary of the insights of Section 6 here, or moving Section 6 forward. Otherwise it’s difficult to follow insights that have not been presented, yet.\n- Sentence repetition in abstract lines 17-19\n- Minor: line 49 —> missing period; line 265 —> model not subscripted\n\n[1] G. Dulac-Arnold, N, Levine, D. J. Mankowitz, J. Li, C. Paduraru, S. Gowal, and T. Hester. \"An empirical investigation of the challenges of real-world reinforcement learning.\" arXiv, 2020." }, "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 line 098 the claim “Our model learns open-loop control utilizing a slow hardware and low attention, and hence also low energy.” is made. Why should that be true?\n* In line 150 the claim that online planning is not feasible in robotics or the brain is made. In my experience this is not true. How did the authors come to this conclusion?\n* What is meant by “reducing actor sample complexity” in line 028 in the abstract? Sample complexity usually refers to the amount of transitions needed for training.\n* In the paragraph starting in line 050, the authors compare the performance of RL agents on MuJoCo tasks with the performance of humans on complex tasks and relate to reaction times. Why would these settings be comparable as the embodiment, action representation, and tasks are completely different?\n* In table one the < sign in the last row should be >.\n* What is meant by online planning in the paper? And how is it separate from MPC? It is not clear to me what online planning means.\n* In line 054 there seems to be a typo, a redundant “the shows”.\n* In line 169 the statement that MPC is limited to systems that have already been modeled accurately is made. However, algorithms like TD-MPC2 use MPC successfully with (partially) learned models.\n* The next paragraph states that MPC requires very short timesteps to work. However, that many robotics papers use high frequencies does not mean that it is strictly required. For example, [1] does MPC over skills which has a very low frequency.\n* The training curves in the appendix are plotted on top of each other such that it is almost impossible to make out some of the algorithms. Would it be possible to improve these plots?\n* Would it be possible to visualize the action sequences predicted by SRL compared to those of SAC?\n* In line 516 the claim “In deterministic environments, a capable agent should achieve infinite horizon control for tasks like walking and hopping from a single state.” is made. This is infeasible in environments with complex, possibly unstable dynamics since errors do not disappear but can instead compound and blow up. I would recommend revising this paragraph.\n\n[1] Shi, Lucy Xiaoyang, Joseph J. Lim, and Youngwoon Lee. \"Skill-based model-based reinforcement learning.\" arXiv preprint arXiv:2207.07560 (2022)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper relates problems that robotics researchers encounter in a creative way with insights from neuroscience. The issue of high computational demands of modern RL and imitation learning algorithms is quite relevant for the robotics community. The writing is furthermore clear and easy to understand. The proposed algorithm is simple, and easy to understand and implement." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper argues that the application of current reinforcement learning (RL) methods to real-world problems is hampered by the assumption that the actor can be executed at a high frequency (matching that of the actuators). Depending on the architecture of the policy, this can lead to a high computational burden. The authors take inspiration from action chunking observed in the brain to propose an RL algorithm that directly predicts action sequences (or macro actions). After training, it can thus run at slower frequencies than the environment MDP. Experiments on several simulated continuous control benchmarks demonstrate good performance when predicting macro actions at frequencies lower than that of the MDP, as well as for variable stochastic time steps. The paper furthermore discusses connections of macro actions to neuroscience." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper shifts its motivation from real-time control for robotics to biological plausibility over the course of the paper. The requirements of robotics and biological plausibility are not aligned well, however. Running policies at high frequencies is not necessarily an issue in robotics as long as the computational resources are sufficient. Furthermore, RL algorithms like Soft Actor Critic (SAC) usually do not aim at being biologically plausible, and the use of backpropagation through time in SRL could be viewed as not biologically plausible.\n\nIn particular, the small MLP policies trained with SAC should run at about 1 ms on modern hardware (or thereabouts). What is more, even when running them with an action repeat corresponding to a policy execution frequency of 50 Hz, the resulting policies usually perform fine [1]. Hence, the computational load of the experimental set up does not pose a challenge to modern hardware and is therefore not well aligned with the motivation. \n\nWhat is furthermore absent from the discussion of the results is actual inference times. According to the abstract, reducing the time it takes to produce an action is a main motivation for the paper. However, it is entirely unclear why computing the first action of the macro action should be faster with the proposed algorithm than with vanilla SAC. Hence, it is unclear which advantage the algorithm in its current form offers in a real-world scenario where primitive actions have to be supplied at a fixed frequency to the actuators. It might be possible to run at a higher frequency if a delay is introduced to account for the time needed to calculate the first action of a macro action. Discussing this trade off would be interesting.\n\nThe motivation of the paper moreover hinges on the claim that existing methods cannot deal well with low control frequencies in the presence of multiple action dimensions (line 197). In my opinion, this claim is unsubstantiated for two reasons: (i) There is a complete lack of baselines targeted at low control frequencies in the experiments. At the very least showing results for SAC with action repeat is an absolute necessity to put the results into perspective, in my opinion. (ii) The two prior works mentioned (FiGAR and TempoRL) do have experiments on continuous control environments with multiple action dimensions in which they do achieve competitive performance. They should also be included as baselines (TempoRL makes an appearance in the appendix and performs well but is not shown in the main text). The evaluation of SAC is furthermore entirely unfair as it was trained without any action repetition but is then evaluated at a substantially lower frequency. This cannot work and is not a relevant baseline (unlike SAC with action repeat, training with randomized time steps, FiGAR, and TempoRL). For these reasons, I think the experimental results lack relevant baselines and cannot be evaluated properly without them. I would suggest adding these baselines (in the main text).\n\nThe field of hierarchical RL is only mentioned in passing in the paper but offers many relevant approaches to easing the computational requirements for control. A higher level (or manager) modulates the lower level (or worker) which produces primitive actions. As the worker may have a simpler architecture, it can run at a higher frequency, while the manager runs at a lower frequency. The worker can be both open loop or closed loop. In practice, simple PD controllers are often used in conjunction with low-frequency policies in robotics and often provide good results. Discussing the hierarchical RL literature to some extent therefore seems necessary. \n\nIn line 055 the claim “When RL agents are constrained to human-like reaction times, even state-of-the-art algorithms struggle to perform in simple environments.” is made. This needs a citations or experiments to back this up.\n\n[1] Allshire, Arthur, et al. \"Transferring dexterous manipulation from gpu simulation to a remote real-world trifinger.\" 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce an algorithm that achieves competitive continuous control at extremely slow control frequencies using action sequences" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024overcoming,\ntitle={Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w3iM4WLuvy},\nnote={under review}\n}" }, "abstract": { "value": "Reinforcement learning (RL) is rapidly reaching and surpassing human-level control capabilities. However, state-of-the-art RL algorithms often require timesteps and reaction times significantly faster than human capabilities, which is impractical in real-world settings and typically necessitates specialized hardware. Such speeds are difficult to achieve in the real world and often requires specialized hardware. We introduce Sequence Reinforcement Learning (SRL), an RL algorithm designed to produce a sequence of actions for a given input state, enabling effective control at lower decision frequencies. SRL addresses the challenges of learning action sequences by employing both a model and an actor-critic architecture operating at different temporal scales. We propose a \"temporal recall\" mechanism, where the critic uses the model to estimate intermediate states between primitive actions, providing a learning signal for each individual action within the sequence. Once training is complete, the actor can generate action sequences independently of the model, achieving model-free control at a slower frequency. We evaluate SRL on a suite of continuous control tasks, demonstrating that it achieves performance comparable to state-of-the-art algorithms while significantly reducing actor sample complexity. To better assess performance across varying decision frequencies, we introduce the Frequency-Averaged Score (FAS) metric. Our results show that SRL significantly outperforms traditional RL algorithms in terms of FAS, making it particularly suitable for applications requiring variable decision frequencies. Additionally, we compare SRL with model-based online planning, showing that SRL achieves superior FAS while leveraging the same model during training that online planners use for planning. Lastly, we highlight the biological relevance of SRL, showing that it replicates the \"action chunking\" behavior observed in the basal ganglia, offering insights into brain-inspired control mechanisms." }, "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": [ "Decision Frequency", "Action Sequence Generation", "Model-Based Training", "Model-Free Control", "Efficient Learning", "Reinforcement 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/aae5ad8300548a00ec85c7d2467398a3f85c3e8d.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/693e01b4048cb114c669781fe0dfe21980ff288f.zip" }, "title": { "value": "Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control" }, "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": 1 }, "primary_area": null, "questions": { "value": "I think most of my questions were listed in the weakness section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I want to preface this by saying that I'm not an expert in PDEs nor in deep-learning based approaches for solving PDEs. Nonetheless, I am a big fan of the proposed approach. The proposed architecture is something I have never seen before and is very nifty! I am also very impressed by the experiments section! Specifically, I LOVE that the authors look at multiple different metrics which allows one to look at the results from multiple different angles, which is very important." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces physics informed multi-scale recurrent learning (PIMRL), which is a novel architecture for learning both micro and macro time scales commonly present in real world data. Specifically, PIMRL introduces a micro and macro architecture, along with corresponding training schemes, that provides an implicit bias for the network to learn both micro and macro timescales." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I again want to preface this section by saying that I am not an expert in this field. I hope that a non-expert's viewpoint will provide criticism that will make this paper stronger.\n\nThere are two big weaknesses that I think limit the paper from being truly great. The first is the writing. There are A LOT of interesting things in this paper that are not adequately explained nor properly described. Moreover, there is a lot of text and figures that are confusing and don't add much to the text; they could be removed to allow for the other aspects of the paper to shine. I will list my qualms with the writing below\n\n- To start, the authors begin by stating that \"direct numerical simulation requires a deep understanding of physics\" but it is not clear what this means. Do the authors mean that one needs an understanding of physics to come up with PDEs that describe physical phenomena? This is independent of numerical simulating equations as a solver does not need to know what physical phenomena an equation is trying to describe. Moreover, since the name of the proposed approach is physics-informed multi-scale recurrent learning, does this not imply that a deep understanding of physics would also be required to effectively use PIMRL?\n\n- Next, Figure 1 takes up a lot of space but does not add much. Personally, given its vagueness it confused me on what PIMRL is. Figure 2 does a much better job explaining the method. I would recommend removing figure 1.\n\n- A big part of PIMRL is PerCNN, as the micro-scale architecture is exactly the PerCNN architecture. To me this is not an issue as science is built upon previous works, but I would have preferred the authors explicitly stated this, i.e., 'We introduce PIMRL, which combines the micro-scale power of PerCNN, with a novel macro scale architecture to...\". Next, it is also not clear why PerCNN needed an extension in the first place (this evidence is lacking both in the text and in the experiments section). In the related works section, the authors state the following about PerCNN \"However, the model suffers from error accumulation during long rollout predictions. To address this issue, a straightforward approach is to reduce the frequency of iterations by increasing the time stepping. Nonetheless, for hard-embedded methods like PeRCNN, accuracy is limited by the time stepping, making simple modifications unfeasible\". This doesn't make any sense to me. In PerCNN, one can choose both the number of iterations in a the $\\Pi$ block as well as the time step, $\\delta t$; thus, it seems it would be rather straightforward to train PerCNN on macro-scale data. Moreover, the comparison to PerCNN doesn't seem fair. First, similar to what the authors did for FNO, it seems straight forward to also include a PerCNN coarse and fine scale. One could even just train one PerCNN on both the micro and macro scale data, where $k$ is dataset dependent. \n\n- To me the most interesting part of the approach is how the macro and micro modules interact. Sadly, this is not given time to shine in the paper! For instance, why was this interaction between the macro and micro module chosen? What role does the mirco-module serve (i.e., does it serve as error correction for the initial condition?)? There are so many interesting design choices here that I think are fascinating but sadly are missing.\n\n- In the paper, the authors use the term physical embedding where they state \"... physical embedding methods, where explicit physics knowledge is embedded into the model to fully leverage physical principles...\". They also state that PIMRL uses this but the entire approach is data-driven. If by physical embedding they mean the *Optional* physical-based FD conv then they should explicitly state this. Also, I think the fact this is *optional* weakens the statement considerably.\n\n\nThe next weakness is the experiments section. While the results are compelling and I love that they look at the results from multiple different angles, I think there are some major things missing.\n- There are no details on how the datasets are constructed. What solver was used? How were the initial conditions chosen?\n- The authors state that they create a micro and macro dataset for training PIMRL. For the other baselines, it seems like they were either trained on the micro or macro dataset. On a first read, it seems like PIMRL was trained on more data. Was the dataset size equalized for the baselines?\n- Error bars are missing in the error propagation plots and Table 2!\n- The ablation study is missing a lot of details. For instance, what does the NoConnect model mean? Is it just purely the micro module or the macro module?\n- Without knowing the solver, the inference time figure is a little underwhelming." }, "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": "1. Despite Figure 2, it remains challenging to understand the inference process. I suggest clarifying the figure or including a detailed description in the manuscript.\n2. Is the RMSE computation performed on the micro- or macro-scale data?\n3. Is model training robust to the number of training trajectories? Exploring the model’s scalability in relation to both the micro and macro modules would be insightful.\n4. Does the model require consistency in the underlying physical parameters during training ? said otherwise, is it robust to out of distribution trajectories ?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper presents a clear and well-formulated motivation that connects nicely with existing literature.\nExperimental results are well-discussed and appear conclusive, demonstrating notable improvements over comparable methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces a framework for handling multi-timescale data by modeling dynamics across two time scales. A micro module manages the high-frequency observations, while a macro module captures long-term dependencies. The micro module operates directly on physical states, leveraging PerCNN, whereas the macro module learns in a latent space and uses a ConvLSTM block for temporal propagation. Initially, the micro module is pretrained, and later fine-tuned jointly with the macro module using mean-squared error (MSE) supervision. The experiments appear to yield promising results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The model seems to require aligned micro- and macro-scale training data, which may limit its applicability.\n- While the model presentation is well-structured, its technical novelty is somewhat limited, as it could be considered a specific implementation of Rubanova et al. [1].\n\nWould the authors consider working with unaligned datasets, such as those with only micro-scale or only macro-scale trajectories?\n\n[1]: Rubanova et al., Latent ODEs for Irregularly-Sampled Time Series" }, "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": "- It is not clear how the boundary conditions are encoded in each baseline: is boundary padding also used? If not it seems like an unfair advantage as it has access to additional information\n- why is the Physics operator optional? Does it depend on the dataset? If not, it should be added to the ablation study." }, "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 beautifully illustrated, making the methods easy to read understand and the results easy to read and understand\n- the proposed method outperforms the state of art on a variety of PDEs" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a data-driven model to simulate a spatiotemporal system. It proposes to leverage multi-scale data by embedding physical knowledge into a micro-scale module and employing a data-driven approach for its macro-scale module. \nThe method is then tested on various fluid dynamics and reaction-diffusion systems equations, reaching impressive results on a challenging dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some method details lack clarity, especially regarding the physics operator.\n- the related work and baselines seem to lack a few strong methods. For example, there is extensive literature on improving FNO (including some published here last year: [https://openreview.net/pdf?id=tmIiMPl4IPa]. Is there a reason for using FNO and not its extensions?\n[https://arxiv.org/abs/2204.11127] also seems a like a strong baseline\n- The ablation study lacks depth: more details on each module's contribution, the micro-scale module's pre-training, and some parameters would help understand the contribution 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": 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": "Why did the authors name the micro module as “physics-informed”? Actually, I think the formalization of Eq. (4) is just a convolution neural work, which is far away from the conventional definition of “physics-informed neural networks”." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper focuses on an important problem: the temporal multiscale property, which is mainly overlooked in previous methods.\n\nThe experiments are relatively comprehensive. The authors test PIMRL in five typical tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the temporal multi-scale property of spatiotemporal systems and proposes a multiscale recurrent learning framework, named PIMRL. In this model, the micro-scale module learns to embed multiscale data and the macro-scale module captures the data-driven evolution. PIMRL performs well in five different tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tAbout the writing.\n\nI believe that this paper needs a significant revision to ensure a clear presentation. Here are some examples.\n\n(1)\tIntroduction: There are many important concepts without explanation. For instance, the authors claim that the micro-scale is “pretrained” in the abstract and introduction. However, in the method part, there are no descriptions of how this pretraining be implemented.\n\n(2)\tRelated work: The authors fail to provide a clear organization of the related work. For example, ConvLSTM is just one classical (maybe old) baseline in spatiotemporal learning. There are extensive papers that weren’t included, such as PredRNN [1] or TrajGRU [2].\n\n[1] PredRNN: a recurrent neural network for spatiotemporal predictive learning\n\n[2] Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model\n\n(3)\tMethod: I think Eq. (2) and (3) do not provide a clear and right formalization for the overall framework. The descriptions of the prediction protocol are self-contradictory. Given that each micro module can conduct a \\delta t stepsize transition and the macro module is \\Delta t, why the output of Eq.(3) is just t+k\\delta t. Besides, as they presented in Figure 2, I cannot tell which part is “history information” or if the micro-scale module is applied to the model prediction or not. A flowchart or pseudocode can be helpful.\n\nThe low quality of writing seriously affects the contribution of this paper.\n\n2.\tCompare with more advanced baselines.\n\nAs the authors mentioned in Lines 125-133, there are many advanced Neural Operators such as MWT and U-NO, which should be compared.\n\n3.\tAbout the novelty.\n\nAs the proposed method is more like a multiscale ensemble method, I cannot score high for the novelty." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a new multi-scale framework calling physics-informed multi-scale recurrent learning (PIMRL) framework to effectively utilize multi-scale time data." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024pimrl,\ntitle={{PIMRL}: Physics-Informed Multi-Scale Recurrent Learning for Spatiotemporal Prediction},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w3rbBVJ9Jg},\nnote={under review}\n}" }, "abstract": { "value": "Simulating spatiotemporal systems governed by partial differential equations is widely applied in biology, chemistry, aerospace dynamics and meteorology. The classical numerical methods require small time stepping to generate predictions, leading to high computational costs. Although machine learning has reduced computational costs, they are limited in terms of stability and accuracy for long-term predictions, especially in cases of insufficient data or varying time scales. They often overlook how to effectively utilize multi-scale data, leading to poor robustness of prediction. To this end, we propose a novel multi-scale framework, termed the Physics-Informed Multi-Scale Recurrent Learning (PIMRL) framework to proficiently harness temporal multi-scale data for spatiotemporal dynamics prediction. This framework consists of two modules: the micro-scale module embeds physical knowledge into neural networks through pretraining, while the macro-scale module employs a data-driven approach to learn the temporal evolution of physics in the latent space. The PIMRL framework consistently achieves state-of-the-art performance across five benchmark datasets ranging from one to three dimensions, demonstrating improvements of over 9\\% in both RMSE and MAE evaluation metrics, with maximum enhancements reaching up to 80\\%." }, "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": [ "PDEs", "physics encoding", "data-driven modeling" ] }, "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/d99f586d62f12b197a3d1d625c932cb13c756460.pdf" }, "presentation": null, "primary_area": { "value": "learning on time series and dynamical systems" }, "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": "PIMRL: Physics-Informed Multi-Scale Recurrent Learning for Spatiotemporal Prediction" }, "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. Could the authors clarify the purpose of the SOT and EOT tokens?\n2. In Section 4, line 260, how are the starting and ending positions determined?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This work addresses a significant issue by reducing the generation of NSFW content.\n2. Extensive visualizations illustrate the effectiveness of GIE." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes GIE, a method for eliminating NSFW content in diffusion models. The authors leverage the attention map of target concepts, reweighting them to synthesize \"growth inhibitors\" that represent undesired content. These reweighted attention maps are then injected into the diffusion process. Additionally, an adaptor is trained to determine the suppression scale dynamically. GIE outperforms eight baseline methods and effectively removes implicit concepts like \"nude.\"" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing is somewhat difficult to follow in certain sections. See questions.\n\n2. There is a lack of ablation studies on each component of GIE, such as the role of the adaptor.\n\n3. The paper does not discuss or compare with related methods like the one proposed in [1], which also targets implicit concept removal in diffusion models.\n\n[1] Implicit Concept Removal of Diffusion Models, ECCV 2024.\n\nI will consider increasing the score if my concerns are addressed." }, "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) Proposed method contradicts with the statement in Line 252: \"A naive approach to erasure is to weaken tokens whose attention maps are similar to the extracted features. Unfortunately, this method might not work because...\". The reason is the following:\n- First, original Attention Map Groups are injected with the Growth Inhibitor I.\n- Second, the attended representation MV is calculated based on injected Attention Map Group M'.\n- As a result, Final output of Cross-Attention is calculated with M'V.\n- In SD v1.4, Value (V) is a linear projection of token embeddings c.\n- Considering all of these mentioned properties, the final output reduces to calculating Final Output = (M_[SOT] x V) + ... + (M*_[target_concept] x V) + (M_[EOT] x V)\n- As this final output calculation demonstrates, (M*_[target_concept] x V) directly corresponds to weakening every single token embedding -- since V is a function of token embedding c.\n\nIt would be good to clarify this. \n\n2) What happens when the concept prompt has more than 1 tokens? For example, when erasing the concept of Van Gogh, the token embeddings has the form <c_[SOT], c_[Van], c_[Gogh], c_[EOT] >. How to reduce c_[Van], c_[Gogh] into one, unified representation? Is Simultaneous Suppression of Multiple Concepts strategy described in Section 4.3 applied here?" }, "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 method doesn't require any fine-tuning concentrated on cross-attention layers -- which is the general methodology in the recent works in concept erasing literature, therefore, this work distinguishes itself from the other works.\n2) By training a Suppression Scale Adapter, the proposed method does not require additional training/finetuning for each different concepts.\n3) As shown in some qualitative experiments, after concept erasing operation, most of the spatial details are preserved -- especially when compared to other methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors focuses on a concept erasing problem which is caused by explicit NSFW prompts and importantly implicit unsafe prompts. Authors propose to inject \"Growth Inhibitors\", which is a slice of attention map group of target concept, to current attention map group. The re-weighting scheme of \"Growth Inhibitors\" are provided by proposed Suppression Scale Adapter." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) There is no clear declaration of why injected Growth Inhibitors I is inserted right before [EOT]. There is no ablation study on the position of Growth Inhibitors. More clarifications and motivations behind this, and ablations will help us to understand the proposed method better.\n\n2) No quantitative experimentation on multi-concept erasing. Furthermore, no qualitative results for more than 2 concept erasing. This makes it difficult to assess the real performance of this method.\n\n3) Qualitative results shows that, proposed method can only erase semantic concept up to a certain extend. For instance, in Figure 7 Bottom row, when erasing Van Gogh and Nude concepts, the resultant image still preserves the style of Van Gogh depicted by the color schema. This problem is also seen in Figure 13." }, "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": "- In line 415 (416?), how could the CLIP model be leveraged to determine whether specific objects exist in the generated images?\n- Do the NSFW removal rates used as a metric refer to the ratio averaged over test prompts calculated using NudeNet?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The main strength of the proposed method is that it does not require fine-tuning, which is highly beneficial in terms of efficiency. Additionally, the method is widely applicable to architectures similar to Stable Diffusion.\n- Most of the proposed methods are based on observations that support the validity of the approach.\n- Compared to the baselines, GIE erases the harmful and other concepts exceptionally well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new method GIE that injects growth inhibitors into the attention map to erase certain harmful concepts. The method can be applied without fine-tuning, with the injection level controlled by an auxiliary network. Experimental results show the effectiveness of GIE in erasing nude concepts while preserving the image quality." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While most methods are based on observations, I feel the paper lacks *sufficient intuition* or theoretical justification for each proposed method. How does inserting an attention map for concept erasure lead to the removal of specific concepts? What is the meaning of these new attention maps in M_replace? How does this reweighted attention map function as a growth inhibitor? It is good that you avoid the naive approach of simply suppressing the most related attention map, but since your solution deviates from previous attention modulation methods, while we can see that it works, it is challenging to understand why it works so effectively.\n- It appears that the adapter learns a function to map intermediate values to suppression values. However, intermediate values vary significantly with different text prompts. Although the authors demonstrate the generalizability of learned prompts, it is questionable whether training the adapter on a limited dataset will generalize effectively when the test set encompasses a broader range of concepts.\n- The model's applicability is currently limited to Stable Diffusion 1.4. Since the method relies on the *specific architecture*, I believe additional experiments on different Stable Diffusion models would better showcase its effectiveness.\n- This also relates to weakness #3: although GIE is effective and efficient, it requires a *high level of engineering*, such as designing an adaptive model and correctly injecting features. The optimal range of weights (w), suppression levels, and injection positions might vary depending on the target concepts or diffusion models. It would have been beneficial to demonstrate that the same method, with similar configurations, performs well on 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": 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": "see weakness" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The writing is clear and easy to follow.\n- The discussed topic and motivation are both innovative and significant.\n- The proposed method can erase inappropriate concepts from text-to-image diffusion models without the need for fine-tuning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel approach for erasing inappropriate concepts from text-to-image diffusion models without the need for fine-tuning. The authors identify that these models, while capable of generating sophisticated images, can inadvertently produce content that is ethically and legally problematic, such as NSFW material and copyright-infringing styles. The proposed method GIE addresses this by injecting growth inhibitors based on identified features related to inappropriate concepts during the diffusion process. An adapter is also trained to infer the suppression scale, accommodating varying degrees of inappropriateness. The paper claims that GIE effectively captures subtle word manifestations at the image level, enabling precise erasure of target concepts without affecting other concepts or image quality." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some previous works on LLMs for enhancing diffusion models, such as [1][2][3][4], should be included in the related work section.\n- I'm curious about the explicit or implicit damage to the model itself when inappropriate concepts are removed. Generally, inappropriate concepts may not be entirely independent of some \"appropriate concepts.\" This analysis needs to be considered.\n- What is the cost of using GPT-4o for data processing in this paper? For instance, the expenses and time involved. Such information should be listed to provide valuable insights to the community.\n\n\n\n\n\n[1] SUR-adapter: Enhancing text-to-image pre-trained diffusion models with large language models\n\n[2] PromptCrafter: Crafting Text-to-Image Prompt through Mixed-Initiative Dialogue with LLM\n\n[3] LLM Blueprint: Enabling Text-to-Image Generation with Complex\nand Detailed Prompts.\n\n[4] LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024growth,\ntitle={Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w4C4z80w59},\nnote={under review}\n}" }, "abstract": { "value": "Despite their remarkable image generation capabilities, text-to-image diffusion models inadvertently learn inappropriate concepts from vast and unfiltered training data, which leads to various ethical and business risks. Specifically, model-generated images may exhibit not safe for work (NSFW) content and style copyright infringements. The prompts that result in these problems often do not include explicit unsafe words; instead, they contain obscure and associative terms, which are referred to as *implicit unsafe prompts*. Existing approaches directly fine-tune models under textual guidance to alter the cognition of the diffusion model, thereby erasing inappropriate concepts. This not only requires concept-specific fine-tuning but may also incur catastrophic forgetting. To address these issues, we explore the representation of inappropriate concepts in the image space and guide them towards more suitable ones by injecting *growth inhibitors*, which are tailored based on the identified features related to inappropriate concepts during the diffusion process. Additionally, due to the varying degrees and scopes of inappropriate concepts, we train an adapter to infer the corresponding suppression scale during the injection process. Our method effectively captures the manifestation of subtle words at the image level, enabling direct and efficient erasure of target concepts without the need for fine-tuning. Through extensive experimentation, we demonstrate that our approach achieves superior erasure results with little effect on other normal concepts while preserving image quality and semantics." }, "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": [ "Stable Diffusion", "Text-to-Image Generation", "Concept Erasure" ] }, "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/842a5bece4fc90836813be8278e61288cc493a4e.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": "Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion 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": 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": "- As selecting appropriate values for topN and alpha is crucial for achieving the desired behavior shift, how can this be optimally chosen for different archs, datasets and tasks? This can also be a drawback?\n- For the intervention, is the adjustment applied to all neurons or limited to those in the final layers? Visualizing the number of neurons impacted and identifying the specific layers they belong to would provide valuable insights\n- Also could you please explain what is meant by \"spatial\" differentiation in this context" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper tackles a highly relevant problem in the field of LLMs—understanding and controlling the behaviors of memorization and generalization.\n- The paper is clearly written, with a logical flow. The research questions and objectives are well-defined, making the study’s purpose and approach easy to follow.\n- The authors construct specialized datasets that distinguish between memorization and generalization. This setup provides a good base for analysis" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores memorization and generalization in large language models (LLMs), inspired by the functional specialization observed in the human brain. The authors want to determine whether LLMs show differentiation among neurons when performing memorization or generalization, then use techniques to predict which behavior the model is likely to perform based on activations. And, lastly, implement inference-time interventions to direct models towards memorization or generalization as needed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The hypothesis that certain neurons control specific behaviors based on brain function? While inspired by brain functionality, the paper doesn’t fully substantiate or utilize the correlation to neuroscience. \n\n- The method relies on having a specialized dataset to differentiate between behaviors, which may limit practical applicability. Furthermore, the study only considers a single task; in real-world applications, models are often fine-tuned across multiple tasks, which may affect behavior control.\n\n- Is the behavior shift strictly binary, meaning that applying a shift immediately moves the model to the other state (e.g., from memorization to generalization)? \n\n- The focus on neuron-level interventions may be too granular. As this requires identification of neuron behavior using custom datasets. Exploring higher-level interventions, such as prompting or input changes to toggle memorization or generalization, might be more relevant? Thoughts?" }, "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": "Could you clarify the experimental procedure? In particular, I am confused about: \n \n1. “The training process involved presenting each instance in its entirety to the LLM rather than as a QA pair. “ Are you finetuning with a causal language modeling objective? Could you give some concrete examples?\n \n 2. “During training, we continuously monitored the model’s ability to perform memorization and generalization on a test set and preserved the model once it demonstrated both behaviors. “ Could you explain why you use this behaviors instead of (1) train until convergence (2) train until performance has reached a maximum on a validation set. The current choice seems an unnatural one compared to the usual practice." }, "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 aims to address an important question — whether the model engages in memorization or generalization. The question is of great importance for developing trustworthy AI and has broad applications, such as privacy-sensitive LLMs. \n\n2. The analysis goes beyond simply “passively” detecting distinct neuron activation patterns; it also includes “actively” steering the model toward specific behaviors. By combining these approaches, the authors effectively differentiate the mechanisms LLMs use when engaging in memorization versus generalization, offering a more comprehensive understanding of these behaviors." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper uses two synthetic dataset to detect whether a model engages in memorization or generalization behaviors. The analysis is done in three stages: (1) determining if neurons in LLMs differentiate spatially for memorization versus generalization, (2) predicting these behaviors based on hidden states, and (3) influencing LLMs to switch between these behaviors through real-time interventions. They find that deep layers are responsible for the two distinct behaviors and show that model behaviors can be controlled." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The study primarily focuses on relatively small models (GPT-2 and GPT-2-medium) and small synthetic datasets. For a paper aiming to establish an empirical pattern, a more comprehensive analysis across a range of model sizes and a broader set of tasks—ideally including less synthetic, real-world tasks—would strengthen the findings and increase confidence in the generalizability of the results.\n\n2. The authors designed two synthetic datasets to observe distinct behaviors in models, serving as indicators of memorization versus generalization. However, the patterns identified appear limited to these specific datasets. To effectively demonstrate (a) whether LLMs exhibit spatial differentiation of neurons for memorization and generalization, and (b) the predictability of these behaviors based on internal representations, it would be important to show that the observed spatial differentiation generalizes across tasks (pattern in task A generalize to task B). From the current experiments on two tasks with two different models, the main unifying insight seems to be that behavior differentiation occurs in the later layers, which may not fully establish the robustness of the findings across varied contexts.\n\n3. The method of detecting differences in hidden layer activations is relatively straightforward, and as noted in the first point, the findings may be limited in their novelty and broader applicability. It remains unclear whether the paper offers substantial new insights or practical implications at scale for advancing our understanding or control of LLM behaviors." }, "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. Did you train the GPT-2/GPT-2-Medium from random initialization? Or just fine-tune the model with designed datasets?\n2. Since the re-phrase the input prompts still introduce additional variables, did you try forward the same inputs multiple times and observe whether the language models behave differently (i.e., memorize or generalize)?" }, "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-writen and very easy to follow.\n2. the key research question in the paper \"whether the generalization behaviours and the memorization behaviours of language models (given similar inputs) correlates to some specific regions of neuron values inside the models?\" is quite important and interesting as well.\n3. the experiments conducted in the paper are multi-faceted, cross-validating the presented results of each part." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work investigates the research question: whether the generalization behaviours and the memorization behaviours of language models (given similar inputs) correlates to some specific regions of neuron values inside the models? With GPT-2 and two specific task settings, the authors manage to identify such regions, leverage them to predict models' behaviours (generalizing versus memorizing) and control models' behaviour through inference-time intervention methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiments (regarding the two synthetic dataset settings, the language model scales) is quite limited, making the reviewer question the generality of the conclusions derived in the paper.\n2. The technical contribution of the paper is also limited (e.g., in the third experiment part, ITI is mostly from Li et al., 2024). The causal analysis part is also adopted in many previous works. Though authors leverage them to investigating this new research question, this point stil weaken the overall contribution of 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": 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 model’s overall performance after training on the two datasets? Does the introduction of memorization patterns influence the model performance?\n\n- The paper computed the NMD for each neuron, have the authors considered the influence of the cluster or interactions between neurons in the behavior change?\n\n- What does “other” mean in Table 1 and table 2?" }, "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 challenging and interesting problem in LLM research.\n- The paper has a very clear research question and a well-defined study domain, enabling a focused investigation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the differentiation between memorization and generalization mechanisms in LLMs. By designing specific datasets, the authors observe the neurons in LLMs show different behaviors associated with memorization and generalization, and they analyze this differentiation in view of neuron-wise mean difference. In addition, the authors build classifiers to categorize memorization behavior and generalization behavior, enabling controlled adjustments (towards memorization or generalization) to the LLM’s output during inference." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the definition of memorization is clear in this study, the definition of generalization remains somewhat ambiguous. The author defined in the paper “generalization involves generating outputs through correct reasoning”, but how to define “correct reasoning”? For example, if a response shows partial memorization, does it still in the case of generalization? The authors distinguish between memorization and generalization only based on two simple datasets, making the scope of the study limited. \n\n- The paper assumes a strict distinction (as they design a binary classifier) between memorization and generalization based on the LLM’s output. However, in reality, LLMs often generate answers based on both mechanisms. In addition, the construction of the dataset may lead the model to overfit memorization patterns, which increases the bias of the model.\n\n- The rate of the behavior change in response to the designed pairs are relatively low (11% for in-context inference and 8.5% for arithmetic tasks). Although the author mentioned that one \"could continuously generate different test instances to collect the desired pairwise representations\". However, dose this still support the results? or if it introduces the risk of random guessing. \n\n- The paper uses \"neuron-wise mean difference\" to show neurons behavior change according to different pairs, however, a detailed analysis is missing, such as how the values correlate with memorization or generalization mechanism. Does the behavior change specific to models and dataset or rather applicable towards general models and datasets?\n\n- In addition to the point above, the paper only evaluate each dataset on one model, it is not sure the results and the intervention are model-specific or broadly applicable.\n\nMinor:\n- The font in Figure 3 and Figue 4 is too small to read." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We explore memorization and generalization in LLMs, showing neuron-wise differentiation and successfully predicting and controlling these behaviors through specialized datasets, classifiers, and interventions." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024think,\ntitle={Think or Remember? Detecting and Directing {LLM}s Towards Memorization or Generalization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w4gkS9RsWh},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we study fundamental mechanisms of memorization and generalization in Large Language Models (LLMs), drawing inspiration from the functional specialization observed in the human brain. Our study aims to (a) determine whether LLMs exhibit spatial differentiation of neurons for memorization and generalization, (b) predict these behaviors using internal representations, and (c) control them through inference-time interventions. To achieve this, we design specialized datasets to distinguish between memorization and generalization, build up classifiers to predict these behaviors from model hidden states and develop interventions to influence the model in real time. Our experiments reveal that LLMs exhibit neuron-wise differentiation for memorization and generalization, and the proposed intervention mechanism successfully steers the model's behavior as intended. These findings significantly advance the understanding of LLM behavior and demonstrate the potential for enhancing the reliability and controllability of LLMs." }, "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": [ "LLM", "generalization", "memorization", "neuron differentiation", "behavior identification", "inference-time intervention", "behavior control" ] }, "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/4f522fcadeff8a9497091fdfec4ba847a70b64d8.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": { "value": "/attachment/d9cc11a872598cca306eee65d9e3aad6cdab8892.pdf" }, "title": { "value": "Think or Remember? Detecting and Directing LLMs Towards Memorization or Generalization" }, "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": "The paper claims that its main advantage is being training and optimization free. However, the proposed method is quite simple, and has many alternatives to try out. For example, using cross image attention, anydoor, and many others to pass appearance in the correct location. \nThere are many design choices that are not clearly justified. \nFirstly, why train a simple segmentation model instead of using one off-the-shelf, or training an efficient network with a teacher-student approach from an off-the-shelf model? Indeed, there are many artifact that seem to be from a bad mask, such as residuals of the original source garments.\nMore importantly, the generation of the source garment is agnostic of the target environment, making it incompatible in terms of lighting etc. There is a short discussion on the matter, but as an adhoc solution that does not address the main issue. \nLastly, the source deformation is rather limited, meaning that different poses would be hard to satisfy. Experiments with different poses should be added, and compared to alternatives." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Simple and efficient method\nMixes image-space deformation with neural generation, which is refreshing\nTraining free method" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a method for training-free virtual try-on. The input is a target image of a person and a source garment image, and the output should be the same target identity and environment, but with the new garment. The main insight is injecting features from the garment to the target image while it is being generated with a diffusion model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Results are underwhelming\nLimited innovation\nNo user study\nCan handle simple cases with simple background and lighting" }, "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": "- It would be great if the authors could provide a comparative analysis of accuracy, computational efficiency, and resource requirements for their proposal of a lightweight CNN against off-the-shelf segmentation methods like SAM, SAM2 or SCHP. \n- Many papers perform a user study as this shows the actual use case of whether customers like the images they are being shown by this method vs state-of-the-art. For e.g. the user study in MM-VTON [4] display results from different methods to the users and ask them to \n> either select the best result or opt for “hard to tell.” \n\nAdditionally, users can also be asked for perceived realism of the results compared to other methods. \n- I believe SSIM/LPIPS being slightly lower/higher is not an issue (in my humble opinion) as the main argument of this paper is resource-constrained setup. However, results need more backing. Specifically, comparisons to different off-the-shelf segmentation methods. A table on accuracy vs resource trade-offs with different setups (e.g. Precise Apparel Localization, attention maps, SAM, SAM2, SCHP ) will help the reader understand what makes their proposal more robust.\n\n[4] https://arxiv.org/abs/2406.04542" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The strength of this paper lies in their argument for a resource constrained setup to achieve virtual try-on." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a virtual try-on method which includes a lightweight CNN to predict masks and then infusing this information with DDIM inversion." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In my humble opinion, the biggest drawback lies in the assumption this paper makes about not using accurate off-the-shelf segmentation methods. SDXL base model which is likely used in this paper has around 3.5 billion parameters, while a SAM-B [1] has around 94.7 million parameters and SAM2 [2] Hiera Tiny has around 38.9 million parameters. These segmentation methods are significantly less computationally intensive as the base diffusion model. In my understanding of the proposal here, the segmentation requires a single forward pass, the outputs of which is used in the different steps of the diffusion model. \n\nThe paper has several further drawbacks:\n - Using an off-the-shelf segmentation method bypasses training anything completely. Comparing the lightweight CNN against SAM, SAM2 or SCHP [3] would help show the advantage of the proposal. \n - Comparing the accuracy of masks in Figure 2 is done against attention maps. It would be insightful to the reader if the comparison was against more reliable segmentation methods like SAM, SAM2 or SCHP inspite of their additional resource requirements. \n - Lacks specific details on many information (I have marked them as questions in the next sections). \n - Which method was used in 2-clustering mentioned in Eq. 7? \n - Reshaping of the masks shown on the top-right of in Figure 3c. Why and how is this reshaping done? It is mentioned in the paper that \n> By reshaping the contents of Bg to the size of Bm and applying appropriate perspective transformations to simulate image rotation, we align the target garment with the position and size of the clothing in the model image. \nIt is unclear how this transformation was achieved. Why not use the M^m directly since thats the only place that needs to be replaced. \n - Following the previous question, M^g is used in Eq. 8 which is the garment mask. It seems unclear why M^g is used since it does not align with the model image.\n\n[1] https://arxiv.org/abs/2304.02643\n[2] https://arxiv.org/abs/2408.00714\n[3] https://arxiv.org/abs/1910.09777v1" }, "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": "No ethics review is needed." }, "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. We observed background distortion and altered color without restoration. However, according to Eq. 8, the background and foreground are treated equally, does this mean that the clothing is equally likely to experience severe distortion? There is no discussion in the article concerning this issue." }, "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 addresses the significant training cost issue of existing virtual try-on models and investigates low-cost methods for conducting try-ons, a valuable subject. The authors explore a viable approach to perform try-ons through the replacement of latents. Unfortunately, the method seems somewhat rudimentary, and there is room for improvement in the quality of the results.\n\n2. The paper is well-articulated, and the data presented is credible." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper reviews the state of current virtual try-on technologies and identifies a key issue: high-quality and realistic image generation rely on resource-intensive models trained on extensive datasets, with existing diffusion-based methods involving up to 800 million parameters. This results in considerable training costs and computational demands that may deter users from trying on clothing with regular poses. Current methods also require diverse inputs that may be burdensome for non-professional users. In contrast, the paper introduces DiffusionTrend, a new approach that simplifies the virtual try-on process by forgoing extensive training on large datasets and avoiding complex preliminary input processing steps. The proposed method uses a lightweight CNN for feature-level clustering, producing effective masks without intensive training, and applies DDIM inversion to integrate detailed garment features into model images with background coherence." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The statement in line 161 \"Traditional segmentation models (He et al., 2017; Kirillov et al., 2023) are typically employed to generate masks. However, in environments where numerous users engage in online consultations simultaneously, computational efficiency becomes crucial. For instance, Segment Anything (Kirillov et al., 2023), while effective, incurs a significant computational cost , which can be impractical for real-world applications, especially when scalability and cost-effectiveness are paramount\" is inaccurate. In fact, input model images and input garment images can be automatically preprocessed through existing pretrained human-parsing models to obtain denpose, openpose, parsing images, agnostic images, etc. Please refer to the code [1][2]. Large segmentation model such as Segment Anything is NOT used. In addition, the computational overhead of image segmentation is very small compared to the diffusion denoising process.\n\n2. There are many existing warping methods, yet this work has chosen to use a simple network to observe the mask and employ perspective transformations to deform the clothing. This does not simulate the myriad of clothing deformations that occur in a real try-on process, often leading to ill-fitting try-ons. In fact, since this work is divided into the warp cloth and diffusing process parts, it's necessary to show and discuss the effects of the warp cloth and beneficial to compare it with explicit warping models.\n\n3. The inference speed of this method should be much slower than typical methods (possibly three times the duration), if I have not misunderstood, as the DDIM inversion takes nearly the same number of steps as the DDIM sampling.\n\n4. The qualitative comparison in Fig. 4 appears to show that the results produced by DiffusionTrend have noticeably oversaturated colors.\n\n[1] https://github.com/sangyun884/HR-VITON/issues/45 \n\n[2] https://github.com/levihsu/OOTDiffusion" }, "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": 3 }, "primary_area": null, "questions": { "value": "Please see weaknesses above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "(+) The writing is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a simple approach for VTON that does not require significant training in diffusion models. The main idea is to train a simple mask generator and replace the corresponding garment region in a given human image with a given garment image in the diffusion process. The experimental results are worse than the state-of-the-art methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(-) The motivations are defective. i) Most SOTA methods only fine-tune the diffusion models and are not too computationally extensive. ii) The densepose/segment map/clothes-agnostic representations/keypoints are not additional inputs, but intermediate results from existing tools, and therefore do not complicate use at all. iii) There are many existing human parsing models (no need to use SegmentAnything) that are powerful and lightweight, which can be obtained by googling the keyword \"human parsing model\".\n\n(-) There is little technical novelty. The training of a simple segmentation network (Sec. 3.2) is trivial and as mentioned above, unnecessary. The editing through masking & copying strategy (Sec. 3.3) is also trivial in diffusion-based image editing (VTON is a niche of general image editing). \n\n(-) The results are much worse than the state-of-the-art methods and are not meaningful. Specifically, due to the masking and replication strategies used in the proposed method, the clothes are not really \"worn\" on the body, but rather look like images attached to the body. For example, in Fig. 4, row 1: the wrinkles of the clothes are lost; row 3: the edge of the shorts are not removed and break the shape of the dress; row 4: the result ignores the body shape of the lady.\n\n(-) The evaluation is limited. The types of clothes tested are much fewer than SOTA. All the examples have a simple background.\n\nTherefore, the contributions of this paper are poor." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024diffusiontrend,\ntitle={DiffusionTrend: A Minimalist Approach to Virtual Fashion Try-On},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w5Q3r8Jq3v},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we introduce DiffusionTrend, a pioneering approach for virtual fashion try-on that forgoes the need for training diffusion models, thereby offering simple, conventional pose virtual try-on services with significantly reduced computational overhead. By leveraging advanced diffusion models, DiffusionTrend harnesses latents rich with prior information to capture the nuances of garment details. Throughout the diffusion denoising process, these details are seamlessly integrated into the model image generation, expertly directed by a precise garment mask crafted by a lightweight and compact CNN. Although our DiffusionTrend model initially demonstrates suboptimal metric performance, our exploratory approach offers several significant advantages: (1) It circumvents the need for resource-intensive training of diffusion models on large datasets. (2) It eliminates the necessity for various complex and user-unfriendly model inputs. (3) It delivers a visually compelling virtual try-on experience, underscoring the potential of training-free diffusion models for future research within the community. Overall, this initial foray into the application of untrained diffusion models in virtual try-on technology paves the way for further exploration and refinement in this innovative field." }, "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 Model; Image Editing" ] }, "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/4394c7a55589f3bbee1cb95967c7d62262b0bad4.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/41ebb0c3e40b374936d0baa983d55fbfe07c01f9.zip" }, "title": { "value": "DiffusionTrend: A Minimalist Approach to Virtual Fashion Try-On" }, "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": 3 }, "primary_area": null, "questions": { "value": "See 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 proposed method is novel and practical in RAG scenarios.\n- This manuscript is clearly written and easy to follow.\n- The experimental results are well conducted, showing advantages in terms of accuracy and efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel framework to enhance the performance of NLI models in verifying retrieved evidence within retrieval-augmented generation settings. The paper addresses the issue of performance drop in out-of-domain inputs. To alleviate this problem, the authors propose an automatic generative domain adaptation method to fine-tune NLI models in an unsupervised manner. In the proposed method, this framework considers both diversity and quality by using a sequential augmentation technique and optimizing a distribution-matching objective in the data generation process. Experimental results demonstrate that the NLI model fine-tuned with the proposed method achieves performance closer to that of LLM models without sacrificing efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper would benefit from a more detailed analysis to clearly demonstrate the robustness of the proposed method across various domains. In real-world scenarios, domain boundaries are often ambiguous, and it’s common to encounter mixed or overlapping domain data. By evaluating the model in a multi-domain or domain-mixing setting, the authors could provide stronger evidence of its robustness and practical applicability in complex, realistic cases.\n\n- It would be valuable to include an analysis of how the model performance depends on the quality of the initial synthetic data generation. If initial data is inaccurately generated, it might negatively influence the model’s performance. An examination of whether the model can correct or adapt to potential errors in this initial phase would clarify the method’s resilience to suboptimal synthetic data.\n\n- The paper’s selective objective function appears complex, suggesting that the model could be highly sensitive to hyperparameter choices. Observing large variations in hyperparameter values for each dataset implies that tuning these parameters may be challenging. Providing further insights into the model’s hyperparameter sensitivity and offering guidelines for tuning could improve the approach's usability and reliability in diverse 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": 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": "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 proposed Automatic Generative Domain Adaptation (Auto-GDA) offers a novel approach to unsupervised domain adaptation, effectively generating high-quality synthetic data to fine-tune NLI models tailored for specific RAG contexts.\n- Empirical results demonstrate that models fine-tuned with Auto-GDA significantly outperform weak teacher models and achieve performance levels comparable to those using human-labeled data, indicating its effectiveness in improving NLI model accuracy.\n- The paper is well written and easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the issue of hallucinations in Large Language Models (LLMs) used in retrieval augmented generation (RAG) applications, where verification of generated outputs through natural language inference (NLI) models is essential. The authors propose Automatic Generative Domain Adaptation (Auto-GDA), an unsupervised framework that generates high-quality synthetic data to fine-tune lightweight NLI models for specific RAG domains. Key contributions include formalizing the unsupervised domain adaptation problem, developing the Auto-GDA framework for efficient sample selection, and demonstrating that models fine-tuned with Auto-GDA outperform weak teacher models and approach the performance of human-labeled references, all while achieving significantly lower latency than LLMs. This work presents a promising solution to enhance NLI model performance in real-time RAG applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The effectiveness of the Auto-GDA framework relies heavily on the quality of the synthetic data generated. Poorly generated data could lead to suboptimal fine-tuning and negatively impact NLI model performance.\n- Although the framework aims to address domain mismatches, (in my own opinion) there may still be challenges in generalizing to highly diverse or previously unseen domains, potentially reducing the model's effectiveness in broader applications. I'm curious how well the proposed method performs under such extreme conditions?" }, "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. What is the time of running each step in Auto-GDA (steps 1-2-3 and NLI model training), for different datasets (with their sizes)?\n2. One of the arguments for the necessity of domain adaptation in the introduction is that “inputs may follow a specific format due to the RAG prompt template” (line 79). Why not pass evidence and claims to NLI models without this template, if it reduces domain shift and hence improves performance?\n3. What criteria is used to select optimal hyperparameters using optuna? Is it performance on some dataset (which one)?\n4. How do you tune hyperparameters of unsupervised domain adaptation baselines?\n5. Are there particular reasons why Vectara-2.1 outperforms AlignScore on RAGTruth and vice versa on other datasets? E.g. due to some specific training data or algorithm specifics.\n6. Do you plan to release open-source code for the proposed approach?\n\n\nComments\n* Due to the high amount of notations, it may be hard to follow the method sometimes, e.g. trying to remember what a particular notation means. \n* Line 43: “even when the most capable LLMs are used with RAG, hallucination rates of 20 – 30% persist (Santhanam et al., 2021).” too old reference\n* Figure 1: “RAG task performance (ROC-AUC).” Unclear x label: it seems that it is RAG performance (unclear how measured with ROC-AUC), but it is NLI performance for RAG as far as I understand\n* Line 233: “we also generate realistic initial samples using LLMs”. Use a more specific term than “samples”\n* Line 384: “Optimizing the objective for a subset Q”. What does Q refer to here?\n* Line 429: better define “complex” category" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* A relevant research direction of adapting RAG components to user domains\n* Comprehensive related work section\n* Detailed description of the proposed approach, used datasets, baselines, and experimental details\n* The proposed method is compared to a series of existing NLI solutions on several datasets, and inference time is also compared for various methods" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper considers the problem of domain adaptation of the natural language inference models (NLI) which are often used in retrieval-augmented generation (RAG) to judge the entailment of the generated response from the retrieved context. The paper proposes Auto-GDA, a method for generating synthetic data for a given domain, that can be used for finetuning an NLI model to improve its performance in this domain. The proposed method is iterative and consists of three steps: (1) seed data generation using an LLM; (2) data augmentation e.g. using paraphrasing or sentence deletion; and (3) data filtering, to minimize their proposed enhanced distribution matching objective; steps 2 and 3 can be repeated iteratively. The method is tested on several datasets that provide human labels for NLI, and compared versus existing off-the-shelf systems and several ablations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presented Auto-GDA method is rather complex (involving multiple steps and components, including heuristic augmentation techniques) and has several important hyperparameters, such as $\\lambda_d$ and $\\lambda_u$ in eq. (2), population sizes M and K, or the number of iterations. Hyperparameter tuning requires running the proposed approach 50 times (line 446), including training of the NLI model on the generated data (from my understanding). At the same time, improvements over simply using LLM-generated data are quite modest (row 4 vs 2 in Table 2). \n - Furthermore, the high cost of running hyperparameter optimization, augmentation, and data selection in the proposed approach, motivates the substantial increase of data points in the simple baseline of using LLM-generated data, to make their computational costs similar. This would make the baseline stronger and reduce its difference versus the proposed approach even further.\n - Performance gains versus out-of-the-box NLI models are also rather small, e.g. comparing the best performing (domain-adapted) Auto-GDA versus the best performing “complex” method out-of-the-box (83.7 vs 80.5, 86.7 vs 85.4, 92.5 vs 90.4, 88.3 < 89.4), or Auto-GDA (Flan-T5) vs MiniCheck-T5 (75.6 \\approx 75.4, 68.7 < 74.1, 82.4 vs 79.1; only one high improvement 78.3 vs 64.0). \n2. I would expect more ablations of the proposed approach, e.g. testing the removal of each of the terms in eq. (2), or on the contrary using only one of these terms for filtering. \n3. The motivation for the proposed approach is to improve RAG pipelines in domain-specific scenarios, however no experiments with domain-specific RAG are presented to demonstrate these improvements. For example, domain-specific RAG scenarios considered in [1] could act as potential testbeds, e.g. RobustQA [2].\n\n[1] Dongyu Ru et al. RAGChecker: A Fine-grained Framework for Diagnosing Retrieval-Augmented Generation. NeurIPS 2024\n\n[2] Rujun Han et al. RobustQA: Benchmarking the Robustness of Domain Adaptation for Open-Domain Question Answering. Findings of ACL 2023" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present an automatic strategy to adapt NLI models for grounding verification to challenging inputs encountered in realistic RAG systems using synthetic data." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024autogda,\ntitle={Auto-{GDA}: Automatic Domain Adaptation for Efficient Grounding Verification in Retrieval Augmented Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w5ZtXOzMeJ},\nnote={under review}\n}" }, "abstract": { "value": "While retrieval augmented generation (RAG) has been shown to enhance factuality of large language model (LLM) outputs, LLMs still suffer from hallucination, generating incorrect or irrelevant information. One common detection strategy involves prompting the LLM again to assess whether its response is grounded in the retrieved evidence, but this approach is costly. Alternatively, lightweight natural language inference (NLI) models for efficient grounding verification can be used at inference time. While existing pre-trained NLI models offer potential solutions, their performance remains subpar compared to larger models on realistic RAG inputs. RAG inputs are more complex than most datasets used for training NLI models and have characteristics specific to the underlying knowledge base, requiring adaptation of the NLI models to a specific target domain. Additionally, the lack of labeled instances in the target domain makes supervised domain adaptation, e.g., through fine-tuning, infeasible. To address these challenges, we introduce Automatic Generative Domain Adaptation (Auto-GDA). Our framework enables unsupervised domain adaptation through synthetic data generation.\nUnlike previous methods that rely on handcrafted filtering and augmentation strategies, Auto-GDA employs an iterative process to continuously improve the quality of generated samples using weak labels from less efficient teacher models and discrete optimization to select the most promising augmented samples. Experimental results demonstrate the effectiveness of our approach, with models fine-tuned on synthetic data using Auto-GDA often surpassing the performance of the teacher model and reaching the performance level of LLMs at 10 % of their computational cost." }, "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": [ "domain adaptation; NLI; RAG; document-grounded; NLP;" ] }, "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/cda65b6e1349ba78e184d3c7352037cbc5bd3546.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/7d11bc86c137056477fbce4e594ea934dbdaadec.zip" }, "title": { "value": "Auto-GDA: Automatic Domain Adaptation for Efficient Grounding Verification in Retrieval Augmented 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": null, "code_of_ethics": null, "comment": { "value": "We sincerely thank you for your constructive and detailed feedback.\n\nW1. Regarding novelty\nThank you for bringing these papers to our attention. While these works make valuable contributions to k-estimation, we believe our approach offers unique advantages:\n- Linear time complexity (compared to quadratic)\n- Scalability to longer sequences\n- Fewer distributional assumptions\nWe will revise our introduction to better position our work within this context.\n\nW2. Regarding performance metrics\nWe appreciate your careful reading of our results. The performance differences between BoW and TF-IDF are indeed subtle but meaningful. We will enhance our presentation with:\n- Confidence intervals\n- Statistical significance tests\nThis will help readers better understand the practical implications of these differences.\n\nW3. Regarding parameter stability\nThank you for this excellent suggestion. While our extensive experiments demonstrate robustness across various parameters, we agree that a more detailed analysis would be valuable. We will add comprehensive sensitivity studies while noting that default parameters often perform well.\n\nW4. Regarding failure cases\nWe greatly appreciate this suggestion and will add detailed failure case analysis. This will help practitioners better understand when to apply our method.\n\nMinor issues:\nWe are grateful for your attention to detail and will address all formatting and citation issues." }, "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": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "We are grateful for your detailed and thoughtful review.\n\nW1. Regarding silhouette coefficient justification\nThank you for pushing us to better justify our choice of the silhouette coefficient. While our theoretical analysis in Section 4 supports this choice, we agree that expanding our explanation would strengthen the paper. We plan to:\n- Enhance the theoretical analysis\n- Add empirical validation\nWe believe these additions will better demonstrate why silhouette coefficients are particularly well-suited for our vector representations.\n\nW2. Regarding empirical evaluation\nWe appreciate your suggestion for broader evaluation. While our current evaluation focuses on our core contribution (linear-time k-estimation), we agree that additional comparisons would be valuable. We will:\n- Add direct SPF comparisons\n- Include detailed runtime analysis\nWe would welcome suggestions for additional comparisons that maintain linear time complexity.\n\nW3. Regarding performance metrics\nThank you for this suggestion. While our metrics directly address k-estimation accuracy, we agree that adding standard clustering metrics would enable better comparison with existing literature. We will add NMI and ARI metrics.\n\nW4/W5. Regarding related work and citations\nWe appreciate your careful attention to the citations. We will revise the related work section accordingly." }, "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": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "We sincerely thank you for your thorough review. Your feedback will help improve our paper.\n\nW1. Regarding technical depth\nWe greatly appreciate this concern and would like to clarify our technical contributions, which we believe are substantial:\n- A novel theoretical framework with rigorous proofs showing why silhouette coefficients succeed on our symbolic representations while failing on raw time series\n- A mathematically grounded optimization framework for joint parameter selection\n- An innovative adaptation of TF-IDF that preserves temporal characteristics\nWe would be happy to expand these sections to better communicate the technical depth of our work.\n\nW2. Regarding comparison methodology\nThank you for this insightful observation. Our vector space transformation enables principled comparisons across different distance measures, though we agree this could be better explained. We plan to:\n- Add formal analysis of distance metric properties\n- Include empirical validation of our choices\nWe believe this will help readers better understand our methodological decisions.\n\nW3. Regarding baselines\nWe appreciate your suggestions for additional baselines. Our current selection focused on methods commonly used in production systems, though we agree some additions would be valuable. We plan to add:\n- K-means objective function optimization\n- Davies-Bouldin Index\nWe would be grateful for specific suggestions of other baselines that maintain linear time complexity.\n\nW4. Regarding citation issues\nThank you for catching these issues. We will certainly fix the citation formatting." }, "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": 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": "W1. Lack of technical depth\n\nThe paper combines existing ideas for solving this problem. Therefore, the technical depth is low, even though the combination of these ideas might be novel.\n\nW2. Unclear how different methods/distances can be compared\n\nIt's unclear how this comparison is meaningful when we need to compare methods relying on different distances. The paper does not clearly articulate how such distances affect the results and it mainly shows results for SAX variants (so inherently for euclidean distance)\n\nW3. Missing potential baselines\n\nSimple baselines, like assign the objective functions of k-means like algorithms are missing. Also there are tons of variants for internal clusteirng validation. Why Silhouette ?\n\nW4. Duplicate references or wrong references\n\nMany references are duplicates. Other references does not exist\n\nduplicates\nXiaosheng Li, Jessica Lin, and Liang Zhao. Linear time complexity time series clustering with\nsymbolic pattern forest. In IJCAI, 2019a.\nXiaosheng Li, Jessica Lin, and Liang Zhao. Linear time complexity time series clustering with\nsymbolic pattern forest. IJCAI, 2019b.\n\nduplicates\nJaewon Yang and Jure Leskovec. Patterns of temporal variation in online media. In Proceedings of\nthe Fourth ACM International Conference on Web Search and Data Mining, 2011a.\nJaewon Yang and Jure Leskovec. Patterns of temporal variation in online media. In Proceedings of\nthe fourth ACM international conference on Web search and data mining, pp. 177–186, 2011b.\n\nit's wrong\nJohn Paparrizos, Paul Boniol, Themis Palpanas, Ruey S Tsay, Aaron Elmore, and Michael J\nFranklin. Fast and exact time series motif and discord discovery in trillions of data points. The\nVLDB Journal, 31:1079–1101, 2022." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1. Timely and important problem especially due to the rise of IoT applications and the need for unsupervised data exploration\nS2. Simply and intuitive ideas\nS3. Results support the overall claims in the paper" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes SPF, a methodology that identifies the number of clusters for time-series data, often a critical parameter for subsequent routines and clustering methods. The idea combines concepts such as SAX, TF-IDF vectors over SAX representations and relies on the Silhouette coefficients to calibrate the number of clusters. Experimental results on several UCR datasets demonstrate the potential of this solution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. Lack of technical depth\nW2. Unclear how different methods/distances can be compared\nW3. Missing potential baselines\nW4. Duplicate references or wrong references" }, "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": "N/A" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "S1. The paper addresses the relevant problem of automatically determining the number of clusters.\nS2. The empirical evaluation makes use of a large number of benchmarking datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript presents an extension of the symbolic pattern forest (SPF) algorithm for clustering of time series data. Using bag-of-words on the symbolic representation, TF-IDF vectors are constructed. The best clustering is selected as the one that maximises the silhouette coefficient (SC)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. The method assumes that silhouette coefficient is a suitable metric for finding the best number of clusters, without justifying this choice. This is a major concern as the silhouette coefficient considers (Euclidean) distance to cluster centres, which is not aligned with the clustering objective of the SPF method. The paper should provide justification for using the silhouette coefficient, or discuss potential limitations of this choice given the SPF method's clustering approach. Moreover, the silhouette coefficient is a well-known metric, so it is unclear what the novelty should be.\nW2. The empirical evaluation does not consider the SPF method, but only weak baselines constructed from the proposed method, meaning that the empirical evaluation does not allow assessment of the performance of the proposed method with respect to state of the art. It is important to compare directly to SPF in the experiments, in order to demonstrate improvement over state of the art.\nW3. The empirical evaluation only considers performance metrics accuracy and near-miss-rate, different from other work in the field, and in the SPF paper (e.g. NMI), making it impossible to compare with those works directly.\nW4. The discussion of related work is overly brief, and fails to present clear assessment of the suitability of existing methods and metrics. E.g. Davies-Bouldin Index and its perceived suitability for the task. Also, there is a large body of work on similarity assessment of time series or clustering of time series, e.g. Keogh et al 2005, Rakthanmanon et al 2012, Paparrizos et al 2015. The paper should discuss these, and explain differences and similarities with the proposed method.\nW5. On the other hand, references UTSAD and STGAT seem out of context, as they do not address clustering of time series. The paper should clarify the relevance of UTSAD and STGAT to the proposed work, or remove these references if they are indeed not directly related.\nW6. The paper contains several redundant sections, such as the description of SAX.\nW7. There are some minor issues, such that Li et al 2019 appears twice in the references, there is a typesetting error in the definition of pi_i(T_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": "(Q1) How would you modify the paper to address the issues regarding related work? How does the proposed method compare to other k-estimation approaches on time series data?\n\n(Q2) What effect do the properties of the chosen UCI datasets have on the performance of the k-estimation using the proposed technique, and how does the clustering performance change on them depending on the chosen k? Is the performance of the SPF algorithm with the ground truth k always the best one, or could other parametrizations outperform it?\n\n(Q3) How impactful are the parameters of the proposed method?\n\n(Q4) Is there any advantage to using BoW over TF-IDF (or the other way around)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(S1) Incorporating BoW and TF-IDF with the concepts of the SPF algorithm sounds like a very sensible approach. Both are a good choice for term-based similarity evaluation and are still commonly used in other settings.\n\n(S2) Aside from minor issues, the submission is well-written and easily understandable while providing an extensive overview of the formulas related to the problem.\n\n(S3) The problem setting is significant as k-estimation is a significant part of clustering in general, which also applies to the setting of time series clustering. The usage of SPF is well-founded due to its low complexity. Introducing k-estimation to the approach helps mitigate one of its weaknesses." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The submission proposes an extension to the SPF algorithm, a clustering approach for clustering time series with linear complexity. The extension allows for the automatic determination of the number of clusters. It is done by performing optimization on the silhouette score using either Bag of Words or TF-IDF." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(W1) Novelty: The abstract of the submission makes the claim that there are no time series clustering methods capable of working without the specification of cluster number k. However, such methods exist already:\n\na) “Spectral Clustering for Time Series” by Fei Wang and Changshui Zhang (2005) is able to discover the optimal number of clusters based on the eigenstructure using a threshold on the value of the eigenvalues.\nb) “Clustering Time Series with Hidden Markov Models and Dynamic Time Warping” by Tim Oates et. al. (1999) also provides a way to estimate the number of clusters based on Dynamic Time Warping. However, even if the submission is not the only method that does k-estimation on time series, it is still a valid and useful direction. It also appears to be the only method that does so for the Symbolic Pattern Forest algorithm.\nc) The paper “Trendlets: A novel probabilistic representational structures for clustering the time series data” by Johnpaul C I et al. (2020) uses the Silhouette Score for cluster number analysis for time series as well, though it does so based on hierarchical clustering methods. This paper should be explicitly covered in related work or even a competitor.\n\n(W2) Despite TF-IDF being considered the better of the two proposed strategies, there is no actual description of the performance metrics outside of the graph and the overall relative performance value. Similarly, near misses should be added to the text for BoW. The results of both BoW and TF-IDF are the same in the Tables in the supplementary files, though Figure 1 claims that TF-IDF performed slightly better.\n\n(W3) As the method works by optimizing the silhouette score, both the values for the score and the actual clustering performance with the given parameters should be indicated. While the cluster numbers match, the detected clusters may not necessarily correspond to the actual ground truth clusters, which could further mean that different cluster numbers may lead to a better performance. Furthermore, an analysis of the stability of the parameters should have been performed, especially as the method has multiple parameters, which themselves include an upper and lower bound. Additionally, an intuition behind choosing the parameters should be given if they strongly affect the performance.\n\n(W4) Regarding the actual experiment, a better analysis of the behavior should be done, considering under what conditions the k-estimation of each of the three approaches failed and whether or not a reason behind it could be established. The section on Relative Improvement is redundant as it only recontextualizes prior results, and the space could be used to do a more in-depth result analysis instead. Similarly, the remaining 2 pages could have been used for this.\n\n(W6) Neither the parameter w nor the alpha ranges seem to be specified anywhere. The code is unavailable, though it should be possible to reimplement given the information provided. Still, this hampers the reproducibility of the results.\n\n(W7) There should be citations for TF-IDF and BoW. Other papers also do not consistently do it, so it is not a major issue. Nonetheless, it would have been better if it had been done. Furthermore, UCI should be cited upon first mention outside of the abstract, not just at a later point.\n\nMinor Issues:\n* Linear time complexity time series clustering with symbolic pattern forest by Li et. al., is cited twice as 2019a and 2019b despite referring to the same paper \n* The formatting appears to be broken for lists, as they are just written in a line without comma separation (see line 291 and lines 314-315) \n* A similar issue happened with the variables for the optimization problem, as they are also not properly separated in line 305\n* The subscript on several equations appears to be broken (see (22)/319 and (23)/321)\n* The near miss metric should probably be more dynamic based on the ground truth cluster number, as claiming 2 clusters for a 3-cluster setting seems more problematic than claiming 70 for 71 true clusters. The chosen datasets generally only have a few clusters, so the current definition isn’t problematic for the submission. It may be relevant for the extension to the full UCI database, however. \n* The formulation for Near Misses, as currently given, would also include all correctly determined cluster counts but does not do so in the evaluation." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper introduces the first unsupervised time series clustering method that automatically determines the optimal number of clusters by applying the Silhouette Coefficient to SAX-based vector representations." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024on,\ntitle={On the Convergence of Symbolic Pattern Forests and Silhouette Coefficients for Robust Time Series Clustering},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w5h443GIGo},\nnote={under review}\n}" }, "abstract": { "value": "Clustering algorithms are fundamental to data mining, serving dual roles as exploratory tools and preprocessing steps for advanced analytics. A persistent challenge in this domain is determining the optimal number of clusters, particularly for time series data where prevalent algorithms like k-means and k-shape require a priori knowledge of cluster quantity. This paper presents the first approach to time series clustering that does not require prior specification of cluster numbers. We introduce a novel extension of the Symbolic Pattern Forest (SPF) algorithm that automatically optimizes the number of clusters for time series datasets. Our method integrates SPF for cluster generation with the Silhouette Coefficient, computed on a two-stage vector representation: first transforming time series into Symbolic Aggregate approXimation (SAX) representations, then deriving both bag-of-words and TF-IDF vectors. Rigorous evaluation on diverse datasets from the UCR archive demonstrates that our approach significantly outperforms traditional baseline methods. This work contributes to the field of time series analysis by providing a truly unsupervised, data-driven approach to clustering, with potential impacts across various temporal data mining applications where the underlying number of clusters is unknown or variable." }, "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": [ "Data Mining", "Time Series", "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/b9caba4cf344044787f213a194366d2b7c8f621c.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": "On the Convergence of Symbolic Pattern Forests and Silhouette Coefficients for Robust Time Series Clustering" }, "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.\tIt looks like the design of the function of sorts isn't very new, is it possible to make a better argument for the difference between that scenario and other approaches in other scenarios? What kind of innovative design was done for this scenario? Is there any similar approach in other scenario?\n2.\tThe title of the article is NOISE-ROBUST PREFERENCE LOSSES FOR DEEP REGRESSION MODELS, why is only the data for the airlines used? Are there any specific scenarios where PLAI outperforms or falls short of these methods?\n3.\tCan the code links be given for validation and reproduction of results." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThe paper is well-structured and clearly written. The authors have successfully communicated complex concepts in a manner that is accessible to readers.\n2.\tThe paper raises an interesting question. How does real-world analyst intervention (or ‘analyst influence’) affect the effectiveness of model training, and how does adjusting to the level of quality of the training samples relative to the model output mitigate the impact of rough training examples.\n3.\tThe structure and iconography of the paper is clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a novel method called Preference Learning from Analyst Influence (PLAI), which introduces a weighted loss function that accounts for the relative quality levels of training samples compared to model outputs. The paper includes a detailed theoretical analysis, the formal definition of relative quality, and the proposal of three PLAI loss implementations: Sigmoid PLAI loss, Focal PLAI loss, and Clip PLAI loss. The experiments validate the effectiveness of PLAI in improving model performance and alignment with analyst preferences." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe paper is not sufficiently experimental. For example, the richness of the data is insufficient. The paper compares PLAI with several baseline methods, but it could benefit from a comparison with other state-of-the-art methods or recent advances in robust regression techniques. This would provide a more comprehensive. understanding of PLAI's performance relative to the current research landscape.\n2.\tThe background research for the article was insufficient and it is suggested that the literature review section should be expanded to cover more existing work relevant to the research topic. This includes recent research developments, classic papers, and high-quality work that is widely recognized in the field.\n3.\tThe link to the experiment is not given, are the results is not verifiable." }, "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 meaning of Figure 2? Figure 2(a) shows market constraints are the most popular methods, but how does this affect the proposed method? Figure 2(b) shows the influenced price is consistently larger than the raw price, so how does this affect the results?\n\n2. In Equation(1), the author assumes that analyst influence on price follows a Gaussian distribution. Please give some justification for this assumption.\n\n3. The author proposed three variant losses based on Equation 8. Please provide some analysis of these losses, especially how they affect the optimization process.\n\n4. Referred to weakness, more details of experiments should be given." }, "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 investigates a real-world problem and models the problem well. The problem setting is interesting.\n\n2. The proposed method utilizing the analyst influence data is well-motivated in this setting. The PLAI fits the practical problem.\n\n3. Experiments on industry data are well-designed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates a real-world problem: dynamic pricing for airline revenue management. The authors propose the PLAI method, which is simple and easy to follow. Based on the analyst influenced data, this paper discovers the relative quality for deep regression tasks in dynamic pricing. Training with data with analyst influence, this paper conduct extensive experiments to verify the effectiveness of PLAI method,." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The problem is orginated from practical problem, as a result, the background should be stated clearly. The background is not so easy to understand in section 3.3. It is not so clear to interpret Figure 2.\n\n2. The contribution of this proposed method is not so obvious. The proposed method seems to be a continual learning for deep regression tasks. The distinction and novelty of this paper should be emphasized.\n\n3. Baselines only cover MAE, MSE, et al., and some basic losses; more recent works in this field should be compared to make the work more sound.\n\n4. Experiments on open datasets should be conducted to ensure the reproducibility of this work. Table 1 should show statistical significance." }, "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": "The paper reads more like an industrial report than an academic paper. The paper is not ready for publication at ICLR." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is well-organized, and the proposed loss function includes some theoretical analysis" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a weighted loss function for deep regression models. The weight of the loss is determined by the quality of the so-called ‘analyst influence’ on the training data. If the coarsely adjusted analyst data performs well on the training sample, the corresponding loss weight is higher than for those samples where the fit is poorer. The quality of the analyst’s influence is assessed using the existing model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1: The proposed method lacks novelty. Weighted loss functions have been extensively studied, and the potential application and contribution of the proposed PLAI loss function are limited.\n\nW2: The authors spend half a page explaining \"airline revenue management\" and \"Bid price prediction,\" which is a digress from the main subject and does not interest most readers.\n\nW3: There is no comparison with state-of-the-art models; only different loss functions were compared. Additionally, the proposed PLAI loss does not show significant improvement in influence accuracy. Compared to MAE loss, the proposed PLAI improves influence accuracy by only 2%." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024noiserobust,\ntitle={Noise-Robust Preference Losses for Deep Regression Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w5pErXbwQl},\nnote={under review}\n}" }, "abstract": { "value": "Deep regression models are widely employed for tasks such as pricing and forecasting. In industrial applications, it is common for analysts to adjust model outputs before they are deployed in commercial products. These adjustments, which we name \"analyst influences\", not only ensure the quality of the final products but also provide training data to improve model performance over time. However, due to the huge volumes of data, analyst influences can be applied broadly and can lack precision, hindering training effectiveness. To resolve the issue, we propose a novel framework Preference Learning from Analyst Influence which creates a weighted loss function that explicitly accounts for the relative quality levels of the training samples in comparison to model outputs. This approach effectively mitigates the impact of coarse training instances. Our extensive experiments on real-world data drawn from airline revenue management demonstrate that the proposed framework not only enhances pricing stability but also improves alignment with analyst influences compared to baselines." }, "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": [ "Regression", "Robustness", "Alignment" ] }, "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/167cf81901cac6a98b537d9248709d7e3e8e5b56.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": "Noise-Robust Preference Losses for Deep Regression 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": 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." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "S1. The paper is well-written and mostly clear.\n\nS2. The proposed view-wise sampling algorithm is interesting and novel.\n\nS3. Exploiting autoregressive generation to preserve global geometry is reasonable. \n\nS4. The experiments are extensive, especially including various tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a transformer-based diffusion field model to better capture global structures and long-context dependencies. It does that by introducing a view-wise sampling algorithm and incorporating autoregressive generation. The proposed method is a general framework that can be applied to multiple modalities, such as video, 3D and game. Extensive experiments are conducted to validate the effectiveness of the proposal." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. As autoregressive generation is typically slower than parallel generation due to its sequential nature, the authors are encouraged to discuss the inference time of the proposed method and baseline methods.\n\nW2. As shown in Table 1, the proposed method achieves better performance against baseline methods on image and video, but worse FID and LPIPS scores on 3D generation task. The authors are encouraged to discuss this phenomenon." }, "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 weakness. I'm gald to increase my scores if my concerns can be addressed." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. With the view-wise sampling strategy, this method can scale up to high-resolution inputs\n2. By introducing long context conditioning, cross-view consistency can be avoided to some extent" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a novel method for unified video, 3D, and game generation, by learning a DiT based mapping from metric space to signal space, which is able to process high-resolution inputs by the proposed view-wise sampling strategy, as well as maintaining global struture with introduced inductive bias such as text prompts. Results have demonstrated the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty is limited. From my perpective, the method proposed in this paper simply alters the sampling strategy of existing approaches through a straightforward design change, which trades off a reduced number of sampled views for a higher input resolution. Though the introduction of long context conditioning can compensate the global structure, this operation is common, and i don't this operation is powerful enough to recover the information lost during the process of view-wise sampling.\n\n2. Since the method amis to learn a mapping from input coordinates to output properties, i think some other methods should also be compared, such as SIREN[1], and the difference between them should be clarified.\n\n3. I'm wondering that whether the proposed method can be applied to more complex scenes generation, instead of simple objects in the task of 3D novel view synthesis.\n\n[1] Implicit Neural Representations with Periodic Activation Functions, NeurIPS 2020." }, "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": 1 }, "primary_area": null, "questions": { "value": "I am not an expert in Diffusion Probabilistic Fields, and the writing of this paper makes me even more confusing. I hope the authors could improve the writing and explain more background and related work. In addition, most of my concerns are about the explanation of the method and motivation. Please refer to weaknesses for more details." }, "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 method is a unified framework for various modalities. This is a relatively new task that might benefit the community.\n\n2. Extensive experiments and ablation studies demonstrate the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on unifying data generation across various modalities, including images, videos and 3D geometry. It introduces a view-wise sampling algorithm along with autoregressive generation to improve the performance. The proposed framework can handle various modalities with good performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **The motivation is unclear.** From the introduction section, the main motivation to use Diffusion Probabilistic Field (DPF) is handling various modalities together with a unified model. As for the unified models, what I understand is a single model that can generate different modalities. However, from the description in the method and experiment sections, each modality has different coordinate-signal pairs and the models are trained for each modality separately. If so, such a method cannot be regarded as a unified framework in my view.\n\n2. **Comparison with conventional diffusion models.** In Line 142-144, when comparing DPF with conventional diffusion models, the main difference is that DPF can be applied to sparse observation of fields. However, in the view-wise sampling subsection (Line 244), each time sample the tokens in n views, which is a dense modeling instead of sparse sampling. \n\n3. **Comparison with DPF.** In my view, the main contribution of DPF is the context query pairs sampling and optimization. However, in Line 502, this paper mentions that the context query pairs are not used, which confuses me about the training objective in this paper. Does this paper use the diffusion optimization objective like epsilon-prediction or velocity-prediction? If so, the method is almost the same with DiT.\n\n4. **Limited performance.** I do not see a part describing the dataset and hyperparameters used for training. So I assume the model is trained on each benchmark. If so, the performance is far from satisfactory since the compared methods are generalizable ones instead of fitting to a benchmark." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Simple Diffusion Transformer on Unified Video, 3D, and Game Field Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w6YS9A78fq},\nnote={under review}\n}" }, "abstract": { "value": "The probabilistic field models the distribution of continuous functions defined over metric spaces. While these models hold great potential for unifying data generation across various modalities, including images, videos, and 3D geometry, they still struggle with long-context generation beyond simple examples. This limitation can be attributed to their MLP architecture, which lacks sufficient inductive bias to capture global structures through uniform sampling.\nTo address this, we propose a new and simple model that incorporates a view-wise sampling algorithm to focus on local structure learning, along with autoregressive generation to preserve global geometry. It adapts cross-modality conditions, such as text prompts for text-to-video generation, camera poses for 3D view generation, and control actions for game generation.\nExperimental results across various modalities demonstrate the effectiveness of our model, with its 675M parameter size, and highlight its potential as a foundational framework for scalable, modality-unified visual content 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": [ "Diffusion Probabilistic Fields", "World Model" ] }, "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/9dbe2df9ab2ed7168f54174a089561f2a964cd23.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": "A Simple Diffusion Transformer on Unified Video, 3D, and Game Field 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": 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. What makes this model a foundation model? I was of the opinion that this term is generally reserved for extremely large models trained on many datasets.\n2. Using MLPs seems very costly in number of parameters — have the authors considered using e.g., CNNs with temporal convolutions?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed architecture outperforms relevant baselines on the used benchmark.\n2. The paper is well written and easy to understand.\n3. A stochastic variant of the loss is introduced, which lowers the potentially prohibitive cost of the set-theory inspired loss." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce Roll-AE, an autoencoder designed for micro-electrode array data, which uses a set-theoretical inspired loss to obtain shift invariance. Latent embeddings of this model can be used to predict various metrics with better accuracy than latent embeddings of an autoencoder trained with a standard MSE loss." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. All the quantitive evaluation is done “indirectly”, based on a decoder trained on latent embeddings of the different architectures. \n\n\t1.1. How do the models compare in terms of reconstruction loss? It might also be nice to show example reconstructions for the different architectures in the supplementary.\n\n\t1.2. It would be useful to include the statistics of memory and computation time (like Fig. 8) for the baseline architectures.\n2. The baselines were not used in the Treatment Clustering and in the Neural Metrics Credentialing sections. It seems like they could also be used for each (but maybe with lower performance).\n3. In Fig. 5 means are reported, adding error bars might be sensible." }, "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": "Can the authors explain how Roll-AE's computational efficiency scales with larger datasets and suggest additional strategies to reduce computational load without losing performance? Have they tested different binning sizes to better capture detailed neural metrics? How could Roll-AE be adapted for other time-series data like ECG or sensor data? How does Roll-AE handle noise and missing data in real-world MEA recordings? can the authors provide more insights into the interpretability of Roll-AE's embeddings, such as whether specific dimensions correspond to particular neuronal behaviors?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper introduces Roll-AE, a novel autoencoder designed to manage spatiotemporal invariance in MEA recordings, moving beyond traditional methods that rely on hand-crafted features or standard autoencoders. It innovatively addresses temporal shifts and spatial permutations due to electrode symmetries, which are common challenges in in vitro MEA data analysis. The methodology is well-executed, with comprehensive experiments on synthetic and biological data providing strong evidence for Roll-AE's effectiveness. Roll-AE's ability to capture complex neuronal firing patterns and biologically relevant phenotypes without predefined metrics holds potential for advancing research in disease modeling, drug discovery, and understanding neuronal networks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The study introduces Roll-AE, an autoencoder designed to extract consistent features from neuronal activity recordings, particularly from stem cell-derived neuronal cultures. Traditional MEA data analysis is often limited by information loss, dependence on arbitrary settings, and challenges with missing data. Roll-AE overcomes these issues by incorporating invariance to temporal shifts and spatial electrode permutations. It processes sets of cyclic permutations of spike trains and uses an aggregation function to ensure consistent embeddings. The authors show that Roll-AE outperforms standard autoencoders in mimicking various neuronal firing patterns on synthetic datasets. Additionally, Roll-AE effectively captures meaningful phenotypes in neuronal cultures treated with siRNA, excelling in classification tasks, treatment clustering, and predicting traditional neural metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Processing all cyclic and electrode permutations for shift and spatial invariance can be computationally heavy, especially with large datasets or longer spike trains. The authors use stochastic shift-invariance to ease this but could further explore efficiency versus performance trade-offs. The paper lacks a detailed discussion on Roll-AE's limitations, such as potential information loss from permutation aggregation and sensitivity to specific patterns. It would also be helpful to see scenarios where Roll-AE might not excel. While Roll-AE could be applied to other spatiotemporal invariant data, there's no evidence or discussion on adapting it to other contexts like biomedical signals. Additionally, the impact of hyperparameters, such as the shift-sampling rate, on performance and computation could be more thoroughly examined to understand the model's 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": 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": "- Did the authors investigate visualizations (via UMAP etc.) of the latent representations obtained for different temporal shifts of the input?\n\n- Can the authors elaborate why is the proposed Roll-AE architecture declared as a \"foundational model\"?\n\n- Given the computational overhead of Eq (1), training load of such a model should be quantitatively elaborated in a table with comparisons.\n\n- There are currently no comparisons to any existing invariant representation learning methods designed for autoencoding models. Furthermore, invariant representation learning from neurophysiological data [1-3] has also been extensively studied based on various loss- or model- regularization techniques in different contexts. This work lacks such discussions or any comparisons of its methodology with existing methods.\n\n[1] \"Learning time-invariant representations for individual neurons from population dynamics\"\n\n[2] \"Capturing cross-session neural population variability through self-supervised identification of consistent neuron ensembles\"\n\n[3] \"Deep site-invariant neural decoding from local field potentials\"" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Empirically the proposed method appears more effective than vanilla data augmentation based AEs to enforce shift invariance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a representation learning architecture that enforces shift-invariance to time-series spike data during training of a standard autoencoder architecture. The goal of the model is to learn reliable embeddings for different temporal rotations of the input, by performing a set-to-set mapping, through a linear assignment loss. Experiments are performed on microelectrode array data to demonstrate its effectiveness in extracting spatiotemporally invariant features from electrophysiological recordings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Unique contribution of the work is not very clear. It appears like the methodological ML contribution is limited, and is essentially similar to existing 2D/3D orientation-invariance approaches being adapted to spike-train shift invariance.\n\n- Presentation and writing could also be improved. Figures with results are not very clear to read with very small text font sizes. Literature coverage is shallow." }, "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": "### Questions:\n- In the abstract, the authors state the deep learning MEA analysis has mostly been focused on in vivo recordings and not in vitro recordings. Why can't the same methods be applied to in vitro recordings? \n- What are excitability phenotypes (line 32)?\n- Is the loss of resolution temporal or spatial (line 45)?\n- Why would the compression of neural measures impact the quality of phenotype and disease models (line 48)?\n- Why is it important for MEA recordings to be invariant to changes in the orientations of the electrodes (line 71)?\n- What is \"the disease\" in line 391?\n- What are the treatment similarities (line 433)? The authors state that the Roll-AE embeddings can be used to characterize treatment similarities, but I could not find them in the main text.\n- Is the electrophysiological dataset publicly available?\n- Is the source code publicly available?\n\n### Additional feedback to improve your paper:\n- I would reference the Figures in the main text in the order they appear, otherwise, the reader has to jump back and forth between the pages (minor point).\n- The introduction seems very long. Although I appreciate the various backgrounds, it makes it harder to decipher what the paper is trying to address (minor point).\n- Fig. 4a y-label might be wrong. Missing metric used for the scores (e.g. MSE or %).\n- Unclear what Fig. 4b labels are? Also, I would include a colorbar to denote what the colors in the matrices mean.\n- Fig. 5 and 6 font size is really small and it is hard to read. I would increase the font size.\n- Consider using different colours in the bar plots in Fig. 5 as it is hard to see the light-coloured bars against the white background." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The proposed autoencoder model seems novel and the goal of extracting meaningful features from neural recordings is important. The model shows superior performance compared to the other autoencoder models on predicting the neural measures on the synthetic dataset. Furthermore, the clustering of the genes in Fig. 6 is interesting and may prove useful in the development of new therapies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors develop an autoencoder model to analyze the firing patterns of neurons recorded in vitro, to overcome the shortcomings of using standard neural metrics (e.g. firing rates, synchrony, bursts). They compare their model to a standard autoencoder and an autoencoder trained using data augmentation on a synthetic dataset and electrophysiological recordings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the paper is interesting, I have noted several shortcomings that question whether the author's intended goal of improving upon standard neural measures is achieved.\n\n1. **Stated weaknesses of neural measures are not addressed**\nIn the introduction the authors state that neural metrics depend on manually picked hyperparameters, but I don't see how the autoencoder fixes this problem as its embeddings are also dependent on particular hyperparameters (e.g. number of units). I also did not find any experiments exploring the effect of hyperparameters on the autoencoder model. The authors state the autoencoder is a foundation model. Does this mean it can be trained once and used on different datasets without re-training from scratch? If so, this should be explored in the paper.\n\n2. **Unclear how the autoencoder is better than neural measures** The authors explore predicting neural measures from the embeddings in the autoencoder on the synthetic and electrophysiological datasets. However, this does not tell us much regarding the interpretability of the autoencoder embeddings (or how they are better than neural measures). The authors employ PCA to analyze the autoencoder embeddings in Fig. 6 to explore the relationship between genes. However, it is unclear if the same relationships could not be found when employing PCA on the neural measures. Besides, the authors state that the use of PCA and related methods (line 37-40) is a shortcoming of conventional MEA analysis, but then proceed to use it in their own analysis. It is unclear what is gained by using the autoencoder.\n\n2. **Unclear if the proposed autoencoder model is an improvement** The reported results do not include error bars or statistical analysis making it difficult to assess if the proposed autoencoder model is indeed better. Error bars should be available for Fig. 5. Also, the authors did not state why the simpler autoencoder models outperform their autoencoder model on various genes in Fig. 5 - this should at least be discussed.\n\n3. **Missing comparisons.** The authors state that point process models (line 76) can address invariance issues in MEA data, but they do not compare to these models in their analysis. Furthermore, self-attention is known to possess invariance properties [1] and it would be useful to compare these models too.\n\n[1] Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S. and Teh, Y.W., 2019, May. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning (pp. 3744-3753). PMLR." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Roll-AE, a new autoencoder, extracts spatiotemporally invariant features from MEA recordings and outperforms standard autoencoders in classification tasks, while characterizing electro-physiological traits in iPSC-derived neuronal cultures." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024rollae,\ntitle={Roll-{AE}: A Spatiotemporal Invariant Autoencoder for Neuronal Electro-Physiology},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w6mjerkePG},\nnote={under review}\n}" }, "abstract": { "value": "Micro-electrode array (MEA) assays enable high-throughput recording of the electrophysiological activity in biological tissues, both in vivo and in vitro. While various classical and deep learning models have been developed for MEA signal analysis, the majority focus on in vivo experiments or specific downstream applications in vitro. Consequently, extracting relevant features from in vitro MEA recordings has remained largely dependent on particular curated features known as neural metrics. In this work, we introduce Roll-AE, a novel autoencoder designed to extract spatiotemporally invariant features from in vitro MEA recordings. Roll-AE serves as a foundational model that facilitates a wide range of downstream tasks. We demonstrate that 1) Roll-AE's embeddings outperform those from standard autoencoders across various classification tasks, and 2) Roll-AE's embeddings effectively characterize electrophysiological phenotypic traits in induced Pluripotent Stem Cells (iPSC)-derived neuronal cultures." }, "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": [ "autoencoder", "electrophysiology", "self-supervised 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/e0ce4aeaa2e1fc55d1540e4170bc9e0ef5128989.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": "Roll-AE: A Spatiotemporal Invariant Autoencoder for Neuronal Electro-Physiology" }, "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": 4 }, "primary_area": null, "questions": { "value": "1. What are the considerations the authors took to categorize the datasets into the proposed categories?\n2. How does the paper's findings hold across LLMs?\n3. Minor: \n * What are the 5 representative models used in Figure 3 right?\n * What do different colors in Figure 4 represent?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. CoT is a popular and often over-used strategy in LLM based applications. This paper dives deep into where it actually is useful and where it isn't - contributing to a better overall understanding of the technique and lay foundations for future improvements. \n2. The paper is well-written, detail experiments well and supports its key insights with exhaustive empirical analysis. \n3. The extensive analysis of contemporary literature is quite interesting and one of a kind." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this work, the author study the effect of CoT on a wide variety of problems and try to identify categories of problems for which it work v/s ones for which it doesn't. Using their meta-analysis of results from various recent papers and their own analysis spanning models and datasets, they conclude that CoT is more useful for math and logic based tasks, while providing minimal gains for other task tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the paper comes up with some interesting analysis, it doesn't really propose alternatives - it claims that CoT is a general approximation to logical solvers but proposes using them during inference. Arguably, it is hard to come up tool augmented neuro-symbolic techniques that fit most symbolic reasoning tasks well -making CoT especially appealing. Even though, this doesn't bring down the paper's contributions, it would have made the work more interesting for me.\n2. The paper distinguishes the 264 datasets on which it bases its main position with but doesn't specify how it comes up with the categorization. A detailed explanation of these exact considerations can help better utilize the paper's recommendations in practice. \n3. Language models are not a monolith and different Language Models behave differently depending on the fine-tuning dataset, preference alignment etc. Trends in Fig 6 already highlight the variation across LMs. The paper makes broad claims without often specifying the LM over which the result were found- almost assuming that they will generalize across models. A few examples:\n\t* Figure 3: What are the 5 representative models?\n\t* Line 320-321: Analysis of the impact of few-shot CoT v/s direct ?\nBesides, the paper lacks key model-wise analysis of its findings." }, "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": "> [323] Not much. Free response capabilities may be hindered by pre-planning or reasoning about the correct response.\n\nI'm unsure what this means. Where is the hinderance to free-response capabilities that is being referred to? I might have missed something, but it was unclear from the writing.\n\n> Informally, we believe many readers of the literature to hold the following view: [...]\n\nThis may not be a really substantive question, but is there a reason for framing this as a belief held by readers, rather than a specific reference to some prior paper(s) as is the norm?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The paper presents a meta-analysis of work in ML and NLP that is appropriate (provides evidence of where chain-of-thought helps without re-running every single experiment)\n - This is not a commonly used approach in ML work, so I cannot be entirely certain, but the methodology of the meta-analysis seems mostly sound.\n- The paper reinforces the findings of the meta-analysis with direct comparisons of multiple models across a number of tasks, both complementing and extending the findings of the meta-analysis\n - This comparison seems sound and thorough, particularly with attention to detail such as ensuring answers appear earlier in the response for direct answering as compared to chain-of-thought, which is a simple and effective way to measure that a model does produce a chain-of-thought\n - Identifying the \"=\" sign as a crucial determiner of gains in MMLU is also a cleverly designed experiment with informative findings\n- The design of the study to separate execution and planning is an interesting and useful approach, and uses prior work like PAL in a clever way to set up a controlled study\n - The discussion of how this relates to early work on chain-of-thought (Nye et al., 2021) provides useful context" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a comprehensive analysis of prompt-based chain-of-thought across a variety of tasks, specifically as compared to direct answer prompting. The analysis proceeds in three parts.\n\n## Part 1\nThe paper presents a meta-analysis of performance gains due to prompt-based chain-of-thought reported in the literature. They choose a set of 110 papers from recent NLP and ML venues that explicitly compare direct answer prompting to chain-of-thought. They manually extract results from these papers. They they group tasks into 14 categories, and analyze the average performance gain due to chain-of-thought in each category. They find that the top 3 categories – symbolic reasoning, math, and logical reasoning – benefit substantially, while other categories see only a small benefit. There is some discussion of exceptions to this rule, but the trend from the literature suggests that chain-of-thought largely benefits only mathematical or symbolic reasoning tasks.\n\n## Part 2\nThe paper presents direct comparisons of direct answering, zero-shot, and few-shot chain-of-thought on 14 LLMs on 20 datasets, broadly categorized as Commonsense, Knowledge, Symbolic, Mathematical, and Soft Reasoning. Here too, they find that performance gains are primarily in tasks that require some form of symbolic or mathematical reasoning, with trends being fairly stable across models. Even gains on Commonsense, Knowledge, and Soft Reasoning datasets can be tied back to mathematical reasoning, like in a mixed dataset such as MMLU (where they find that a majority of the gain is in problems that involve the \"=\" sign).\n\n## Part 3\nThe paper then examines why chain-of-thought might be so particularly helpful for mathematical or symbolic reasoning. For this, the paper identifies the chain-of-thought process as consisting of two processes – planning and execution. In the planning phase, the problem is mapped to a plan, and in the execution phase, the plan is executed to get an answer. The paper presents a way to compare the influence of these stages by having models generate a plan in code/logic, and then comparing the execution of the plan with direct LLM answering, execution with chain-of-thought, and execution with a symbolic engine. The paper argues that most of the gains of chain-of-thought appear to be from more reliable execution, rather than superior planning.\n\nGiven these, the paper concludes that:\n- prompt-based chain-of-thought is not an effective method for reasoning construed broadly\n - The paper does highlight that methods that leverage more compute or search, might be more promising, as might methods that train models to use chain-of-thought\n- in tasks where chain-of-thought does provide an advantage in better executing complex plans, there may be more reliable symbolic execution options that could do even better" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There is no discussion of model size in the meta-analysis. Work that proposed chain-of-thought (such as Wei et al. (2022) and Kojima et al. (2022)) explicitly discuss how size is a crucial factor in the effectiveness of chain-of-thought. In fact, if we consider Table 26 of Kojima et al. (2022) that compares chain-of-thought and direct answering across a range of model sizes, of the 4 models tested, only 1 shows a substantial gain with chain-of-thought. Given that the results presented aggregate across paper and category, and that models of different sizes are not distinguished in the meta-analysis figures, it is unclear how much results of chain-of-thought on smaller models (even as large as 7B (Kojima et al., 2022; Table 26)) is the reason why the method appears to be ineffective.\n - This concern is mostly addressed in the direct comparisons, which show similar findings across a range of sizes, and hence I don't think this significantly bears on my rating, but if something can be said about the role of size in the meta-analysis it might further bolster the argument.\n- Section 5 ends with \"When possible, LLMs should be paired with symbolic solvers at inference time when solving symbolic tasks to achieve consistently better performance over direct answer **and** CoT.\" While it is an important takeaway that in some cases, the benefits of chain-of-thought can be reaped to a greater degree with exact symbolic execution, one aspect of the problem that doesn't seem to be discussed is the difficulty of getting models to produce the inputs to symbolic execution engines. These engines are powerful, but brittle, and minor errors in producing a program to be executed can lead to complete failure. Chain-of-thought on the other hand allows the model to be more flexible and robust. This is directly visible in the results, for example, for FOLIO. Models with chain-of-thought or plan+chain-of-thought execution perform _better_ than the plan+tool solver counterparts, which could be largely due to the relatively high unparseability rates. This suggests that the program writer-solver pair of (LLM, LLM) is better than the pair of (LLM, tool solver) in some cases (i.e. the LLM is a better executor of the plans it writes than a tool). If what matters to us is the final answer, should we not consider a solution that makes it more likely to get the correct final answer (setting compute requirements aside for a second)?\n - I'm not suggesting that we should declare chain-of-thought the winner here, just that I think this adds some nuance to what is a definitive recommendation made in the paper. I don't think this would change my recommendation, but I'd be glad if the authors would engage about 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": 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 some tasks, the use of CoT reasoning can decrease performance. What do the authors consider to be the key reasons for this? Will this phenomenon continue to be an issue in future LLMs?" }, "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 offers both quantitative performance metrics and qualitative insights into why CoT is effective in certain tasks, enhancing the depth of understanding." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a comprehensive analysis of the CoT prompting technique. Through a meta-analysis of existing literature and extensive experiments on various datasets and models, the authors demonstrate that CoT is particularly effective for mathematical and symbolic reasoning tasks but offers minimal benefits for other types of reasoning. The study provides valuable insights into the conditions under which CoT excels and suggests a shift towards more advanced computational paradigms for LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I believe the core findings of this paper are already well-known among many researchers in this field. Although the authors conduct a comprehensive investigation, the core contributions appear to be limited.\n2. The authors define symbolic and non-symbolic reasoning but categorize experimental datasets into multiple types, which may lead to confusion." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We characterize Chain-of-Thought’s performance across over 100 papers in the literature, 20 datasets, and 14 models, showing that it helps primarily on tasks involving math, symbolic, and algorithmic reasoning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024to,\ntitle={To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w6nlcS8Kkn},\nnote={under review}\n}" }, "abstract": { "value": "Chain-of-thought (CoT) via prompting is the de facto method for eliciting reasoning capabilities from large language models (LLMs). But for what kinds of tasks is this extra \"thinking\" really helpful? To analyze this, we conducted a quantitative meta-analysis covering over 100 papers using CoT and ran our own evaluations of 20 datasets across 14 models. Our results show that CoT gives strong performance benefits primarily on tasks involving math or logic, with much smaller gains on other types of tasks. On MMLU, directly generating the answer without CoT leads to almost identical accuracy as CoT unless the question or model's response contains an equals sign, indicating symbolic operations and reasoning. Following this finding, we analyze the behavior of CoT on these problems by separating planning and execution and comparing against tool-augmented LLMs. Much of CoT's gain comes from improving symbolic execution, but it underperforms relative to using a symbolic solver. Our results indicate that CoT can be applied selectively, maintaining performance while saving inference costs. Furthermore, they suggest a need to move beyond prompt-based CoT to new paradigms that better leverage intermediate computation across the whole range of LLM applications." }, "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", "Large Language Models", "Textual Reasoning" ] }, "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/4d86839ee19c3ae5b4d5dd7d21aab9c706b437e7.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": "To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning" }, "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": "- \"we sample sequences x_c from the model conditioned on $[c, q, x^∗]$\"(L.246): Could you explain the rationale behind including $x^*$ in the sampling of $x_s$ ? what happens if the sampling is only conditioned on $[c, q]$ without the target ? (which better corresponds to Eq.4).\n- I'm concerned about the context generation process, you mention that you use GPT-4 however, it somehow defeats the purpose since GPT-4 can be prone to hallucinations due to its training data becoming obsolete. I'm really not convinced by the prompt that you used: *\"Please help me generate five complete statements as [context]s according to the semantics of incomplete facts '{prompt}' and '{target}'.\"*. In fact, if GPT-4 hallucinates, it will provided non factual contexts, hindering the optimization process and potentially incurring further hallucinations. Consider adding a 'Limitations' section that addresses the potential risks of using GPT-4 for context generation, including the possibility of hallucinations or outdated information. This section could also explore potential mitigation strategies or alternative approaches for context generation.\n- From L.431, could you provide a clear definition of what 'temperature' refers to in this context ?\n- In Algorithm 1, only $L_\\text{ICE}$ is shown in the optimization process. If $L_\\text{FT}$ is also used, could you update the algorithm to reflect this? Additionally, an ablation study on the effect of $\\lambda$ would provide valuable insights into the relative contributions of $L_\\text{ICE}$ and $L_\\text{FT}$ to the overall performance.\n\n*Typos:*\n- Figure 1: for the fine-tuning part, it should be $p_{\\theta_s}(x | q)$ in the FT loss (and not $p_{\\theta_s}(x)$)." }, "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 clear, well-motivated and the idea is novel as far as I know.\n- Compared to other baselines, this method is the only one capable of effectively editing knowledge continually." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents ICE, a regularization loss that aims at addressing the limitations of the traditional fine-tuning loss to update knowledge. Experiments on the KnowEdit dataset show its effectiveness to update the model's knowledge especially in the continual editing setting compared to other baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The pipeline is quite heavy, relying on sampling at every optimization step and GPT-4 for augmented contexts." }, "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": "Questions\n\n\n1. I don’t understand the direct optimization problem stated in L238 - why can't the loss be backpropped jointly through both distributions? Anyway, the solution of freezing the target distribution makes sense, but then I'm confused by the $s$ vs. $s+1$ in Equation 7 – wouldn't it be more accurate that both subscripts should be $\\theta_{s}$ except we do not backprop through the left term of the KL term? That is what is stated (and drawn in the figure) but having $s+1$ in the subscript feels wrong because that’s in the future? \n\n2. How are stop/end-of-sequence tokens controlled in the output sequence, and how is fluency measured with respect to that? In the examples shown in D.6, all of the methods start to generate more (possibly unrelated) text after giving the correct answer. How does that unrelated text get factored into the various metrics?\n\n3. I’m confused by “Observation 1” and the subsequent proof because empirically, this is the same as row 4 in the ablations table 4, and it looks like it is substantially better than the fine-tuning baselines. Actually, I’m confused whether 3.1 and 3.2 are actually represented empirically anywhere so we have a sense of where it stands relative to other methods.\n\n4. This was mentioned above, but to ask more directly mostly out of curiosity: what is the cost (in terms of training time + prompting) for this loss? Is it substantially slower or faster than FT-M, and what about after factoring any hyperparameter tuning?\n\n## Other minor comments not affecting the judgement:\n\nIn the implementation details, GPT-2 is mentioned but it isn’t mentioned elsewhere in the paper. It is in the appendix, and so the mention of GPT-2 would be less confusing if it was moved entirely to the appendix.\n\nThroughout the paper, `` should be used instead of '' to start quotations.\n\nI'd also suggest renaming the abbreviation ICE to CICE, as ICE was already used to refer to in-context editing in [Cohen et al., 2023](https://arxiv.org/pdf/2307.12976), which is also cited in the paper as [6].\n\nFinally, contemporary work worth knowing about and citing in a future draft: [Rozner et al., 2024](https://arxiv.org/abs/2406.09920)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The method is novel for knowledge editing by identifying an issue with prior approaches for targeted knowledge editing. While there has been prior work on fine-tuning and naive work on in-context (prompt-based) knowledge editing, the combination of distilling the in-context editing directly into the parameters has not been done.\n\n2. The empirical results, while not perfect on all metrics and datasets, show promise across the baseline methods presented and on the standard metrics and perplexity.\n\n3. Ablation studies emphasize the importance of both dynamically updating the target distribution during training and on the importance of the context. In particular, an in-depth analysis of the static target distribution shows that dynamic targets actually lead to better convergence.\n\n4. More analysis shows that as the model is edited (updated) more, the degradation of the model is less prominent than the other baseline methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents an auxiliary loss for finetuning an LLM which steers it towards new knowledge and applied in the knowledge editing domain. This is through “in-context editing” (ICE), which minimizes the distances of the output *distribution* of the original model without the new knowledge to that of a model which is conditioned on the new knowledge in the prompt. Besides accuracy and fluency, other facets of generation are evaluated, like whether unrelated knowledge is affected and whether the learned new knowledge generalizes to related knowledge. Compared to prior methods (ROME, MEMIT, FT-L, FT-M), ICE performs relatively well on most metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The method is similar to knowledge/context distillation or gisting, and so a connection should be drawn there. Still, applying this method appears novel for knowledge editing. However the lack of references to KD/gisting makes it hard to place how related (or not) this idea is to that line of work.\n\n[Snell et al., 2022](https://arxiv.org/abs/2209.15189) - Context Distillation\n\n[Mu et al., 2023](https://arxiv.org/abs/2304.08467) - Gisting\n\n2. The paper advocates for conditioning on “context” to generate target token distributions. It isn’t clear based on the data that this “context” is what makes the ICE method good, as opposed to the training objective. \n\n The method requires GPT-4 outputs to generate the context, while the other baselines being compared against are not allowed any access to this (external model) context. If the method is instead interpreted as knowledge distillation, it is less clear how whether this is possible without this context from a stronger LLM.\n\n Concretely, the effect of context should be isolated from the distillation-like training objective. The latter targets the one-hot problem motivated by the introduction of the paper, but the former is discussed heavily by this paper.\n\n If I understand the ablations table correctly, the rows with x in “Context” gives us a view of what the distillation-like objective could do for model editing. It looks competitive (or even better) than the baselines presented in Table 3, and so I wonder what the context actually adds? \n\n3. In addition, one way to explore this would be to measure an “upper bound” of how good the model could be if it were perfect with context, i.e. evaluate the model corresponding to $p_\\theta(x, | c, q)$ both before and after training. \n\n4. The baselines included do not seem comprehensive and are a little misleading. In particular, MEND and SERAC are other method mentioned by Zhang et al., 2024 [survey] that achieve strong results only slightly worse than FT-M. The omission of those results make it seem like ICE is tied with FT-M, both much better than other methods. In reality, ROME/MEMIT are relatively weak baselines compared to the other methods.\n5. Another limitation of the method that should be acknowledged or perhaps even addressed directly is the number of modified parameters and cost to make the edits. MEMIT/ROME are local, while FT-M, FT-L, and ICE are full-model. But my understanding is that the latter 3 are actually similar in terms of training cost and modified parameters\n\n[Mitchell et al., 2021](https://arxiv.org/abs/2110.11309) - MEND \n\n[Mitchell et al., 2022](https://proceedings.mlr.press/v162/mitchell22a/mitchell22a.pdf) - SERAC." }, "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 hear the thoughts of the authors about how much the base in-context ability of the models impact the success of the method.\nThe discussion section could be expanded with such consideration. It would be interesting especially since the paper does not experiment with different base models" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper proposes an interesting methods that is likely to be useful for future applications. \nThe paper does a good job at demonstrating the usefulness of the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a method for knowledge editing that is based on leveraging the in-context ability of the model. To learn a new fact by finetuning, the methods proposes to train the model with supervision coming from itself when it can answer the query correctly based on contextual information. \n\nWhat is interesting is that since the method relies and benefits from the in-context adapting ability of the model to have better update, better models would benefit even more from finetuning with this method [1, 2].\n\nOverall, this is a good contribution with interesting potential for future work. I think the author could include a discussion about how much the base in-context ability of the models impact the success of the method.\n\n[1] Yu et al 2023 Characterizing mechanisms for factual recall in language models.\n[2] Monea et al. 2024. A Glitch in the Matrix? Locating and Detecting Language Model Grounding with Fakepedia" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper only studies one base model, it is not clear how this generalizes to other model. In particular, I suspect that the size of the model and their base in-context capabilities might play an important role in the success 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": 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": "What is the value of lambda used? Was a different value used for different datasets? How was this value found?\n\nWhy is the WikiBio dataset the only one missing the Portability score?\n\nPerplexity is reported as an evaluation metric but is that the perplexity of the trained model itself or the perplexity of some other reference model on the generated text? \n\nIt is said that one of the baselines \"demonstrated nearly the best performance\" in a survey. Why was the best model not reported as a baseline?\n\nThe information about which layers are updated by the proposed method should be in the main paper, not hidden deep in the appendix. The main paper currently only mentions this information for \"other baselines\" and the proposed model does not qualify as a baseline.\n\nGiven the very imminent US elections, the example used throughout the paper should probably be updated." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The idea is interesting and makes sense, as the probability distribution provides more detailed information compared to just a one-hot target.\nEvaluation is performed on several different datasets and with a number of different metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method of updating an LLM to incorporate some specific piece of new information.\nThe LLM prompt is prefixed with this new information and the output distribution is used as the target for training the same LLM without that additional context.\nThis loss is used as a regulariser, along with the main fine-tuning loss.\nResults show good performance compared to the chosen baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "A crucial baseline is currently missing. There needs to be an evaluation of the exact same proposed model but with lambda set to 0. Using only the fine-tuning (FT) training loss.\nThis is important to understand what effect the proposed regularising objective has on the model. As far as I can see, this has not been reported in the paper at the moment.\nFT-M and FT-L are reported but these differ from the vanilla FT objective and update only 1 specific layer in the model. \nIt is mentioned in the appendix that the proposed ICE model is trained by updating the same layers as MEMIT, which would be 5 layers.\nThere needs to be a baseline that updates the same layers as the proposed model using the same FT objective, with the only difference being that the proposed ICE loss component is turned off (lambda = 0).\n\nThe clarity of the paper could be improved.\nFor example, Section 3.2 seems to propose a method but the actual mechanics or motivation are unclear to me. And then it is said that actually this method is equivalent to the vanilla method described in the previous section anyway.\n\nThe novelty of the method is somewhat overstated in the paper. The technical solution is essentially the same as previous work such as Snell et al (2022), the main thing changed is the content of the prompt. That paper is indeed referenced but only among a list of different directions. The particular novel aspect of the paper should be made clear and previous work should be attributed accordingly. As far as I can see, the novel aspect is the application of this method to updating facts in the LLM." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose Consistent In-Context Editing, an approach for tuning language models through contextual distributions, overcoming the limitations of traditional fine-tuning methods that learn towards one-hot targets." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024incontext,\ntitle={In-Context Editing: Learning Knowledge from Self-Induced Distributions},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w6rHCuN3YG},\nnote={under review}\n}" }, "abstract": { "value": "In scenarios where language models must incorporate new information efficiently without extensive retraining, traditional fine-tuning methods are prone to overfitting, degraded generalization, and unnatural language generation. To address these limitations, we introduce Consistent In-Context Editing (ICE), a novel approach leveraging the model's in-context learning capability to optimize towards a contextual distribution rather than a one-hot target. ICE introduces a simple yet effective optimization framework for the model to internalize new knowledge by aligning its output distributions with and without additional context. This method enhances the robustness and effectiveness of gradient-based tuning methods, preventing overfitting and preserving the model's integrity. We analyze ICE across four critical aspects of knowledge editing: accuracy, locality, generalization, and linguistic quality, demonstrating its advantages. Experimental results confirm the effectiveness of ICE and demonstrate its potential for continual editing, ensuring that the integrity of the model is preserved while updating information." }, "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": [ "Knowledge Editing", "In-Context Learning", "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/e51b2ac111576428ff9ce45f2ba2cd04db46f898.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/1530d60c117b0581ee723e548b75df7fcc7f2563.zip" }, "title": { "value": "In-Context Editing: Learning Knowledge from Self-Induced Distributions" }, "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": 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": "Please refer to my questions above." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper studies an important problem, i.e., automatic financial trading with LLM-enhanced reinforcement learning.\nThe author conducted extensive experiments to validate the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel approach to algorithmic trading that combines Deep Reinforcement Learning (DRL) with Large Language Models (LLMs). The author introduce \"Stock-Evol-Instruct,\" a new instruction generation algorithm that helps optimize trading strategies by incorporating LLM-driven insights from financial news. The study examines six different LLMs, including LLaMA-2, LLaMA-3, Mistral-7B, Falcon-7B, OpenELM, and GPT-4o, integrating them with DQN and DDQN reinforcement learning methods. Testing their approach on Silver (SLV) and JPMorgan (JPM) stocks, the paper found that LLM-enhanced trading strategies often outperformed traditional RL methods alone, with some models achieving significant improvements in Sharpe Ratio and ROI. The fine-tuned Mistral-7B model, in particular, showed strong performance with ROIs of 53.15% and 48.36% for JPM and SLV respectively. The paper demonstrates the potential of combining LLMs with reinforcement learning to create more sophisticated and effective algorithmic trading systems." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The literature review is not thorough, many relevant papers in LLM/AI for finance and trading area are missing. Some paper I found relevant are listed below, but I believe there are more to be included.\n\n2. The writing of the paper can be improved. However, the main problem is the novelty of the work is really limited due to that the author did not conduct a thorough literature review for the field of RL/DL/LLM for trading. There are so many important and related work in this field that should at least be included as baselines for comparison. I recommend authors first conduct a comprehensive literature review in this field, then conduct comprehensive and fair baseline comparisons.\n\n\nYu Y, Li H, Chen Z, et al. FinMem: A performance-enhanced LLM trading agent with layered memory and character design[C]//Proceedings of the AAAI Symposium Series. 2024, 3(1): 595-597.\n\nYu Y, Yao Z, Li H, et al. FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making[J]. arXiv preprint arXiv:2407.06567, 2024.\n\nYuan Z, Liu J, Zhou H, et al. LEVER: Online Adaptive Sequence Learning Framework for High-Frequency Trading[J]. IEEE Transactions on Knowledge and Data Engineering, 2023.\n\nYuan Z, Liu H, Hu R, et al. Self-supervised prototype representation learning for event-based corporate profiling[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2021, 35(5): 4644-4652." }, "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. Would the authors consider adding other baselines in future work?\n\n2. Given the importance of risk management in financial trading, could the authors elaborate on any plans to integrate risk-sensitive Reinforcement Learning (RL) approaches?" }, "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 paper presents an innovative combination way of LLMs with RL algorithms and it expands the applicability of LLMs in quant trading. Compared to the previous work of applications in LLM and RL in financial trading, like FinMem, FinGPT and FinRL, which focuses on either financial LLMs—using NLP techniques to interpret market sentiment and to output trading decision—or DRL models designed for financial decision-making based purely on quantitative market data, this study bridges these two areas.\n\n2. The study is thorough in its design and evaluation, offering clear descriptions of methodologies, metrics (ROI, Sharpe Ratio), and prompt types (zero-shot, instruction-based, and exemplar-based).\n\n3. The paper is well-structured, with distinct sections dedicated to prompt design, trading environment, evaluation metrics, and empirical results.\n\n4. The significance of this work lies in its contribution to enhancing algorithmic trading strategies through a novel combination of LLMs and DRL. Though still exploratory, this combination highlights a useful direction for future research into more nuanced decision-making systems in finance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the integration of LLMs with RL to enhance algorithmic trading strategies in financial markets. By leveraging LLMs such as GPT-4o, LLaMA, and Mistral alongside DQN and DDQN models, the study demonstrates how LLM-driven insights from financial news can improve trading agents' decision-making.\n\nIt evaluates the effectiveness of various LLM-RL combinations on two case-study stocks, Silver (SLV) and JPMorgan (JPM), measuring performance through metrics like Return on Investment (ROI) and Sharpe Ratio (SR). Results indicate that LLMs integrated with RL strategies can outperform traditional RL alone by providing contextual market insights.\n\nIn the experiment, LLaMA-3 showed higher accuracy in trade prediction, while Mistral-7B excelled in profitability. It also emphasizes the role of prompt design in optimizing model outputs and addresses the limitations of prompt diversity, suggesting future expansion." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Absence of Baseline Comparisons: While the paper demonstrates that LLM-RL combinations outperform traditional RL models, it lacks comparisons to other established baselines or hybrid approaches in financial NLP or multi-modal models for trading, such as FinRL, FinMem." }, "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": 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": "How do you plan to address the inconsistencies in the metrics in your evaluations?\n\nWhat criteria were used for selecting the LLMs? How do they compare in performance beyond their general strengths in NLP?\n\nGiven the limited prompt variety, do you foresee exploring automated prompt generation or using techniques like semi-supervised learning to enhance prompt diversity?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper presents a forward-thinking approach by merging LLMs and deep reinforcement learning, bridging the strengths of LLMs in handling unstructured data and deep RL's adaptability in a dynamic environment.\n\nThe authors outline their contributions well, particularly the development of Stock-Evol-Instruct.\n\nTesting on real-world data (Silver and JPMorgan) is a strong point, as it validates the model's performance in practical scenarios.\n\nThe paper incorporates different prompt designs for LLM predictions, which offers a robust testing ground for prompt optimization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores a new approach to algorithmic trading by integrating LLMs with deep reinforcement learning to analyze and predict stock market trends.\n\nIt contributes meaningfully to algorithmic trading research by proposing an innovative model that combines LLMs and deep reinforcement learning. With refinements in prompt diversity, a deeper analysis of metrics, and further exploration of model limitations, the research could have a significant impact on the field of financial technology." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There is an inconsistency in performance metrics (SR and ROI), which the authors highlight. The authors could delve deeper into explaining the contributions under which these inconsistencies occur and propose adjustments.\n\nThe study relies on a relatively small set of prompts for instruction generation, which may restrict the model's generalization across various market conditions.\n\nThe Stock-Evol-Instruct method includes in-depth evolving steps, which may introduce redundant complexity. Simplifying this process or providing a clearer rationale for each evolution step could improve understanding and replicability.\n\nA more systematic analysis of how different LLMs align with specific trading objectives would strengthen the case for the chosen 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": 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": "During evaluation, does the RL agent have access to the news information or LLM output? Need to be more clear about what input features used during evaluation.\n\nCould you include a visual representation of the trading trajectories in your experiments? It’s a common approach in stock trading papers and helps readers better understand the performance over time.\n\nWhat was the rationale behind choosing JPMorgan and Silver for testing? \n\nMore detailed explanations in the text and clearer figures could improve the reader's understanding in the relationship between LLM and RL agent." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors tested a wide range of LLMs with various prompt types, and the writing is clear and concise in most sections." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a method leveraging Large Language Models (LLMs) to enhance reward mechanisms in a reinforcement learning (RL) trading algorithm, specifically a Double Deep Q-Network (DDQN). Traditionally, RL trading algorithms derive rewards solely from investment returns. However, this paper explores using LLMs to modify or replace conventional rewards in certain scenarios, aiming to help the RL agent trade more like a human. Additionally, the authors use prompt-based techniques to create a novel stock market dataset for fine-tuning the LLM. The paper evaluates the performance of different LLMs, using various prompts, on two stocks (JPMorgan and Silver), showing that RL agents incorporating LLM-generated rewards outperform those using traditional reward structures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It's standard practice in stock trading papers to show trading trajectories in charts, but this was missing here.\nThe paper lacks a clear justification for choosing JPMorgan and Silver for testing. Since LLMs rely heavily on textual data, similar studies typically select stocks based on the amount of news coverage during a given period.\nWhile the paper evaluates each LLM and prompt, it fails to include common baselines like \"buy and hold\" strategies. Additionally, it does not compare results with a baseline using only LLMs for stock trading, which would help show the added value of the RL trading algorithm.\nThe relationship between the RL algorithm and LLM is difficult to follow. Often times, LLM or RL agents alone can be used for trading. Therefore, it is not obvious to user how does their combination work. This paper uses LLMs to refine rewards in addition to price-based rewards for the RL trading algorithm. This could be explained more clearly in both the text and accompanying figure (figure 1)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A novel approach to algorithmic trading by integrating LLMs with Deep RL, for optimizing stock market trading strategies through the development of the Stock-Evol-Instruct algorithm." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024advancing,\ntitle={Advancing Algorithmic Trading with Large Language Models: A Reinforcement Learning Approach for Stock Market Optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w7BGq6ozOL},\nnote={under review}\n}" }, "abstract": { "value": "In the fast-evolving landscape of financial markets, effective decision-making tools are essential for managing complexities driven by economic indicators and market dynamics. Algorithmic trading strategies have gained prominence for their ability to execute trades autonomously, with Deep Reinforcement Learning (DRL) emerging as a key approach for optimizing trading actions through continuous market interaction. However, RL-based systems face significant challenges, particularly in adapting to evolving time series data and incorporating unstructured textual information. In response to these limitations, recent advancements in Large Language Models (LLMs) offer new opportunities. LLMs possess the capacity to analyze vast volumes of data, providing enhanced insights that can complement traditional market analysis. This study proposes a novel approach that integrates six distinct LLMs into algorithmic trading frameworks, developing Stock-Evol-Instruct, an innovative instruction generation algorithm. This algorithm enables RL agents to fine-tune their trading strategies by leveraging LLM-driven insights for daily stock trading decisions. Empirical evaluation using real-world stock data from Silver and JPMorgan demonstrates the significant potential of this approach to outperform conventional trading models. By bridging the gap between LLMs and RL in algorithmic trading, this study contributes to a new frontier in financial technology, setting the stage for future advancements in autonomous trading systems." }, "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": [ "Algorithmic trading", "Stock market", "Large language models", "Deep reinforcement 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/4fb3de25b973a9d15a365a151f079fb68cbe0506.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": "Advancing Algorithmic Trading with Large Language Models: A Reinforcement Learning Approach for Stock Market Optimization" }, "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": 3 }, "primary_area": null, "questions": { "value": "See previous." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The algorithm is certainly interesting. The PIED framework certainly looks practical. The motivation behind and justification of each step is presented, and the experimental results look good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel algorithm for solving the inverse problem in PDEs - that is, estimating the parameters of a PDE governing the dynamic characteristics of a system. In particular, the paper assumes that (a) obtaining (x,y) observations of the system is expensive, so we must select our observation points carefully; and (b) observations require an initial setup that we cannot (practically) repeat, so we must specify our test points up-front and not dynamically as in e.g. Bayesian Optimization.\n\nTo tackle this problem the paper suggests physics-informed experimental design (PIED) that uses two sets of PINNs to select appropriate test points for a given system of PDEs (with a-priori unknown parameters). The general approach uses a set of PINNs operating as forward simulators to generate functions satisfying the PDEs, sampling these for a set of points X, then using PINNs as inverse solvers to estimate the PDE parameters from the observations. The efficacy of the points X is measured as the difference between the \"real\" PDE parameters (used in the forward simulators) and the corresponding estimated estimated PDE parameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One point that needs to be address in the paper is that of computational cost. After all, one of the motivations of using PINNs in parallel is computational efficiency, so it would be good to have a comparison in terms of same." }, "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": "- I would suggest to describe the PIED framework as described in algorithm 3 more clearly in the main text of the paper\n- After the experimental design is complete and sensor placements $X$ are obtained: Are the obtained sensor placements supposed to be universally informative and optimal for various downstream tasks of the practitioner, also those that involve classical simulators, or is the intention to always perform inference of the parameter $\\beta$ with the trained inverse PINN (which may be inaccurate compared to a slower classical solver). \n- Related to the previous question: Is it possible that the PINN memorises information about the PDE system at arbitrary evaluation points through the shared initialisation $\\theta_{SI}$, such that the sensor locations $X$ are not the most informative to make predictions about the system, but only the most informative when we additionally have access to a PINN with parameters $\\theta_{SI}$. In other words, could it be that the optimal experimental design is different when one wants to use classical simulators + MCMC to solve the inverse problem?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Overall, the paper is well written and seems methodologically sound.\n- The paper addresses a challenging experimental design problem with little existing approaches.\n- There is sufficient experimental results to support to proposed approach.\n- The paper combines an interesting set of ideas: \n\t- Meta learning and shared weights for more efficient PINN training. \n\t- Approximate training dynamics through of the inverse solver to make back-propagation through the inverse solver feasible." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper suggests PIED, a method for optimal experimental design for PDE inverse problems via Physics-informed neural networks (PINN). The PIED framework consists of three steps: 1. A PINN to learn a forward simulator that maps parameters $\\beta$ to solutions of the PDE $u_\\beta$. 2. An observation selector that returns noisy predicted observations $\\tilde Y= u_\\beta(X) + \\epsilon$ at a fixed number of sensor placements $X$. 3. A PINN inverse solver that predicts the ground truth parameter $\\hat \\beta(X, \\tilde Y)$ from $X$ and $\\tilde Y$. 4. The optimal sensor placements $X$ are found by minimising the MSE between $\\beta$ and $\\hat \\beta(X, \\tilde Y)$ which requires back-propagation through the inverse problem solver. The paper claims the following contributions and innovations: \n- The weight initialisations of all PINN-based components are shared and from a pretrained model which stabilises and accelerates training.\n- The forward simulator for various choices of $\\beta$ are learned in parallel\n- The mean square error between $\\hat\\beta(X, \\tilde Y)$ and $\\beta$ is approximated by learning the inverse solver with a smaller number of training steps (FIST criterion for ED)\n- The mean square error between $\\hat\\beta(X, \\tilde Y)$ and $\\beta$ is approximated through a linearisation of the PINN training dynamics (MoTE criterion for ED)\n\nThe framework is tested on optimal experimental design for sensor placement on Eikonal, Wave, and Navier-Stokes equations as well as groundwater and cell growth dynamics, and benchmarked against Bayesian experimental design, random sensor placements, and sensor placements on a grid." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The approach is mostly motivated by the comparison with classical simulators, but other types of neural surrogate models for the forward simulator are not mentioned. This is probably because most such approaches like Fouier Neural Operators don't investigate the ED downstream task. Nevertheless, this made me wonder how much this approach actually relies on Physics-informed neural networks. For example, the processing in parallel threads would be typical for all approaches that target ED with a neural network.\n- The description of the algorithm/framework is a little inconsistent. The PIED framework in Figure 1 suggests that the PIED framework is trained end-to-end, while Algorithm 3 in the appendix clarifies that training proceeds in three phases that are not interleaved: Optimising for initial weights, training to define ED criteria, optimisation of ED criterion. I wish that this was clearer from the main text." }, "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 Fig. (c) and Fig. (d), the training loss and test error at final stage are similar for both methods. How can this be used to highlight the advantages of the proposed method.\n2. How to define the value of μ in Eqs. (17) and (18). It is a very important parameter to determine the difficulty for solving the NS equations.\n3. How to select the noise variance for different experiments. The noise variance may have a great influence for the prediction results. Can the method still be available when the noise variance is very large." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The theoretical analysis of PINN is thorough." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors developed a physics-informed experimental design approach for the inverse problems to find the optimal and most informative design parameters. The paper is well-organized." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The effect of some experiments is not significant." }, "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. What's the relationship between the 'PINN for ED' and the 'adaptive-sampling PINN for inverse problem'?\n2. Can a trained PINN generalize to a new ED problem with new $\\beta$? Can the model (including the meta-learning part) generalize to a new distribution of $\\beta$？" }, "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 joins many advanced techniques together to solve the problem.\n2. most of the representation is clear, with extensive experiments and details." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper solves the one-shot experiment design using PINNs in both forward and inverse problems. It overcomes computational bottlenecks by parallelism and meta-learning of initialization. The experiments on both synthetic and real-life datasets show the performance improvements." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The overall algorithm flow is unclear in the main text. Figure 1b is too general and missing almost all of the details of the technique.\n2. Does not compare with existing NN-based experiment design methods.\n3. Almost all techniques are adopted from existing literature." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "An experimental design framework for PDE-based inverse problems that uses PINNs and its training dynamics, in a fully differentiable architecture to perform continuous optimization of design parameters." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024pied,\ntitle={{PIED}: Physics-Informed Experimental Design for Inverse Problems},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w7P92BEsb2},\nnote={under review}\n}" }, "abstract": { "value": "In many science and engineering settings, system dynamics are characterized by governing partial differential equations (PDEs), and a major challenge is to solve inverse problems (IPs) where unknown PDE parameters are inferred based on observational data gathered under limited budget. \nDue to the high costs of setting up and running experiments, experimental design (ED) is often done with the help of PDE simulations to optimize for the most informative design parameters (e.g., sensor placements) to solve such IPs, prior to actual data collection. This process of optimizing design parameters is especially critical when the budget and other practical constraints make it infeasible to adjust the design parameters between trials during the experiments.\nHowever, existing experimental design (ED) methods tend to require sequential and frequent design parameter adjustments between trials. Furthermore, they also have significant computational bottlenecks due to the need for complex numerical simulations for PDEs, and do not exploit the advantages provided by physics informed neural networks (PINNs) in solving IPs for PDE-governed systems, such as its meshless solutions, differentiability, and amortized training. \nThis work presents Physics-Informed Experimental Design (PIED), the first ED framework that makes use of PINNs in a fully differentiable architecture to perform continuous optimization of design parameters for IPs for one-shot deployments. \nPIED overcomes existing methods' computational bottlenecks through parallelized computation and meta-learning of PINN parameter initialization, and proposes novel methods to effectively take into account PINN training dynamics in optimizing the ED parameters. \nThrough experiments based on noisy simulated data and even real world experimental data, we empirically show that given limited observation budget, PIED significantly outperforms existing ED methods in solving IPs, including for challenging settings where the inverse parameters are unknown functions rather than just finite-dimensional." }, "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": [ "Physics-Informed Neural Network", "PINNs", "Experimental Design", "AI For Science" ] }, "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/e8c1e5c5e1313873bb8c6e8897f593fc210d2600.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/ff8544849d822a16eb2fa940e5a28377d7c7ffc0.zip" }, "title": { "value": "PIED: Physics-Informed Experimental Design for Inverse Problems" }, "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": "Additional comments/questions: \n* [L042-044] This is a bold claim....what about the field of XAI? Pearl's ladder of causality has been quite influential in deep learning (including XAI), so I would assume that any methods aimed at counterfactual explanations would address these questions to some extent.\n* I find the abstract a bit strange. Causal inference is a broad field attempting to answer (from a statistical perspective) the how and why questions, but it is not explicitly mentioned. Also, \"existing CBMs\" are mentioned in passing but the authors don’t say what they are or how they work. Maybe it would be better to more precisely situate the paper in the context of recent attempts to integrate causal inference with deep learning. \n* [Sec 2] I don't insist, but it would be nice to see a citation acknowledging early efforts of using VAEs for causal inference [https://proceedings.neurips.cc/paper/2017/hash/94b5bde6de888ddf9cde6748ad2523d1-Abstract.html]\n* [Eqn 1, L123] I don't understand why sampling a counterfactual concept c' does not also require sampling a fresh set of covariates x'. It seems to me that for concepts that matter in the real world, you would need to update the observed features as the concept is manipulated. To ground this in the x-ray example, if the concept is changed from \"narrow joint space\" to \"bone spurs\", would the distribution over pixels not also need to be updated? Clarifying this point seems important to justify why the $p(x|z)$ term is ignored in the optimization [L152].\n* [L169] how are specific values for $\\lambda_i$ [Table 6] chosen? Was a validation set used for tuning? Is the method sensitive to these values? I'm not satisfied with \"these details are in the released code\" [L850]\n* [L260-265] The experiments inherit some datasets from prior papers, such as Koh et al 2020. Is there a reason why the OAI x-ray dataset from that paper was not used? X-rays are used as a working example throughout the paper [Fig 1, Fig 2, Sec 3.2] so this omission surprised me.\n* [Sec 4, Sec 5] How does the method do when asked to simulate concept pairings that are OOD w.r.t. its training data? I would imagine for the CUB dataset there will be combinations of bird characteristics that are not represented by real-life birds; is it possible to examine the predictions for these concepts?. The strength of human counterfactual reasoning comes in part from OOD generalization, e.g. imagining a pink elephant. Replicating this using ML seems quite difficult. In the end I still have questions around whether what is being done here is meaningfully different from structured generative/discriminative modeling [L503-509] with interpretable features that can be adjusted on the fly. I think this is certainly related to counterfactual reasoning, but may not encompass it completely.\n* [L516] If I express the counterfactual as E[Y|do(X)=X', Y=Y'] and the interventional as E[Y|do(X)=X'], then it would seem that counterfactual explainability methods can answer the \"how\" question by just marginalizing out the specific observation, i.e. E[Y|do(X)=X'] = E_{Y'}[ E[Y|do(X)=X', Y=Y']. You could replace the observation being a label (Y=Y') with the observation being a concept or other covariate as needed. Is there something wrong with this line of reasoning? Modeling counterfactuals seems more challenging than modeling interventions, so I am wondering if the authors have thought of how to go from one to the other in a post-hoc manner without designing a method that explicitly models all three levels of Pearls' ladder." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors tackle an extremely challenging problem (counterfactual reasoning using deep learning); making meaningful progress along this direction would have a big impact.\n\nThe authors show how training using a structured objective allows for test-time manipulations that shed light on how predictions are made\n\nThey extend CBM to include counterfactual sampling, which in principle allows a single model to estimate quantities at all three levels of Pearl’s ladder of causation.\n\nThe proposed method relies on concept labels, which may be difficult to collect in realistic settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors extend the concept bottleneck model framework to allow simulation of counterfactual concepts and labels. The method is based on a variational approach that models concepts as latent confounders, and approximates this posterior using concept labels at training time. This extension opens the door to explaining prediction at a finer level of detail, focusing on intermediate concepts rather than pixel values." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The training objective is fairly complex and introduces many hyperparameters. The authors do not discuss how to tune these.\n\nThere is a disconnect between the working example of x-ray imaging and the types of datasets considered in the experiments.\n\nI felt that details of the SCM design and subsequent training (e.g. ignoring p(x|z)) were not 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": 3 }, "primary_area": null, "questions": { "value": "Could you provide additional visual examples of counterfactuals generated by your model and baseline methods for each dataset? Visualizing these comparisons would be highly intuitive and beneficial for readers, helping them clearly see the improvements." }, "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 straightforward and intuitive, combining two existing approaches, counterfactual reasoning and concept bottleneck modeling with moderate novelty. The experiments provide solid evidence of its effectiveness.\n2. This paper is well organized, and abstract and introduction effectively set the stage for the paper's major content.\n3. The method enhances the interpretability of counterfactuals, making it potentially valuable for many real-life settings like medical applications, as the authors suggest, beneficial for clinical decision-making processes." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a method that combines counterfactual and concept bottleneck that can answer \"What?\", \"How?\" and \"Why not?\" all at once. It leverages latent variable models to generate counterfactuals via variational inference. It achieves comparable classification accuracy to standard CBMs and black-box models and outperforms in generating interpretable, concept-base counterfactuals." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the paper uses 3 datasets, they are relatively simple. In this paper, authors use examples of medical images (fig 1) but did not actually experiment with any medical dataset. Using more diverse and complex real-life dataset (e.g., CIFAR-100 or MIMIC-CXR) would strengthen the paper and demonstrate broader applicability.\n2. The paper lacks visualizations of counterfactuals. For example, when discussing improvements in counterfactual quality, it would be helpful to show visual comparisons with baselines. Providing real counterfactual comparisons would better substantiate the claims and help readers understand the model's advantages." }, "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": "- Please refer 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 paper is easy to read, and the figures are well described and informative.\n- Explanations seem to be easy to use.\n- The method is tested against many baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a concept bottleneck model with the ability to generate counterfactual explanations. The paper aims to answer three questions: what, how and why not. The what questions are answered by showing the concepts used to make predictions. The how questions are answered by changing the concepts and observing how the prediction changes. Finally, why not questions can be answered by providing an alternative outcome and observing how the concepts change to understand what-if situations. The method also enables what the authors call “task-driven” interventions that allow users to correct wrongly predicted concepts. The models are tested on three datasets: dSprites, MNIST add CUB, and tested on various metrics. In general, based on the experiments, the proposed model demonstrates good performance across multiple tasks and datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I am concerned regarding the number of hyperparameters that need to be set for the loss function. It is not clear how important they are, and how much tuning is needed. How easy is it to tune these hyperparameters? How do they affect the results? Is it easier to have separate systems rather than one big as the one proposed?\n- The method requires handcrafted concepts, and it might not be easy to come up with these concepts. Further, there might be a lot of labor involved in creating these concepts. Moreover, the predictive performance depends on how well these concepts describe the input features.\n- Figure 1: Is it only for illustration or is this actually the result from trained models on x-ray images?\n- I might have missed it, but I cannot find the citation for the dataset from which the x-ray images come. If I did not miss it, please cite it so it becomes easier for readers to find it. Moreover, are there any restrictions on showing these images, that is, what is the license of these images?\n- I believe the x-ray images to be from the Osteoarthritis Initiative dataset. It is used to motivate the paper, but why is it not used in the experiments? Are there any weaknesses with the method making the dataset unsuitable for experimentation and showcasing results?\n- Line 156: Why does z' need to be conditioned on both c and y? Don't we already know y if we know c?\n- Line 224: How do we know Equation 5 gives us sparser explanations? Is it because the posterior is moved towards the prior distribution? But this is defined in the latent space. How can we know that the concepts c will be sparse even though the posterior of z is close to the prior?\n- How are the models trained? Are certain parts frozen, or are all components trained jointly?\n- Is it possible to show readers the explanations themselves for the datasets used and not only Figure 1 for a dataset not used? We have seen many quantitative results, but part of XAI also involves qualitative qualities. How do these explanations perform qualitatively with end-users? Are they useful to human end-users? And who are the intended end-users for these explanations?\n- Although we do not need to run post-hoc searches, new black-box components are introduced into the model. How does that affect the overall interpretability of the model? I believe that can also have negative side effects and not only positive influences." }, "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": "- 5.1 I find it important to have the results of this question also in the main paper, e.g., average values, as it is an important one. I would additionally suggest the authors change the wording of the conclusion. It seems that on average CF-CBMs are slightly lowerd in predictive performance. While I don’t think this is a drawback of the proposed emthod, I think stating the method is on par is slightly overstated. That CF-CVM is not necessarily on par is an important information for the reader/user. Again stating this won’t decrease the contribution. \n\n- 5.2. I don’t understand the intuition behind why CF-CBMs should be less prone to confounding factors? What mechanism in their learning is beneficial for this? Maybe an ablation study would be intersting to underscore these findings.\n\n- Please provide code to the paper, e.g., an anonymized github link.\n\n- Lastly, this [1] might be an interesting paper to add to the Introduction as it tackles learning concepts (e.g., for CBMs) focussing on learning inspectable concepts, e.g., via counterfactual querying. Of course [1] focuses on counterfactual concepts, whereas this work focuses on counterfactuals for a model decision.\n\n[1] Wolfgang Stammer, Antonia Wüst, David Steinmann, and Kristian Kersting. “Neural Concept Binder.” In: Advances in Neural Information Processing Systems (NeurIPS) (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 very well written and structured. It is well motivated and the methodological contribution as far as I can tell is clear and important." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces the framework of counter factual CBMs, aimed specifically to increase the variability and depth of inspection question that can be asked towards a CBM. In other words, CF-CBM aims to increase the level of interpretability and interactability of standard CBM models without losing performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I just have a few minor remarks which I have added to the questions section." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "CounterFactual Concept Bottleneck Models predict, simulate, and generate alternative scenarios in a single step, without needing post-hoc analysis." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024counterfactual,\ntitle={Counterfactual Concept Bottleneck Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w7pMjyjsKN},\nnote={under review}\n}" }, "abstract": { "value": "Current deep learning models are not designed to simultaneously address three fundamental questions: predict class labels to solve a given classification task (the \"What?\"), simulate changes in the situation to evaluate how this impacts class predictions (the \"How?\"), and imagine how the scenario should change to result in different class predictions (the \"Why not?\"). The inability to answer these questions represents a crucial gap in deploying reliable AI agents, calibrating human trust, and improving human-machine interaction. To bridge this gap, we introduce CounterFactual Concept Bottleneck Models (CF-CBMs), a class of models designed to efficiently address the above queries all at once without the need to run post-hoc searches. Our experimental results demonstrate that CF-CBMs: achieve classification accuracy comparable to black-box models and existing CBMs (“What?”), rely on fewer important concepts leading to simpler explanations (“How?”), and produce interpretable, concept-based counterfactuals (“Why not?”). Additionally, we show that training the counterfactual generator jointly with the CBM leads to two key improvements: (i) it alters the model's decision-making process, making the model rely on fewer important concepts (leading to simpler explanations), and (ii) it significantly increases the causal effect of concept interventions on class predictions, making the model more responsive to these changes." }, "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": [ "Concept Bottleneck Models", "Concept Based Model", "Counterfactuals", "Explainable AI", "Interpretability" ] }, "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/f8f33bbceaac9bce963805ab29d16aff0c15beb2.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": { "value": "/attachment/5b8121ff87d8537b37423c36a8a217dd5df58be3.zip" }, "title": { "value": "Counterfactual Concept Bottleneck 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": 2 }, "primary_area": null, "questions": { "value": "please could you:\n(1) indicate in what extent the stationarity index is new ?(if not, please refer to the literature on KS test for stationarity measurement)\n(2) explain in details how you use the stationarity index that you propose and finally choose the way you decompose the time-series.\nmost importantly:\n(3) could you explain how you use the so-called constraint module and provide a complete algorithm." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper tackles non-stationary times-series forecasting, a classical but still crucial in many applications. They propose to use the Koopman operator machinery to adress this issue. Leveraging random Fourier methodology for the measurement space, they can propose an efficient estimation of these operators that they learn in the context of a decomposiiton of the time-series into a stationary and non-stationary part.\nThe paper comes with a new (but relatively direct) result on the estimation of eigenfunctions of a Koopman operator when leveraging the finite dimensional reproducing Kernel Hilbert space induced by the Random Fourier Features.\nA secondary contribution is the proposition to use a Sinkhorn loss to measure the distance between the distribution of forecasts and the ground truth." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper builds on Koopman operator methodology to address non-stationary time series forecasting in an efficient way based on Random Fourier Features. The authors consider the Reproducing Hilbert space built from Fourier Features as the measurement function space, and report approximation results on Gram matrix as well on the Koopman operator eigenfunctions. In the section dedicated to learning, they assume a decomposition of the time series into a stationary and no stationary part and learn the corresponding global and locals Koopman operators following the steps of (Liu et al. 2023). A statistical test (KS) is proposed to measure the stationarity of the data from which helps to define the decomposition in stationary/non-stationary signals. Finally they introduce an alignment loss that measures how the forecasted distribution differs from the ground truth assuming Gaussianity of the state variable distribution. Experimental results include a few competitors and feature different time-series benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While very promising this paper, this paper seems to suffer from incompleteness in its presentation and should be re-writtent to make it reproducible.\n\nThe learning of the global and local koopman operator is not clearly posed but are direclty inherited from (Liu et al 2023) : the loss function is even not mentioned, here likely the square loss given the equations. A constraint is evocated but never described in the context of learning: when is it used?\n\nThe decomposition into stationary and non-stationary parts which is crucial here is only discussed in the appendix, but it remains very unclear how the stationarity index is used (probably a typo on S_V / S_alpha ?). I also find disturbing that the authors refer to a Python code and not mathematical equations. \n\nFinally it is important to refer to previous papers on:\n- Random Fourier Features for Koopman operator learning: Salam, T., Li, A. K., & Hsieh, M. A. (2023, June). Online Estimation of the Koopman Operator Using Fourier Features. In Learning for Dynamics and Control Conference (pp. 1271-1283). PMLR.\nother attempts to work with finite dimensional measurement spaces \n- (available in arxiv in 2023)\nMeanti, G., Chatalic, A., Kostic, V., Novelli, P., Pontil, M., & Rosasco, L. (2024). Estimating Koopman operators with sketching to provably learn large scale dynamical systems. Advances in Neural Information Processing Systems, 36. \n\n- in the experimental part, the authors refer to loss terms they never introduced - they also do not mention how they choose the number of random Fourier feature no provide an analysis of it." }, "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. Could the authors provide more detail on how the spectral kernel method integrates within the Koopman framework and specifically outline its role in enhancing efficiency?\n2. Why is multiresolution DMD, which appears closely related, not discussed as part of the literature? Could the authors clarify the differences and potential benefits of KokerNet over multiresolution DMD?\n3. Could the authors include more insights into how they ensured their model's robustness to real-world shifts in data distribution, especially for rapidly changing non-stationary components?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The application of Koopman operators for time series forecasting represents a significant advancement. The spectral kernel method for measurement functions is a commendable innovation that simplifies the computational process. The creation of an index to facilitate the decomposition into stationary and non-stationary components is particularly valuable. This not only enhances model interpretability but also provides insights into the dynamics of the data. Empirical results show KokerNet's superiority, underscoring its practicality and effectiveness in real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents KokerNet, a Koopman kernel network for time series forecasting. It addresses critical issues in existing methods, such as the high computational costs and the challenges posed by data distribution variations. By employing spectral kernel methods to construct a measurement function space, the authors achieve a notable reduction in computational burden. Furthermore, the decomposition of time series into stationary and non-stationary components is used to enhance interpretability. The global and local Koopman operators are effectively utilized to predict future behaviors of these components." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The manuscript does not clearly explain the integration of the kernel method within the Koopman network framework. Readers are left without a clear understanding of how the spectral kernel functions operate within the Koopman operator or how these methods combine to provide computational advantages. More detailed explanations and illustrative diagrams would significantly aid comprehension and convey the novelty of the approach.\n\nThe manuscript also lacks citations and discussion of closely related and classical methods, particularly multiresolution DMD, which already employs a similar architecture for handling non-stationary time series. KokerNet’s model appears to be a special case of this approach. Expanding the literature review and explicitly contrasting KokerNet with multiresolution DMD would clarify the unique aspects of the method.\n\nThe theoretical components presented are largely standard and do not demonstrate clear, novel benefits for KokerNet’s performance or interpretability. The paper would be strengthened by a clearer, more thorough explanation of how the theory specifically benefits the model’s forecasting ability or contributes new insights to the field." }, "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": "Since the purpose of the questions is to try and alter the reviewer's mind, it would be good if you could address those main claims raised in the \"weakness\" section if possible. In particular,\n\n - Why did you feel it was required to show a half-page proof of Bochner's Theorem, and do you feel that there is strong value in presenting a Chernoff-based bound of the methodology given that the bound is not really serving much purpose (to my own opinion), and that it leads more to the obfuscation of the text rather raise its intrinsic value? \n\n - How come you didn't try to compare the proposed methodology to other Kooopman operator based forecast methods that are actively used in industry?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper showcases an innovative development at the intersection of (1) Kernel integral transforms theory, (2) Koopman operator theory. An interesting decomposition is presented in order to tackle the non-stationarity of the time series which can appear in practical settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes learning the Koopman operator in a proposed low dimensional space by using cosine functions, motivated by a use of Bochner's theorem and kernel integral transforms. It performs several studies in relation to time series forecasting to assess the merits of the proposed methodology." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The majority of the paper doesn't present much perceivable novelty in my opinion. For example there is an attempt in Appendix C.1, Lemma C.1 to *prove* an obvious application of Bochner's theorem which anyway has already been done in the famous \"Random Fourier Features\" paper. This is already known several times over and thus there should be a *compelling* reason to present its proof and make it appear like its your own innovation (I am not against re-writing well-known proofs but it must be for a good reason). There is no apparent reason in the paper why this should be shown as a line-by-line proof and this leads me to the suspicion of there being an artificial mathematical padding of space, unfortunately, in order to try and obfuscate the reader. \n\nThis is amplified when considering Lemma C.2 and Theorem C.1, since the Chernoff-based exponential probability bound (C.7) proven doesn't appear to serve any intrinsic purpose and appears to be a standard matrix probability theorem slightly re-arranged so as to obfuscate the readability of the paper. Such bounds *are* useful don't get me wrong, but usually then there needs to be a stronger commentary and appeal to something like Computational Learning Theory. One can argue that it is useful in a sense for then proving Theorem C.1 in particular for equation C.16 but then ultimately what does Theorem C.1 intend to say? That ultimately with a high enough number of samples there is a convergence between a kernel approximator and a mean estimator of a set of random matrices? This is quite an obvious conclusion and doesn't need 1+ pages of proof. Of course it *can* be interesting if the bound intends to be studied in some sufficient manner (as per Computational Learning Theory), but then it is not made any reference to in the actual paper other than \"here is a complex bound\". An appeal to Law of Large Numbers in a random matrix form would be much more simpler, familiar, and appropriate than what is currently presented for example, just to say \"it works\" rather than \"how it works\" according to sampling number complexity\n\nAnd for example Theorem C.2 (ii) is something that should be known to the reader *if* they are quite familiar with Koopman Operator Theory in practice and thus no proof is really required yet one is presented which essentially overlaps the intended purpose of the previous proofs / lemmas, which ultimately all (over the course of 3+ pages) serves to raise the point \"it works in large sample rates\", which unless there is good reason to belief this should not be the case, is an obvious conclusion. Ultimately to a reader not well versed in mathematics this would appear to do nothing but introduce of a lot of new notation to intentionally obfuscate the readability of the paper, in my opinion. \n\nMathematics aside, there does not appear to be ample quantitative results to show that the proposed method actually works out in practice and thus should be used as an alternative to other competing Koopman operator based methods. For example for the purpose of forecasting why should someone be motivated to use this method over the standard DMD-based methods which have been shown to be enjoying absolute plethora of use cases in the engineering fields which experience massive levels of non-stationarity and large dimensionality? The authors are suggested to read for example \"Higher Order Dynamic Mode Decomposition and Its Applications\". If the proposed methodology is to be used seriously as good alternative then it would be worth explicitly analyzing why it might do better than higher order dynamic mode decomposition and such, which are currently actively used in everyday engineering industries for problems such as spatio-temporal fluid dynamics *without* an explicit need for training. Essentially, because \"Koopman\" is the name of methodology I would expect much more \"Koopman\" based comparisons and analysis." }, "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. In the design of the measurement function, the encoder utilizes a mathematical function in the form of a Reproducing Kernel Hilbert Space (RKHS), while the decoder adopts a neural network in the form of a Multi-Layer Perceptron (MLP). What is the rationale behind this asymmetrical structure?\n\n2. Regarding the segmentation of non-stationary components, what criteria are used to determine the segmentation length? Is the final performance sensitive to this parameter?" }, "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 is overall well written\n\n2. KokerNet enhances efficiency and interpretability by using a spectral kernel method for dimensionality reduction and decomposing time series into stationary and non-stationary components with a KS test-based index, outperforming state-of-the-art models.\n\n3. The idea of integrating a distribution constraint into the forecasting process is sound." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Koopman Kernel Network (KokerNet), a novel approach for time series forecasting that addresses the computational and interpretability challenges of existing Koopman-based methods. It uses a spectral kernel method to create a low-dimensional feature space, reducing computational costs. KokerNet decomposes time series into stationary and non-stationary components using a Kolmogorov-Smirnov (KS) test-based index, providing a more interpretable decomposition. It also includes a distribution module to handle time-varying distributions. Experiments show that KokerNet outperforms state-of-the-art models, enhancing both efficiency and interpretability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors claim that decompositions in previous works often rely on empirical determinations of component composition and proportions, which lack interpretability. They propose to utilize the Kolmogorov-Smirnov (KS) test to guide the decomposition of time series into stationary and non-stationary components. I believe that a comparative analysis with these methods through ablation experiments is necessary. For instance, consider employing the approach based on Fourier transformation coefficients from reference [1] to differentiate between stationary and non-stationary components, replacing the KS test index selection in the proposed model.\n\n[1] Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024kokernet,\ntitle={KokerNet: Koopman Kernel Network for Time Series Forecasting},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w7vn6ah0Qg},\nnote={under review}\n}" }, "abstract": { "value": "The Koopman operator has gained increasing attention in time series forecasting due to its ability to simplify the complex evolution of dynamic systems. However, most existing Koopman-based methods suffer from significant computational costs in constructing measurement functions and struggle to address the challenge posed by the variation in data distribution. Additionally, these approaches tend to empirically decompose time series or distributions into combinations of components, lacking interpretability. To tackle these issues, we propose a novel approach, **Ko**opman **ker**nel **net**work (**KokerNet**), for time series forecasting. On one hand, we construct a measurement function space using the spectral kernel method, which enables us to perform Koopman operator learning in a low-dimensional feature space, efficiently reducing computational costs. On the other hand, an index is designed to characterize the stationarity of data in both time and frequency domains. This index can interpretably guide us to decompose the time series into stationary and non-stationary components. The global and local Koopman operators are then learned within the constructed measurement function space to predict the future behavior of the stationary and non-stationary components, respectively. Particularly, to address the challenge posed by the variation in distribution, we incorporate a distribution module for the non-stationary component, ensuring that the model can make aligned distribution predictions. Extensive experiments across multiple benchmarks illustrate the superiority of our proposed KokerNet, consistently outperforming the state-of-the-art 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": [ "Spectral kernel", "Koopman operator", "Time series" ] }, "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/b5b2b41f4c2ba772d9cc2953462fc11ed3eeade8.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": "KokerNet: Koopman Kernel Network for Time Series Forecasting" }, "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": 4 }, "primary_area": null, "questions": { "value": "* The panel labels in Figure 2 are badly formatted - please make these clearer." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "* Authors clearly lay out their proposed framework describing the 3 level of uncertainty they they aim to model.\n* The results are presented such that each level of uncertainty is evaluated - this makes it easy for the reader to better understand the Authors important contribution.\n* Authors provide analysis of the estimates of uncertainty, and demonstrate that the aleotoric uncertainty corresponds better with coordinate prediction error at the first level.\n* The figures in the paper are very useful for visualizing the variations that arise from sampling the fitted transform - they demonstrate the fundamental problem with biomedical image registration in that the results from a downstream task are highly dependend on the transformation paramaters, but the uncertainty can be successfully quantified." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors present a method to estimate the uncertainty in medical image registration at 3 different stages: (1) on the estimate of the distribution of the deformation field at each voxel (assuming this is Gaussian and gathering a mean and standard deviation per point), (2) on the distribution of the fitted transformation, and (3) on the distribution of possible outcomes on the downstream task. The uncertainty from the first level is used in the transformation fitting step weighting down contributions from less certain pixels. By drawing samples of the fitted transformation, a distribution of results for the same downstream task is generated which is used to estimate the 3rd level of uncertainty. Authors demonstrate their approach in the context of registration-based Brain image segmentation where the given input image is deformed to match the standard MNI atlas and the labels from the atlas are propagated to the deformed image." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The addition of uncertainty in the fit transform did not impact the result from the Demons transformation model. Authors explain that this is likely because the loss from the Demons method is used in the segmentation loss and that leads to the average coordinates landing close to the optimum. This is unclear - please can Authors expand on this and also elaborate on whether they would expect to see improvements if they used an alternative to Demons?" }, "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": { "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": "1. Why is uncertainty estimation important in neuroimaging registration? What specific benefits does it bring, and why is it necessary?\n2. For modeling epistemic uncertainty, why did the author choose to use Monte Carlo dropout, and how exactly was it implemented?\n3. Given that the deformation field (non-linear transformation) is a high-dimensional tensor, how do authors verify the accuracy of the ground-truth labels when collecting them? As I understand, NiftyReg is a registration method—if this method is used to generate ground-truth labels, does that imply the proposed method can only perform as well as NiftyReg at best?\n4. The total training loss includes multiple terms marked as optional in the paper. Are these terms truly necessary? Was an ablation study conducted to assess their impact on performance?" }, "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 addresses an important problem in neuroimaging, aiming to use uncertainty estimation to enhance the accuracy and reliability of deep learning-based registration.\n2. The figures and explanations are clear and well-organized, which makes complex ideas easier to understand.\n3. The empirical findings are insightful. The authors compare their proposed method with other different methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a framework to integrate uncertainty estimation into deep learning-based image registration for neuroimaging. By propagating epistemic and aleatoric uncertainties from voxel-level predictions to transformation models and downstream tasks, the framework enhances registration accuracy and reliability. Experiments show that this uncertainty-aware approach can boost brain MRI registration performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presentation could be improved. The background, motivation, and knowledge gap are not clearly explained, making it challenging to follow the paper’s purpose and direction. Although the study is about why uncertainty estimation matters in brain registration, it doesn’t make a strong case for why this is important. There isn’t enough support or reasoning behind this focus, which leaves readers unsure about the value or impact of the work.\n2. The experiments are insufficient. This paper only compares its method to a single approach, RbR. Although RbR was released in 2024, it is still unpublished and available only on arXiv, which makes it less robust as a baseline for comparison. Including more established and published baselines would be essential to provide a solid foundation for evaluating the effectiveness of the proposed approach.\n3. From the experimental results (e.g., Table 1), the introduction of uncertainty estimation does not appear to provide a statistically significant improvement to the model. This makes it difficult to assess its effectiveness and raises questions about the practical value of incorporating uncertainty estimation into the approach.\n4. Most of the latest work in brain registration focuses on unsupervised learning, largely due to the high cost of collecting accurate transformation ground-truth for high-dimensional images (e.g., 3D MRI). However, this paper still stay on a supervised learning approach, making it seem less aligned with current trends and potentially less practical for real-world applications where labeled data is limited.\n5. The evaluation criteria lack validity. This paper collects ground-truth transformations using NiftyReg, a tool introduced around 15 years ago. Since then, more state-of-the-art methods have been shown to outperform NiftyReg, indicating that the ground-truth it provides may be inaccurate. This undermines the reliability of the results presented in the paper and raises concerns about the credibility of its findings.\n6. The novelty of this work is limited. The main contribution appears to be the addition of an uncertainty loss term, while the other loss terms and network architecture are based on existing work. This incremental improvement does not significantly advance the field, as much of the framework relies on previously established 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": "- In Figure 2, what does the error correspond to exactly?\n- Could the authors comment on the run time cost of their approach or other considerations regarding its practical application?\n\n===\n- The authors already mention many related works but they could also consider:\n - Chen J et al. \"A survey on deep learning in medical image registration: New technologies, uncertainty, evaluation metrics, and beyond.\" arXiv preprint arXiv:2307.15615 (2023).\n - Zhang X et al. \"Heteroscedastic Uncertainty Estimation Framework for Unsupervised Registration.\" International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024.\nAnd comment especially on the second one." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The framework proposed is flexible and can be applied to a broad range of registration techniques.\n- The authors performed diverse experiments to validate their approach.\n- The results are mostly convincing.\n- The paper is well written and related works well introduced." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a registration framework that includes uncertainty estimation at three different levels (network output, transform models' parameters and downstream tasks). They show that uncertainty estimates correlate with registration errors and that estimating uncertainty can improve registration performance. The approach is applied to brain MRI registration." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The ground truths considered originate from automatic registration and segmentation approaches but it is not mentioned whether quality control was performed and so to which extent they can be relied on.\n- The experiments regarding the downstream task are limited to qualitative results and the fact that the segmentations are not overlaid on the T1w image only allows assessing differences between the segmentation maps but not whether one matches better the anatomy than another one." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a principled way to propagate uncertainties (epistemic or aleatoric) estimated at the level of spatial location by these methods, to the level of global transformation models, and further to downstream tasks." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024hierarchical,\ntitle={Hierarchical Uncertainty Estimation for Learning-based Registration in Neuroimaging},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w8LMtFY97b},\nnote={under review}\n}" }, "abstract": { "value": "Over recent years, deep learning based image registration has achieved impressive accuracy in many domains, including medical imaging and, specifically, human neuroimaging with magnetic resonance imaging (MRI). However, the uncertainty estimation associated with these methods has been largely limited to the application of generic techniques (e.g., Monte Carlo dropout) that do not exploit the peculiarities of the problem domain, particularly spatial modeling. Here, we propose a principled way to propagate uncertainties (epistemic or aleatoric) estimated at the level of spatial location by these methods, to the level of global transformation models, and further to downstream tasks. Specifically, we justify the choice of a Gaussian distribution for the local uncertainty modeling, and then propose a framework where uncertainties spread across hierarchical levels, depending on the choice of transformation model. Experiments on publicly available data sets show that Monte Carlo dropout correlates very poorly with the reference registration error, whereas our uncertainty estimates correlate much better. % with the reference registration error. Crucially, the results also show that uncertainty-aware fitting of transformations improves the registration accuracy of brain MRI scans. Finally, we illustrate how sampling from the posterior distribution of the transformations can be used to propagate uncertainties to downstream neuroimaging tasks." }, "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 registration", "uncertainty estimation", "medical image analysis" ] }, "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/7857488e0068acac3bf9c85fceaa727191e3b603.pdf" }, "presentation": null, "primary_area": { "value": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, 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": "Hierarchical Uncertainty Estimation for Learning-based Registration in Neuroimaging" }, "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": "- What is the input to tSNE in fig.1(b)? Current caption is quite confusing, and I don't understand why the visualization shows cost is sparser.\n- What are the $\\mu$ and $\\hat{Q}$ in eq.(8)?\n- How do you get the true Q-values in table 1? E.g., what is the policy for MC simulation? how many simulations do you get for the Q computing? And are the state-action pairs for evaluation in \"dataset\" in table 1 the same as the VPA training dataset?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Offline-to-online method is important to the practical application of safe RL.\n- The two challenges pointed out by this paper (i.e., Q value estimation error and Lagragian multiplier mismatch) are insightful.\n- The experiment is extensive, and the proposed method outperforms the compared baseline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to solve two issues in offline-to-online safe RL, Q value estimation error and Lagragian multiplier mismatch, and proposes a new algorithm. The proposed method first re-trains Q function with an online objective to correct Q estimation and then uses an adaptive PID control to update Lagrangian multiplier. The authors run experiment on several safe RL tasks and claim their method outperforms other baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some concerns on VPA:\n - If you believe the pretrained Q functions from offline learning is not accurate, why not directly learn a new Q function by online objective or eq.(8)&(9) instead of training based on the pretrained Q functions? I believe it is a very straightforward baseline which should be compared.\n - The error in Q value estimation is not only from the mismatch between offline and online learning objective, but also from the distribution shift issue due to the online policy update. Can the proposed method mitigate error from this aspect?\n\n- Although the experiments show that adapative PID Lagrangian updater can stablize the cost performance, its effectiveness is still questionable. The adaptive PID introduces three more hyperparameters $\\alpha, \\beta, \\gamma$. Along with original $K_p, K_i, K_d$, there are **six** hyperparameters for the updater in total. However, the adaptive PID only shows a marginal improvement over PID in fig.5. Meanwhile, the original hyperparameters of PID are not well tuned. The authors select $K_p=10^{-4}, K_i=10^{-5}, ...$, which are not only very different from the original PID Lagrangian paper [1], but also different from the recent safe RL benchmarks using PID update [2][3], which provide more stable performances of PID update than the reported one in fig.5. \n\n- The experiment results show that the proposed method outperforms other offline-to-online baselines or naively starting from offline policy. However, in the experiment, we can find those offline-to-online methods are even worse than purely online learning (i.e., learning from scratch). Therefore, I believe those baselines are not strong enough to illustrate the effectiveness of new method. Meanwhile, the performance of purely online learning is also significantly under-reported: SAC-Lag can quickly achieve ~700 reward with even smaller cost in Ballcircle according to [3]. It can also achieves > 2500 reward in Halfcheetach-velocity in [2]. \n\nminor issues:\n- I don't think PPO (\"Schulman et al 2017\", the first citation in line 93) is a safe RL algorithm, it is a standard RL algorithm.\n\n[1] Responsive Safety in RL by PID Lagrangian Methods\n\n[2] OmniSafe: An Infrastructure for Accelerating Safe Reinforcement Learning Research, https://www.omnisafe.ai/en/latest/\n\n[3] FSRL, https://fsrl.readthedocs.io/en/latest/index.html" }, "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) Can you clarify what the x-axis “steps” means in Figure 1 and Figure 2? Does it mean the optimizer update times?\n\n(2) Can you compare your method to pure offline safe RL baselines such as CDT [1] and pure online safe RL baselines such as PPO-Lag and SAC-Lag?\n\nReference:\n\n[1] Zuxin Liu, et al. \"Constrained decision transformer for offline safe reinforcement learning.\" International Conference on Machine Learning. PMLR, 2023." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(1) Interesting topics: it focuses on an interesting topic: safety in reinforcement learning, specifically exploring offline-to-online adaptation to enable safer and more efficient policy learning—a crucial yet underexplored area in RL." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the challenges of safe RL by proposing Marvel, a framework that bridges offline and online RL to achieve safer and faster policy learning. Based on the authors’ claim, Marvel tackles these issues with two key components: Value Pre-Alignment, which adjusts Q-functions for accurate value estimation, and Adaptive PID Control, which optimizes Lagrange multipliers for balancing rewards and safety constraints during online fine-tuning. Experimental results show that Marvel outperforms presented baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) Typo: In Lines 413 and 414, the HalfCheetah, Hopper, and Swimmer in DSRL are from Safety-gymnasium [1], not Bullet-Safety-Gym.\n\n(2) Confusion about experiment results: Why do the results of training-from-scratch look so poor? Based on the report [1], SAC-Lag should achieve 600+ rewards on Ball-Circle and 350+ rewards on Car-Circle.\n\n(3) Concerns about the experiment results. The experiment results shown in Figure 6 indicate that the proposed method can not reach a reasonable reward. For example, in Drone-Circle, the agent should achieve a reward of 600+ using SAC-Lag. However, the proposed Marvel only achieves 150-.\n\nReference:\n[1] https://fsrl.readthedocs.io/en/latest/tutorials/benchmark.html" }, "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. How would the method perform if equipped with a large-scaled model, such as the structure of Guided Online Distillation?\n \n2. How does aPID improve baselines, as the paper claims it can enhance their performance?\n\n3. Could other types of O2O baselines be considered, such as Online Decision Transformers [1], which do not involve Q-functions?\n\n4. What are the base algorithms for Warm Start?\n\n5. A guide for hyperparameter settings would be beneficial, such as for $\\alpha$ and $\\alpha_c$. Is the method sensitive to these hyperparameters?\n\n[1] Zheng Q, Zhang A, Grover A. Online decision transformer[C]//international conference on machine learning. PMLR, 2022: 27042-27059." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The explanatory experiments in the Method section really enhance the reader's understanding of the proposed methods. Improvements in the clarity of some figures could further aid in the visualization of the results.\n \n2. The paper is well-written, with each part of the methods meticulously described and well-motivated, contributing to a coherent and comprehensive presentation of the research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on offline-to-online (O2O) safe reinforcement learning (RL). It identifies two main issues causing inefficiency in previous methods: erroneous Q-estimations and Lagrangian mismatch. To address these issues, the authors propose two methods: Value Pre-Alignment and Adaptive PID Control. These methods aim to improve the standard O2O safe RL framework. Several experiments provide evidence of the effectiveness of these two components." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The algorithms and baselines lack stability in some environments. More repeated experiments may be needed. The improvement over baselines is not significant, showing performance gaps only in two tasks: BallCircle and CarCircle.\n \n2. All figures need to be polished, especially Figure 3. The legend should not be limited to one or two figures but could be placed below all figures. Additionally, the lines representing each method should be improved, as it is difficult to distinguish between methods in some 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": 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 meanings of 'Warm Start', 'From Stractch', and 'PID' are not clear in introduction, which makes the Figure 1 is hard to understand." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The O2O in safe RL is a very important but overlooked issue. This paper systematically explores this issue.\n2. The paper's writing logic is very clear, progressing methodically from problem analysis, to problem formulation, and finally to the solution. So the understanding is relatively easy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This article systematically investigates the challenges of O2O (Offline-to-Online) in safe RL and proposes corresponding solutions. Specifically, the paper 1) employs value pre-alignment to counteract erroneous Q-estimations and 2) utilize adaptive PID control to overcome Lagrangian mismatch. The method introduced has demonstrated outstanding performance across datasets in DSRL." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper studies the problem of Offline to Online finetuning for safe RL. The first challenge can be viewed as a common issue in the O2O (Offline-to-Online) domain. The second challenge is unique to safe RL, but I question its significance. Figure 2 suggests that a good initialization of Lagrange multipliers is more conducive to ensuring a reasonable cost. However, even if the initial values of the Lagrange multipliers are not reasonable, they can be adjusted through Equation (3). The proposed aPID seems to merely expedite this adjustment process, as shown in Figure 5. In addition, all baselines involve aPID, but the performance is still poor. This also indicates that **challenge 2 is not crucial**.\n\n2. Guided Online Distillation is an important baseline of this paper. Regarding the reason \"its usage of large pretrained model leads to an unfair comparison with standard RL frameworks\", I have some disagreements. Guided Online Distillation is based on decision transformer (DT). Despite having more parameters, DT is far away from a large model. Moreover, DT's performance on D4RL is typically weaker than that of standard RL algorithms, such as IQL." }, "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. For Value Pre-Alignment, how do the entropy terms in equations (8) and (9) contribute to achieving the desired optimistic/pessimistic estimation of rewards/costs? Besides, when calculating Spearman’s rank correlation coefficient, how are the \"true\" Q-values obtained or estimated?\n2. For adaptive PID, could the authors explain why the non-linearity, i.e., tanh, is applied only to $K_p$ rather than to all $K_p$, $K_i$, and $K_d$, given that the authors state that the non-linearity can help respond quickly to larger errors while avoiding frequent adjustments and reducing oscillations (lines 351-354)? Additionally, the aPID introduces several more hyperparameters, e.g., the initial $K_p$, $K_i$, $K_d$, and their respective minimum and maximum values; how are these hyperparameters selected in practice, and is the proposed method sensitive to variations in their values?\n3. For evaluation of O2O offline RL methods[1, 2, 3], typically it involves first assessing the performance of the offline pretrained policy (e.g., rewards and costs), followed by online finetuning, and then evaluating the finetuned policy to observe performance changes. To strengthen the analysis, could the authors provide: \\\n a) Performance metrics for the offline pretrained policies prior to fine-tuning. \\\n b) Details on the online finetuning process, including the number of online interaction steps and the cumulative cost, e.g., Figure 2 in [4]. \\\n c) Comparison of the finetuned policy performance with pure online safe RL methods. \\\n d) Analysis of whether the proposed method achieves higher rewards with fewer constraint violations. \\\n[1] https://proceedings.mlr.press/v202/yu23k/yu23k.pdf \\\n[2] https://arxiv.org/pdf/2110.06169 \\\n[3] https://arxiv.org/pdf/2201.10070 \\\n[4] https://proceedings.mlr.press/v202/liu23l/liu23l.pdf" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The motivation for offline-to-online safe RL is clear, focusing on using offline data to accelerate online fine-tuning and reduce risky interactions in safety-critical scenarios.\n2. The design of the method appears reasonable, introducing two key components that seem well thought out.\n3. The results seem to indicate better performance, with evaluations conducted across various robots and tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the warM-stArt safe Reinforcement learning with Value prE-aLignment (Marvel) framework, featuring Value Pre-alignment (VPA) and Adaptive PID Control (aPID). VPA aligns pretrained Q-functions with true Q-values using offline data, promoting active exploration while managing costs. aPID adjusts Lagrange multipliers based on cost violations, ensuring stable online learning. Marvel outperforms baselines, achieving safe, high-reward policies with a few online interactions, and is compatible with many state-of-the-art safe RL methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The method lacks theoretical justification regarding sample efficiency, leaving it unclear how it enables more efficient online finetuning.\n2. The evaluation results are unconvincing, with certain key comparisons missing." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper presents Marvel, a framework for offline-to-online safe reinforcement learning using a pretrained-and-finetuning approach." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024marvel,\ntitle={Marvel: Accelerating Safe Online Reinforcement Learning with Finetuned Offline Policy},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w9bWY6LvrW},\nnote={under review}\n}" }, "abstract": { "value": "The high costs and risks involved in extensive environment interactions hinder the practical application of current online safe reinforcement learning (RL) methods. While offline safe RL addresses this by learning policies from static datasets, the performance therein is usually limited due to reliance on data quality and challenges with out-of-distribution (OOD) actions. Inspired by recent successes in offline-to-online (O2O) RL, it is crucial to explore whether offline safe RL can be leveraged to facilitate faster and safer online policy learning, a direction that has yet to be fully investigated. To fill this gap, we first demonstrate that naively applying existing O2O algorithms from standard RL would not work well in the safe RL setting due to two unique challenges: \\emph{erroneous Q-estimations}, resulted from offline-online objective mismatch and offline cost sparsity, and \\emph{Lagrangian mismatch}, resulted from difficulties in aligning Lagrange multipliers between offline and online policies. To address these challenges, we introduce \\textbf{Marvel}, a novel framework for O2O safe RL, comprising two key components that work in concert: \\emph{Value Pre-Alignment} to align the Q-functions with the underlying truth before online learning, and \\emph{Adaptive PID Control} to effectively adjust the Lagrange multipliers during online finetuning. Extensive experiments demonstrate that Marvel significantly outperforms existing baselines in both reward maximization and safety constraint satisfaction. By introducing the first policy-finetuning based framework for O2O safe RL, which is compatible with many offline and online safe RL methods, our work has the great potential to advance the field towards more efficient and practical safe RL solutions." }, "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": [ "Offline-to-Online Fine-tuning", "Safe Reinforcement Learning", "Constrained Markov Decision Processes", "Reinforcement 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/fced1fa15ce1f51f8f650fa1e9012f5ca03c6cd8.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/0a819861a02a2b53667051e11048199394199f5d.zip" }, "title": { "value": "Marvel: Accelerating Safe Online Reinforcement Learning with Finetuned Offline Policy" }, "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": 4 }, "primary_area": null, "questions": { "value": "On page 1, I was confused by the phrase “tasks are distinct from the training data.” What does this mean?\nAre there any other reasons to account for why human preference data might be preferable to sparse reward functions?\nHow do you actually generate the 6 reward function candidates? Do you randomly sample from the LLM? If so, how?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I appreciated the fact that this paper took steps to optimize their baseline methods within reason. For instance, for Eureka, the authors continued generating candidate reward functions until the LLM had generated 6 executable ones (to make things fair for comparison against their own method).\n\nThe comparison against PrefPPO was strong." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel framework called In-Context Preference Learning (ICPL), which automatically generates dense reward functions by utilizing an LLM capable of querying humans for preference data. The authors find that their method greatly outperforms one baseline, PrefPPO, with respect to sample efficiency (PrefPPO requires far more human preference queries) and task performance. The authors also find performance comparable to that of Eureka, a baseline that also utilizes an LLM for generating dense reward functions but relies upon access to ground-truth sparse reward data rather than human preference data. The authors argue, since ICPL does not require access to a ground-truth sparse reward function, it has a clear advantage for tasks that are less well-defined or require human intuition. Additionally, they argue that training with human preferences will enable greater model alignment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "According to table 1, ICPL performance seems no better than that of Eureka. Furthermore, substituting in the values from table 3, ICPL performance with real human preference queries does not exceed Eureka’s performance on any task except Ant. Since ICPL does not outperform Eureka, ICPL’s benefit relies upon the ease of obtaining human preference queries in comparison with a ground-truth sparse reward function. I’m not convinced that this benefit is significant.\n\nOne argument, from the introduction, is an appeal to the success of preference-based training in other domains. I’m not convinced that this success generalizes to the domain of LLM-generated reward functions. \n\nThe other core argument in favor of preference-based training is that human insight—expressed through preference queries—can better align agent behavior with human preferences. The authors motivate this through their custom HumanoidJump task, wherein the task is “to make humanoid jump like a real human.” They argue that this is a domain in which designing a sparse reward function would be difficult due to the nuances/subjectivity of mathematically defining jumping “like a real human.” In my mind, the paper largely hinges on this argument, however the authors only offer one case-study as evidence of the efficacy of human preference data in this domain.\n\nI could be convinced otherwise, but I think there would need to be a more thorough analysis of human preferences in comparison with sparse reward functions in order to be certain.\n\nAlso, I found section 5.3.2: Evaluation Metric to be very confusing. I wasn’t sure what an “environment instance” was. I also didn’t understand which set of task metric values was used to compute the maximum for the RTS." }, "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": "Please clarify the points mentioned above, that would really help me make a better decision about the paper. In particular explaining why this method would be better in some way that EUREKA\n(Either because ICPL doesn't require some assumption made by EUREKA or it's better in some other way).\n\nMinor:\nWhy GPT-4o for the human experiment only? I'm not sure how much it matters actually, but found it curious." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The idea is generally really interesting and compelling. The idea of having LLMs generate an initial reward function and then iteratively repeat it is really interesting.\n\nThe human study was really compelling and thought out. It's really good that this was actually tried and not just assumed it would work with real human feedback.\n\nPaper really well presented, ideas presented very clearly. Motivation clear and compelling." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces ICPL, a method for iteratively improving reward functions for RL problems with human feedback. The method has LLMs generate reward functions specified by code, trains and executes these rewards, and then ranks the final trajectories with human feedback to then update the reward functions again." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am struggling to figure out what the compelling advantage is of this method over the baseline Eureka. As far as I understood reading the paper, Eureka operated from the same set of assumptions about the environment as did ICPL. And in the non-human experiment performed very similarly. In the related work it says that EUREKA requires humans to give feedback in text, whereas ICPL only requires ranked preferences. During the description in 5.2 it also says that sparse rewards are used to select the best candidate reward function. Does that mean that this is additional assumptions EUREKA needs. There was also not a comparison to EUREKA in the human study. Was that because it would not work without these other assumptions? I think it's possible I'm just misunderstanding here, so if authors could clarify this point it would really help me understand the paper and potentially improve my rating.\n\nIt's stated in the intro and conclusion that ICPL surpasses RLHF is efficiency, but RLHF is not mentioned anywhere in the experiments. Is this an experimental finding of the paper, or are authors just saying based on known findings about the efficiency of RLHF. Could a direct comparison be made in the first (non-human) experiments since you don't need actual humans and can thus potentially run more. More clarity on this point would really help.\n\nBased on 5 iterations, I'm not sure that you can make the claim that it will monotonically improve much past that point. Did authors try past 5 (10, 20).\n\nOne sort of undiscussed thing here is that, requiring new models to be trained every iteration does mean that loop is pretty slow. Was 5 iterations chosen for that reason (so it wouldn't take multiple days). This should be maybe discussed as a weakness. E.g. for human studies or using humans, doesn't that mean the humans need to wait hours or else get new humans to provide feedback for every iteration?" }, "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": "* What was the demographic data for the human provided feedback? \n\n* It seems like Eureka has very similar performance to ICPL, what would you say is the benefit of your method compared to Eureka? Eureka seems to have some constraints but it would be nice to show in experimentation or come up with a scenario where it would fail.\n\n* It would be good to justify the length of the paper. For example, what sections do you believe require the additional space? You are of course free to use all the space but the readability of the paper could improve by making it more crisp.\n\n* Why did you not use PEBBLE as a basline given that you made use of BPref? Also, did you consider any other baselines as there are more recent works [1,2,3] (To name a few)? It would be great if you discuss how you determined which baseline to use and if you considered any others.\n1. Kim, C., Park, J., Shin, J., Lee, H., Abbeel, P., & Lee, K. (2023). Preference transformer: Modeling human preferences using transformers for rl. arXiv preprint arXiv:2303.00957.\n2. Park, J., Seo, Y., Shin, J., Lee, H., Abbeel, P., & Lee, K. (2022). SURF: Semi-supervised reward learning with data augmentation for feedback-efficient preference-based reinforcement learning. arXiv preprint arXiv:2203.10050.\n3. Marta, D., Holk, S., Pek, C., Tumova, J., & Leite, I. (2023, October). VARIQuery: VAE Segment-Based Active Learning for Query Selection in Preference-Based Reinforcement Learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp. 7878-7885). IEEE.\n\nI am more than willing to up the score given reasonable answers to these points." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Demonstrates a substantial reduction in the number of human queries needed for preference-based learning which is sorely needed since human-in-the-loop approaches should ideally just require a handful of preferences.\n\n* It's appreciated that the evaluations of the method is done both in synthetic data and with real humans.\n\n* The paper is well written and the provided method is explained well. Even tho generating reward functions from LLMs is not novel, the way the iteratively make use of human preferences to update their prompt is." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes In-Context Preference Learning (ICPL), a method using LLMs for more efficient preference-based reinforcement learning. ICPL uses LLMs, such as GPT-4, to synthesize reward functions based on environment context and task descriptions. These generated functions are refined through human feedback iteratively. The approach shows significant efficiency gains, requiring fewer preference queries compared to traditional RLHF and achieving comparable or superior performance to methods using ground-truth reward functions. The experiments validate ICPL's performance across various simulated and real human-in-the-loop RL tasks, showcasing robustness in complex, subjective environments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* As with all works that uses LLMs to generate reward functions from human feedback I question how well it will perform with more complex tasks which is one of the big reason for using human feedback.\n\n* The synthetic experiment uses completely noiseless preferences while the standard in these kind of control environments are typical a noise of let's say 10%. What is the rationale for using noiseless preferences and what would be the effect of noisy preferences for your method? \n\n* While the authors uses B-Pref from Kimin et al for some reason they use only the PPO version even tho the repository is more associated with PEBBLE the SAC version. Why is SAC not used as well?\n\n* 6 participants are very low for a study with humans. Still, it is better than some papers that run their method with just the authors feedback. It would be nice with some more information about the experiment like demographic data as well as discussing the limitation of a smaller sample size when it comes to generalizability.\n\nMinor things:\n* You introduce the same abbreviation on multiple occassions.\n\n* To make the related work more complete, there is another paper using LLMs with preferences.\n1. Holk, S., Marta, D., & Leite, I. (2024, March). PREDILECT: Preferences Delineated with Zero-Shot Language-based Reasoning in Reinforcement Learning. In Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction (pp. 259-268)." }, "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": 3 }, "primary_area": null, "questions": { "value": "(1) Why is it necessary to use pair-wise human feedback (a good example and a bad example) if RTS is available? Why not just use all the reward functions with their RTS as prompt (maybe together with other information like reward trace, differences, etc) to generate reward functions?\n\n(2) Could you please explain the counter-intuitive results in Table 2? It seems the more prompt components you remove (from w/o RT, to w/o RTD, to w/o RTDB), the better performance it gets (w/o RT wins 2 tasks, w/o RTD wins 3 tasks, and w/o RTDB wins 4 tasks), but adding all the components back, i.e., ICPL(Ours), it wins all the tasks." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "It replaces the implicit reward model in traditional RLHF with an LLM and its output reward function. This enhances the interoperability and capacity of the reward design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a reward design method. It uses LLMs to generate reward functions to calculate the reward, and the prompt of the LLM is learned through human feedback of the policy rollouts and other historical information in the loop.\n\nIt replaces the implicit reward model in traditional RLHF with an LLM and its output reward function. This enhances the interoperability and capacity of the reward design." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) ICPL involves human labor, but does not show any significant gain over Eureka, which doesn't require any human feedback.\n\n(3) For challenging tasks, true human feedback does not work better than proxy human feedback. This undermines the necessity of involving humans.\n\n(2) For challenging tasks, like humanoid jump task, ICPL does not have any solid comparisons with other baselines." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose In-Context Preference Learning, a method that address the challenges of reward design in reinforcement learning through the integration of large language models and human preferences." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024fewshot,\ntitle={Few-shot In-context Preference Learning using Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=w9tS6NRmxX},\nnote={under review}\n}" }, "abstract": { "value": "Designing reward functions is a core component of reinforcement learning but can be challenging for truly complex behavior. Reinforcement Learning from Human Feedback (RLHF) has been used to alleviate this challenge by replacing a hand-coded reward function with a reward function learned from preferences. However, it can be exceedingly inefficient to learn these rewards as they are often learned tabula rasa. We investigate whether Large Language Models (LLMs) can reduce this query inefficiency by converting an iterative series of human preferences into code representing the rewards. We propose In-Context Preference Learning (ICPL), a method that uses the grounding of an LLM to accelerate learning reward functions from preferences. ICPL takes the environment context and task description, synthesizes a set of reward functions, and then repeatedly updates the reward functions using human feedback over videos of the resultant policies over a small number of trials. Using synthetic preferences, we demonstrate that ICPL is orders of magnitude more efficient than RLHF and is even competitive with methods that use ground-truth reward functions instead of preferences. Finally, we perform a series of human preference-learning trials and observe that ICPL extends beyond synthetic settings and can work effectively with humans-in-the-loop." }, "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", "Preference-based RL", "Reinforcement Learning from Human Feedback", "Reward Design" ] }, "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/3067d58f5cdf6fa5658e28108631bc6d9fec285d.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": "Few-shot In-context Preference Learning 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": 4 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. This work effectively incorporates quantization-aware training into co-training and significantly reduces the performance gap between multi-task co-trained models and their 4-bit quantized counterparts.\n2. The authors design task-specific learnable multi-scale activation quantizer and SLLD to solve the issues of naive integration.\n3. From the experimental results, it appears that the author's techniques are effective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The study finds that directly applying co-training to existing QAT methods significantly degrades performance. The main issue identified is the inadequacy of activation quantization scales in the co-training framework. To address this, the authors propose a Task-Specific Scales Quantization method suitable for multi-task co-training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In table 1, quantitative results for super-resolution tasks are shown. But I am still curious about the results of deraining and denoising tasks.\n2. How about the parameters change of your method?\n3. Can you provide the some comparisons with some SOTA single task methods to further demonstrate the superiority of your 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": 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": "As listed in Weaknesses, there are two questions:\n1. Could the authors provide the computational cost of training and inference compared to existing QAT methods?\n2. Could the authors report the variance of the results of the proposed SLLD method in the ablation studies?" }, "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 provide a comprehensive evaluation of the challenges of the task, i.e., directly integrating multi-task co-training with QAT. This helps clarify the bottleneck of existing QAT methods and motivates the proposed TSQ-MTC method to address the performance degradation issue.\n\n2. The proposed TSQ-MTC method introduces two novel components, TLMAQ and SLLD, to enhance the representational ability of shared features across different tasks and preserve the information from full-precision features in deeper layers of the Transformer model. \n\n3. The experimental evaluation across the main text and appendices provides detailed insights into the effectiveness and versatility of the proposed TSQ-MTC method in two co-training data scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Task-Specific Scales Quantization for Multi-Task Co-Training (TSQ-MTC) to address the performance degradation issue of existing quantization-aware training (QAT) methods when integrated with co-training. The proposed method introduces a task-specific learnable multi-scale activation quantizer (TLMAQ) to enrich the representational ability of shared features across different tasks and a structure-based layer-by-layer distillation (SLLD) to ensure that the quantized features effectively preserve the information from their full-precision counterparts. Extensive experiments on two co-training data scenarios demonstrate the effectiveness of TSQ-MTC, which achieves a 4-bit quantized low-level visual foundation model based on IPT with a PSNR comparable to the full-precision model and a 7.99× compression ratio in the ×4 super-resolution task on the Set5 benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper referred the challenge of computational and memory overhead of co-trained models (Lines 047-049), but the computational complexity and efficiency of the proposed TSQ-MTC method are not discussed in detail. \n\n2. The authors detailedly analysis the proposed SLLD in Table 4 and Figure B. However, according to the ablation studies provided in Table 3, Table D, Table E, the proposed SLLD has minor improvements in performance given the potential for minor fluctuations in experimental 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": 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": "Apart from the points (2) and (3) outlined in the \"Weaknesses\" section, the following are additional concerns:\n(1) While the model design is good and interesting, the model seems a little complex. Would it be possible to provide some comparative analyses regarding the model's speed and complexity?" }, "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 method effectively addresses the scale discrepancies among multi-task features, aligning with intuitive expectations.\n(2) The novelty is good, the introduction of task-aware quantization into multi-task learning represents an innovative approach.\n(3) Method is architecture-agnostic, featuring a wide range of applicability.\n(4) originality, quality, clarity, and significance: This paper demonstrates good originality, is well-written, and clearly communicates ideas. Furthermore, the application of quantization to enhance model efficiency holds significant importance for multi-task learning" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel method for multi-task network quantization-aware training, addressing performance degradation caused by model quantization. The approach incorporates two techniques: (1) Task-Specific Learnable Multi-Scale Activation Quantizer (TLMAQ): Addressing the scale conflict in quantizing diverse task features, improving the quantized model's representational capabilities across tasks. (2) Structure-based Layer-by-Layer Distillation (SLLD): Strengthening full-precision model supervision over quantized models, reducing information distortion from quantization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) There is a minor error in Section 2.1 - regarding the order of description for the \"single-modal task-related data scenario\" and the \"multi-modal data scenario\" are misaligned with the order of \"former scenario\" and \"latter scenario\". \n(2) The paper only considered the low-level multi-task scenarios; however, However, it lacks effective exploration of high-level visual tasks. Specifically, Can the proposed method still maintain effectiveness when processing high-level visual multitasking data such as NYUD-v2, Pascal Context and Cityscapes.\n(3) The experiment involved a comparison of model quantization techniques, including LSQ (2019) and PAMS (2020). Nevertheless, a comparison with the most recent Quantization-Aware Training (QAT) methods is conspicuously absent." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024efficient,\ntitle={Efficient Low-Bit Quantization with Adaptive Scales for Multi-Task Co-Training},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=wA2RMD2AFq},\nnote={under review}\n}" }, "abstract": { "value": "Co-training can achieve parameter-efficient multi-task models but remains unexplored for quantization-aware training. Our investigation shows that directly introducing co-training into existing quantization-aware training (QAT) methods results in significant performance degradation. Our experimental study identifies that the primary issue with existing QAT methods stems from the inadequate activation quantization scales for the co-training framework. To address this issue, we propose Task-Specific Scales Quantization for Multi-Task Co-Training (TSQ-MTC) to tackle mismatched quantization scales. Specifically, a task-specific learnable multi-scale activation quantizer (TLMAQ) is incorporated to enrich the representational ability of shared features for different tasks. Additionally, we find that in the deeper layers of the Transformer model, the quantized network suffers from information distortion within the attention quantizer. A structure-based layer-by-layer distillation (SLLD) is then introduced to ensure that the quantized features effectively preserve the information from their full-precision counterparts. Our extensive experiments in two co-training scenarios demonstrate the effectiveness and versatility of TSQ-MTC. In particular, we successfully achieve a 4-bit quantized low-level visual foundation model based on IPT, which attains a PSNR comparable to the full-precision model while offering a $7.99\\times$ compression ratio in the $\\times4$ super-resolution task on the Set5 benchmark." }, "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": [ "Low-Bit Quantization", "Multi-Task Learning", "Co-Training", "Quantization-Aware Training", "Quantization Scale" ] }, "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/63ce18b55aaa5bdebb47260c4711bd8f5c76408a.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/1b752698bd52c700732a1305a8976fb0b18f282f.zip" }, "title": { "value": "Efficient Low-Bit Quantization with Adaptive Scales for Multi-Task Co-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": 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": "Questions are asked in the weakness section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "•\tThe paper’s approach is innovative in building on the NeuroAI benchmarking with dynamic visual responses rather than using static scene responses, which are commonly evaluated by current image model. It’s the first investigation of benchmarking many models in response to naturalistic videos of human actions.\n\n•\tThe paper gives a comprehensive model evaluation experiments conducting with this dataset. Spanning from video, language, image models over 350+, including a variety of architectures and objectives.\n\n•\tThe paper is fully public with all data, code, model accessible. Further, it provides an interesting direction that Human-aligned DNNs may be a promising direction for dynamic social perception, and suggests that developing models that can handle relational and temporal elements essential for social scene understanding." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the author extends a dataset of natural videos describing human action interactions by providing human-annotated sentences for each video and investigates the limitations of over 350+ models to predict human behavioral ratings and neural responses to dynamic social scenes. It concludes that language models predict action and social ratings better than image and video models but perform poorly at predicting neural responses in the lateral stream and provides insights into how well current AI systems replicate human social vision. More importantly, the author highlights the gap in current models' ability to understand dynamic social interactions and suggests potential directions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "•\tLimited Coverage or advanced video and language models: Although the dataset has been tested in 350+ models. The majority of them are obsoleted and cannot fully present the overall performance on the state-of-the-art image, video, and language models, like MViT, Co-DETR, DINO, GPT4o, LLAVA, Llama, etc. The most recent model on the paper’s list is up to 2021.\n\n•\tThe author conducts experiments on a dataset, consisting of 250 3-sec videos, which is relatively small and less representative for training and evaluating deep learning models for making a significant claim on social action recognition. The data limitation might reduce the generalizability of getting conclusions when trading this complex multi-agent interactions task. \n\n•\tWhile the paper make a claim that language models are successful in predicting behavioral response but not in neural responses since providing language captions is insufficient for achieving neural alignment, it encourages the society to make a better connection between neural responses and models with a more well-designed approach. Instead, the conduct of experiments are insufficient in pinpointing precise architectural or training factors that contribute to performance differences by given task." }, "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) I am unfamiliar with the usage of fMRIs in machine learning. But I imagine the variance from individual to individual must cause difficulties in predicting brain responses. Does it not make more sense to condition fMRI prediction on the baseline fMRI readings before the video viewing? Is the text or video input enough?\n2) Figure 7 and Figure 8 mention image-language models being evaluated - but to my understanding, all models only take one modality. Are there models that take both images and language?\n3) Action datasets deal with label subjectivity - is there any disagreement across annotators concerning the dataset being trained over? I imagine this is particularly a problem in the domain of social action." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) Action datasets contain both physical and social actions, and are not focused on the exclusive modeling of either action. Exploring models trained over social actions exclusively provides very valuable insights on the difficulty of this domain of action.\n2) The evaluation is very broad - 350 models is very impressive. The claims in the discussion are well supported." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The primary contribution of this work is an analysis of 350 models over an existing dataset of social actions. The labels range from user ratings to fMRI images for brain regions engaged in the watching of videos of social action. The authors find that networks trained over images outperform language/video modalities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The size of the dataset is very small (200 training videos), and the results are definitely impacted by this. Models trained over video datasets in particular must be large due to the variation over the time dimension. The models (and modalities) that perform well might largely be because of the size of the dataset.\n2) Audio as a modality is missing, but I would argue is just as valuable as the sequence of images alone across many of the subjects (e.g. valence, arousal). Audio provides less benefit in action datasets focused on physical actions, but might be just as important as the visual modality w.r.t. social actions.\n3) There is a lack of discussion around pre-training of the different models. This bears more importance than the model architecture (CNNs vs Transformers) especially due to the small size of the dataset. The video models may or may not be trained over datasets that include social actions (like Kinetics)." }, "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": { "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": "see weaknesses" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors extend a human social video dataset with human annotated text descriptions.\n2. The authors benchmark 350 image, video and language models for behavior rating and neural response prediction.\n3. The authors highlight gaps in alignment of these models and compare their performance along different axes like architecture, training objective etc." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. The authors benchmark hundreds of image, video and language models for behavior ratings and neural responses based on human social videos and their captions.\n2. The authors compare these models based on their predictions of the behavior rating and neural responses.\n3. The authors present several conclusion about the compared models and highlight gaps in their alignment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Just mentioning broad non-exhaustive categories like \"self/category supervision, multi-modal and convolution vs transformer\" is not a systematic approach to model selection. The authors should first select dimensions of model categorisation they want to compare like supervision type, generative or discriminative, model size, modality etc. and then choose models representing each type. The entire process needs to be detailed to make sure there is no bias in model selection that might affect conclusions downstream. The authors should provide their detailed model selection method.\n\n2. While the authors benchmark several vision models, the vast majority of them are trained for image classification. Models trained with other objectives like object detection (only 2 used as far as I can tell), segmentation (none used as far as i can tell), masked reconstruction (none used as far as i can tell) etc. should also be benchmarked. More generally, vision based models should be categorised based on the different objectives used to train them and compared to provide insights into how different training objectives affect performance/alignment.\n\n3. While there are several generative language models benchmarked, the vision language models are primarily discriminative. Generative vision models like diffusion based (like stable diffusion), GAN based (like VQ-GAN) and VAE (like VQ-VAE) based models should also be benchmarked for insights into the performance gap between generative and discriminative representations for both language and vision models.\n\n4. The comparisons between the benchmarked models need to be fair in terms of the number of parameters. It looks like the authors have compared the models regardless of the size. The authors should investigate and report whether there exists trends between prediction performance and model size for insights into how model size affects representation quality/alignment." }, "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": "- Please include additional details about the experiments.\n- It seems to remain unclear whether developing a human-aligned model would enhance existing AI models, why not test some social behavior tasks used in the computer vision community, such as the social interaction tasks in [Ego4D [CVPR2022]](https://arxiv.org/abs/2110.07058) ? (not required to conduct such experiments)\n\nMoreover, given my limited expertise in this specific topic, I recommend that the significance of this paper be assessed by reviewers with specialized knowledge in this area." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The relationship between AI models and the human brain is an important area of study.\n- This paper presents a highly comprehensive benchmark for examining how AI models—spanning image, video, and language models—respond to social interactions compared to human brain responses in similar scenarios.\n- The limitations and discussions offer valuable insights that can inform and guide future research developments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Unlike previous neuroscience AI studies that focus on deep learning models' responses to static images, this paper examines models' responses to human social interactions in dynamic videos. The authors create a small benchmark and evaluate an impressive number of models (over 350 image, video, and language models) on this dataset. These comprehensive experimental results could offer valuable insights for researchers in this field." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The technical contribution is limited. While I understand the overall effort and workload invested in this paper, it does not address the question of how to develop a human-like AI model.\n- Questions regarding the evaluation method: Why is a linear mapping applied between the extracted features and human data, such as fMRI responses? The rationale or assumptions behind this choice of linear mapping are not clear to me. Furthermore, is it appropriate to apply the same linear mapping approach across all models, despite their significant differences? If the evaluation method is not well-justified, the overall efforts should be reassessed carefully." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This study benchmarks 350+ AI models against human behavioral and neural responses to videos of social actions and highlights significant gaps in AI's ability to model dynamic social vision." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024modeling,\ntitle={Modeling dynamic social vision highlights gaps between deep learning and humans},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=wAXsx2MYgV},\nnote={under review}\n}" }, "abstract": { "value": "Deep learning models trained on computer vision tasks are widely considered the most successful models of human vision to date. The majority of work that supports this idea evaluates how accurately these models predict behavior and brain responses to static images of objects and scenes. Real-world vision, however, is highly dynamic, and far less work has evaluated deep learning models on human responses to moving stimuli, especially those that involve more complicated, higher-order phenomena like social interactions. Here, we extend a dataset of natural videos depicting complex multi-agent interactions by collecting human-annotated sentence captions for each video, and we benchmark 350+ image, video, and language models on behavior and neural responses to the videos. As in prior work, we find that many vision models reach the noise ceiling in predicting visual scene features and responses along the ventral visual stream (often considered the primary neural substrate of object and scene recognition). In contrast, vision models poorly predict human action and social interaction ratings and neural responses in the lateral stream (a neural pathway theorized to specialize in dynamic, social vision), though video models show a striking advantage in predicting mid-level lateral stream regions. Language models (given human sentence captions of the videos) predict action and social ratings better than image and video models, but perform poorly at predicting neural responses in the lateral stream. Together, these results identify a major gap in AI's ability to match human social vision and provide insights to guide future model development for dynamic, natural contexts." }, "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": [ "NeuroAI", "vision", "fMRI", "deep learning", "social perception" ] }, "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/0b67ecea9bbcc2e43a013b9a067b792ca76a526d.pdf" }, "presentation": null, "primary_area": { "value": "applications to neuroscience & cognitive science" }, "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/6c653e961cc521971f7179ba26e079cd49a52253.zip" }, "title": { "value": "Modeling dynamic social vision highlights gaps between deep learning and humans" }, "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": "1. What does CSR represent? The authors use SCR to represent the Constructed Samples Rate in Section 4.2 but use it to represent the Clean Samples Rate in Section 4.4.\n2. The reviewer is confused about how the second objective in Eq.(4) is satisfied.\n3. How do you obtain the datasets $D_a$ and $D_c$? What if the samples in $D_c$ have similar features to the unlearned data? Please add discussions in the main text." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is the first to identify the privacy threats posed by the exposure of both unlearned data and unlearning intentions during machine unlearning processes. The idea is novel.\n2. The experiments are extensive and show the significant superiority of OUbL over SOTAs in terms of both privacy protection and unlearning effectiveness.\n3. The paper is overall well-structured. The narrative is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the problem of implementing machine unlearning without exposing erased data and unlearning intentions to the server. The authors propose an Oblivious Unlearning by Learning (OUbL) approach to protect both unlearned data and unlearning intentions during machine unlearning processes. OUbL constructs a new dataset with synthesized unlearning noise and achieves unlearning by incremental learning. Comprehensive experiments demonstrate the effectiveness of OUbL." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Issues with the writing.\n - Lack of clarification on notations. E.g., the mapping of the functions $\\mathcal U(\\cdot)$ and $\\mathcal A(\\cdot)$ in the Problem Statement part, the definition of $\\nabla$ and the function $\\ell$ in Section 3.1, and the definition of $\\mathcal I$ in Eq.(3), the description of $I$ on line 225.\n - Some typos, e.g., missing $\\nabla$ on line 221.\n - The $H_\\theta^{-1}$ in Eq.(3) denotes the inverse of the Hessian matrix evaluated at $\\theta$ **on the dataset $D\\backslash D_u$**.\n - The reviewer suggests using another symbol like $\\theta_o$ instead of simply $\\theta$ to represent the original trained model for reducing ambiguity since $\\theta$ seems to be a variable in Eq.(4).\n - The reviewer suggests using the command \"\\citep\" instead of \"\\cite\" when the references are not the objectives in the sentence.\n - Presenting Algorithm 2 with a detailed description in the main text would be better. \n2. Issues with the figures. \n - Both Figure 1 and Figure 2 describe the main process of OUbL, with Figure 2 being more detailed. Therefore, the reviewer thinks that Figure 1 is unnecessary.\n - Incomplete compilation of notation $\\mathcal I_{D_u}$ in Figure 2.\n - The Problem Statement of OUbL on page 3 doesn't include the phase of Figure 2 c), the incremental learning step with clean dataset $D_c$.\n3. Inconsistencies between Eq.(4) and Eq.(8). The second objective in Eq.(4) is evaluated on the dataset $D\\backslash D_u$ (size $N$) and Eq.(8) is evaluated on the dataset $D_a$ (size $P$).\n4. The reviewer thinks that the comparisons between centralized unlearning methods and federated unlearning methods are unfair since the settings are different. The author should discuss the performances of OUbL applied in federated unlearning 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": 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": "Refer to the weakness section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The idea of oblivious unlearning–do not let the server know which records the user wants to delete in the first place is a crucial concern in unlearning, and has not been addressed in previous work.\n2. The proposed solution is a novel application of gradient matching in the field of machine unlearning. Overall the idea makes sense and the solution is very elegant. I do have some questions regarding the practicality of the proposed solution and I will consider raising my rating if all are resolved. Please refer to the weakness section." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for machine unlearning through learning. The idea is to synthesize an auxiliary dataset, whose gradients match the model updates caused by the data to be unlearned. In this way, the updates due to the auxiliary dataset will cancel out the updates due to the data being deleted. In the meantime, the server does not know which records the user wanted to delete in the first place, providing an extra layer of privacy protection (avoiding secondary injury when curing the initial injury)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The setup that the server chooses to continue to learn on the provided auxiliary data is not motivated. E.g., the server may stop training before the user realizes that she/he wants to delete some data from the model, and the unlearning algorithm is never run. Hence, it is not entirely correct that the intention of unlearning can be hidden. In addition, if the server himself is malicious, then he sees the data to be deleted from the very beginning of the training process. Machine unlearning also does not help, since privacy is breached before unlearning happens. If the server is benign, and other end users of the model are the malicious ones, then there is no need for oblivious unlearning (since the server is benign).\n\n2. How to estimate the model change caused by the data being deleted seems to be a difficult problem, without the knowledge of the training algorithm and the training dataset. For the training dataset, the authors assume that they have access to some training data. Is there any constraint on the size and distribution of this small dataset, e.g., how large it is compared to the dataset to be deleted, and the whole training dataset, so that the proposed solution could work? In addition, if the training algorithm is unknown (maybe also the learning rate and optimizer are not known), then how to ensure the overall changes to be applied (caused by the constructed auxiliary dataset) is approximately the same as the change caused by the data to be deleted. E.g., if the learning rate is large and causes the auxiliary data’s gradients to overshoot the original changes to be canceled out, then the privacy of the data to be deleted cannot be preserved.\n\n3. The idea of gradient matching is similar to a previous paper in the field of private ML, Deep Leakage from Gradients by Zhu et al. I am curious what would happen if the user directly applies their method to construct the auxiliary dataset. Would the method be no longer oblivious (the constructed dataset may look weird)?\n\n4. From the server’s perspective, how to distinguish a benign user, who wants to delete their data from a malicious user, who wants to upload noisy gradients to sabotage the model performance?\n\n5. Can you provide some examples (high-resolution preferred) of the original auxiliary dataset and their noisy counterparts?" }, "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": "The model performance drops by approximately 5% compared to other benchmarks when USR is set to 1% in the experiment. This suggests that the method may impair model performance, especially as the USR increases. Any thoughts on addressing this issue?" }, "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 introduces an interesting perspective, highlighting that current unlearning methods often require informing the server about the data to be forgotten, which could pose privacy risks.\n2. The proposed approach is clearly explained and well-motivated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses a limitation in current unlearning methods, which require notifying the server to apply the unlearning algorithm—potentially leading to privacy leakage. To address this, it introduces a new method called OUbL. OUbL constructs a noisy dataset that enables the model to unlearn specific data using its original algorithm, without explicit server intervention. Experimental results indicate that this method performs well in preserving privacy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Existing work addresses privacy issues for the forgetting set using membership inference (MI) attacks as a common metric. Comparing your approach with these methods using MI attacks could strengthen the paper.\n\nReferences: Kurmanji, M., Triantafillou, P., Hayes, J. and Triantafillou, E., 2024. Towards unbounded machine unlearning. Advances in neural information processing systems, 36.\n\n2. The method requires continuous access to the training set. It is unclear how it would function if access to the training set were unavailable.\n\n\nPresentations:\n1. In-text citations lack parentheses. For example, on line 42, it should be written as (Thudi et al., 2022; Hu et al., 2024b) instead of Thudi et al. (2022); Hu et al. (2024b).\n2. Some abbreviations are not clearly explained. For instance, it is difficult to understand what \"USR\" refers to on line 373 until it is defined on line 511." }, "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 line 409-411, how are these privacy epsilon values for OUbL being concluded? It’s not clear to me from the figure. I think it would be best to explain how these privacy metrics are being derived, especially since there is no theoretical guarantees on the privacy parameter.\n\nSuggestions\n\nParenthetical citations should be cited with \\citep{}, seems like all of the citations are done as in text citations using \\citet{}\n\nline 374: sate -> rate\n\nline 157: is it better to use \\approx instead of \\equal? it doesn’t seem that exact equality is being achieved" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The same learning algorithm can be used for unlearning - this is particularly impactful because the current scope of unlearning algorithms is significantly limited." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper formulates the oblivious unlearning by learning (OUbL) problem - users strategically adds noise to an auxiliary dataset such that incrementally training the model based on the auxiliary dataset recreates the effect of unlearning the deleted data. The auxiliary noise is calculated using Hessian based approximation of the unlearning influence and then taking gradient steps which minimize the unlearning objective and the training loss on auxiliary dataset. The performance of OUbL is experimentally compared with federated unlearning approaches and differential privacy approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It seems unrealistic that each user in a dataset would have access to clean samples from the data distribution, could you clarify how a user would receive clean samples in practice?\n\nFurthermore, as more unlearning requests come in, the distribution over the dataset is changing, so these clean samples would be coming from a changing distribution, making it even more difficult to obtain clean samples. How would this distribution shift be handled?\n\nUsers uploading their data to an ML server which already has access to their data during a deletion request does not seem like a major privacy concern. In what specific scenarios will this create a privacy concern?\n\nThe paper lacks theoretical guarantees on the privacy of users after unlearning, for example, in terms of the typical privacy epsilon parameter" }, "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. Can the authors provide a clearer description of the threat model, particularly defining the knowledge and capabilities that a user needs to perform unlearning?\n\n2. In what real-world scenarios can a normal user have white-box knowledge of the model, and how can the proposed approach be applied in such settings? Could the authors discuss practical situations where this would be feasible?\n\n3. How much computational cost is incurred by users when requesting unlearning, and is this burden affordable for typical users? Would it be possible to provide quantitative analysis or benchmarks to showcase the overhead for users?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The concept of using synthesized unlearning noise is innovative and offers a new perspective on achieving oblivious unlearning.\n- The paper addresses a timely and important problem in the field of machine learning, with implications for privacy-preserving model updates.\n- The authors’ efforts in designing the mechanism to generate unlearning noise demonstrate creativity and technical insight." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an Oblivious Unlearning by Learning (OUbL) approach to address the challenge of implementing unlearning without revealing erased data or unlearning intentions to the server. The approach involves constructing a new dataset with synthesized unlearning noise to achieve this goal. While the idea is novel and inspiring, there are several concerns regarding the threat model, practicality, and computational burden on the user that require further clarification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of a clear threat model: The manuscript does not clearly describe the threat model, making it difficult to assess the security and assumptions under which the proposed approach operates. Specifically, there is no detailed explanation of the knowledge and capabilities a user must have to estimate unlearning updates and construct the synthesized auxiliary dataset.\n\n- Practicality concerns: The proposed method appears to require white-box access to the model, particularly for the noise synthesis step through gradient matching. This may limit the real-world applicability of the approach, as most users may not have such access or capability in common unlearning scenarios.\n\n- Computational burden on the user: The approach seems to shift a significant amount of computational cost from the server to the user, particularly in generating unlearning noise. The authors do not provide an analysis of the computational cost for users, leaving unanswered whether it is feasible for normal users to request unlearning in practice." }, "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": "My main questions are with my concerns raised in the weakness part.\n\n1. Can you provide justifications in regard to weakness one?\n\n2, Can you provide an answer to the applicability of this method in weakness two?\n\n3, How does the proposed method differ from existing works?\n\n4, Can the server easily defend against the proposed unlearning method?\n\n5, Can the proposed method work for other data types?" }, "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 is well-written and easy to understand.\n\n2, A new unlearning method is proposed, and it solves the existing method’s limitations of uploading the unlearned data to the server.\n\n3, Extensive experiments on several datasets and model architectures validate the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the machine unlearning problem. The motivation for this paper is the observation that most existing works require the user to upload their original data for unlearning, which might disobey the purpose of machine unlearning. Thus, the authors propose a new machine unlearning method by learning. The model owner updates the target model on synthetic unlearning data, and the user does not need to request for unlearning. Extensive experiments validate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1, Some important justifications are missing from the attack model. First, the idea of the proposed attack is motivated by the claim that users do not want to leak the unlearning intention because of potential unlearning attacks. Then, unlearning is hidden by the server updating the model. In this case, my question is why the server would update the model as the user wishes. I mean, the server decides the update itself. I did not see how this works. Second, even if the server implements the update, what if the server filters out unlearning samples? From the current version, it seems the authors miss justifications on the two points.\n\n2, The proposed method’s applicability is in question. For example, how does the learning rate affect the effectiveness of this method? If the server keeps the learning rate as a secret, can the proposed method work?\n\n3, There are several existing works discussing unlearning without seeing the original unlearning data. For example, the following papers. The authors should have a clear discussion on them.\n\n[R1] Chundawat, Vikram S., et al. \"Zero-shot machine unlearning.\" IEEE Transactions on Information Forensics and Security 18 (2023): 2345-2354.\n\n[R2] Golatkar, Aditya, Alessandro Achille, and Stefano Soatto. \"Eternal sunshine of the spotless net: Selective forgetting in deep networks.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\n\n[R3] Tarun, Ayush K., et al. \"Fast yet effective machine unlearning.\" IEEE Transactions on Neural Networks and Learning Systems (2023).\n\n4, The proposed method does not use the forget data but seems to use retain data. Is this reasonable? Why the server can access and retain data? If the server can access the retained data, then it mean the server kept the original training data? If so, it seems to violate the original assumption.\n\n5, The experiments are conducted on only image datasets. Can the proposed method be used for other data types? For example, text or tabular data." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose an Oblivious Unlearning by Learning (OUbL) method, ensuring unlearning effectiveness, once the server incrementally updates the model (incremental learning) using the users' new dataset (synthesized with unlearning noise)." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024oblivious,\ntitle={Oblivious Unlearning by Learning: Machine Unlearning Without Exposing Erased Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=wAemQcyWqq},\nnote={under review}\n}" }, "abstract": { "value": "Machine unlearning enables users to remove the influence of their data from trained models, thus protecting their privacy. However, it is paradoxical that most unlearning methods require users first to upload their to-be-removed data to machine learning servers and notify the servers of their unlearning intentions to prepare appropriate unlearning methods. Both unlearned data and unlearning intentions are sensitive user information. Exposing this information to the server for unlearning operations conflicts with the privacy protection goal. In this paper, we investigate the challenge of implementing unlearning without exposing erased data and unlearning intentions to the server. We propose an Oblivious Unlearning by Learning (OUbL) approach to address this privacy-preserving machine unlearning problem. In OUbL, the users construct a new dataset with synthesized unlearning noise, ensuring that once the server continually updates the model using the original learning algorithm based on this dataset, it can implement unlearning. The server does not need to perform any tailored unlearning operation and remains unaware that the constructed samples are for unlearning. As a result, the process is oblivious to the server regarding unlearning intentions. Additionally, by transforming the original erased data into unlearning noise and distributing this noise across numerous auxiliary samples, our approach protects the privacy of the unlearned data while effectively implementing unlearning. The effectiveness of the proposed OUbL method is evaluated through extensive experiments on three representative datasets across various model architectures and four mainstream unlearning benchmarks. The results demonstrate the significant superiority of OUbL over the state-of-the-art privacy-preserving unlearning benchmarks in terms of both privacy protection and unlearning effectiveness." }, "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": [ "Machine unlearning", "Privacy Preserving" ] }, "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/ab380e5ce68e57faeb3135879e8f71ac1f42ad1b.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": "Oblivious Unlearning by Learning: Machine Unlearning Without Exposing Erased Data" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]