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": 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": "Some questions and details that came to my mind while making the revision of the manuscript are:\n\nQ1: Some comments are added in the review of SOTA methods and background about pre-processing and how problematic this is. What do pre-processing steps mean for the authors? and which method is it referring to?\n\nQ2: Paragraph in L212: What are the real implications of the use of compact Matern kernel for RBFs. Are we limited on the type of covariance functions used, right?\n\nQ3: The nugget for numerical stability… but what is actually the dimension/order of this one?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "It is always nice to see contributions to the scalability of GPs and particularly GPR. Despite the fact that I do not share some of the thoughts/ideas around the comments on the SOTA methods, I think the paper has a valuable point in certain directions. The Mercer's Theorem is well-known in the GP community and has been considered since the beginning of the interaction between GPs and ML. Perhaps, the main strength is the combination of the HS-SVD together with the pseudo-inverse taken from Pozrikidis (2014) and later the Sylvester determinant theorem. Even if there are some details or deeper analyses missing in such directions, I think that putting these methods together is a powerful message to look back to in these times when scalability matters a lot." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method for Gaussian process regression (GPR) that reduces the classical problem of cubic and quadratic complexity in GPs. In particular, the work exploits the well-known Mercer decomposition of kernels, such that the covariance function can be expressed as an infinite sum of eigenfunctions. Under this condition, it is possible to use a special type of SVD to obtain an usable decomposition of the covariance matrix. Under the assumption of compact Matern kernels and the truncation of eigenvalues, the performance is compared on 4 different simulations with synthetic data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My gut feeling with the submission is that it has yet limitation and details that are not clear enough to validate the quality of the contribution. Some thoughts:\n\n[W1] - Mercer's theorem is a well-known method, as well as the inverse approximation Sherman–Morrison–Woodbury formula which is basically a special case of the Woodbury matrix identity. The use of Sylvester's determinant is also common in the community, as far as I remember. In this regard, the HD-SVD is the only method that is kind of new to me here, but I am not yet sure if the combination of these together makes a super strong contribution.\n\n[W2] - The criticism + comments around the whole literature of GPs are somehow vague, in the sense that so much detail and effort has been put on them, and are kind of discarded due to high-level opinions and not much precision on their issues and limitations. In this sense, GPR and its scalability is super-explored since 20y ago.. Then, what concerns me is that near-zero analysis is derived around the quality of the inverse for example, the effect of assuming the compact Matern kernel or the first paragraph in section 3.2 around the noise term and how reasonable is that. To me, this last detail is the worst one so far of the paper, in the sense that it is not super scientific (what is the order of the noise term, everyone can add to any ML model some noise to the data and observe nicer properties and it is easily modeled by an isotropic Gaussian right?)\n\n[W3] - I do not buy the technique of truncating the eigenvalues in the HS-SVD and the justification that it stabilizes everything. How many of them are truncated? Little details on this point are added in my opinion.\n\n[W4] - The fact that only synthetic-data simulations are used in the submission tends to be a bad sign in the ML/GP community, even more if the main motivation is scalability and large-scale 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": 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": "Along with little experimental support, the paper also doesn't provide any theoretical support in terms of bounding the mean squared error for using the HS-SVD of the compact Matern instead of the full kernel. Can anything be said about the value of the approximate MSE due to using HS-SVD with respect to the true MSE if the full kernel were used?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The generalization of the 1D compact Matern kernel to higher dimensions is interesting. It will be useful to fully investigate and understand the properties of this kernel. In the experiments provided, the use of this kernel shows considerable promise. However, there are quite a few weaknesses of paper as pointed out below." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new technique for Gaussian process regression based on the Hilbert-Schmidt singular value decomposition (HS-SVD) of the compact Matern kernel. The compact Matern kernel proposed in the paper is a generalization of the 1-D compact Matern proposed in Cavoretto et al., 2015 to higher dimensions. The eigenvalues and eignevectors of the compact Matern kernel can be computed easily via simple closed-form expressions and thus, it is possible to compute the Mercer decomposition of this kernel easily, unlike commonly used kernels like Radial Basis Function (RBF), Matern, and exponential kernels. Thus, one can find the HS-SVD of the compact Matern kernel by keeping only the first m eigenvalues and corresponding eigenvectors. Instead of computing and storing the complete $n \\times n$ Kernel matrix, the storage cost is reduced to $O(nm)$. Operations like computing the inverse and log determinants of the kernel also become cheaper and the time complexity is reduced to $O(nm^2)$ from $O(n^3)$. Finally, some experiments using synthetic datasets are performed to show the effectiveness of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Insufficient experiments/evidence: The main value proposition of the paper seems to be the introduction of the generalized compact Matern kernel whose HS-SVD can be computed effectively. Some experiments are performed for nonlinear function approximation but much more evidence is required to accurately judge how well this method is truly outperforming. There is also no theoretical support showing how well the proposed approach is performing (see questions).\n\nApart from the points mentioned above, the paper doesn't clearly mention how to choose the parameter m (the dimension of the SVD) which is crucial to the method. In general, is there an empirical rule of thumb or some known theoretical bound that could guide the choice of m? The same point holds for the other parameters of the compact Matern kernel $\\rho, \\alpha, \\beta$.\n\n2) Missing related work: There have been works that use SKI and take time almost linear in m where m is the number of “inducing points” and also doesn't scale exponentially with the input dimension r. See for example https://arxiv.org/pdf/2305.14451. Specifically, in the caption under Table 3, the authors note that methods like SKI cannot be sued since the the number of grid points grows exponentially with r. However, it seems that the above paper tackles this very problem.\n\n3) The messaging of the paper seems to be confusing in some places. Most of the ideas presented in the paper like using an approximate Mercer decomposition of kernels have been widely known and studied. The main contribution of the paper seems to be the construction of the generalized compact Matern kernel whose HS-SVD can be computed effectively. But, for example, in the discussion section 5, the authors claim that they introduced the HS-SVD method and constructed the compact Matern kernel as an illustrative example. However, as pointed out above, HS-SVD (which basically follows from the Mercer decomposition) has been pretty well known. So, to claim that they introduced the HS-SVD method itself is a bit misleading I feel. Especially since they don't provide any algorithm for obtaining the HS-SVD of any other kernels.\n\nOverall, I feel that this paper needs to be improved before it can be considered for publication at ICLR. Hence, I'm recommending a reject." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In your approximation of page 4, you simply do truncation of the representation. Do you have a theory for the approximation? Or small m can make the approximation not trustworthy.\n2. You spend a great extent discussing the smoothness of the compact Matern kernel, is the result new?\n3. Could you add some real-world datasets in the experiment parts?\n4. Have you compared your method with the method in 'Sample and Computationally Efficient Stochastic Kriging in High Dimensions'? As far as I know, this work also states that they achieve the SOTA. You can include a comparison with this method in experiments, or explain why such a comparison was not included." }, "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 novel, using HS-SVD to help reduce computational cost." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the computational issue in the GP area, which is widely known as a bottleneck in large-scale GP applications. THe paper proposes a method based on the Hilbert-Schmidt singular value decomposition that obtains a low-rank decomposition“for free”, reducing both time complexity and space complexity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Writing: many sections use \"fast\" as the title. It may lead to confusion about what the \"fast\" really means. You should add something related to complexity in it to make it clearer.\n2. Experiment: I can not see the experiment on the real dataset. Can you explain why you do not do experiments on the real dataset?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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 addition to the weaknesses, the following questions need to be addressed:\n\n**Q1.** In Definition 2.4, the authors state that they use a truncated form of the HS-SVD for computation. However, reference [1] suggests that truncating the series may result in deviations from the standard kernel space span. How do the authors determine the truncation order, and what measures are taken to ensure that the truncated form remains within the kernel space span?\n\n**Q2.** Following Q1, the authors also mentioned that \"the truncation strikes a balance between computational efficiency and the accuracy of the low-rank approximation\". Can the authors provide a quantification or analysis of this trade-off? The authors may refer to reference [2] for a similar analysis.\n\n**Q3.** In Section 3.1, the authors introduce the compact Matérn kernel. How does this kernel differ from the iterated Brownian bridge kernel discussed in reference [1]? In addition, [1] also mentioned that it is possible to extend their kernel to high-dimensional case by using tensor products, how does this compare to the proposed kernel in Definition 3.2?\n\n**Q4.** In Section 3.2, the authors claimed that their proposed method can reduce the computational complexity of the matrix inversion from $O(n^3)$ to $O(m^3)$. However, this appears similar to existing low-rank approximation methods, such as the Nyström approximation, which also achieves the same complexity. Can the authors clarify how their approach is advantageous compared to these established methods? It is suggested to add a table comparing the proposed method with existing methods, in terms of the time and space complexities.\n\n**Q5.** In Section 3.3, the authors claimed that truncating the small eigenvalues improves stability by removing ill-conditioned parts of the kernel matrix. Can the authors provide theoretical justification for this claim? The reviewer also did not find any experiments supporting this claim.\n\n**Q6.** The authors use runtime (in seconds) and RAM/VRAM usage as metrics for performance evaluation; however, these are hardware-dependent metrics. Could the authors also provide hardware-independent metrics such as FLOPS and arithmetic intensity?\n\n**Q7.** In Section 4, the authors claimed that \"by integrating the strength of dimension reduction and variational inference methods with HS-SVD’s low-rank decomposition, we could achieve the best of both worlds—enhancing stability while maintaining fast runtime and low memory costs.” Can the authors perform an experiment/ablation study to support this claim?\n\n### References\n[1] Cavoretto, R., Fasshauer, G.E. and McCourt, M., 2015. An introduction to the Hilbert-Schmidt SVD using iterated Brownian bridge kernels. *Numerical Algorithms*, 68, pp.393-422.\n\n[2] Griebel, M., Rieger, C. and Zwicknagl, B., 2015. Multiscale approximation and reproducing kernel Hilbert space methods. *SIAM Journal on Numerical Analysis*, 53(2), pp.852-873." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The main strengths of the paper are as follows:\n\n**S1.** Proposes a method for obtaining low-rank approximations of kernel matrices based on HS-SVD, that requires minimal tuning or preprocessing and does not depend on GPU resources.\n\n**S2.** Discusses the numerical advantages of the proposed method, particularly in reducing both computational and memory cost.\n\n**S3.** Empirically demonstrates that the proposed method is superior to the baselines on simulated large-scale datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the scalability issue of Gaussian process regression (GPR), which suffers from cubic time complexity and quadratic space complexity when dealing with large datasets. The authors propose a scalable framework based on the Hilbert-Schmidt singular value decomposition (HS-SVD) to achieve a more efficient low-rank approximation of the kernel matrix, thereby reducing the time complexity to $O(nm^2)$ and space complexity to $O(nm)$. Empirical results demonstrate that the proposed framework outperforms existing methods in terms of runtime and memory usage on simulated large-scale datasets, while requiring minimal preprocessing and being accessible without GPU resources." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are several weaknesses of the paper:\n\n**W1.** The novelty of the proposed method is questionable, as the HS-SVD has already been introduced and thoroughly studied in [1]. Many definitions, lemmas, and theorems in this work are derived from [1]. For instance, Definitions 2.4 and 3.4 are adapted from [1] without proper citations.\n\n**Suggested action:** The authors should clarify their contributions in relation to [1], and ensure proper attribution for all borrowed sections, including definitions and theorems.\n\n**W2.** The paper lacks adequate theoretical analysis and justification regarding several key aspects:\n- There is no clarification on how the authors determine the truncation order (see Q1), nor is there any quantification or analysis of the corresponding trade-off (see Q2). Specifically, it is unclear how the authors chose the truncation order and whether they have tested other values to obtain the results presented in Tables 1-4.\n- There is no justification for how truncation enhances stability by eliminating ill-conditioned parts of the kernel matrix (see Q6).\n- The authors do not explain how the proposed method can be integrated with existing dimension reduction and variational inference methods, nor how this integration could result in faster runtime and lower memory costs (see Q8).\n\n**Suggested action:** The authors should consider conducting an ablation study to analyze the effect of different truncation orders, and/or provide a theoretical bound on the approximation error.\n\n**W3.** Most of the baseline methods used in experiments are relatively outdated, with many published within 5-10 years ago. The authors should consider including more recent state-of-the-art methods to evaluate their method. Some relevant works are as follows:\n- Wu, K., Wenger, J., Jones, H.T., Pleiss, G. and Gardner, J., 2024, April. Large-scale Gaussian processes via alternating projection. In *International Conference on Artificial Intelligence and Statistics* (pp. 2620-2628). PMLR.\n- Allison, R., Stephenson, A. and Pyzer-Knapp, E.O., 2024. Leveraging locality and robustness to achieve massively scalable Gaussian process regression. *Advances in Neural Information Processing Systems*, 36.\n- Li, K., Balakirsky, M. and Mak, S., 2024, April. Trigonometric Quadrature Fourier Features for Scalable Gaussian Process Regression. In *International Conference on Artificial Intelligence and Statistics* (pp. 3484-3492). PMLR.\n- Noack, M.M., Krishnan, H., Risser, M.D. and Reyes, K.G., 2023. Exact Gaussian processes for massive datasets via non-stationary sparsity-discovering kernels. *Scientific reports*, 13(1), p.3155.\n\n**Suggested action:** The authors should consider incorporating more recent works as baselines (e.g., the ones listed above) and discuss how their method compares, both theoretically and empirically, to the updated baselines.\n\n**W4.** The experimental validation is insufficient. While the proposed method shows superior results on simulated datasets, it lacks evaluation on real-world datasets. Additionally, most simulations only use $n = 10,000$ samples, which is not convincing enough to support the claim that their method is suitable for large-scale settings. It is recommended that the authors include experiments on larger real-world datasets to strengthen their results. Please refer to the experiments in the references mentioned against W3.\n\n**Suggested action:** The authors should conduct additional experiments using real-world datasets, such as those from the [UCI Machine Learning Repository](https://archive.ics.uci.edu), which contain millions to tens of millions of samples, to better evaluate the scalability of their 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": 3 }, "primary_area": null, "questions": { "value": "Suggestions: \n- Describe exact approaches to scale GPs in the literature review.\n- Explain better why your approach is novel, given that applying decomposition of the kernel is not.\n- The Simulations should not focus on a particular synthetic function in low dimensions. It should be the MLE, RMSE, CRPS (...) of the approximation for a variety of machine-learning-relevant datasets. One of them could be a synthetic function. In the simulations, you might also compare to the Vecchia approximation (which should be presented in the intro). The Vecchia approximation is the state-of-the-art in statistics. \n\nQuestions:\n- How would the method change for non-stationary kernel designs? It is my assessment that recent applied studies increasingly try to take advantage of non-stationarity. \n\n- Are you confident about the correctness of your MSE scores? They seem large for the given function." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The manuscript is clearly written. \n- While not novel in itself (in my opinion), the content might very well be of interest to readers. \n- Scalable GPs are a hot topic.\n- The method is easy to implement, so it has a good chance to be applied." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript proposes a methodology to scale up Gaussian processes, focusing on Gaussian process regression. The core idea is to decompose the covariance matrix via Hilbert-Schmidt singular value decomposition that, in principle, can obtain a low-rank representation for free. The method is explained and a pseudocode is offered. The method is then tested in 5 different simulations. \n\nOverall, I enjoyed reading this paper. It is well-written and easy to read. However, there are some significant shortcomings regarding novelty, the literature survey, and the test results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(A) Major:\n(1) Novelty: The idea of singular value decomposition for GPs is not new and was even described in the GP cookbook (Williams and Rasmussen). The principle has been applied in various papers, including [Sivaram Ambikasaran, Daniel Foreman-Mackey, 2015] and [Drineas, P. and Mahoney, M. W. (2005)]. I fail to recognize the novelty of the proposed approach, since, in general, decomposing the kernel into eigenvalues and eigenfunctions is well-known (as is even stated in the manuscript). Please clarify what part of the \"Method\" section is in fact novel. One piece of potential novelty might come from the kernel which I address in the next comment. \n\n(2) Methods for scaling Gaussian processes have recently moved away from the limitations to a particularly (often stationary) kernel design. The community has come to the revelation that flexible kernel designs are what makes GPs powerful function approximation tools. Dictating one or the other kernel design is therefore counter-productive. Please describe how the method could be extended to non-stationary kernels. \n\n(3) Simulation Experiments. The simulations all use one particular function (in 1d or 2d) with points on a grid or randomly distributed. At the very least the method has to be tested on some real higher-dimensional datasets. There are so many to choose from and really any will do, but one particular synthetic function is not sufficient to demonstrate performance. Also, I am concerned about the MSE error scores (MSE is not optimal but more on that later). The MSE seems very high for all methods. The function is approximately bounded by [-1,1] in 1d and [-.2,.2] in 2d; an MSE of 0.1 seems high, especially for large datasets. I might have missed something here but this was a concern. Next, assuming gridded data is pretty unrealistic and those tests don't have too much value. My suggestions for improvements are (a) new test datasets from various fields (topography, weather data, Housing data (available through scikit-learn)), robot-datasets, and so on, (b) add the CRPS and the Negative log predictive density scores to the results. \n\n(4) Literature overview: The introduction focuses on prior work in approximate GPs. Recent work has resulted in methods to scale up exact GPs. \nExact Gaussian Processes on a Million Data Points\nKe Alexander Wang, Geoff Pleiss, Jacob R. Gardner, Stephen Tyree, Kilian Q. Weinberger, Andrew Gordon Wilson\n\nExact Gaussian processes for massive datasets via non-stationary sparsity-discovering kernels\nMM Noack, H Krishnan, MD Risser, KG Reyes \n\nIn addition, it seems that the Vecchia approximation is neither discussed nor compared to. \nPlease add a short discussion to the manuscript and a reason why those methodologies were not compared. \n\n(B) Minor:\n(1) The MSE alone is not a good score to judge the performance of a GP. It should be augmented by the CRPS. \n(2) Deep Kernel Learning is mixed in the intro with methods for scalability. That seems to be coming a little bit from left field, since the main purpose DKL is a flexible way to achieve non-stationarity, not scalability. \n(3) The comments on ill-conditioning surprised me. With a realistic noise model and a PSD kernel, ill-conditioning should never be an issue." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024scalable,\ntitle={Scalable Gaussian Process via Hilbert-Schmidt Singular Value Decomposition},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xUHL8mtSUL},\nnote={under review}\n}" }, "abstract": { "value": "Gaussian process regression is widely used for its flexible mean predictions and inherent uncertainty quantification. However, its scalability is limited by cubic time complexity, $O(n^3)$, and quadratic space complexity, $O(n^2)$, making it infeasible for large-scale datasets. Although recent advances have introduced approximate methods with time complexity $O(nm^2)$, where $m\\ll n$ is a tuning parameter, these methods each have their own bottlenecks, such as requiring a relatively large $m$ or involving expensive preprocessing steps. Moreover, for extremely large datasets with millions of samples, the space complexity $O(n^2)$ becomes another significant bottleneck. In this paper, we present a novel method based on the Hilbert-Schmidt singular value decomposition that obtains a low-rank decomposition ``for free\", reducing both time complexity to $O(nm^2)$ and space complexity to $O(nm)$, with no preprocessing overhead. We used simulated large-scale datasets to demonstrate the performance of our method compared to state-of-the-art approaches." }, "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": [ "Scalability", "Gaussian process regression", "Hilbert Schmidt singular value decomposition", "compact Mat\\'ern" ] }, "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/da401bb55ca4ae8c76df31c5c3cdda3541cd6e7d.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": "Scalable Gaussian Process via Hilbert-Schmidt Singular Value Decomposition" }, "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": "Are there anything you can prove about the node-level structural entropy values?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Overall, the proposed algorithm is intriguing due to its general empirical performance and some of the novel ideas it introduces. The concept of node-level structural entropy, which quantifies the contribution of an individual node to the global structural entropy of a graph, could be of independent interest. The experiments indicate that the proposed method outperforms existing selection methods in many common learning tasks. Furthermore, ablation studies validate the contribution of each module (node-level structural entropy and importance-biased blue noise sampling) to the overall effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies sample selection which aims to extract a small, representative subset from a larger dataset. The authors introduce a novel sample selection scheme, termed Structural-Entropy-based Sample Selection (SES), which uses and extends the concept of \"structural entropy\" (Li and Pan, 2016), which assesses how nodes and edges within a graph are hierarchically organized to form multi-level communities. The proposed scheme seeks to address the limitations of existing selection methods, which often prioritize local information and neglect the broader, global context of the samples.\n\nThe algorithm begins by constructing a k-NN graph G for the dataset based on similarity. It then calculates the structural entropy value for each node in G (referred to as node-level structural entropy). While structural entropy was originally defined for an entire graph rather than individual nodes, the authors extend this concept using the Shapley value method. Roughly speaking, node-level structural entropy calculates the average increase in structural entropy when a node is added to all potential subgraphs of G. After that, each node is assigned an importance value, which is the product of its structural entropy and training difficulty. An importance-biased blue noise sampling method is then used to select samples. instead of sampling solely based on importance scores, this sampling process prevents the selection of overly similar samples, thus maintaining diversity within the selected subset.\n\nThe effectiveness of the SES scheme is demonstrated through experimental studies and compared to other selection methods in various tasks such as supervised learning, active learning, and continual learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I believe the results could benefit from stronger theoretical justification. While the high-level ideas are reasonable, the mathematical properties are not mentioned in this paper. A deeper discussion on these aspects should be provided. Especially I am interested in what are the mathematical properties of the node-level structural entropy." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 construction method of the Encoding tree is challenging. The author mentioned the Huffman tree construction; how does this affect the conclusions of this article?\n\n2, Equation (5) in the paper calculates the overall importance score of the node. Is this importance score node-wise or point-wise sampling? A data point may appear in multiple nodes, does this affect the proposed sampling method?\n\n3, Have you considered other ways of combining global structural entropy and local training difficulty indicators, rather than through the multiplication method (S(u) = Se(u) * St(u))?\n\n4, Can you provide more theoretical explanations or proofs to demonstrate why structural entropy, as a global indicator, is helpful for considered tasks (e.g. supervised, active, and continual learning)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1, The method utilizing the encoding tree effectively measures local structural information across different scales, providing valuable insights. Various experiments on different learning tasks demonstrate SES’s effectiveness.\n\n2, Structural entropy is decomposed from the overall graph level to individual nodes, resulting in a node-level definition of structural entropy. By leveraging properties of the Shapley value, authors show that the Shapley value of a node can be computed in linear time relative to the number of edges." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a Structural-Entropy-based sampling (SES) method for data selection. This approach integrates global structural information, quantified using structural entropy, and local information (training difficulty) to choose informative and representative samples for training. The authors show that incorporating global structural information can improve the quality of sample selection. Traditional methods often focus on local data properties, such as training difficulty, while ignoring broader connectivity patterns. SES constructs a sample graph using k-nearest neighbor (kNN) relationships to model sample similarity and applies structural entropy at a node level to evaluate each sample’s significance in preserving the overall data structure." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1, The overall idea is relatively simple and incremental, mainly based on the concept from Li and Pan 2016. Also, the proposed method is very heuristic without solid theoretical validation. For example, in line 212-213, \"Given our emphasis on ..., we only use....\". Can you give more detailed explanation? In my opinion, this is not a serious claim for a research article. In line 352, you say \"10X speedup\" because you use only 10\\% of dataset. Can you provide the realistic experimental time for this claim? Do you count the construction time of your sample? \n\n2, While the Shapley value of a node can be computed in linear time relative to the number of edges, the edges in a kNN graph is O(k|X|). It is still time consuming to calculate the Shapley values for all nodes of the encoding tree, especially for large dataset.\n\n3, The author does not clearly explain why preserving the global structure of the dataset is beneficial or provide theoretical guarantees regarding its impact on performance in tasks such as supervised, active, and continual learning. Similarly, the rationale for prioritizing samples with high local information (i.e., training difficulty) is insufficiently justified.\n\n4, The blue noise sampling method effectively promotes diversity, yet the balance between selecting challenging samples and maintaining sample diversity could be further clarified. A comparative analysis with methods that explicitly optimize for diversity, such as clustering-based selections, would help to clarify the advantages of SES.\n\n5, The method relies on the quality of the embedding used to construct the kNN graph. If the embedding representation is poor, structural entropy may not accurately capture the global structure of the data.\n\n6, Some experimental parts are not sufficient. For example, in the continual learning part, they only consider three baselines and two memory sizes (100, 200). Moreover, do you consider both Class-Incremental Learning and Task-Free Continual Learning?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.The writing of this paper is good and easy to follow.\n\n2.This paper has a certain degree of innovation, introducing a structural entropy as a global metric for sample selection." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel sample selection method that can capture both local and global information of data simultaneously. Specially, first, they use a $k$-NN to construct the sample graph of original data. Second, they employ structural entropy to measure global information and use training difficulty to capture the local information of the graph. At lats, they utilize the importance-biased blue noise sampling to select a set of diverse and representative samples. \n\nThe main contribution of this paper is to propose a new global metric for sample selection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Hypergraphs [1] can capture more complex relationships, and I believe that if the authors had used hypergraphs instead, the results would have been even better. I would be delighted to see the authors add new experiments to verify my hypothesis.\n\n2. How would the authors define informative and representative samples? They mention these concepts multiple times, but do not provide a detailed explanation.\n\n3. Taking sample as a node is not very reasonable, as it treats the sample as a whole and ignores the local information contained in the sample itself. Perhaps using a Hyper-class representation [2] would yield better results. I would be delighted to see the authors conduct new experiments to verify my hypothesis.\n\n4. The experimental results do not show the variance, so I hope the authors can make up the variance.\n\n5. Just a little confused, why most of the comparison algorithm's experimental results are not as effective as random selection? So what's the meaning of those comparison algorithms? Or are those comparison algorithm chosen by the author appropriate? Are the parameters of the those algorithms not optimally tuned?\n\n6. The effect of blue noise sampling (BNS) is not as good as message passing (MP) when the sampling rate is greater than or equal to 20%, why not directly use MP? Simply because BNS does not require hyperparameter tuning, it sounds far-fetched.\n\n7. As can be seen from Figure 4, (b) has better boundary division, while (c) has several categories of samples grouped together. It seems that (b) is better. How can we understand this phenomenon?\n\n[1] Learning with hypergraphs: Clustering, classification, and embedding.\n\n[2] Hyper-class representation of 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": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The idea is interesting and reasonable.\n2. The experiments are sufficient and the experimental results are good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a sample selection method with structural entropy. It first introduces the structural entropy of a graph into the sample selection task, and then applies the Shapley value to decompose the structural entropy of a graph to the node-level structural entropy. At last, it designs a score for sample selection with the node-level structural entropy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main techniques used in this paper are all existing methods, e.g. the structural entropy and Shapley value. The paper seems an application of these techniques in the sample selection task. The paper should clarify its novelty and technical contributions.\n\n2. The paper derives Eq.(3) from the Shapley value, and then removes the second term, leading to Eq.(4), which is used in the method. However, the motivation is unclear. Why should we remove the second term? What is the advantage of removing this term? The ablation study of this term should be conducted in the experiments. The paper only claims that they want to design a global metric and the second term is a local metric. I do not think it's convincing enough. What is the advantage of only considering global information? Moreover, in Eq.(5), the method combines the global score S_e and the local score S_t. It seems a little contradictory. In Eq.(4), they do not want the local term but in Eq.(5) they need the local term. It seems strange and not well-motivated.\n\n3. Eq.(5) needs more justification. Why should we multiply S_e and S_t? What is the advantage? Why not just sum S_e and S_t, or combine them with other forms? I think more ablation study is needed to justify the effectiveness of Eq.(5).\n\n4. In the experiments, the experimental setup should be introduced in more detail. For example, what is the difference between the supervised setting and the active learning setting? They both need to select some samples to train the model and then is used to predict the test data." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024structuralentropybased,\ntitle={Structural-Entropy-Based Sample Selection for Efficient and Effective Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xUMI52rrW7},\nnote={under review}\n}" }, "abstract": { "value": "Sample selection improves the efficiency and effectiveness of machine learning models by providing informative and representative samples. Typically, samples can be modeled as a sample graph, where nodes are samples and edges represent their similarities. Most existing methods are based on local information, such as the training difficulty of samples, thereby overlooking global information, such as connectivity patterns. This oversight can result in suboptimal selection because global information is crucial for ensuring that the selected samples well represent the structural properties of the graph. To address this issue, we employ structural entropy to quantify global information and losslessly decompose it from the whole graph to individual nodes using the Shapley value. Based on the decomposition, we present $\\textbf{S}$tructural-$\\textbf{E}$ntropy-based sample $\\textbf{S}$election ($\\textbf{SES}$), a method that integrates both global and local information to select informative and representative samples. SES begins by constructing a $k$NN-graph among samples based on their similarities. It then measures sample importance by combining structural entropy (global metric) with training difficulty (local metric). Finally, SES applies importance-biased blue noise sampling to select a set of diverse and representative samples. Comprehensive experiments on three learning scenarios --- supervised learning, active learning, and continual learning --- clearly demonstrate the effectiveness of our method." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Sample selection", "graph", "structural entropy", "blue noise sampling" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/0ecccdff385bd6e7217224ca1dd20f835c9d5c3f.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": "Structural-Entropy-Based Sample Selection for Efficient and Effective Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "See above" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The strategy is rather simple and can easily be applied to diffusion models and flow-based models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes the Curriculum Consistency Model (CCM), which stabilizes and balances the learning complexity across timesteps. It defines the distillation process as a curriculum and introduces Peak Signal-to-Noise Ratio (PSNR) as a metric to quantify the difficulty of each step in this curriculum. Extensive empirical studies are performed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper writing is disastrous. There are too many expression and grammar errors. In particular, the paper tile emphasizes consistency models but the flow-based models are highlighted in the main context. \n\n- The multi-step iterative generation method has been explored by the literature, see [1]. The GAN trick is not motivated. \n\n- The paper should compare to more recent consistency models like improved CD, multi-step CM, etc. The arguments on the issues of CMs in this paper are partially problematic. \n\n[1] SCott: Accelerating Diffusion Models with Stochastic Consistency Distillation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 CIFAR-10 experiment, CCM is only applied to FM models. Was there a reason for not applying it to diffusion-based models?\n- Prior works have reported that training of CM is difficult in terms of both stability and time complexity. Can you report those aspect of CCM as well?\n- Have you tried *fine-tuning* a pretrained DM or FM with the proposed training procedure? According to ECT it reduces the required training time and increases stability of training. \n- Does the proposed value of $T_\\text{SNR}$ also generally work well with other models / datasets? If not, \"our method is not very sensitive to PSNR\" might not be a valid claim." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The paper presents concrete motivation for each component of the proposed method. \n- The proposed method shows some promising empirical results.\n- The paper conducted various ablation studies to justify the importance of each proposed component." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel training procedure for consistency models, introducing: 1) a dynamically adjusted training objective based on learning difficulty, measured by PSNR, and 2) a distillation target acquired through multi-step generation. The resulting model, named the Curriculum Consistency Model (CCM), demonstrates performance that is comparable to, if not better than, existing baseline models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The dynamical adjustment, which is one of the major contribution of this paper, resembles **continuous-time training schedule** of ECT [1]. \n- Multi-step iterative generations will increase the time complexity of training. \n- Performance on CIFAR-10 is very promising, however, on large scale datasets such as ImageNet 64x64 or CoCo2017, the gain is either marginal or none. If the proposed method do increase time complexity of training, this result is not promising enough.\n\n\n(Minor remarks)\n- Personally, I believe Chapter 2 is not so kind for readers especially if they are unfamiliar with flow models.\n- $G$ was not defined on line 151.\n- The (0, 1) or (0, T) convention of FM and diffusion models seem to be opposite of one another, making it hard to read.\n- Assigning $x_\\text{target}$ is missing in Algorithm 1.\n\n\n[1] Geng et al. \"Consistency Models Made Easy\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "**Q0 |** Please address the concerns and questions in Weaknesses.\n\n**Q1 |** PSNR is computed between the student and teacher outputs. If I understand correctly, and the teacher output corresponds to the $u$->1 mapping, then PSNR and the CM loss are quite similar. Why not simply use the CM loss as is?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**S1 |** The paper addresses an important problem of improving the training of consistency models.\n\n**S2 |** The proposed learning strategy is reasonable and well-motivated. \n\n**S3 |** Compared with standard CD, CCM accelerates convergence and improves overall performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel training method for CMs that adaptively selects consequentive timesteps to ensure that the difficulty of the learning targets is maintained throughout the training process across different samples. The method is validated on CIFAR10, ImageNet64 and T2I generation, demonstrating strong performance when combined with adversatial training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**W1 |** Recent works [1,2] have also addressed the training complexity and instabilities in CMs and proposed various related training techniques to improve convergence. However, these works are neither discussed nor compared with.\n\nI believe the presence of [1,2] largely limits the main contribution and that CCM needs to be carefully compared against the techniques from these works, omitting the effect of the GAN loss.\n\n**W2 |** CCM has been applied only to FM models on CIFAR10 and ImageNet64. It makes comparisons with the original CD, CTM, iCT[1], and ECM[2] difficult. Could the authors evaluate CCM on the corresponding diffusion models with and without GAN loss?\n\n**W3 |** For T2I generation, the improvements seem negligible. FID, especially at 5K, is an unreliable metric, as noted in [3, 4]. Therefore, I believe the demonstrated FID gains may be unrepresentative or marginal. Could the authors conduct a human preference study or consider adding FID30K and alternative metrics, e.g., CMMD[3], ImageReward[5], and PickScore[6]? \n\n**W4 |** Inconsistent terminology and notation regarding the student, teacher and target models/outputs. Figure 1 denotes a teacher model that maps $u$ to 1, yet in Section 2, the teacher maps $t$ to $u$.\n\nL210-211: $x_{est}$ - teacher output, $x_{target}$ - student output. L252: $x_{target}$ - teacher output. Algorithm 1: $x_{est}$ is a student output and $x_{target}$ is missing.\n\n**W5 |** The related work section can be largely extended. I recommend citing and discussing CM-related works [1,2,7], as well as other distillation methods [8,9,10,11,12,13]. A comparison with these works would be highly beneficial.\n\n**W6 |** The experiment in Figure 7 needs more details.\n\n---\n[1] Song et al. Improved Techniques for Training Consistency Models, 2023\n\n[2] Geng et al. ECT: Consistency Models Made Easy, 2024\n\n[3] Jayasumana et al. Rethinking FID: Towards a Better Evaluation Metric for Image Generation, 2024\n\n[4] Podell et al. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis, 2023\n\n[5] Kirstain et al. Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation, 2023\n\n[6] Xu et al. ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation, 2023\n\n[7] Salimans et al., Multistep Distillation of Diffusion Models via Moment Matching, 2024\n\n[8] Berthelot et al. TRACT: Denoising Diffusion Models with Transitive Closure Time-Distillation, 2023\n\n[9] Luo et al. Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models, 2023\n\n[10] Yin et al. One-step Diffusion with Distribution Matching Distillation 2023\n\n[11] Yin et al. Improved Distribution Matching Distillation for Fast Image Synthesis, 2024\n\n[12] Zhou et al. Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation, 2024\n\n[13] Kim et al. PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher, 2024" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Given the potential for other similarity metrics, such as SSIM or LPIPS, to assess similarity in pixel space and thus learning complexity, have the authors explored these options?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1 - The method is simple yet effective, demonstrating significant performance improvements over the base model in CIFAR and ImageNet settings.\n\nS2 - The experiments are thorough, covering both Diffusion and Flow models and spanning unconditional, class-conditional, and text-conditional generation tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a straightforward and effective approach for optimizing the training time grid for discrete consistency models, resulting in competitive performance across various image generation benchmarks. By defining a notion of \"learning complexity\" using PSNR, the work provides insights on the training dynamics of consistency models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1 - The concept of \"learning complexity\" lacks theoretical justification. In Sec 3, the authors propose that high complexity leads to confusion, while low complexity reduces learning efficiency, but these observations appear mostly empirical. Including more theoretically motivated explanations, as seen in [1] regarding sampling time grid optimization, especially their discuss on the relation between time grid and the KL upper bound, would add depth to the discussion.\n\nW2 - The authors should cite [2] and discuss its model, as it is closely related and also focuses on improving training efficiency in consistency models. Additionally, experiment results with NFE = 2 should be included for a more comprehensive comparison, as it's almost standard practice in most consistency model papers.\n\nW3 - The notation $G_\\theta$ is not clearly defined and is inconsistently used between Section 2.1 and Algorithm 1. In Section 2.1 the last variable in $G_\\theta$ seems to be the target time step, while in Algorithm 1. it is replaced with condition $c$. \n\n[1] Sabour, Amirmojtaba, Sanja Fidler, and Karsten Kreis. \"Align your steps: Optimizing sampling schedules in diffusion models.\" arXiv preprint arXiv:2404.14507 (2024).\n[2] Geng, Zhengyang, Ashwini Pokle, William Luo, Justin Lin, and J. Zico Kolter. \"Consistency Models Made Easy.\" arXiv preprint arXiv:2406.14548 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024see,\ntitle={See Further When Clear: Adaptive Generative Modeling with Curriculum Consistency Model},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xVOMtecrAS},\nnote={under review}\n}" }, "abstract": { "value": "Significant advances have been made in the sampling efficiency of diffusion models, driven by Consistency Distillation (CD), which trains a student model to mimic the output of a teacher model at an earlier timestep. However, we found that the learning complexity of the student model varies significantly across different timesteps, leading to suboptimal performance in consistency models.\nTo address this issue, we propose the Curriculum Consistency Model (CCM), which stabilizes and balances the learning complexity across timesteps. We define the distillation process as a curriculum and introduce Peak Signal-to-Noise Ratio (PSNR) as a metric to quantify the difficulty of each step in this curriculum.\nBy incorporating adversarial losses, our method achieves competitive single-step sampling Fréchet Inception Distance (FID) scores of 1.64 on CIFAR-10 and 2.18 on ImageNet 64x64.\nMoreover, our approach generalizes well to both Flow Matching models and diffusion models. We have extended our method to large-scale text-to-image models, including Stable Diffusion XL and Stable Diffusion 3." }, "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": [ "adaptive curriculum learning", "noise schedule", "flow matching", "consistency 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/af3a26bd58d23df6e2c815987a84d1eb38cb0273.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/3f839082fdfa91e0d0d5e40e4b79d1aff0b579d8.pdf" }, "title": { "value": "See Further When Clear: Adaptive Generative Modeling with Curriculum Consistency Model" }, "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 see the weaknesses. Some other (related) questions are as follows.\n1. In the experiments involving TimeSiam, which encoder was used? The TimeSiam paper proposes both PatchTST (CI properties) and iTransformer (CD properties) as encoders. Since they have different characteristics, specifying the encoder used is essential. Additionally, since TimeSiam is utilized for classification tasks in your paper, it should be categorized appropriately, and the performance of CM + TimeSiam in classification tasks should also be evaluated.\n2. The CI framework does not attend between channels but attends with timestamps or patches in each channel. Then, how is the formulation of A in equation (1) justified as the identity matrix in equation (1) indicates the channel relationship? For example, how does PatchTST match to this formulation?\n3. Is the CD ratio consistent across different models for the same dataset? To establish the CD ratio as a reliable metric for measuring dataset CD, it should be verified whether CD ratios computed using different models (e.g., iTransformer vs. UniTS) yield consistent results.\n4. Since the CM has been validated on datasets with significant CD, it would be beneficial to test its performance on synthetic datasets with uncorrelated channels to verify that CM yields a low CD ratio and does not introduce unnecessary dependencies.\n5. Time series data often exhibit non-stationarity, which can affect correlation measures. How does your method handle non-stationary data, and does the CM adjust for changes in channel dependencies over time?\n6. Have you considered other measures that can capture nonlinear or more complex dependencies between channels, such as mutual information? This could potentially enhance the CM's ability to model complex inter-channel relationships." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The paper critiques the limitations of using correlation coefficients for measuring inter-channel relationships and proposes a new way to measure channel dependence through the CD ratio. The CM introduces very few parameters (α and β from domain parameters) yet effectively learns the implicit inter-channel relationships, leading to performance improvements in time series models.\n\n2. The authors validate their approach through various experiments, including few-shot and zero-shot settings, and provide thorough ablation studies and analyses to demonstrate the effectiveness and suitability of the CM structure and its relationship with CD." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the limitation that time series models often consider only explicit heterogeneity among datasets, such as varying sequence lengths and numbers of channels, while overlooking implicit heterogeneity like channel dependencies within the data. The authors propose a module called the Channel Mask (CM) to enable models to reflect channel dependence (CD) by adjusting the degree of CD based on dataset characteristics, achieving Partial Channel Dependence (PCD). By integrating CM into existing time series models, they demonstrate improved performance across various time series tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper lacks theoretical grounding or in-depth analysis to explain why the proposed method leads to performance improvements. A deeper understanding of the underlying mechanisms would clarify and strengthen the contribution.\n1.1. For example, how can (or how should) we define the concept of “implicit heterogeneity”? Why do we need this concept? While there are studies on channel dependence in time series machine learning, is the concept in this work different from existing studies? Furthermore, how is the rigorous definition of implicit heterogeneity connected to the proposed CM method? What is the rationale behind the use of correlation information between different real-world time series datasets in different domains to achieve better performance?\n1.2. For another example, while the use of correlation matrices primarily captures linear relationships between channels, the application of a sigmoid function in the CM introduces nonlinearity to the model. The authors mention that static correlations (global CD) are reflected through the CM, while dynamic, local correlations are captured by the attention mechanism. This design may help address some aspects of nonlinearity and temporal variation in channel dependencies. However, it remains a question whether this approach is fully sufficient to reflect the complex and changing characteristics inherent in time series data. A clear explanation of the rationale behind how the CM models these dynamic changes, or further investigation into its effectiveness in this regard, would strengthen our understanding. These considerations might be related to the rigorous definition of implicit heterogeneity in time series.\n1.3. Meanwhile, the authors state that “However, most previous works have focused on the model architecture to either capture or disregard CD, often overlooking the potential differences in CD across datasets.” However, they do not show specific theoretical analysis results on why and how existing studies on the model architecture are limited in capturing the differences in CD across datasets.\n\n2. The technical contribution is also limited.\n2.1. Following the above comment, the claim that existing CD models overlook differences in CD between datasets may not be fully substantiated. Since models like Crossformer, iTransformer, and TimeSiam learn attention patterns specific to each dataset, it's unclear whether they truly neglect dataset-specific CD differences.\n2.2. While the authors introduce PCD as an intermediary concept between channel independence (CI) and channel dependence (CD), the structure of CM, which uses a channel correlation matrix, cannot be applied to CI models. The study applies CM to CD models (iTransformer, UniTS) and shows performance improvements but does not verify whether applying CM to CI models can enhance performance. Thus, the proposed PCD cannot be extended to CI settings, limiting its utility in models that assume channel independence.\n2.3. The CM seems to be applicable only to Transformer models that apply attention along the channel axis and cannot be applied to models that apply attention along the time axis. For instance, existing studies like ST-MEM [1] and UniTST [2], which learn CD by applying attention along the time axis after channel flattening, cannot utilize CM. This limitation reduces the general applicability of the proposed method to other architectures.\n- [1] Na, Y., Park, M., Tae, Y., & Joo, S. (2024). Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram. arXiv preprint arXiv:2402.09450.\n- [2] Liu, J., Liu, C., Woo, G., Wang, Y., Hooi, B., Xiong, C., & Sahoo, D. (2024). UniTST: Effectively Modeling Inter-Series and Intra-Series Dependencies for Multivariate Time Series Forecasting. arXiv preprint arXiv:2406.04975.\n\n2.4. The CD ratio proposed in the study is calculated after combining CM with the time series model and training, and its value may vary depending on the model used. The paper does not provide sufficient evidence to demonstrate that the CD ratio is consistent across different models for the same dataset, raising concerns about its adequacy as a metric for measuring channel dependence.\n2.5. Meanwhile, although the authors started their argument from the emergence of time series foundation models (TSFMs), they do not provide sufficient validation of this work for different TSFMs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Can you provide comparisons of your method against the same architecture with full channel dependence and full channel independence?\n- What theoretical guarantees or analysis can you provide to show when partial dependence would outperform the extreme cases?\n- How does your method handle scenarios where dataset boundaries are ambiguous or when data comes from multiple unknown sources?\n- Can you quantify the computational overhead and memory requirements of your approach compared to full dependence and independence cases?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Introduces a novel methodology for handling channel relationships specifically designed for pretraining foundation models in time series\n- Proposes a systematic framework for incorporating dataset-specific channel dependencies into the pretraining process\n- Demonstrates superior performance in challenging scenarios (few-shot and zero-shot learning)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel concept called Partial Channel Dependence (PCD) to address implicit heterogeneity in time series data, specifically focusing on varying dependencies between channels. The authors propose a channel mask mechanism that combines correlation matrices (for relative dependencies) with learned domain parameters (for absolute dependencies). The approach is evaluated across multiple time series tasks (forecasting, classification, imputation, and anomaly detection) in both few-shot and zero-shot settings, demonstrating its versatility across different foundation models and single-task models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Dataset Identification Requirement**: The method requires knowing exactly which samples belong to which dataset during pretraining. This is a strong assumption that may not hold in real-world applications where data sources might be mixed or unclear.\n- **Lack of Fixed vs. Variable Dependency Analysis**: The paper assumes variable channel dependencies are necessary but doesn't justify why a simpler fixed dependency structure wouldn't work equally well. Without this comparison, the added complexity of variable dependencies might be unnecessary.\n\n- **Missing Critical Baseline Comparison**\n\n - No comparison against the same architecture with full channel dependence\n - No comparison against the same architecture with full channel independence" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1.\tCould the authors apply a similar masking approach to CI models like PatchTST or PITS, and compare their performance across different datasets? If possible, please provide performance comparisons for the original settings w CM and w/o CM." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The method is straightforward, and the motivation is clearly articulated.\n* Extensive experiments across various scenarios validate that PCD enhances the performance of Transformer-based multivariate time series models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Partial Channel Dependence (PCD), a method designed to capture the varying dependencies between channels across different datasets. PCD achieves this by applying channel-wise attention multiplied by the corresponding dataset mask. Experimental results demonstrate that PCD yields an average performance improvement across various time series tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The results for iTransformer and PatchTST presented in Table 3 differ significantly from those reported in the original papers. Additionally, the ETTh2, ETTm1, and ETTm2 datasets are well-known multivariate time series datasets. Could you please provide a comprehensive comparison of the same baselines on these datasets across all prediction lengths?\n* While PCD does enhance the performance of Transformer-based models for multivariate time series forecasting, I recommend including recent MLP-based and CNN-based models in the baselines, such as RLinear and ModernTCN. Additionally, GNN-based models like CrossGNN are also adept at capturing multivariate relationships. Including these would strengthen the performance comparisons.\n* The length of the input time series is an important factor influencing experimental results. Therefore, I would like to see the impact of varying sequence lengths on the results w CM and w/o CM.\n* There are several literature on channel dependency modeling, such as [1-2]. Detailed discussion is needed.\n\n[1] Zhao et al., Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators, ICLR 2024\n\n[2] Qi et al., Enhancing Multivariate Time Series Forecasting with Mutual Information-driven Cross-Variable and Temporal 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": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This work focuses on a highly valuable research direction and highlights the importance of Partial Channel Dependence in Time Series analysis.\n2. This work proposes a concise approach to adjust the correlation matrix obtained from prior knowledge." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces the concept of partial channel dependence (PCD) to address implicit heterogeneity in time series (TS) data. By utilizing a channel mask that incorporates a correlation matrix to encode relative dependencies between channels and domain parameters to learn dataset-specific absolute dependencies, the authors refine the correlation matrix for better channel dependency adjustments. The effectiveness of PCD is validated across four TS tasks—forecasting, classification, imputation, and anomaly detection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The concept of Partial Channel Dependence (PCD) is already discussed in the CCM[1] where the authors employ a clustering approach to capture latent PCD. The current content lacks a detailed comparison and discussion with this work.\n2. It is important to note that not all foundation models are based on attention mechanisms (TTM [2]), and not all methods that utilize attention mechanisms effectively capture the attention between channels like UniTS (TimesFM[3], Timer[4], MOIRAI[5], MOMENT[6]). As a plugin for foundation models, the generality of the CM method is insufficient.\n3. The paper does not specify how the correlation matrix mentioned was constructed, and references [7] and [8] primarily focus on capturing correlation relationships with lag properties. In contrast, this work does not explore lag properties but instead utilizes the complete sequences. The construction method of the correlation matrix needs to be explained in detail.\n4. Domain parameters highly correlated with the construction method of the correlation matrix, If use other methods, they may not be effective. The effectiveness of domain parameters in the experiments lacks further validation.\n5. In the experiments, there is a lack of comparison with other plugins, such as LIFT[8], and each task only validates the plugin's improvement on a few models, making its generality difficult to confirm. More importantly, a large number of foundation models have not been considered in the experiments, such as [2-6]. The existence of these issues raises serious doubts about the effectiveness and generality of the plugin. \n\n[1] From Similarity to Superiority: Channel Clustering for Time Series Forecasting.\n\n[2] Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series.\n\n[3] A decoder-only foundation model for time-series forecasting.\n\n[4] Timer: Transformers for Time Series Analysis at Scale.\n\n[5] Unified Training of Universal Time Series Forecasting Transformers.\n\n[6] MOMENT: A Family of Open Time-series Foundation Models.\n\n[7] Vcformer: Variable correlation transformer with in-herent lagged correlation for multivariate time series forecasting.\n\n[8] Rethinking channel dependence for multivariate time series forecasting:Learning from leading indicators." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In conclusion, while the paper presents an extensive array of experimental evidence, the motivation for addressing channel dependency heterogeneity is not entirely convincing in terms of its necessity for building a foundational model. Although channel dependencies may be relevant to specific tasks in the time-series domain (e.g., forecasting), the underlying rationale and its influence on design choices are not fully explained in the manuscript. Additionally, certain key experimental results appear inconsistent, which is problematic given the paper's emphasis on empirical performance gains over theoretical justification. Nevertheless, the proposed architecture is straightforward and appears tailored to address varying channel dependencies, and it has the potential to be impactful if the dependencies are better elucidated and some of the reported results are clarified.\n\nIf the authors address these points and resolve potential experimental issues in the main results table, I would be inclined to raise my rating." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper presents a straightforward algorithm with clear explanations.\n- The manuscript covers a broad range of experiments across various time-series tasks (anomaly detection, forecasting, imputation, etc.) in different settings (supervised, few-shot/zero-shot domain/task) with extensive analyses (e.g., robustness to missing values) that support its claimed advantages.\n- The newly introduced experiment, \"masked channel prediction,\" is particularly novel and promising, though it is only briefly discussed in the manuscript." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes a time-series foundation model designed to capture varying channel dependencies across different datasets. By using a correlation matrix as a prior, the model incorporates a learnable domain parameter to construct a domain-specific channel mask, effectively capturing global channel dependencies. In conjunction with local channel dependency masks (as introduced in prior works, such as iTransformer), the proposed model accommodates diverse channel dependencies across datasets, leading to performance improvements in tasks such as forecasting, imputation, and anomaly detection (in full-shot, few-shot, and zero-shot settings)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Despite the demonstrated performance gains across multiple tasks and settings, some of the motivation and design choices remain insufficiently addressed in the manuscript. Additionally, certain reported experimental results differ from what appears in the paper, and key experimental details are not fully explained.\n\nFirst, I would appreciate clarification from the authors regarding the principles underlying channel dependency in the proposed time-series foundation model:\n\n- Why is it crucial to account for varying channel dependencies? To what extent is this heterogeneity in channel dependency necessary for constructing a robust foundation model for time-series data? Prior research has highlighted its importance for forecasting tasks both empirically [1][2] and theoretically [3], yet the rationale for its role in the foundation model is not fully studied. What is the anticipated impact of the time-series foundation model, and how does it address channel dependencies? (Simply stating \"inherent heterogeneity\" may be insufficient in this context.)\n- The global mask encapsulates correlation across various domains during training, which is then reused during testing without further adjustment. Is it sufficient to rely on the global correlation matrix alone? Since local correlations may vary over time, the correlation matrix at test time might differ significantly from that during training. How do you ensure its stability under these conditions?\n- Is correlation an appropriate metric for capturing channel dependencies? Some studies in forecasting suggest a \"causal relationship\" between channels, while others tackle spurious correlations (where channels appear correlated but are not causally related).\n\nFurthermore, several design choices would benefit from clarification:\n\n- What motivated the choice to use domain-specific global attention while sharing local attention across multiple domains? What specific roles do global and local attention play in the model's functioning?\n- Is the \"pair-wise dot product\" between global and local attention sufficient to achieve the intended effect? Under extreme test-time scenarios where variables v1 and v2 exhibit low global correlation (approaching zero) but high local correlation, this design might yield low attention scores, potentially failing to capture abrupt correlation increases under certain test-time conditions.\n\nIn addition, some minor experimental inconsistencies and essential experimental details require clarification:\n\n- Some scores of baselines appear lower than what has been reported in the original paper. In Table 3, the MSE/MAE score on iTransformer is way higher than it has been reported (Appendix F, Table 8 in [4]) Have these results been reproduced, and if so, what could account for the discrepancies?\n- How is the global correlation matrix defined? Does it only consider the correlation matrix over the training period?\n- In zero-shot experiments on new datasets, how is the domain parameter handled? As the domain parameter cannot be directly learned for unseen domains, is it substituted with that of a similar domain?\n\n[1] Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators (ICLR 2024) \n\n[2] Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series (arXiv 2024) \n\n[3] Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning (ICML 2024)\n\n[4] ITRANSFORMER: INVERTED TRANSFORMERS ARE EFFECTIVE FOR TIME SERIES FORECASTING (ICLR 2024)" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce the concept of partial channel dependence (PCD) to partially adjust the channel dependence (CD) captured by the model through the proposed channel mask (CM), which contains dataset-specific information." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024partial,\ntitle={Partial Channel Dependence with Channel Masks for Time Series Foundation Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xVU6rY37X9},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in foundation models have been successfully extended to the time series (TS) domain, facilitated by the emergence of large-scale TS datasets. However, previous efforts have primarily focused on designing model architectures to address explicit heterogeneity among datasets such as various numbers of channels, while often overlooking implicit heterogeneity such as varying dependencies between channels. In this work, we introduce the concept of partial channel dependence (PCD), which enables a more sophisticated adjustment of channel dependencies based on dataset-specific information. To achieve PCD, we propose a channel mask that captures the relationships between channels within a dataset using two key components: 1) a correlation matrix that encodes relative dependencies between channels, and 2) domain parameters that learn the absolute dependencies specific to each dataset, refining the correlation matrix. We validate the effectiveness of PCD across four tasks in TS including forecasting, classification, imputation, and anomaly detection, under diverse settings, including few-shot and zero-shot scenarios with both TS foundation models and single-task 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": [ "Time Series", "Foundation Model", "Channel Dependence", "Transformer" ] }, "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/3536f3f5f0c48d18414fd9b73b6cc97c02ffa3b9.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/147c0c2589209dfe9e6c47d6775bd84dbdacb2e8.zip" }, "title": { "value": "Partial Channel Dependence with Channel Masks for Time Series Foundation Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In the text is not mention what g could be or some options. It could be like the utility function used in Natural Evolutionary Strategies? because it seems to assign more weight/importance to high ranking solutions.\n\nI think it would be very interesting in a future work to explore more the latent diffusion version of the algorithm, and techniques to understand which parameters could be ignored or which parameters should be perturbed together." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "By connecting methods from different fields the work proposed an interesting new evolutionary algorithm that can be further improved using variations and techniques found in diffusion models.\n\nThe text is very clear and the math derivation from diffusion to an EA algorithm is easy to follow. Figure 1 really helps to visualize how the population of the diffusion EA spreads to the higher fitness regions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors try to connect Diffusion Models and Evolutionary computation by arguing that both processes do iterative refinements through an update rule plus some perturbation. In evolutionary algorithms the update comes in the form of natural selection, in diffusion is the denoising phase. The perturbation corresponds to mutation for evolution, and the diffusion phase for diffusion models. \n\nFrom the above connection an evolutionary algorithm that performs in its iteration a diffusion process is proposed. Instead of aiming to recover some distribution of the data, the goal is to turn the random initial population points towards a distribution centered around the optimized function optimum." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Maybe I missed it but there is no discussion on how to choose the mapping to a density function g() or which g() was used for the experiments. It seems is important for the search to work properly to have a mapping that makes it clear which points should be paid more attention by assigning to them bigger weights. Maybe there is some connection to common selection strategies in evolutionary algorithms that could be discussed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "As mentioned in the appendix, different Alpha and noise schedule settings are tested, what about the detailed experimental results? Is there any analysis about the results of different settings, which could be useful for the users?" }, "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 easy to follow. The proposed method Diffusion Evolution is well-motivated, with clear explanation and illustration." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new evolutionary algorithm inspired by diffusion models. It views the evolution processes as the denoising process of diffusion models, and designs an evolutionary algorithm named Diffusion Evolution, which is based on the denoising framework of the famous diffusion model DDIM. During evolution, it takes each individual in the population as a solution in the denoising process, and updates it with similar updating rules in DDIM. To facilitate better performance, it further follows previous works to optimize in the latent space. Experiments on several benchmark functions and a simple cart-pole controlling problem show that compared with classic baseline methods include CMA-ES, OpenES and PEPG, the proposed Diffusion Evolution can obtain more diverse solutions with good fitness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "As there has been a variety of evolutionary algorithms with various inspirations, the most important issue when proposing a new method should be clarifying its strength compared with previous methods, from the perspectives of theoretical analysis and extensive experiments. However, the proposed Diffusion Evolution in this paper is lack of theoretical analysis. Meanwhile, the experiments are much too simple. For example, only five synthetic functions and a simple cart-pole controlling problem are included. As the diversity of solutions seems to be the strength of the proposed Diffusion Evolution, except for methods like CMA-ES, which are not specified for diversity, comparison with other kinds of evolutionary algorithms like Quality-Diversity (QD) are necessary." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 the experiments section, the authors select three different evolutionary strategies for performance comparison. However, the rationale for choosing these specific methods is not clearly articulated. I would appreciate more insight into this selection process, particularly since each of these methods was introduced over five years ago." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The idea of regrading the diffusion model as an evolutionary algorithm is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript introduces an innovative perspective by interpreting the diffusion model as an evolutionary algorithm, highlighting the mathematical similarities between the diffusion and evolutionary processes. It proposes a Diffusion Evolution Method that employs iterative denoising to heuristically optimize solutions within the parameter space, drawing an analogy between the generative process of the diffusion model and the selection and mutation mechanisms in biological evolution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The mathematical demonstration process in this manuscript is insufficient. The authors conceptualize the evolution process as a transformation of the probability density function, deriving the diffusion evolution algorithm from the Bayesian formula. However, this derivation overlooks key complexities inherent in evolutionary dynamics, such as genetic drift and gene recombination. Additionally, the application of the Bayesian formula assumes conditional independence, which may not be held in the context of evolutionary processes due to the interactions and competition among individuals. Furthermore, the manuscript does not provide a clear explanation of how to effectively map a complex fitness function into a probability density function, which is essential for the practical implementation of the proposed model.\n2. Since the author claims that the diffusion model functions as an evolutionary algorithm, the experimental section must reflect this comparison appropriately. Specifically, in scenarios involving multiple tasks and comparisons, aspects such as speed, performance metrics (including best, worst, median, and average results), and the stability of performance should be examined. Additionally, the model's excellence should be illustrated through the mean variance of the results. Currently, the scope of experiments conducted is insufficient and does not adequately support the claims made in the manuscript.\n3. There are several grammatical errors and unclear expressions throughout the article. For example, the definitions of certain terms lack clarity, the derivation process for the formulas is inadequately explained, and the labeling of the charts is inaccurate.\n4. The authors categorize diffusion models as evolutionary algorithms on the basis that both methods perform distribution transformation. However, this classification may be overly broad. For instance, semantic segmentation models also involve transforming distributions—from real images to pixel-level segmentation results. Should we, therefore, classify these segmentation models as evolutionary algorithms as well? This comparison seems inaccurate. A more robust argument would establish that the Markov process within a non-equilibrium thermodynamics framework (diffusion) function as an unconstrained parameter optimization technique (evolution). If this argument cannot be substantiated, I recommend revising the title to better reflect the content.\n5. The manuscript presents two primary sets of experiments: multi-objective evolution and latent space diffusion evolution. However, the multi-objective evolution experiments are limited to a two-dimensional parameter space and a simplistic fitness function, which fails to demonstrate the proposed approach's efficacy in high-dimensional parameter spaces or with more complex fitness functions. Similarly, the latent space diffusion evolution experiments focus solely on the CartPole task, lacking validation across a broader range of reinforcement learning tasks. Furthermore, the experimental results do not include any statistical significance tests, making it challenging to determine whether the observed improvements in algorithm performance are statistically significant. This omission significantly undermines the credibility of the findings presented." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 work potentially contributes new insights and methods for both diffusion model and evolutionary algorithm communities, even if the practical impact in specific real-world applications remains to be validated. Can the authors address this?\n\nCan statistical tests be added?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "he core idea of treating diffusion models as evolutionary algorithms is innovative. It extends diffusion models to broader applications by framing them as tools for evolutionary tasks, potentially contributing new knowledge to both evolutionary biology and AI communities.\n\nTechnically: The paper provides thorough mathematical grounding for the equivalence between diffusion and evolution, specifically explaining diffusion as a probabilistic denoising process analogous to evolutionary mechanisms like mutation and selection.\n\nExperimentals: Comprehensive experiments benchmark Diffusion Evolution against traditional algorithms (e.g., CMA-ES, PEPG) ascross various fitness landscapes, demonstrating its strength in maintaining diversity and achieving multiple optima.\n\nApplications in High-Dimensional Problems: The adaptation of Diffusion Evolution to high-dimensional tasks via latent space diffusion showcases its practical viability in reinforcement learning contexts, with empirical results suggesting its potential in complex environments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the theoretical and practical parallels between diffusion models and evolutionary algorithms, proposing the \"Diffusion Evolution\" approach, which adapts diffusion models for evolutionary tasks. This new method, Latent Space Diffusion Evolution, enhances the ability to find diverse and optimal solutions in high-dimensional parameter spaces, particularly within reinforcement learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Complexity of Explanation: The theoretical connections between diffusion and evolution, while compelling, are highly complex, and the paper occasionally sacrifices clarity for depth. This might limit accessibility to a broader audience.\n\nAlthough promising, the methodology may have limitations in scenarios requiring open-ended evolution, a challenge acknowledged briefly. More thorough discussion could help set realistic expectations for potential users.\n\nWhile the Diffusion Evolution algorithm demonstrates superiority in finding diverse solutions, certain comparisons (e.g., high-fitness solutions) against traditional evolutionary algorithms lack statistical significance or detailed analysis of computational cost versus benefit, particularly in high-dimensional settings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024diffusion,\ntitle={Diffusion Models are Evolutionary Algorithms},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xVefsBbG2O},\nnote={under review}\n}" }, "abstract": { "value": "In a convergence of machine learning and biology, we reveal that diffusion models are evolutionary algorithms. By considering evolution as a denoising process and reversed evolution as diffusion, we mathematically demonstrate that diffusion models inherently perform evolutionary algorithms, naturally encompassing selection, mutation, and reproductive isolation. Building on this equivalence, we propose the Diffusion Evolution method: an evolutionary algorithm utilizing iterative denoising -- as originally introduced in the context of diffusion models -- to heuristically refine solutions in parameter spaces. Unlike traditional approaches, Diffusion Evolution efficiently identifies multiple optimal solutions and outperforms prominent mainstream evolutionary algorithms. Furthermore, leveraging advanced concepts from diffusion models, namely latent space diffusion and accelerated sampling, we introduce Latent Space Diffusion Evolution, which finds solutions for evolutionary tasks in high-dimensional complex parameter space while significantly reducing computational steps. This parallel between diffusion and evolution not only bridges two different fields but also opens new avenues for mutual enhancement, raising questions about open-ended evolution and potentially utilizing non-Gaussian or discrete diffusion models in the context of Diffusion Evolution." }, "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 learning", "evolutionary computation", "Evolutionary Algorithms", "Diffusion Models", "Optimization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/51cbb736f950b3c3bd50484d1ceeefb5f1521ccc.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": "Diffusion Models are Evolutionary Algorithms" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "N/A" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper seems to be significantly altered for format or been generated by an automated program. Specifically:\n1. Figures 1, 2 have captions with notably smaller font size. On the other hand, there are many extra spacing in the paper that could have been utilized: Figures 1, 2 use only 1/3 of the width, so do Tables 1-9; Lines 83-90 are empty sections. Table 1 is not even referenced in the text.\n2. The citations are weirdly repeated numerous times: Line 50-75, notably Hu et al 2020 has been cited for 5 times. It happens multiple times throughout the paper: at lines 137-141, 145-148, 185-194, to name a few.\n3. Even if one ignores all these unusual things, the numbers reported in the tables looks indistinguishable from the alternatives." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "See summary" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Minors: \n1. Line 152, manifold $M$ not $\\mathcal{M}$.\n2. $J$ is loss function, $f$ is objective function.\n3. Line 218 $x$ vs $x_{i-1}$ that comes from lazy typing to include all $grad$, $f$ and $x_{i-1}$ inside the command mathrm. \n4. Again, inconsistent notations: retraction map $R$ and $\\mathcal{R}$\n5. $\\phi^{\\lambda_i}$: $\\lambda_i$ is the power???\n6. No definition of $C_{i+1}$, it seems that the reader needs to obtain it from similar formula for $C_{i-1}$.\n7. Introducing Barzilai-Borwein method with notation $\\omega_{i-1}$, $\\zeta_{i_1}$ and $\\zeta_{i_2}$ without explanation. \n8. Algorithm 1: Compute $\\zeta_i$ according to equation (7), but equation (7) is an inequality.\n9. Theorem 4.2, no definition of $x^{\\star}_{i+1}$.\n10. Line 249: $C_{i-1}$ is a convex combination of $C_{i-1}$ and $f(x_{i-1})$?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper proposed a method for optimizing on real manifolds. \nSome theories are provided with support from empirical results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides an analysis of optimization on manifold. It proposes a Reset method to improve the convergence and model stability when compare with other mentioned methods. More particular, section 4 recall important tools in Riemannian manifold and give summary of some gradient descent methods. Section 4.2 presents the Reset method that contains three steps to update, coming from $x_{i-1}$ to $x_{i+2}$. There are two adjusting steps using gradient direction and function $B_{x_i}$. Theorem 4.1 proved that the accumulation points is also the stationary point. Theorem 4.2 give some upper bound for the difference between consecutive gradient updating using SGD, Adam and AdamW. Experiments is carried out on CIFAR-10 and CIFAR-100 for image generation task and cluster constrat task." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is not clear about the insight/motivation of each mentioned/proposed method presented through out the paper. The first one are derivations in inequalities (6) and (7). There is not picture or figure to illustrate clearly the advantage of method. I could only imagine that the first order of Taylor approximation is not good enough, thus we could obtain a better one via Armijo search, when some certain conditions satisfied. That goes through an interpolation between some bounds and expect that the interpolation will help to achieve a better bound. The work also mentioned the Barzilai-Borwein method but there is no explanation about reason for computing the correlation between two vectors. \n\nExperiment results: For Image generation task, in Table 2, when adding your method with different type of manifold (in fact it is not clear for me why we have different type of manifold here), the proposed method is at best with spv, but does not work better with \"e\", \"fr\", \"o\", etc. In Table 3, there are some mixed performances between those method when comparing with each other except the \"spv\". Is there any explanation for the performance of each method?\n\nFor image generation task, there is no picture shown, rather than the number. Since average precision is already high, is the improvement noticeable in the pictures?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "+ What is the precise definition of $B_{x_i}(x_i)$ including what are its inputs, outputs, its purposes, and the motivation for introducing this function?\n\n+ What is the usage of $\\zeta$'s? Seems like they have never been used in the pseudocode or the algorithm in Eq. (5).\n\n**References:**\n\n[1] Brendan O’donoghue and Emmanuel Candes. Adaptive restart for accelerated gradient schemes. Foundations of computational mathematics, 15:715–732, 2015.\n\n[2] Bonnabel, S. (2013). Stochastic Gradient Descent on Riemannian Manifolds. IEEE Transactions on Automatic Control, 58(9), 2217-2229.\n\n[3] Yun, J., & Yang, E. (2023). Riemannian SAM: Sharpness-Aware Minimization on Riemannian Manifolds. In Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS 2023)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The restart method was first introduced by O'donoghue & Candes [1] in the Euclidean settings. This paper suggests an approach to extend this method to the Riemannian setting. If revised properly, this could be an interesting perspective for improving generic optimizers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel Riemannian optimizer that incorporates an additional \"reset\" step to adjust the learning rate." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Presentation**: Firstly, several serious representation issues make the paper incoherent and difficult to follow, obscuring the key messages and contributions. Key issues include:\n\n+ *Related Works and Preliminary Sections*: The related works and preliminary sections are placed in the appendix. To improve clarity, many sections could be condensed or moved to the appendix, creating space for these key areas—for example, explanations of SGD, Adam, AdamW, and most experimental results in the tables.\n\n+ *Main Algorithm Presentation*: The algorithm presentation has gaps in clarity and completeness. For instance, variables such as $ \\zeta_{i_1} $ and $ \\zeta_{i_2} $ appear in the pseudocode but are unused in subsequent steps. In Eq. (5), the process of obtaining $ \\alpha_{i+1} $ is unclear. Additionally, the line \"Set $ R_{x_{i+1}}(-\\alpha_{i+1} \\nabla f(x_{i+1})) \\leftarrow B_{x_i}(x_i) $\" is confusing, as we cannot \"set the retraction\" to a specific value.\n\n+ *Function $ B_{x_i}(x_i) $*: The definition and role of $ B_{x_i}(x_i) $ are unclear and potentially redundant. What are its inputs and outputs? Although it’s described as a \"step size correction function,\" suggesting it outputs the step size, in Eq. (5), it produces $ x_{i+1} $, implying it represents the next model. Furthermore, it is stated that $ B_{x_i} $ is selected from SGD, Adam, or AdamW—all Euclidean optimizers—suggesting $ x_{i+1} $ might not lie on the manifold, making it impossible to compute $ x_{i+2} $ directly from $ x_{i+1} $. Additionally, the use of $ B_{x_i}(x_i) $ as notation is confusing since it appears to take only a single input $ x_i $. The reasoning for adding a correction step with operator $B $, which is supposed to be a primary contribution, is also not discussed.\n\n+ *Experimental Results*: The experimental results were provided on many real manifolds. However, it is unclear if the improvements are attributed to the Reset method or the manifolds. In particular, it would be more fair to compare the proposed method with other Riemannian optimizers, such as RSGD. \n\n+ *Minor Formatting Issues*: Multiple minor issues are present, such as inconsistent font sizes in table captions, unusual word choices (e.g., “contenting” in the pseudocode), inconsistent font styles (e.g., $x_i$ in the text but not italicized in Eq. (5)), table formatting inconsistencies, and citation format errors (e.g., \\cite{} vs. \\citep{}).\n\n**Soundness:** In both theory and experiments, it is critical to compare the Reset method with other Riemannian optimizers, such as RSGD [2] or Riemannian-SAM [3], in addition to Euclidean baselines like SGD, Adam, and AdamW.\n\nThe theoretical contributions are relatively limited. Theorem 4.2 only demonstrates the gradient deviation of the Reset method when compared with SGD, Adam, and AdamW, and it includes an unavoidable term $\\epsilon_0^2 $, indicating a potentially unfavorable trait of the Reset method. Additionally, comparing the Reset method, a Riemannian optimizer, to Euclidean methods may be unfair from a theoretical perspective.\n\nThe experimental results are also not particularly insightful. Specifically, all results on the CIFAR-10, CIFAR-100, STL-10, and SVHN datasets achieve accuracies of at least 98%, making it challenging to discern performance differences. On the Market-1501 and DukeMTMC-reID datasets, variations of the Reset method applied to the same base optimizer yield considerable performance differences, suggesting that the method may be sensitive to variant choices, making tuning more challenging.\n\n**Contribution:** To my understanding, the restart method was proposed by O'donoghue & Candes [1]. This work seems to be a trivial extension of this work to the Riemannian setting, which makes its contributions limited. Moreover, the motivation and intuition of the proposed algorithm are unclear, mostly due to the incohesive presentation." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Based on the theory of real surface optimization, we propose a new optimization method, named the Reset Method." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024reset,\ntitle={Reset Method based on the Theory of Manifold Optimization on Real Manifolds},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xVw8YNEtH3},\nnote={under review}\n}" }, "abstract": { "value": "Manifold optimization is prominent in the fields of applied mathematics, statistics, machine learning, and in particular, deep learning. By leveraging the intrinsic geometric properties of manifolds, constrained optimization problems can be transformed into unconstrained optimization problems on certain manifolds. An innovative method, Reset Method, is introduced that combines manifold optimization and standard methods (SGD, Adam and AdamW), aiming to enhance the improvement of precision. The efficacy of our proposed method is corroborated by extensive deep learning experiments, providing visible higher precision." }, "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": [ "Manifold Optimization", "Real Manifolds", "Method", "Deep 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/915ece546feb0284b8d4c61dd77d3e759323fd5a.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": "Reset Method based on the Theory of Manifold Optimization on Real Manifolds" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "See the 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 easy to follow" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a work on aggregating contextual information into time series forecasting. Specifically, the authors propose to use a universal context encoder to encode the contextual information as embedding and boost the time series forecastor with the encoding. Experimental results suggest that the contextual information boost the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The contribution is limited. This paper mainly discuss about contextual information in time series forecasting. However, it is well known that the contextual information is helpful for time series. What really matters is how to make use of it. In this paper, the authors list three reasons about why it is hard to make use of the information, such as multi modality and non-uniformity across domains. However, when really talking about how to make use of the context, the authors just simply concatenate the category variables with continuous variables. This design is unrelevant with the challenges mentioned. What if some out-of-domain dataset contain some variables that are unseen on the training data? Also, the design is well-justified. Why not align the metadata as one modality (like text) and then convert it to embeddings with one shared encoder? Do we need to align the timely metadata (like news with timestamp) with the speific timestamp before forwarding them all into the cross attention?\n\n2. The theoretic analysis is useless but takes 1.5 pages. All the theoretic results are trivial corollary of conclusions from introductory undergraduate-level text book of information theory and machine learning. I suggest that the authors are just trying to decorate their papers with some theoretic analysis\n\n3. Experimental results are not comprehensive enough. We prefer experiments on more datasets (like ETT) and baselines (like TimesNet).\n\n4. The authors list 4 reasons why they prefer fine-tuning in Sec 5.2. But there is lack of empirical supports. I did not see training curves that reflect the unstable training of all-from-scratch. \n\n5. The bitcoin-news dataset description is too short. This could be the most insightful part about how to collect contextual data. But there is no details. How and where did you get the data? Are the data filtered? Are the collection based on time or keywords? Some many details are missing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "The presentation and problem setup are clear and easy to follow. However, I have two major concerns regarding the theoretical analysis of the benefits of adding exogenous variables and the overall technical contribution of the paper.\n1. By definition, mutual information is non-negative. Even in the worst-case scenario where A and C are completely independent, the increase in mutual information is 0. A higher mutual information value does not necessarily translate to higher accuracy or better performance. Noisy or even misleading data can also increase mutual information as long as some degree of dependency exists. Could the authors elaborate on this issue?\n2. Adding exogenous variables to time series forecasting models is not a novel concept. It is a natural extension of general time series forecasting models and is a dominant approach for work in some domains, such as financial market predictions. For example, older transformer-based models like Autoformer can be adapted to include exogenous variables and are often used as baselines in other papers. Moreover, newer transformer-based models, such as Timexer, also support this approach natively. There are also attempts to use large language models (LLMs) to incorporate exogenous market information for stock movement prediction, such as Plutos. Therefore, I believe the overall contribution of the work is somewhat weak, as it does not introduce a new concept.\n3. It is expected that the authors should at least compare the proposed framework with other models that can accept the same input. Otherwise, it is difficult to justify the performance of the proposed framework." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The presentation is clear.\n2. The problem studied is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a framework designed to incorporate exogenous variables into time series forecasting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The contribution is limited.\n2. There is a lack of comparison with SOTA models." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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": "Refer to \"Strengths And 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 starting point of this article is good. Exogenous contextual features can indeed serve as key auxiliary information to influence time series forecasting.\n2. The paper is mostly well-presented and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This article proposes ContextFormer, a novel plug-and-play framework that utilizes cross-attention blocks to integrate multimodal metadata, including categorical, continuous, time-varying contexts, and even textual information, into existing any context-agnostic base forecasters. The author selects two SOTA models of PatchTST and iTransformer as the base forecasters and validates ContextFormer framework on seven real-world datasets spanning energy, traffic, environmental, and financial domains. Experimental results confirm the effectiveness of the framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Introduction part summarizes three challenges in incorporating metadata into forecasting models, but this article does not provide a detailed explanation of why ContextFormer can solve these problems. For example, the first challenge mentions the importance of aligning time series history and multimodal metadata. However, there is a lack of evidence on how ContextFormer ensures that the two have been aligned, the alignment effect, and the impact of aligned representation on forecasting results. Although the effectiveness of ContextFormer is reflected in the final MSE/MAE metrics, I mainly focus on these intermediate results and suggest the author to supplement these explanations.\n\n2. In related works part, the author lists some methods for forecasting with covariates. As far as I know, it also includes methods such as TFT, TSMixer, and TimeXer, etc., but this article does not compare with these methods. It is suggested that the author supplements these comparison experiments.\nSome references:\nTFT: https://doi.org/10.1016/j.ijforecast.2021.03.012,\nTSMixer: https://arxiv.org/pdf/2303.06053,\nTimeXer: https://arxiv.org/pdf/2402.19072.\n\n3. In Table 2 caption, it is mentioned that \"The best results for each base model in each row are highlighted in bold.\". However, the results of the Retail dataset (horizon=96, MSE metric) do not match this statement. It is recommended to revise the wording.\n\n4. Doubt about lines 491-492: \n 1) The explicit motivation for comparing with Chronos and whether the comparison is fair? \n 2) Are the results in Table 2 sufficient to support this conclusion? I agree with the example provided by the author (486~487), but for other datasets, I have the following questions: (a) Is it still necessary to compare the context-aware model with Chronos for the two datasets of Air quality and Electricity, as the context-agnostic model is already better than Chronos? (b) For the Retail and Bitcoin datasets, Chronos performs the best. I hope the author can provide detailed explanations to alleviate my concerns." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The authors can refer to the weakness listed above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. As a plug-in solution, ContextFormer can be integrated with various predictive backbones for a diverse array of applications. \n2. The two-stage fine-tuning methodology guarantees that the lower bound of ContextFormer is at least equivalent to that of context-agnostic approaches.\n3. The extensive experiments look hopeful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces ContextFormer, a method designed to incorporate multimodal contextual information into existing time series forecasting models. Traditional models typically rely solely on historical data, neglecting external influences such as news, weather, and market trends that can significantly impact predictions. ContextFormer effectively addresses this limitation by integrating diverse types of metadata—including categorical, continuous, time-varying, and textual information—using cross-attention mechanisms. The experiments demonstrate that ContextFormer can enhance forecasting performance by up to 30% compared to state-of-the-art models across various datasets in fields such as energy, traffic, environment, and finance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Inadequate Presentation.** The organization of the paper lacks rationality. Excessive emphasis is placed on the importance of external context, which is well-known. The methods section is overly succinct. The descriptions of Metadata Embedding and Temporal Embedding are overly concise, leaving the application dimensions and the output shape ambiguous. Furthermore, the embedded interaction in the Information Fusion component is similarly unclear, too. Its applicability to both a single embedding for one variable (iTransformer and PatchTS) and a single embedding for one timestamp (Informer and Autoformer) across two distinct architectures remains uncertain. In summary, the workflow of the algorithm is perplexing.\n\n2. **Limited Innovativeness.** Although the author presents a case in the introduction regarding the influence of news on stock prices, the treatment of this unstructured external text information is only mentioned in Appendix C.7 cursory and lacks corresponding experimental results on BITCOIN-NEWS. The models and data discussed throughout the paper primarily rely on structured auxiliary information, despite incorporating both continuous and categorical variables. The prior research has focused on integrating structured external information to enhance prediction accuracy, including exogenous variables [1] and timestamps [2]. The paper's core contribution remains ambiguous. Moreover, The three challenges outlined in the introduction are unpersuasive.\n\n3. **Unrigorous Experiments.** As stated in Weaknesses 2, there has been some prior work aimed at integrating structured external information to enhance prediction accuracy. This paper should be compared with these frameworks that utilize external information, rather than solely with models that lack auxiliary information.\n\n**References**\n\n[1] 2024, TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables\n\n[2] 2024, Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-Local Fusion Perspective" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024context,\ntitle={Context Matters: Leveraging Contextual Features for Time Series Forecasting},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xW4J2QlqRx},\nnote={under review}\n}" }, "abstract": { "value": "Time series forecasts are often influenced by exogenous contextual features in addition to their corresponding history. For example, in financial settings, it is hard to accurately predict a stock price without considering public sentiments and policy decisions in the form of news articles, tweets, etc. Though this is common knowledge, the current state-of-the-art (SOTA) forecasting models fail to incorporate such contextual information, owing to its heterogeneity and multimodal nature. To address this, we introduce ContextFormer, a novel plug-and-play method to surgically integrate multimodal contextual information into existing pre-trained forecasting models. ContextFormer effectively distills forecast-specific information from rich multimodal contexts, including categorical, continuous, time-varying, and even textual information, to significantly enhance the performance of existing base forecasters. ContextFormer outperforms SOTA forecasting models by up to 30% on a range of real-world datasets spanning energy, traffic, environmental, and financial domains." }, "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": [ "Time series forecasting", "Contextual features", "Predictive 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/670e33ae98e00b23c0a071c911c512036a765b10.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": "Context Matters: Leveraging Contextual Features 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": 3 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See Above." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The writing in this paper is clear and easy to follow.\n- The paper advances MoE research by providing a fully open-sourced, state-of-the-art MoE architecture, which is beneficial for the research community.\n- The paper presents a thorough analysis of key design choices in MoE, offering valuable guidance on building high-performance MoE models.\n- The analysis is insightful, with discussions on phenomena such as router saturation and expert co-activation providing fresh perspectives and meaningful implications for the field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces OLMoE, a fully open, state-of-the-art language model built on a sparse Mixture-of-Experts (MoE) architecture. The authors conducted extensive experiments to validate the effectiveness of the proposed method, including evaluations after pre-training and adaptation phases. Additionally, they explored key design choices within the MoE framework, examining factors like expert granularity, routing strategies. Their analyses provided valuable insights into MoE, including router saturation, expert co-activation, and domain/vocabulary specialization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have a question regarding the experimental results: were the model parameters quoted directly from the original paper for the results shown in Table 2? For instance, in the original paper, OpenMOE’s activation parameter count is reported as 2.1B, whereas Table 2 shows an activation parameter count of 2.9B for OpenMOE. I recommend that the authors carefully verify the accuracy of these values." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "The work is well presented and possible suggestions for improvements are addressed in the future work section." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1) Strong empirical results with state-of-the-art performance for 1B active parameters.\n2) Good exploration of the MoE design space which forms a good guide for MoE model design.\n3) Novel analysis of routing behavior in MoE models during training and inference.\n4) This is the only MoE model where the model weights, code, data and checkpoints are openly available and thus the work is entirely reproducible." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a mixture-of-experts (MoE) LLM model called OLMoE that has 1B active parameters and 7B total parameters. The OLMoE model Pareto dominates many state-of-the-art models in the performance vs. active parameters space. The paper explores and presents insights on what is optimal in the design-space of MoE parameters and present analysis of routing behavior in MoEs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Other state-of-the art MoE models in related works are not exactly in the same parameter count configuration (1B/7B) so an exact comparison cannot be made to this model's performance.\n2) Most of the design choices and training choices are based on prior work and the novelty is more in the design space exploration and analysis of routing behavior." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 4 }, "primary_area": null, "questions": { "value": "1) What do you think about the necessity of expert parallelism? This model used dropless MoE, so it anyway will be unbalanced when using expert parallelism during training and inference. Without expert parallelism, it is still okay when the model is small. However, if we are aiming at a very large model, which has very large experts even if we are using the \"fine-grained MoE\", the expert parallelism would still be required? So how can we handle the token drop problem in this case?" }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1) There is no doubt that training MoE LLMs is challenging. This work offers a couple of important takeaways about how to train a good MoE LLMs, which is very helpful to the community.\n2) The presentation is very clear. For instance, the Table 1 delivers many key designs clearly at the early section of the paper.\n3) The model performance is good as well. As shown in Table 2 and 3, the model performs competitive with dense open models and partially open models (e.g. Qwen, Deepseek).\n4) The Analysis in Section 5 is informative, which greatly help readers and authors to understand how is the model working. This can also greatly speedup the growth of the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work is devoted to sharing the insights, data, and checkpoints of a series of MoE LLMs. The model achieved promising results on various benchmarks as a fully open model family." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Although the model has been relatively large, it is still much smaller than the SoTA MoE LLMs. I understand it is hard to get enough training resource for a fully open projects." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A state-of-the-art Mixture-of-Experts LLM with 1B active and 7B total parameters trained for 5T tokens that is 100% open-source" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024olmoe,\ntitle={{OLM}oE: Open Mixture-of-Experts Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xXTkbTBmqq},\nnote={under review}\n}" }, "abstract": { "value": "We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all available models with similar active parameters, even surpassing larger ones like Llama2-13B-Chat and DeepSeekMoE-16B. We present novel findings on MoE training, define and analyze new routing properties showing high specialization in our model, and open-source all our work: model weights, training data, code, and logs." }, "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", "mixture-of-experts", "open-source" ] }, "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/46823ebaa15350504c20f9133a77c76ccdca4f0b.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": "OLMoE: Open Mixture-of-Experts Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "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": "1. The method utilizes a conditional VAE that integrates information from both the adjacency matrix and attribute vectors during training, resulting in more precise graph generation.\n\n2. The scalability of the method is quite good. The method can be used to generate large-scale graphs, which is quite competitive compared to other auto-regression models. \n\n3. The paper is well-written and easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces CGRAPHGEN, a novel framework for controlled graph generation that allows for fine-grained control over graph topological properties. The authors propose a conditional variational autoencoder (VAE) that, unlike previous approaches, utilizes both the graph adjacency matrix and attribute vectors during training for improved decoder tuning and relies only on attributes during inference. This enables CGRAPHGEN to generate graphs that closely match the specified structural attributes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The baselines and the datasets are quite simple. The authors are recommended to compare with more recent graph conditional generation methods. e.g. [1] [2] [3]\n\n[1] Yang, Carl, et al. \"Conditional structure generation through graph variational generative adversarial nets.\" Advances in neural information processing systems 32 (2019).\n[2] Ommi, Yassaman, et al. \"Ccgg: A deep autoregressive model for class-conditional graph generation.\" Companion Proceedings of the Web Conference 2022. 2022.\n[3] Mo, Zhanfeng, Tianze Luo, and Sinno Jialin Pan. \"Graph principal flow network for conditional graph generation.\" Proceedings of the ACM on Web Conference 2024. 2024.\n\n2. it is unclear how the hyper-parameters are defined. In Figure 5, the performance seems quite stable for different gamma, e.g. there's a drop when gamma = 0.8 on arxiv dataset. \"When γ increases and more information is drawn from the prior pθ, the generation error increases.\" is not always true.\n\n3. No theoretical analysis of how the proposed method can reduce the generation error better than other baseline 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": 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 d(Z_c) in Eq. (6)? There is no explanation for this notation." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. mixing the attributes and graph representation in latent space for VAE is somehow new for controlled generation. \n2. the results for controlled graph generation is seemingly good regarding attribute alignment." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on controlled graph generation that generates graphs satisfying specific topological attributes. It introduces a new scheduling technique, MIXTURE-SCHEDULER, to combines desired attributes with adjacency matrix representations during training for precise graph generation, and it then uses only attributes during inference. Experiments demonstrate that generated graphs have better aligned attributes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The biggest concern is the paper lacks of a rigorous deduction for the VAE model and learning objective. For most VAEs, we generally start from the miminization of log likelihood and use variational inference to factorize it. However, the formulations in this paper are very heuristic. We do not know whether the mixing of attributes and graph representation is valid. Mixing the prior with posterior looks also weird to me. What I expect should be starting something like $P(G|c) = \\int_{Z_G, Z_c} P(G|Z_G, Z_c, c)P(Z_G|\\theta, c) P(Z_c|c) dZ_G dZ_c$.\n2. It seems the graph encoder/decoder can only deal with adjacency matrix, but how about graphs with node features?\n3. The evaluation only measures the attributes, but the validness of the graph in many domains is also important (e.g. for molecules)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "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. This paper proposes a new setting for the controlled graph generation task, which is highlighted by the injection of fine-grained topological control.\n2. The proposed method seems technically sound to me." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a conditional variational autoencoder for graph generation with fine-grained topological control. The proposed model incorporates a scheduling technique to integrate representations from both the adjacency matrix and attribute distribution to enable fine-grained control." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The number of baseline models compared in the experiments appears to be limited.\n2. I'm not sure if it's reasonable to use only the MAD metric to evaluate the generation results based on various topological attributes." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Is there any code for the proposed model?\n- Have you tried other more complex neural network architecture?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The model offers flexibility in controlling multiple structural properties (e.g., graph density, connectivity, clustering coefficient), enabling accurate graph generation across various domains.\n- The mixture-scheduler seems to be novel and it smoothly integrates prior and posterior distributions, improving the quality and stability of generated graphs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes CGRAPHGEN, a novel conditional variational autoencoder framework for generating graphs with fine-grained control over topological attributes. The framework introduces a MIXTURE-SCHEDULER, a scheduling technique to combine structural and attribute-based latent representations. Experiments on multiple datasets show that CGRAPHGEN outperforms baseline models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed model doesn't seem to be much of an improvement compared to GraphVAE-like models. The condition architecture is very common in generative models, and feature/attribute based conditional graph generation seems to be a common trick in most methods. Therefore, I think the proposed model may lack enough novelty.\n- Lack of baselines. I have noticed this paper include the diffusion-based model (EDGE), why not other SOTA graph generative models like DruM, DIGress and so on. For graph generation, I think it is more convincing to compare these models or at least other vae-based models. As far as I know, I believe these models can also incorporate the attribute feature to achieve conditional graph generation.\n- For the mixture-scheduler part, I don't really understand the meaning of regarding the time $t$ as epoch in the training stage. From Figure 4(b), it seems there is no clear effect on whatever the $\\beta(t)$ is.\n- In your ablation, I find the experiments with masked only one attribute, is there any flexibility attributes choice?\n\n[1] Efficient and degree-guided graph generation via discrete diffusion modeling. \n[2] Graph generation with diffusion mixture.\n[3] Discrete denoising diffusion for graph generation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A {VARIATIONAL} {FRAMEWORK} {FOR} {GRAPH} {GENERATION} {WITH} {FINE}-{GRAINED} {TOPOLOGICAL} {CONTROL}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xYquBPHppn},\nnote={under review}\n}" }, "abstract": { "value": "Controlled graph generation is the process of generating graphs that satisfy specific topological properties (or attributes). Fine-grained control over graph properties allows for customizing generated graphs to precise specifications, which is essential for understanding and modeling complex networks. Existing approaches can only satisfy a few topological properties such as number of nodes or edges in output graphs. This paper introduces CGRAPHGEN, a novel conditional variational autoencoder that, unlike existing approaches, uses graph adjacency matrix during training, along with the desired graph properties, for improved decoder tuning and precise graph generation, while relying only on attributes during inference. In addition, CGRAPHGEN implements an effective scheduling technique to integrate representations from both adjacency matrix and attribute distributions for precise control. Experiments on five real-world datasets show the efficacy of CGRAPHGEN compared to baselines, which we attribute to its use of adjacency matrix during training and effective integration of representations, which aligns graphs and their attributes in the latent space effectively and results in better control." }, "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": [ "Controlled Graph Generation" ] }, "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/73a1124bb7f1385fb0c3a2a12f0bddcf2738a898.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": "A VARIATIONAL FRAMEWORK FOR GRAPH GENERATION WITH FINE-GRAINED TOPOLOGICAL 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": 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": 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": "Clarification on Novelty: Needs a clearer explanation of differences from previous work, specifically the distinct contributions beyond those in “Grounded Video Description” (Zhou et al., 2019)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Novel Task, Dataset, and Model: The paper introduces a new task with a manually annotated evaluation set and a large-scale automatically generated dataset.\nDataset build up: Uses LLMs to generate consistent video-level captions from frame-level annotations, improving temporal consistency.\nModel that support vlm to output bbox for video: VideoGLaMM incorporates spatio-temporal adapters and temporal objectness, supporting consistent tracking of objects in video." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a new task (GROC), a manually annotated test dataset, and an automatically generated dataset. It introduces VideoGLaMM, a model designed to generate captions and track object bounding boxes over time. Key features include spatio-temporal adapters, a bounding box decoder, and a temporal objectness head, all aimed at bridging multimodal understanding for video and language." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Limited Model Comparison: It would be better for VideoGLaMM to include comparisons against additional models like InternVL and Qwen-VL, which also produce bounding boxes in text format, to strengthen its evaluation. Since the paper aims to establish a new task, a more comprehensive benchmark would enhance the reliability of its proposed contributions.\n\nDiscussion Scope: Additional discussion is needed on alternative methods for VLM bounding box outputs, specifically addressing why GLaMM was prioritized over other methods like [1,2]. This would help clarify its selection rationale within the broader context of existing approaches.\n\n[1] Bannur, Shruthi, et al. \"MAIRA-2: Grounded Radiology Report Generation.\" arXiv preprint arXiv:2406.04449 (2024).\n[2] Zou, Ke, et al. \"MedRG: Medical Report Grounding with Multi-modal Large Language Model.\" arXiv preprint arXiv:2404.06798 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "1. The GROC evaluation dataset is labeled by human annotators, and the authors only mention their protocols for human annotators in Appendix F. The authors may need to disclose more details regarding the labeling template, the platform they use, the salary of annotators, and other factors to ensure that the labeling process is a fully responsible practice.\n\n2. The collected and planned-to-be-released training dataset is generated fully automatically with the help of LLMs. Safety and bias issues need to be addressed during the generation process, which are ignored in this paper." }, "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": "Most of my major concerns can be seen in the weakness section. Here I leave a few more:\n1. Line 229-230 Sec. ?? -> Reference is missing.\n2. The baselines are insufficient. One simple strategy is to directly feed the frames with grounding boxes into a video LLM. This is a widely used method called visual prompting, which can be tested in this training dataset without any additional structure or model tuning.\n3. The ablation study is not convincing to me. As shown in Figure 3, I think most of the audience would like to know whether including grounded information really helps with video captioning; that is, a comparison of removing and keeping the grounding video encoder." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The released training and evaluation datasets could be useful in the field of video captioning after addressing the necessary ethics concerns." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This submission gives a task, dataset, and model for grounded video caption generation. The main contributions are:\n\n* They define the task as GROunded Video Caption Generation (GROC) and create a manually annotated test dataset specifically for this task.\n \n* They design an automatic annotation pipeline that uses an existing model for grounding and a LLM for frame-level and further video-level captioning. They use this approach to the randomly selected subset of HowTo100M, and generate the final HowToGround training dataset.\n\n* They propose a VideoGLaMM model, trained on the collected HowTo100M dataset, which achieves best performance on grounded video caption generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I would say the most important weakness of this paper is **over-claim, fails to discuss and compare previous published research and show no respect to previous work**. Let's start with the beginning:\n\n1. The authors claim that they propose a new task called grounded video caption generation (GROC). However, I don’t think so at all. Please refer to paper [1] titled \"Grounded Video Description.\" I think that, in vision-and-language research, most researchers use the terms \"video description\" and \"video captioning\" interchangeably (the slight difference may be that \"video description\" sometimes includes more details). Paper [1] is the first to collect grounded video-text datasets and propose a model that uses grounding information to generate better video descriptions. They didn’t even claim that this task is a so-called new task because it is viewed as using alternative guidance to improve video description generation, not as a new task. One quick and intuitive way to compare these two is to check Figure 1 in this submission and Figure 2 in paper [1]. Thus, **I completely reject the claim that this task is newly proposed by the authors of this submission**.\n\n2. The more interesting thing is that the authors actually cited paper [1] but **did NOT mention it at all in the main pages**. Instead, they \"secretly\" placed it in the Appendix. In my opinion, a paper with so much overlap should be carefully discussed in the related work section, in section 3 where the authors claim their novelty, and also at the end of section 5, where the authors describe how they construct their datasets.\n\n3. As for the experiments, the proposed model is also lacking of novelty. Put more relevant papers here, like [2, 3, 4], which are also not properly cited and discussed in this submission. Thus, **I also clearly reject the claim in line 055-056 that \"At the same time, producing natural language video descriptions where the described objects are spatio-temporally grounded with bounding boxes in videos has received much less attention\"**. From my understanding, the authors replace the previsous LSTM networks with large language models. For sure, LLMs can get better results than LSTMs. However, these methods and models are not compared in their experiments. I don't believe that, in the era of LLMs, simply replacing the language module in previous methods with an LLM can be considered a big innovation in the overall model structure.\n\nIn summary, I clearly think this submission needs major revisions and a complete reframing in order to meet the publication standards of a conference like ICLR.\n\nReferences:\n\n[1] Grounded Video Description, Zhou et al. CVPR 2019.\n\n[2] Learning to Generate Grounded Visual Captions without Localization Supervision, Ma et al. ECCV 2020.\n\n[3] Attend and Interact: Higher-Order Object Interactions for Video Understanding, Ma et al, CVPR 2018.\n\n[4] Comprehensive Visual Grounding for Video Description, Jiang et al. AAAI 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": 4 }, "primary_area": null, "questions": { "value": "* L140 mentions \"video selection where ‘interesting’ videos are selected.\" What criteria define a video as \"interesting\"? Please provide details.\n* L262 mentions \"After rejecting videos for which our proposed automatic annotation method failed…\" How were these videos rejected? Was it done manually or through an automatic process? More details are needed.\n* Although the HowToGround dataset is generated from aggregated image-level annotations, this approach may struggle with capturing interactions that require understanding the temporal dimension, such as determining whether someone is walking forwards or backwards. Is this a limitation of your dataset and the resulting model? Quantitative evaluation of actions requiring temporal interactions could strengthen the paper.\n* How does the VideoGLaMM generalize to other datasets, apart from the HowToCaption dataset?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The paper tackles the important task of grounded video captioning, which has numerous applications in downstream tasks such as video retrieval and open-world object tracking in videos.\n* The writing is clear.\n* The paper introduces the HowToGround dataset for grounded video captioning and provides a high-quality, manually annotated test set for the task.\n* The paper proposes an architecture for grounded video captioning, building upon a pre-trained grounded image captioning model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the task of grounded video captioning, where, given a video, a model must generate a caption and provide bounding boxes for the objects appearing in the caption. The authors introduce two datasets for this task: a small, manually annotated test set, and a larger, automatically annotated training set. Finally, they train a grounded video captioning model on this dataset and evaluate it against two baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The first paragraph of the introduction feels more like a related work section, making it hard to follow. I would suggest starting by clearly stating the task you aim to solve and providing a high-level summary of the existing work.\n* The method for automatically creating the dataset uses SVOs, which can lead to important details being omitted from captions, such as the color of the shirt in Figure 2. This omission could hinder downstream tasks. For instance, a model aiming to find videos of orange shirts might miss this information, which is otherwise captured in current captioning models like GlaMM.\n* The paper’s main contribution is the creation of the HowToGround dataset. However, the choices made during dataset creation are not justified, nor does the paper include any ablation study of the creation steps. Including such a study, showing alternative choices for each step, would greatly strengthen the paper. For example, how about using SAM-2 or an open-world object detector to identify object locations? How would VideoGLaMM perform if step 3 were omitted from the dataset creation process? How well does the resulting VideoGLaMM for different train set sizes?\n\nTypos/Errors:\n* L230 - Section reference is broken." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "See Weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- An automated grounded video annotation pipeline was designed, significantly reducing annotation costs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new task: grounded video caption generation, which aims to generate captions for videos while also providing bounding boxes for the objects mentioned in the captions. \nTo achieve this, the authors designed an automatic grounded video annotation pipeline. Based on the dataset constructed using this pipeline, the paper trains a model called VideoGLaMM, which performs well on the grounded video caption generation task." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The authors claim to introduced a new task: grounded video caption generation. However, there is prior research on this task, such as PG-VIDEO-LLAVA[1]. How do the authors justify their assertion of novelty?\n\n[1] PG-Video-LLaVA: Pixel Grounding Large Video-Language Models.\n\n- For videos that contain two or more events, how should the automated annotation pipeline be applied? For example, consider a scenario where a man picks up a knife to chop vegetables and then puts the vegetables into a pot. If the knife and pot appear simultaneously in the frame but the action hasn't progressed to the pot yet, how can we avoid generating a bounding box for the pot prematurely?\n\n- The case studies (e.g., Figure 4) presented by the authors seem overly simplistic. Can VideoGLaMM be effectively applied to videos that involve multiple events?\n\n- There are very few comparative methods included in the study. The authors should compare more methods, such as the integration of video captioning models with object detection models, and PG-Video-LLaVA." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper proposes an interesting and reasonable task of generating captions for videos while simultaneously grounding the corresponding objects.\n\nThe paper presents a technically sound framework to achieve GROC." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a task called grounded video caption generation. It presents a manually annotated benchmark comprising 1,100 video clips, as well as a 50k training set created using an automatic annotation pipeline. GLaMM is adapted with video encoders to achieve GROC. Similar to GLaMM, the authors evaluate VideoGLaMM using a collection of video captioning and object detection metrics. The evaluation results indicate that the proposed method outperforms comparable approaches across various downstream tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am primarily concerned with the automatic data generation pipeline. In this process, the authors used GLaMM to generate frame-wise captions and then converted these into Subject-Verb-Object (SVO) triplets. Since there is not alignment between generated SVOs, if there are two different instances with the same label in different frames, it is possible that the LLM might treat them as a single instance, potentially leading to inaccurate conclusions and relationships between them.\n\nSimilarly, in the tracking by language process, instances are tracking with only textual description of the phrases. This cannot guarantee 2 different instances with same label are mistakenly treated as the single instances. For example, imagining a video that there are 2 different people appearing in different frames but all drinking beers, the process is likely to link them as one person by just considering its textual description. The strategy described in Supp. is not very convincing as there can be 2 different people both with white shirts and drinking.\n\nThe process appears to capture only a portion of the video content, potentially overlooking valuable concurrent events within the video clip.\n\nI am also curious about the visual encoders used in VideoGLaMM. Is there any specific reason or necessity for using two different image backbones separately for captioning and bounding box generation?\n\nthe reference in Line 230 is missing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Did the automatic annotation process take into account the ambiguities mentioned in the weaknesses section? Are there any methods in place to reject annotations for data that the model finds ambiguous?\n\n2. Why was SAM2 not considered in the model's structural design? As far as I know, this model targets video and seems more suitable for your task.\n\n3. I hope to see a comparative analysis between the current vLLM and video grounding models in your task, including their differences in performance, efficiency, and application scenarios. This would help in better understanding the strengths and weaknesses of each model and provide guidance for future research directions." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Propose a new task: grounded video caption generation, which is significant as large-scale spatio-temporal grounding of natural language is a key step in advancing fields such as human-robot interaction and embodied perception.\n2. Propose a manually verified test dataset, which is likely to further drive progress in this field.\n3. The model architecture is both innovative and intuitive, with subsequent ablation studies demonstrating the effectiveness of these design choices." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel task—grounded video caption generation—along with a newly annotated dataset, GROC. The authors introduce a language-based tracking method to create pseudo-annotations for a large-scale instructional video dataset, named, HowToGround, which is then used to train a newly designed model. Finally, they compare the model's performance against still-image grounding baselines, reporting state-of-the-art results in grounded video caption generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There are concerns regarding the technical soundness of the AUTOMATIC ANNOTATION METHOD section. I understand that the grounded video caption generation task requires high precision; however, due to the richness and ambiguity of language, I am skeptical about the consistency of the language-based tracking method in object localization. For example, while the case of the woman drinking a beverage in the authors' paper is unambiguous, many instances will likely present significant ambiguity. If, in the third frame of Figure 2, the image caption still states 'holding a glass,' how can one determine whether 'beverage' and 'glass' refer to the same object?\n\n2. The writing quality is poor, containing a lot of redundancy, and some phrases read awkwardly and lack fluency. For instance, the phrase 'Differently than the image-based task, where bounding boxes...' could be improved. Additionally, some sentence structures need adjustment, and paragraph transitions require attention, particularly when describing the GROC construction process. There are also typos to address, such as in Section ?? on page 5. Overall, the writing and illustrations need further polishing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nkazakos2024grounded,\ntitle={Grounded Video Caption Generation},\nauthor={Evangelos Kazakos and Cordelia Schmid and Josef Sivic},\nyear={2024},\nurl={https://openreview.net/forum?id=xYzOkOGD96}\n}" }, "abstract": { "value": "We propose a new task, dataset and model for grounded video caption generation. This task unifies captioning and object grounding in video, where the objects in the caption are grounded in the video via temporally consistent bounding boxes. We introduce the following contributions. First, we present a task definition and a manually annotated test dataset for this task, referred to as GROunded Video Caption Generation (GROC). Second, we introduce a large-scale automatic annotation method leveraging an existing model for grounded still image captioning together with an LLM for summarising frame-level captions into temporally consistent captions in video. \nFurthermore, we prompt the LLM to track by language – classifying noun phrases from the frame-level captions into noun phrases of the video-level generated caption. We apply this approach to videos from the HowTo100M dataset, which results in a new large-scale training dataset, called HowToGround, with automatically annotated captions and spatio-temporally consistent bounding boxes with coherent natural language labels. Third, we introduce a new grounded video caption generation model, called VideoGLaMM, and train the model on the new automatically annotated HowToGround dataset. Finally, results of our VideoGLaMM model set the state of the art for the new task of grounded video caption generation. We perform extensive ablations and demonstrate the importance of key technical contributions of our model." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Evangelos_Kazakos2", "~Cordelia_Schmid1", "~Josef_Sivic1" ] }, "authors": { "value": [ "Evangelos Kazakos", "Cordelia Schmid", "Josef Sivic" ] }, "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": [ "vision-language models", "VLM", "LLM", "video grounding", "automatic annotation", "pseudo-labeling" ] }, "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": "kazakos|grounded_video_caption_generation" }, "pdf": { "value": "/pdf/7c363c0c76057bc0f14cf8ad8d057a1f6e8bc84b.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": "Grounded Video Caption Generation" }, "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": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The method is tested only in fine-tuning regime. Is there any reason for that? \nTesting the proposed models trained from scratch would strengthen the empirical significance of the proposed method. I am not sure which datasets would be suitable, but datasets like ImageNet and its variants (ImageNet-A [1], and ImageNet-C[2]), and scientific machine learning benchmarks might be good datasets as uncertainty prediction would be critical. \n\n[1] https://arxiv.org/abs/1907.07174\n[2] https://arxiv.org/abs/1903.12261" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The work extends generalization bounds from finite-dimensional parameter spaces to functional spaces, which leads to FHBI, a theoretically grounded Bayesian inference algorithm.\n- The performance improvement by the proposed method is validated through extensive comparisons with previous works." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents an algorithm called Flat Hilbert Bayesian Inference (FHBI), which incorporates Sharpness-aware minimization (SAM) technique in Bayesian inference. Specifically, together with SAM, the authors perform Stein variational gradient descent (SVGD) as the dynamics of the model parameters, which leads the particles (models) to flat and diverse modes. FHBI is tested on VTAB-1K, a collection of various classification tasks, and achieves better performance on average among different Bayesian NN methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I don't see any particular weaknesses in this 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": "Have you compared SVGD+SAM to SGLD+SAM [1, 2]? \n\n\n[1] Van-Anh Nguyen, Tung-Long Vuong, Hoang Phan, Thanh-Toan Do, Dinh Phung, and Trung Le. Flat seeking bayesian neural networks. Advances in Neural Information Processing Systems, 2023.\n\n[2] Yang, Xiulong, Qing Su, and Shihao Ji. \"Towards Bridging the Performance Gaps of Joint Energy-based Models.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Incorporating SAM into SVGD can somehow improve the stability. The idea generally makes sense. \n2. The paper is clearly written overall, and it is easy to capture the motivation.\n3. The empirical results are superior and tested on various datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper combines sharpness aware minimization (SAM) with SVGD in RKHS as the Flat Hilbert Bayesian Inference (FHBI). It extends the proof of [1] to infinite-dimensional functional space.\nEmpirical validations show better performance of FHBI compared to previous Bayesian inference approaches on various datasets.\n\n[1] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Even though I agree that the proposed method may be effective in some conditions, I am not convinced by the theory in this paper. Following are my concerns. \n1. Notations. \n * Empirical posterior should be $\\mathbb{P}\\_{\\theta|S}$ not $\\mathbb{P}\\_S$, which is the data distribution. And the prior distribution should be $\\mathbb{P}_{\\theta}$.\n * $p(\\theta|S)\\propto p(\\theta)p(S|\\theta) = p(\\theta)\\prod_{i=1}^n p(y_i, x_i|\\theta)$\n The exponential average loss should be related to a specific distribution. I do not know how you can obtain this form directly. \n * What is \"general loss\"? I have never heard this terminology. $\\mathcal{L}_{\\mathcal{D}}(\\theta)$ is often called population loss/ true error/ generalization error in different papers or learning theory books. \n2. For Theorem 2, what is the exact definition of $h(1/\\rho^2)$? It should be clearly presented in the main paper. In your derivation, $\\mathcal{O}(\\rho^2)$ is simply ignored, which means $\\rho$ should be very small, which in turn gives a large $h(1/\\rho^2)$.\n3. You are actually deriving an upper bound of the true risk, which is the empirical risk plus some complexity term. However, the sample complexity is not directly reflected in the bounds presented in the main paper. You claim that you can approximate the true posterior $p(\\theta|\\mathcal{D})$. This is impossible if you only have limited samples $n$. \n\n4. Experiments\n\t* How does data augmentation affect the empirical results? Have you used it in all baselines or just your method? \n\t* As shown in Figure 3, the runtime of the FHBI increases fast w.r.t the number of particles. Compared to SVGD, the margin also grows with the number of particles. Where is the computation bottleneck? Can you find any way to reduce it? \n\n### Minor: \n1. Is the template used correctly? There is no line number. \n2. Typo. In the related work section, sharness -> sharpness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- In my understanding, Theorem1 expresses the tradeoff relation between (1) worst-case empirical loss that scales with the size of the neighborhood (radius of the step size) (2) wellness of the upperbound approximation based on the empirical loss. The reviewer is making in his mind a super-rough analogy of \"local linear approximation of a complex function\", whose worst case error scales with the size of the neighborhood, and \"functional ascending step\" is analogous to choosing the direction of worst error in this approximation. \n\nIf this analogy is correct, the neighborhood radius $\\rho$ as well as the update size $\\epsilon$ would be roughly analogous to a stepsize in Euler-Maruyama simulation of the ODE, except that in the context of FBHI the system to be simulated is infinite dimensional. Would the method improve its performance by choosing $\\rho$ and $\\epsilon$ to be small, and running the iterations for greater number of times, as in ODE simulation? Also, with this intuition it feels as if $\\rho$ shall be similar to $\\epsilon$; why are they chosen quite differently in the experimental section? \n\nWhile the reviewer is not too confident of the intuition based on this super-rough analogy, the reviewer also believes that it is crucial that the paper provides some explanation (either numerical or experimental, or at the very least, the heuristic with intuitive explanation) regarding the choice of $\\rho$ and $\\epsilon$, both for the sake of the future practical user of FBHI and the for the sake of the future schemes that will possibly branch out from FBHI.\n\n- In the similar note as above, because the \"wellness\" of RKHS method generally depends on the affinity between the choice of the kernel (and the hence the nature of the continuity of functions in the space) and the dataset. In the case of this research, the choice, I believe, is indirectly related to dataset because RKHS is a space of functions on \"parameters\". Is there any ablation study regarding this choice, or at the very least, a good heuristics that would help the user to choose appropriate Kernel in the applications? \n\n- It is explained in the paper that $\\rho=0$ would correspond to SVGD, and \"in equation\" it seems so. However, because the paper emphasizes its connection to SVGD, the reviewer wishes to see why this comes about in connection to Theorem 1 and the derivation of SVGD. \n\n- Is there comparisons against SAM in the setting of section 6.2?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This research presents a scheme that generalizes both SVGD and SAM in the context of Bayesian Inference. \nThe idea of FBHI itself is very clearly stated and convincing, and its efficacy are validated with ample experiments. \nThe layout of the derivation is very instructive as well. \nThe claimed sharpness advantage (that FBHI would is more sharpness-aware) is also validated in experiments, empirically \nshowcasing the mechanism behind the advantage of the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Flat Hilbert Bayesian Inference(FHBI), a method to compute a posterior inference by integrating along the flow of distributions that converges to the target posterior. By formulating the derivative of the flow with the pushforward function that lives in a vector valued RKHS, the framework naturally establishes the \"flow of infinite number of particles\" which allows one to target the \"general posterior\" as opposed to \"empirical\" posterior in the setting of Bayesian Inference. Importantly, the method differs from Stein Variational Gradient Descent in that it derives the infinitesimal pushforward function that reduces the \"worst upper bound\" of the KL divergence in a way that resembles adversarial training.\n\nThe propsoed paradigm is made implementable by the computable form of $\\lambda p(\\theta | S)$, together with the upper bound of the KL divergence to general posterior defined with the KL divergence to empirical posterior. \nThe Efficacy of the algorithm is verified through experiments and its scaling properties are investigated and compared with respect to SVGD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the relation of FHBI to SAM and the interpretation with spatial and angular repulsive force is insightful, the reviewer was a little confused in the introduction and in the multiple reference to SAM that sounded as if SAM would be the major idea on which to develop FHBI (which, in retrospect, is not \"directly\" so?). Meanwhile, it is true in the algorithm that when m=1, the algorithm agrees with SAM in the end. In the current presentation the connection to SAM seems something a posteriori.\n\nIf indeed the sharpness-aware philosophy is indeed the \"motivation\" of FHBI (which is, unfortunately, not yet clearly conveyed to this reviewer), the reviewer would like to see more analytical connection to SAM's derivation. \n\n- In the similar note, as the reviewer will post in the \"Questions\" section, the reviewer feels that the supposedly the important connections to SVGD and SAM are not clearly explained through objective function and derivations (**more than $\\rho=0$ and $m=1$ )" }, "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": 3 }, "primary_area": null, "questions": { "value": "Major:\n* Something seems amiss in going from the empirical posterior $p(\\theta|\\mathcal S)$ to what you call the general posterior $p(\\theta|\\mathcal D)$. Let $L_n(\\theta) = \\frac{1}{n} \\sum_{i=1}^n -\\log p(y_i|x_i,\\theta)$. Then the empirical posterior is $p(\\theta|\\mathcal S) = \\exp(-n L_n(\\theta)) p(\\theta)$. The population counterpart to this is $\\exp(-n L(\\theta)) p(\\theta)$. But this is *not* your general posterior which no longer has any dependence on sample size $n$. \n\nMinor:\n* Currently, the way the loss $\\ell$ shows up in Section 3 might give the wrong impression that it is something one can freely choose. In fact the loss has to be (proportional) to the negative log likelihood. I think you need to write out explicitly what $\\ell(f_\\theta(x),y)$ is. \n* The sentence above Eqn (5), \"In turn, the solution $\\hat f^*$ that solves the maximization problem above is given by\". Which maximization problem above? There are so many approximations here, please use \\label and \\ref to refer to exact maximization problem. (I also doubt this is a true statement, (5) is not really a solution but an approximation of a solution right?)\n* Would you consider labeling the iterative procedure right below Lemma 1 and making it painfully clear how that turns into Algorithm 1?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Originality: \nI'm uncertain about the level of originality here. Variational families within an RKHS framework have, to my knowledge, been explored previously. \n* Quality: The work appears to be of high quality. The results seem technically correct, and while I haven’t meticulously checked every mathematical detail, the derivations appear consistent with expectations.\n* Clarity: Overall, the paper is well-written. I appreciated the clear, step-by-step walkthrough of the optimization problem and the detailed exposition of various bounds and relaxations necessary to implement the proposed approach.\n* Significance: The proposed method, Flat Hilbert Bayesian Inference (FHBI), is presented as a generalization of Stein Variational Gradient Descent (SVGD) and Sharpness-Aware Minimization (SAM). The experiments demonstrate promising potential for FHBI in applications like LoRA-style fine-tuning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In traditional variational inference, the goal is to approximate an intractable posterior \\( p(\\theta | \\mathcal{S}) \\) by selecting a variational distribution \\( q \\) from a family \\( \\mathcal{Q} \\) that minimizes a divergence, often the KL divergence, resulting in the optimization problem:\n\n$$\n\\arg \\min_{q \\in \\mathcal{Q}} \\, \\text{KL}(q(\\theta) \\parallel p(\\theta | \\mathcal{S})),\n$$\n\nwhere \\( p(\\theta | \\mathcal{S}) \\), termed the \"empirical posterior\" by the authors, is the target distribution.\n\nThis paper introduces a variational family within a Reproducing Kernel Hilbert Space (RKHS) framework and further reformulates the optimization problem to focus on approximating the \"general posterior,\" \\( p(\\theta | \\mathcal{D}) \\), over the dataset \\( \\mathcal{D} \\). The optimization problem thus becomes:\n\n$$\n\\arg \\min_{f \\in \\mathcal{H}^d, \\, \\|f\\| \\leq \\epsilon} \\, \\text{KL}(q_{[I + f]}(\\theta) \\parallel p(\\theta | \\mathcal{D})),\n$$\n\nwhere \\( q_{[I + f]}(\\theta) \\) represents the transformed variational distribution. This reformulation leads to multiple steps of approximation, relaxation, and bounding, culminating in an iterative optimization procedure (detailed below Lemma 1).\n\nThe proposed method, Flat Hilbert Bayesian Inference (FHBI), is designed to enhance generalization in Bayesian inference by leveraging the structure of RKHS." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The general posterior defined, $p(\\theta|\\mathcal D)$, does not appear to me to be the correct theoretical counterpart to the \"empirical\" posterior. See more in questions below. \n* If the paper’s originality hinges partly on targeting the \"general posterior\" $p(\\theta|\\mathcal D)$, I have reservations about its practical benefits. In addition to my concerns about the missing sample size term, targeting this posterior seems unlikely to yield practical advantages and might even be counterproductive. The authors expend considerable effort introducing approximations to recast the optimization problem in terms of the \"empirical posterior,\" which requires potentially loose bounds. The value of this detour would be enhanced by a more thorough discussion of why these approximations are justified or necessary.\n* The paper emphasizes \"improving generalization\" as a primary benefit of FHBI, yet this claim seems tenuous without more foundational support. It's unclear what is meant by \"enhancing generalization in Bayesian inference\" mathematically, as the method doesn’t inherently introduce any features that theoretically boost generalization. Rather, it appears that FHBI, when applied to fine-tuning, yielded improved generalization performance on a benchmark dataset compared to baseline methods. This distinction between observed outcomes and the initial methodological intent would be clearer if the authors clarified that FHBI’s generalization performance was more an empirical finding than a theoretically driven design choice." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024improving,\ntitle={Improving Generalization with Flat Hilbert Bayesian Inference},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xZ2lTzfyFv},\nnote={under review}\n}" }, "abstract": { "value": "We introduce Flat Hilbert Bayesian Inference (FHBI), an algorithm designed to enhance generalization in Bayesian inference. Our approach involves an iterative two-step procedure with an adversarial functional perturbation step and a functional descent step within the reproducing kernel Hilbert spaces. This methodology is supported by a theoretical analysis that extends previous findings on generalization ability from finite-dimensional Euclidean spaces to infinite-dimensional functional spaces. To evaluate the effectiveness of FHBI, we conduct comprehensive comparisons against seven baseline methods on the VTAB-1K benchmark, which encompasses 19 diverse datasets across various domains with diverse semantics. Empirical results demonstrate that FHBI consistently outperforms the baselines by notable margins, highlighting its practical efficacy. Our code is available at \\url{https://anonymous.4open.science/r/Flat-Hilbert-Variational-Inference-008F/}." }, "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": [ "Bayesian Inference", "Sharpness-aware Minimization" ] }, "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/07ef79c33f20124245927a1a00e4c805b551a813.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": "Improving Generalization with Flat Hilbert Bayesian Inference" }, "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. In Fig 1, if a particular hidden state has already been marked for pruning in the previous layer, then in the current layer, the counter should simply increment by 1, and this hidden state should not be pruned again. However, in Fig 2, for layer 2, why is S_{768} still selected for pruning?\n2. In Equation 6, why does it use the number of layer i? Why does it assign greater weights to the deeper layers?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ It targets BERT optimization, which is important.\n+ It proposes a novel metric ICCR, which provides a new way of evaluating the model efficiency.\n+ The flow charts (Fig 1 and 5) and examples (Fig 2) can help understand the paper.\n+ It provides the setting of hyper-parameters and is open-sourced." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents P-BERT which is an optimization of the standard BERT model. It is developed through the integration of three model compression techniques: pruning, quantization, and knowledge distillation. It also introduces a novel metric, the Inverted Computational Complexity Ratio (ICCR), to better capture model efficiency and complexity. Experimental results demonstrate that P-BERT achieves a reduction in computational complexity of at least 60% while maintaining comparable accuracy to the baseline BERT across several natural language processing tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- P-BERT performs well in CoLA, while its performances on other tasks are less competitive.\n- The genetic algorithm is time-consuming and may get stuck in a local optimum. Therefore, giving enough reasons and motivations to choose the generic algorithm would be better.\n- Knowledge distillation can guarantee the accuracy of the model but may be time-consuming.\n- It would be better to provide more evaluations to show that the ICCR metric fits well for BERT optimization." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Is the proposed methodology of using the genetic algorithm to prune and quantize applicable to other models?\n\nIs the computational complexity metric in Eq 6 able to capture the significance of the deep layers? Could you provide some details like layerwise precision and % of parameters pruned of one of the resultant models?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed technique is a generalized methodology to explore pruning and quantization strategies for a given model and task. The Pareto front and the loss curve plots are quite insightful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper combines pruning, quantization and knowledge distillation to reduce the computational complexity of the BERT models. A genetic algorithm is used to hone down on the parts of the layers to prune and the appropriate precision of the layers for quantization. Knowledge distillation is later used to transfer knowledge from the baseline BERT model. They also use Inverted Computational Complexity Ratio (ICCR) as a metric to evaluate model compression factor. Results section shows the trend of decreasing estimated inference time with increasing ICCR for four different benchmarks. There are also comparisons of the performance of P-BERT’s estimated inference time and accuracy against other BERT models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The results presented in the paper fail to emphasize the utility of this approach. Table 5 presents other models with higher complexity ratios and lower inference time than what’s best achieved by P-BERT. Even for CoLA where P-BERT achieves the best accuracy among tuned models, the compression factor doesn’t translate to reduction in inference time.\n\nIf the chosen experimental hardware setup isn’t able to leverage the pruning and quantization benefits, then a different evaluation platform or metric could be selected." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 MSE is chosen over MAE in Eq. 3 and 4? \nWhy do we need to consider L_hard, L_soft, L_hidn, and L_attn in Eq. 5? What's the significance of each? Any ablation study to compare which ones are more important?\nFigure 1, flow chart, why the sum of absolute? Any intuition or reference?\nFor the definition of K_model in Eq. 6, why do we need to include the layer index \"i\"? Does this mean that the last few layers carry more weights?\nWhy is the particular genetic algorithm chosen in Sect. 4.4? \nWhat challenges did you run into in Sect. 5.4 when comparing against those models?\nWould you like to consider palettization as an additional compression technique?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The use of a genetic algorithm to select compressed models based on the evaluation metric and the newly proposed inverted computation complexity ratio (ICCR) is seldom studied. It is an encouraging direction to explore.\nThe overall flow of the paper is easy to follow.\nThe comparison of the proposed P-BERT with other optimized BERT model on the selected evaluation tasks is well summarized." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores the combination of weights pruning, bit quantization, and teacher-student learning via knowledge distillation in compressing BERT models. They used a genetic algorithm to determine the cluster of weights to be pruned and quantized. They proposed a new metric, the inverted compuational complexity ratio, which is defined roughly as the ratio of computation required bewteen a standard BERT and a compressed model, to capture the extent of compression. They evaluated the trade-off bewteen compression and accuracy on several Hugging Face classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main compression techniques employed by the team are all common model compression techniques: pruning, quatization, and KD. Only a new genetic algorithm to search for prunable parameters do not seem sufficiently novel to me. The authors could conduct ablation studies to compare each of the three techniques in the P-BERT framework to help us understand what trade-offs can be made.\nThe ICCR is conveniently defined to describe the extent of compression, but it fails to capture how the compression is achieved (through pruning vs quantization, which layers, etc.). This could have hindereed the authors from interpreting the inconsistenties observed in Tables 1-4.\nThe presentation of the paper is another aspect that can be improved. \n1. The references are cited in a rather unusual manner. Include the references in parentheses will help readers. \n2. Tables and figures should not be bewteen texts (Fig. 1, 2, 3, 4, 7, 8; Table 5).\n3. The authors can help readers understand the results better by interpreting the results by referencing to their figures and tables. For example, Section 5.2.1 only describes the outcome, but the readers are left to understand the significance of those plots. I personally did not get what the authors mean to convey here.\n4. Some hyperparameters are given (without intution or prior knowledge, e.g. Section 5.2, 5.3). Some claims are laid without reference (Section 2, Section 4.3.2).\n5. Appendix can be part 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": 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": "- Is there a specific rationale for combining all three techniques (quantization, pruning, and KD) in the proposed optimization, and would focusing on a single technique be more beneficial?\n- How do the authors justify treating all layers within BERT similarly for quantization and pruning, given the distinct characteristics of attention layers and feed-forward networks (FFN)?\n- What theoretical or empirical justifications are provided for the formulation of the Inverted Computational Complexity metric, especially concerning the inclusion of layer number as a factor?\n- How does the dependence of pruning rate and number of bits on the layer type impact the reliability of the proposed metric, and why was this not addressed in subsection 4.3?\n- Have the authors conducted an ablation study or provided theoretical evidence to demonstrate the credibility of their proposed metric and justify the specific choices made in equation (6)?\n- Given the limited number of observations shown in Figures 3 and 4, how confident are the authors in the generalizability of their results, and could a larger dataset or different experimental setup change these findings?\n- What comparisons can the authors provide to validate the superiority of their metric over FLOPs-based metrics in terms of hardware efficiency, especially considering the specificity of efficiency to hardware architecture and operation types?\n- How does the proposed approach compare with the latest state-of-the-art works in BERT model optimization involving quantization, pruning, or knowledge distillation?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The paper tackles an interesting problem for the ML research community—optimizing large models like BERT for hardware deployment on resource-constrained devices.\n- The discussion of different optimization techniques (quantization, pruning, and KD) is quite engaging and well-explained." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the problem of optimizing Transformer-based models like BERT for hardware deployment. The authors propose a multi-level optimization using pruning, quantization, and knowledge distillation, aiming to make the BERT model more compact while preserving high accuracy on the target task. Furthermore, the paper introduced a novel metric, namely, Inverted computational complexity, to quantify the model’s computation requirements. The optimization process is conducted via a genetic algorithm to explore a predefined set of pruning and quantization parameters for the BERT model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper needs more novelty. The proposed optimization for BERT directly applies previous and well-known techniques (i.e., quantization, pruning, and KD). The authors didn't well explain why leveraging the three techniques altogether and not focusing on exploring one technique (e.g., pruning). Overall, the paper seems more like a direct application of existing methods without further improvement. To enhance the paper's novelty, the authors may consider discussing in details the benefits from each optimization technique or exploring opportunities to develop a new algorithm that more effectively integrates different optimization techniques in a way tailored specifically to the BERT architecture.\n- The paper also lacks proper discussion from an architectural point of view, specifically regarding the type of layers being pruned or quantized (e.g., attention or FFN). Different from CNNs, quantization or pruning in Transformers is not straightforward and needs careful consideration, especially at the attention layer level. However, the authors consider all layers the same and didn't explain why and how quantization/pruning is applied to the BERT layers. To strengthen the paper's contribution, the authors should include a breakdown of the impact of pruning and quantization on attention layers versus feed-forward layers and then explain their rationale for treating all layers uniformly.\n- The authors posit strong claims on their proposed Inverted computational complexity metric without theoretical or empirical evidence. First, the metric is formulated as a product of the layer number, pruning rate, and number of bits, which are also the parameters being explored by the genetic algorithm. What type of information (if any) is being extracted from this product needs to be clarified. For example, what's the utility of the layer number' i' in equation (6)? Second, the pruning rate and number of bits depend on the layer's type, which hasn't been discussed in subsection 4.3. Overall, to demonstrate the utility and credibility of the proposed metric, the authors must (i) theoretically discuss the metric computation in equation (6), (ii) conduct an ablation study (with different combinations of the metric's components), (iii) compare their metric against established hardware performance indicators (e.g., latency and memory) to show its practical relevance, and (iv) Discuss how the proposed metric accounts for different layer types, given that pruning and quantization may affect them differently.\n- The discussion in 4.3.2 needs to be more convincing since the results shown in Figures 3 and 4 cannot be generalized because of the limited number of observations (scatter points). Additionally, while the authors claim their metric is better than FLOPs because quantization is not included in the latter, both metrics are not a good proxy for hardware efficiency estimation [1]. This is because efficiency is specific to the hardware architecture and type of operations. For the authors to justify their claim, an ablation study could be conducted to compare their proposed metric and a FLOPs-aware quantization (where each layer’s FLOPs is multiplied by the number of bits).\n- There’s no comparison with the latest existing works on BERT model optimization with quantization [2], pruning [3], or knowledge distillation [4]. Without a comprehensive comparison with these SOTA works it’s hard to draw any tangible conclusion on the effectiveness and novelty of the proposed approach. Overall, the paper should discuss how P-BERT differs from or improves upon [2, 3, 4]. The authors could add a comparison table that includes their method alongside SOTA approaches, highlighting key differences and improvements.\n\n**References:**\n- [1]: Dehghani, Mostafa, et al. \"The efficiency misnomer.\" arXiv preprint arXiv:2110.12894 (2021).\n- [2]: Shen, Sheng, et al. \"Q-bert: Hessian based ultra low precision quantization of bert.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 05. 2020.\n- [3]: Liu, Zejian, et al. \"EBERT: Efficient BERT inference with dynamic structured pruning.\" Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021. 2021.\n- [4]: Muhamed, Aashiq, et al. \"CTR-BERT: Cost-effective knowledge distillation for billion-parameter teacher models.\" NeurIPS Efficient Natural Language and Speech Processing Workshop. 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": 1 }, "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": "As mentioned in the weaknesses above, it would be helpful to clarify the following questions for the reviewers:\n- How does the proposed combination of three techniques differ from or improve upon previous approaches that have used subsets of these methods? Providing more detail on this could help highlight the originality of the work.\n- Intuition and Mathematical Justification of the Inverted Computational Complexity Ratio:\n - It would be useful to provide more discussion on the design of the K_model in the Inverted Computational Complexity Ratio. Specifically, the ratio is based on a weighted summation of j_i \\times b_i. Is there an assumption that deeper layers contribute more to the efficiency improvement metric? Any intuition or mathematically justification is welcome.\n - The ratio shows a linear correlation with the number of operations. What is the definition of the number of operations in this context? \n - If the authors intend to reflect reduction in computational cost due to low-bit quantization, it would be helpful to display the number of \"quantization bits\" in table 5. Why is this proposed ratio a better metric than using the number of parameters, FLOPs, and quantization bits to measure the efficiency of the compressed model? Adding more discussion is helpful for readers to understand the advantage of this new metric compared to existing traditional metrics. \n- In the result section, it could be helpful if the authors provide a more nuanced discussion of their model's strengths and weaknesses compared to these baselines, particularly in cases where P-BERT underperforms." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Explored the method of combing three techniques to reduce the Bert model to an efficient model.\n- Proposed a metric \"Inverted Computational Complexity Ratio\", by calculating the ratio of weighted summation of quantization bits and remaining number of values in each layer. \n- Showed complete experiments and provided ablation study in details for different ratios on multiple tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, authors proposed a method to combine multiple techniques to reduce the original Bert model to an efficient model for computational restricted hardwares and remain accuracy. The author also proposed a metric \"Inverted Computational Complexity Ratio\" to measure the efficiency of compressed Bert models. The methods shows reasonable results in accuracy and efficiency compared to baseline models, but it doesn't out perform all the baseline models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses:\n\n- Not Sufficient Discussion of Original Innovation:\nThe proposed method is not showing enough discussion of original innovation, as it is essentially a combination of existing methods—pruning, quantization, and knowledge distillation. It would be helpful if authors could clarify the improvements of their proposed method compared to previous approaches. \n\t- The use of unstructured pruning with a genetic algorithm is a well-established technique for model pruning and has been extensively studied in prior works across multiple models. For example, see Multi-Objective Pruning for CNNs Using Genetic Algorithm (Chuanguang, et al., https://arxiv.org/pdf/1906.00399), and Pruning Decision Tree Using Genetic Algorithms(Jie Chen, et al., https://ieeexplore.ieee.org/document/5376632)\n\t- Similarly, quantization and knowledge distillation are applied in a straightforward manner. These techniques have already been proposed and thoroughly investigated for transformer-based models in earlier works, as referenced by the authors in the related works section.\n\n- Missing Design Reasoning of the Inverted Computational Complexity Ratio:\nThe proposed inverted computational complexity ratio could benefit from clearer mathematical reasoning and further justification.\n\t- The ratio is based on K_model, defined as the summation of i \\times j_i \\times b_i, where i represents the layer index. This makes the metric a weighted sum where the layer index serves as the weight. The issue with this definition is that K_model becomes disproportionately sensitive to deeper layers. For instance, Layer 11 contributes more to this metric when it is pruned or quantized, compared to other layers. It would be valuable for the authors to offer more insight and intuition into why deeper layers are weighted more heavily in their metric, and how this aligns with real-world computational overall efficiency improvements. \n\n- Performance is Not Significant Compared to Other Models:\nThe performance conclusion claimed by the authors is not sufficient enough when compared to other models.\n\t- The authors stated that their method achieved “promising results with competitive accuracy, particularly in CoLA”, but the evidence presented does not sufficiently demonstrate accuracy or efficiency advantages over competing models on multiple tasks, particularly when compared to models like TinyBERT and I-BERT. For instance, in Table 5, TinyBERT, despite having a much larger inverted computational complexity ratio, 27.1, outperforms P-BERT on accuracy in all tasks except CoLA. Moreover, I-BERT, which has a inverted computational complexity ratio of 2.9, similar level to P-Bert, still outperforms P-BERT in tasks such as MRPC, STSB, and even CoLA. It would be great to add more discussion of the performance gap when comparing to other baselines. Discussing both the advantages and disadvantages will offer a more balanced view and help readers better understand the specific contributions and limitations of the proposed approach.\n - The authors stated that the method is “hardware-aware optimization”, but the paper does not discuss much about the meaning of \"hardware-aware\" and how their approach is optimized for \"hardware-aware\". Assuming the authors are referring to “low-computational resource hardware,” the paper does not discuss its advantages and disadvantages compared to other baseline methods optimized for hardware. For example, one of the baseline models, I-BERT, which uses “integer-only distillation” shows advantages when deployed on hardware that supports only integer calculations. Providing similar examples of how the proposed approach is optimized for hardwares would strengthen the paper’s claims.\n\n- Typos (minor):\n - In section 5.5, authors mentioned the proposed method is not suitable for more powerful \"high-end systems, such as GPT\", which I guess should be referring to \"GPU\"." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "P-BERT combines pruning, quantization, and knowledge distillation to cut BERT's computational needs by 60%, maintaining accuracy and scores." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024pbert,\ntitle={P-{BERT}: Hardware-Aware Optimization of {BERT} Using Evolutionary Techniques},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xaXvHdH9Y4},\nnote={under review}\n}" }, "abstract": { "value": "Transformer-based models have emerged as the go-to standards in Natural Language Processing (NLP), revolutionizing the landscape of NLP applications. As complex models continue to proliferate, the need for more efficient computational processing becomes increasingly imperative. This has led to the rise of model compression techniques, implemented to target computational inefficiencies. Expounding on this, we propose Pyramid-BERT (P-BERT), the integration of three established model compression techniques to further reduce the computational inefficiency of the standard BERT models, and subsequently optimize BERT under the hardware characteristics. Specifically, the techniques employed are pruning, quantization, and knowledge distillation. The first two aforementioned correlated techniques work simultaneously to remove redundant specifications while leveraging knowledge transfer from baseline models. These techniques enable a substantial reduction in computational cost, making P-BERT highly suitable for portable, low-power devices such as cellphones, wearable devices, and smartwatches, and thus enabling hardware-friendly processing on various computing engines. Additionally, we will be proposing a new metric, the inverted computational complexity to quantify the complexity and efficacy of the model. This metric aims to more accurately capture the hardware-specific performance characteristics. Our experimental results show that P-BERT achieves a remarkable reduction of at least 60\\% in the inverted computational complexity ratio while ensuring comparable accuracy and scores across many downstream tasks compared with the baseline BERT 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": [ "Model Compression", "Large Language Models", "Computation Complexity", "BERT", "Hardware-Aware" ] }, "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/e8670a779d343f72650b9243645fc9ff04aaddfc.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/018cf7ce45b0760eac1c5bc2a869ee94fe1f92fa.zip" }, "title": { "value": "P-BERT: Hardware-Aware Optimization of BERT Using Evolutionary Techniques" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to Weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* This paper focuses on enhancing the robustness of visual imitation learning, addressing a practical and impactful topic. The approach holds significant potential for advancing research in the field of robotics.\n\n* The idea of using diffusion inversion to remove low-level visual variations while preserving high-level scene structures is both novel and intriguing.\n\n* The evaluations are thorough, with experiments conducted in both simulated environments and on real-world robots. The real-world experiments highlight the method's strong generalization capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Stem-OB, a method that enhances visual imitation learning by using image inversion from pretrained diffusion models to reduce low-level visual variations while preserving high-level scene structure. Stem-OB creates a shared representation that is robust to various appearance changes without additional training. Empirical results on several benchmarks demonstrate the effectiveness in challenging environments with lighting and appearance changes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* I found the structure of this paper somewhat difficult to follow, as certain sections lacked clarity. For instance, in the preliminaries, the definition of diffusion inversion seemed to be a combination of both forward and backward diffusion. But in Figure 2, it seems that proposed method uses the noised observations as the policy input. If the proposed method is trained on the inversion-altered space, why using the noised version of the observation as input can improve the performance? How can the authors guarantee that applying forward diffusion to the image improves generalization?\n\n* There seems to be a trade-off when choosing the inversion step, which can not be either too large or too small. Is there any explanation about this phenomenon? \n\n* The theoretical analysis section was also challenging to understand. The authors discuss a loss between two latent variables, $x_0$ and $y_0$. What is the intuition of calculating the loss of two different images? From my understanding, the attribute loss here should refer to the inversed data $\\hat{x}_0$ and its original version $x_0$. Could the authors also clarify the statement, “images with fine-grained attribute changes tend to become indistinguishable sooner than those with coarse-grained modifications under identical diffusion schedules”? \n\n* Section 4.1 reminded me of another study [1], which employs diffusion models to purify noise within noisy demonstrations while preserving the optimal structure. Are there any conceptual similarities between that approach and the proposed method?\n\n* Typos: Line 242: $er$f should be $erf$.\n\n[1] Imitation Learning from Purified Demonstrations, ICML 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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 most significant questions are already listed under weaknesses. In addition, I would be interested to get a clarification on the following points:\n\n- line 272: How many images?\n- line 428: Can the authors explain how they get to this conclusion? The training setting performance of RO in particular looks to me like it was not trained correctly.\n- line 430: Given the result in D2B, I don't think the conclusion can be that there is always a high success rate?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The idea is, as the authors write, 'simple yet effective' and the authors manage to neatly package a quite theoretical idea and a very practically result measured in real-world robotic experiments all in 1 paper\n- While I am less convinced of the theoretical grounding, the idea itself is well explained\n- The authors combine experiments on simulated datasets with actual real-world tests. Real-world tests are incredibly important for visual methods, yet in case of robotic applications very hard to make reproducible. The combination of both is a commendable experimental design.\n- Similarly, the authors conduct real-world user studies to investigate their hypothesis that diffusion inversion progresses along semantic hierarchies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to use diffusion inversion as input preprocessing for visual imitation learning methods in order to improve their generalization to visual changes between demonstration and test setup. The proposed approach is evaluated in real-world robotic experiments as well as on 2 simulated datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I find the derivation in Section 4.1 very hard to follow. Only the first sentence of Section 4.2 makes it clear that the whole derivation is based on the assumption that semantically similar images are closer in latent space. That is a HUGE assumption and the following experiments on intra- and inter-class confusion do not show this over multiple levels of semantic hierarchies, but only for the semantic leafs. The introduction (lines 60-65) however argues that much more than leaf classes, abstract semantic hierarchies are important to generalize over perceptual differences.\nIn the empirical study on diffusion inversion, one might argue that out of the 5 empirical examples, \"bowl\" and \"cup\" are most similar. However, Table 1 shows that the latent representation of 'cup' samples is actually on average closer to any of {'drawer', 'duck', 'faucet'} than to 'bowl'. To me this makes the assumption 'semantically similar images are closer in latent space' quite unbelievable. For sure, this assumption and therefore the formulation of semantic overlap as overlap of latent gaussians is anything but 'intuitive' (line 270) in presence of this data.\n- In relation to the above, I miss a clear definition of what kind of variation should be compensated through diffusion inversion. The abstract and introduction repeatedly claims that diffusion inversion can extract the 'high-level structure' of the scene, suggesting generalization over different object instances, shapes, appearances, relative placements, and robot arms. Section 4.1 considers 'variation', 'fine-grained attribute change', 'coarse-grained modifications', and 'semantic overlap' without defining any of these. The investigation in Section 4.2 is focused on variations within a semantic object class, i.e. a demonstration with 1 cup should be repeated with a different cup. The experiments then consider lighting change and a limited set of object appearance change of sometimes multiple object instances, while locations are fully fixed. The problem is that all of this currently does not fully fit together and it would be good if the authors can define more accurately what kind of variations they expect diffusion inversion to abstract / generalize over, and then design experiments accordingly to show improvement with exactly these variations.\n- There are a coupe of odd aspects about the simulation experiments that raise questions about the soundness of the results:\n - For the benchmarks from ManiSkill and MimicGen, why were not all tasks evaluated? For ManiSkill, plug-charger seems to be specifically excluded and for MimicGen 4 out of 12 tasks were picked without any explanation.\n - While I did not quickly find comparable numbers for ManiSkill, the MimicGen paper reports much higher success rates for the investigated tasks. E.g. for Threading, MimicGen reports around 19% sucess from just 10 videos and 98% success from 1000 videos. For the 500 videos used in the experiments here, the success is below 19% for 3 out of 4 variants, including the proposed method. Why are the achieved success rates so low? And can the proposed method actually improve anything in a more state-of-the-art setting?\n - Why is the RO baseline excluded from the simulation experiments?\n- In all experiments (real and simulated), most of the differences between methods are within the standard deviation, so it is very hard to say if any conclusion can be drawn. Why is the standard deviation so high?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "The current method conducts experiments solely on Stable Diffusion. It would be valuable to understand if the conclusions drawn from Stem-OB are applicable to other generative models, such as Flux or SD 3 (which uses flow instead of diffusion). Demonstrating that Stem-OB generalizes across different models would strengthen the claim of robustness and broaden the impact of this approach beyond a single model framework." }, "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-written, with clear motivation and solid experimental validation, including both real-world and simulated experiments.\n\n2. The central idea resonates well: robotic observations often include excessive low-level details, while effective scene understanding requires capturing high-level structural information. This paper draws on an intriguing insight from Yue et al. (2024): \"Rather than uniformly removing information from different semantic hierarchies, the process brings structurally similar images closer in the early stages of inversion.\" Building on this, Stem-OB leverages this observation to project data into a space that prioritizes inherent structural features." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Stem-OB, a method that uses diffusion inversion to transform observations into a shared, robust representation. This approach enhances robustness to various appearance changes, such as shifts in visual patterns and lighting conditions. During inference, Stem-OB approximates the partial inversion process to maintain efficient inference speed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Stem-OB enhances the generalization capabilities of robotic systems by using diffusion inversion to map observations into a shared representation. However, another promising line of research—self-supervised representation learning—also aims to unify raw observations into a common representation space, reducing low-level visual discrepancies. I suggest that the authors consider benchmarking Stem-OB against these approaches, particularly methods that leverage self-supervised learning (SSL) for robotic representations, to offer a more comprehensive comparison." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "See above. I currently give a weak accept to the paper but am inclined to vote for acceptance should my points above be taken into consideration during the rebuttal." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper relies on a central observation from a previous paper (Yue et al. 2024), which found that during the inversion process of a diffusion model, fine-grained (or high-frequency) details of an image are destroyed first before the low-level semantic concepts. This paper builds on this basic concept and proposes a new preprocessing step that uses this theoretical property of diffusion models as a preprocessing step, which, in the process, improves the generalization of the IL algorithm. \n* The technical principle seems well established at this point, but this paper takes these theoretical findings and applies them to a new problem. While the whole approach is very simple from a technical standpoint, the paper re-introduces the necessary background in section 2 before developing the reader's intuition as to why this preprocessing might work in section 4.2 through theoretical explanation and small experiments. I believe that this paper warrants publication because it shows that diffusion model inversion can be useful for additional tasks. \n* The presentation of the paper is clear and easy to follow overall. \n* The results nicely highlight the drawbacks of previous methods and that the proposed method offers much better generalization capabilities. \n* I appreciated the detailed appendix, with theoretical derivations, additional details for reproducibility, and additional experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Visual imitation learning is a method for robots to learn tasks through visual human demonstrations. While multiple methods exist for preprocessing visual data that increase the generalizability of VIL, these methods are only partially efficient. To address the generalization issue, this paper proposes to use the inversion process of a pretrained diffusion model as a preprocessing step to increase the generalizability of visual imitation learning methods. The proposed method relies on the intuitive insight that images with only small perturbations between them are closer in latent space (and therefore become indistinguishable under diffusion faster), than data that exhibits large differences. The method is tested on real and synthetic data and shows that the proposed preprocessing significantly increases the models generalization capabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I believe that the paper's technical novelty is small, as the main concept has already been introduced in other papers. That being said, I also believe that the paper is worth publishing, as it adds to the growing body of literature showing the usefulness of diffusion model inversion for various tasks, and the empirical validation is well executed.\n\nI have additional suggestions to improve the paper's readability and make it easier for the reader to understand.\n1. Section 2.2 can be cut (title), and the paragraph can be integrated into Section 3.1. \n2. As a reader, I would really like a short problem formulation that properly introduces the problem the paper addresses. This should properly define the inputs and outputs. I would also suggest rearranging the sections/subsections of the paper to something like this:\n* Introduction\n* Related Work\n* Problem Definition\n* Preliminaries\n* Method\nThis would greatly improve the flow of the paper. \n3. While the paper is generally easy to follow, some sentences are should be simplified or run through a writing program. I have added some examples below:\n* Line 85: \"To be specific, our method is as simple as inverting the image for reasonable steps before\" What is reasonable? Maybe rephrase?\n* Line 226: \"Intuitively, given a source image and its variation, distinguishing between them becomes increasingly difficult\" This sentence does not make a lot of sense (although I can guess what is meant).\nThere are many more of these convoluted sentences. Cleaning up the writing a bit would go a long way to improve the paper.\n\nThere" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024stemob,\ntitle={Stem-{OB}: Generalizable Visual Imitation Learning with Stem-Like Convergent Observation through Diffusion Inversion},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xaYlO03tIk},\nnote={under review}\n}" }, "abstract": { "value": "Visual imitation learning methods demonstrate strong performance, yet they lack generalization when faced with visual input perturbations like variations in lighting and textures. This limitation hampers their practical application in real-world settings. To address this, we propose ***Stem-OB*** that leverages the inversion process of pretrained image diffusion models to suppress low-level visual differences while maintaining high-level scene structures. This image inversion process is akin to transforming the observation into a shared representation, from which other observations also stem. *Stem-OB* offers a simple yet effective plug-and-play solution that stands in contrast to data augmentation approaches. It demonstrates robustness to various unspecified appearance changes without the need for additional training. We provide theoretical insights and empirical results that validate the efficacy of our approach in simulated and real settings. *Stem-OB* shows an exceptionally significant improvement in real-world robotic tasks, where challenging light and appearance changes are present, with an average increase of **22.2%** in success rates compared to the best baseline. Please refer to [this link](https://stem-ob.github.io/) for more videos and details." }, "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": [ "Robotics", "Imitation Learning", "Visual Imitation Learning", "Robustness", "Diffusion Model", "Diffusion Inversion" ] }, "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/ae8800e1d43fc51454a8e5c4c43e50a67467bbcb.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/555799937b2fe8bf9844db9b7ed4824619e621db.zip" }, "title": { "value": "Stem-OB: Generalizable Visual Imitation Learning with Stem-Like Convergent Observation through Diffusion Inversion" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Could you explain why lines 43-45 present a contradiction? Some papers do consider product graphs (eg https://arxiv.org/abs/2206.15174). Moreover, addressing temporal and spatial (graph) processing in subsequent steps does not seem to reduce expressiveness (see eg https://arxiv.org/abs/2103.07016).\n- Lines 92-93. Claiming that CNN cannot model spatial information is simply wrong; see e.g. the field of Geospatial Deep Learning.\n- Definition 1. Why is this structure called a graph if the only information is in the node features? Also, why is it called \"hypervariate\" if it has $D$ features, as per the original time series?\n- Sections 3.1 and 3.2. These sections describe well-known facts: a fully-connected graph is regular and each permutation leads to an automorphism. Is there an additional insight in these sections that I may have missed?\n- Definition 2. Similarly to Definition 1, why is this called \"hyperspectral\"? What makes it “hyper”?\n- Section 4.1. As a reference, could you indicate a few state-of-the-art approaches within the \"pure graph paradigm\"? Currently, I only see only FourierGNN being mentioned.\n- Line 212. The elements can only be rearranged as long as they are labeled or indexed by the corresponding frequency-node pair. This means they can be stored in memory in any order, but their indexing cannot be ignored. Therefore, the statement in line 215 does not logically follow from line 212. Is there an argument to support the conclusion in lines 215-217?\n- What is novel in the proposed use of framelets compared to existing literature?\n- Could you report standard deviations alongside the results presented?\n- Figure 6. What does a contour plot represent for discrete variables? Additionally, how can the optimal $k$ be a non-integer value?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The paper addresses limitations of a previously published method, FourierGNN." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a time-series forecasting model that first computes the Fourier transform for individual components of the time series. It then constructs a KNN graph using triples as nodes -- each of which consisting of a time-series index, feature dimension, and frequency. Finally, the model employs framelets to make predictions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The presentation is confusing, often blending obvious claims with unsupported statements and using ambiguous terminology.\n- The paper tackles challenges that seem obvious and already addressed in the literature: (1) the importance of maintaining temporal order in time-series processing, and (2) the unsuitability of fully connected graphs in graph-based processing.\n- As far as I could understand, the paper’s original contribution is limited to constructing a KNN graph from a spectral representation of the original multivariate 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": 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": "Based on the above weaknesses, the following aspects of this work could benefit from additional explanations/extensions:\n- **[Q1] Discrepancies in MSE/MAE performances for different datasets/horizons:** Based on **[W2]**, could you explain why the model for specific datasets performs significantly worse in terms of MSE in comparison to baselines but has the best MAE values. \n- **[Q2] Visualizations for predictions/learned graph dependencies:** To address discrepancies in scores, some visualizations of predictions (true and predicted time series) corresponding to such cases (low MAE, large MSE), but also, in general, are a crucial qualitative way to access performance variations between models/datasets. Based on **[W5]**, could you provide some visualizations on the learned dependencies provided by the graph module of UFGTIME for some real-world datasets?\n- **[Q3] Proper comparison of different graph learning modules:** The paper could improve its discussion of related works by differentiating between pure graph paradigms and methods that produce sparse graphs in the context of GNN architectures.\n- **[Q4] Complexity analysis of GNN-based graph learning baselines:** The paper and results could benefit from a discussion on the complexity of different GNN-based baselines along different graph learning modules beyond the already mentioned FourierGNN.\n- **[Q5] Impact of window size and input variables on memory complexity:** including analysis/ablations that address these factors could provide a better understanding of the scalability and resource requirements of UFGTIME for real-world time series datasets.\n- **[Q6] Misc:** No standard deviations/averaging across multiple runs for the provided performance results are mentioned. Is it possible to demonstrate that the small enhancements observed across multiple cases are statistically significant compared to the best competitor?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Strong points of the presented work are the following:\n- **[S1]:** The authors showcase the importance of spectral graphs for time series forecasting while highlighting the importance of sparsity in the structure.\n- **[S2]:** The proposed framework remains simple in its design.\n- **[S3]:** Experimental results show the very competitive performance of the proposed method in terms of MSE with popular baselines.\n- **[S4]:** Ablation studies highlight the importance of performance when learning a (sparse) graph structure on underlying dependencies in forecasting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present a **GNN-based model tailored to time series forecasting** without an essentially existent aprior graph structure. They focus on overcoming the limitations of a recent state-of-the-art method that considers hypervariate (fully connected) graphs by highlighting the importance of capturing the most *crucial inter- and intra-series correlations* by incorporating some sparsity. They thus propose a method that extracts a KNN-based graph structure built upon representation extracted by the Fourier transform. Their introduced (hyperspectral) graph structure is then processed by the so-called *global Fourier framelet message-passing operator* to capture global patterns from the time series examples. After the **spectral graph embedding module**, the inverse Fourier transform returns the representation in the original time dimension, followed by a two-layer feed-forward network to predict the output vector (on the future steps) combined with the *trend embedding* of the original signal. The authors evaluate the performance of their proposed method against popular graph-based and non-graph-based models in forecasting and presenting properties of the critical components of their model through ablation studies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **[W1]:** Performance improvements achieved by the proposed method are mostly minor compared with FourierGNN for the datasets considered in short-term forecasting.\n- **[W2]:** The proposed method is outperformed by baselines on datasets used for long-term forecasting. Interestingly, in several cases, there is a discrepancy in the performances achieved in terms of MSE and MAE (e.g., MAE scores of the proposed UFGTIME on ETTh2 are the best against baselines, but corresponding MSEs are almost double in value compared to the Autoformer).\n- **[W4]:** The authors focus on the challenges of the pure graph paradigm, yet most graph-based methods for time series before this work [1] were similarly using KNN [3] or differentiable methods (Gumbel softmax) [2] to achieve graph sparsity. Additionally, embeddings were learned based on time series inter-variable correlations or temporal dynamics [2] or as node embeddings [3]. The paper misses an extensive discussion, e.g., in the related work section, on the critical characteristics of the graph modules introduced in the literature to adequately position the proposed UFGTIME's contribution.\n- **[W5]:** The learned graph dependencies are not evaluated qualitatively. Similar to the studies/visualizations for the graph learning module in [3], it would be interesting to assess the validity of learned time series dependencies produced by the proposed graph module.\n- **[W6]:** The proposed method is quite efficient regarding memory/time cost. However, it would be interesting for the discussion to be enhanced with explanations of the complexities of all considered GNN-based baselines. It is unclear to which extent the memory complexity/efficiency of the proposed method is affected by the window size and number of variables in the input, which could provide additional ablation studies.\n\n[1] Yi, K., Zhang, Q., Fan, W., He, H., Hu, L., Wang, P., ... & Niu, Z. (2024). FourierGNN: Rethinking multivariate time series forecasting from a pure graph perspective. Advances in Neural Information Processing Systems, 36.\n\n[2] Shang, C., Chen, J., & Bi, J. (2021). Discrete graph structure learning for forecasting multiple time series. arXiv preprint arXiv:2101.06861.\n\n[3] Wu, Z., Pan, S., Long, G., Jiang, J., Chang, X., & Zhang, C. (2020, August). Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining (pp. 753-763)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "+ For the results in Sec. 3.2, what temporal characteristics were used in constructing the Laplacian matrices? Additionally, what are the attention weights applied in this context?\n+ I am curious about the necessity of this method—what are its key advantages compared to competitive time series models like PatchTST, TimeMixer, and a recent study https://arxiv.org/pdf/2403.14587?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ The work presents a feasible way of addressing the limitations of hypervariate graphs by transforming time series into the frequency domain and constructing hyperspectral graphs. The use of frequency-based graph representation is an interesting twist on how spatio-temporal dependencies are handled.\n+ This paper introduces a well-defined framework with innovative techniques, such as the graph framelet message passing operator. The theoretical analysis and complexity evaluations provide a thorough understanding of the approach.\n+ The technical descriptions of the novel hyperspectral graph and framelet message passing are articulated well.\n+ The work addresses practical limitations in a recent GNN-based time series forecasting method, proposing improvements that could lead to better generalization in various time series forecasting domains, including both short- and long-term applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the hyperspectral graph structure to enhance the ability of GNNs in capturing temporal dependencies for multivariate time series forecasting. The authors proposes UFGTIME, which transforms time series data into a frequency domain representation, preserves sequential information, and constructs a sparse graph with signal similarity using KNN. This graph structure is combined with a framelet message passing mechanism, aiming to improve the ability to capture both global and local temporal patterns without over-smoothing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ The core motivation of this work feels narrow. While improving upon FourierGNN is a valuable contribution, the paper does not clearly establish why the disjoint modeling architecture in existing GNN-based methods fails critically in practice, or why FourierGNN's approach is inherently superior. A stronger emphasis on the overarching challenges of multivariate time series forecasting would better justify the significance of this work.\n+ The storytelling could be improved, particularly in the introduction. The two key challenges in FourierGNN are not clearly articulated, and after reading the introduction, it is unclear if this research addresses a significant problem.\n+ The empirical validation in Sec. 3 is somewhat weak in certain aspects. For instance, the experimental setup for the evaluation in Tab. 1 is unclear, and the performance degradation after permutation looks significant without compared to benchmarking methods, leaving doubt about the strength of the claims. Additionally, the visualization results in Fig. 2 show minimal differences between sparse and fully connected Laplacian patterns without the attention weights. \n+ The construction of the hyperspectral graph closely mirrors that of the hypervariate graph, with the key distinction being that it bypasses direct modeling of sequential information by transforming time series into discrete frequency components.\n+ There is a lack of comparative discussion against competitive time series models, particularly recent approaches like PatchTST and TimeMixer. The advantages of UFGTIME over these models are not sufficiently explored, which reduces the clarity around the unique contributions of this research." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "--" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "* Developing effective deep learning architectures for time series forecasting is an important research direction, and graph-based approaches are appealing in many settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a graph representation for groups of multivariate time series and a time series forecasting model based on this representation. The main motivation behind the method is to address limitations in a specific previous work (FourierGNN), which represents the input time series as a fully connected graph where each node is an observation in time and space. This representation clearly has the drawback of discarding the temporal ordering of the observations. The introduced method attempts this issue, but such a limitation is specific to FourierGNN and does not apply to a broader body of literature operating in similar settings. There are also some issues with the soundness of the method and of the empirical evaluation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are major weaknesses that prevent me from recommending acceptance.\n\n* **Soundness of the approach.** The proposed approach builds a graph representation by considering the K-nearest neighbors of a representation of the input obtained through an FFT along the temporal dimension. However, in such a representation, each node corresponds to a different frequency and entity. I don't understand why one would want to connect nodes that correspond to different frequencies. For example, with such a representation, signals that have all of their spectrum concentrated at totally different frequencies would get connected. This representation does not preserve the structure of the data, similarly to the method the paper is trying to fix.\n* **Poor novelty and weak motivation.** As noted, the limitation the paper aims to address is specific to a single model. Addressing such a narrow issue isn't sufficient when most state-of-the-art approaches do not have this drawback. Many spatiotemporal graph neural networks utilize sound graph representations of the input, and in many cases, the processing of spatial and temporal dimensions is well integrated (e.g., see [1]). Additionally, several signal processing methods operate on pure graph representations of sets of time series. For instance, [2] uses product graph representations. The visibility graph is another principled graph representation of input time series (e.g., see [3]). All of these representations avoid the issues the proposed method attempts to solve and are more sound than the one presented in the paper. Lastly, the novelty of the graph framelet operator is limited, as it appears to be a straightforward adaptation of an existing method.\n* **Soundness of the empirical evaluation.** The empirical evaluation has some issues. Firstly, there is a growing consensus that the ETT benchmarks used for the experiments in Tab 3 are not significant for evaluating deep learning methods for time series forecasting. For example, in [4], a simple autoregressive linear model trained using ordinary least squares achieves better results than those reported in the table in almost all the considered scenarios. Secondly, some of the results in Table 2 involve baselines that require a predefined input graph, which, according to the appendix, is a kNN graph obtained in an unclear way (e.g., which similarity metric was used to create this graph?).\n\nGiven these issues, the paper is far below the acceptance threshold by ICLR standards.\n\nMinor comments\n\n* There must be some reporting errors in Table 3. Looking at the MSE for UFG in ETTH1, the MSE at 192 steps is lower than the MSE at 96, which does not make any sense if the testing is done properly. Differently, the MAE increases as one would expect.\n\n[1] Wu et al., \"Traversenet: Unifying space and time in message passing for traffic forecasting\" TNNLS 2022\\\n[2] Sabbaqi et al., \"Graph-time convolutional neural networks: Architecture and theoretical analysis.\" PAMI 2023.\\\n[3] Lacasa et al., \"From time series to complex networks: The visibility graph\" PNAS 2008\\\n[4] Toner et al., \"An Analysis of Linear Time Series Forecasting Models.\" ICML 2024\\" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Pure Graph Paradigm for Multivariate Time Series Forecasting in the Frequency Domain" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024ufgtime,\ntitle={{UFGT}ime: Reforming the Pure Graph Paradigm for Multivariate Time Series Forecasting in the Frequency Domain},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xaafWdM5jI},\nnote={under review}\n}" }, "abstract": { "value": "Recent advances in multivariate time series forecasting have seen a shift toward a pure graph paradigm, which transforms time series into hypervariate graphs and employs graph neural networks (GNNs) to holistically capture intertwined spatiotemporal dependencies. While promising, this approach faces notable challenges. First, converting time series into hypervariate graphs often neglects essential temporal sequences, which are vital for accurately capturing temporal dependencies. Second, treating the graph as a complete structure can obscure the varying importance of intra- and inter-series connections, potentially overlooking key local patterns. To address these challenges, we introduce a novel hyperspectral graph data structure that embeds sequential order into frequency signals and employs a sparse yet meaningful topological structure. In addition, we propose the \\textsc{Ufgtime} framework, featuring a frequency-based global graph framelet message-passing operator tailored to hyperspectral graphs, effectively mitigating the smoothing issue and capturing global insights through sparse connections. Extensive experiments demonstrate that our framework significantly surpasses state-of-the-art methods, excelling in both short- and long-range time series forecasting while achieving superior efficiency. Our code is available at:~\\url{https://anonymous.4open.science/r/UFGTIME-E352}." }, "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": [ "Multivariate Time Series Forecasting", "GNN", "Pure Graph Paradigm" ] }, "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/bb48ed381c4c01129a121a561a5def019874cb6f.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": "UFGTime: Reforming the Pure Graph Paradigm for Multivariate Time Series Forecasting in the Frequency Domain" }, "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": "See weakness" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Computational Efficiency: By removing the reliance on large pre-trained models, DQ V2 lowers computational costs.\n\n2. Good insight for data augmentation: The pre-trained MAE model is equivalent to a data augmentation method (in introducing prior knowledge and implicit regularization into the training process)\n\n3. The writing is clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the high computational cost of Dataset Quantization (DQ) due to its reliance on large pre-trained models like MAE and ResNet. They propose DQ V2, which removes pre-trained models by using a random CNN-based data augmentation that retains semantic structure by masking objects and replacing backgrounds, enhancing diversity without costly models. The goal of data augmentation (synthesizing) in their pipeline is to enhance data diversity and representation without relying on costly pre-trained models. \n\nEvaluation: Evaluated on ImageNette, CUB-200-2011, Food-101, and ImageNet-30, DQ v2’s performance is compared with DQ’s. DQ v2 achieves comparable or better performance than the original DQ method, showing an average improvement of about 1.57%." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of Quantitative Analysis on Computational Gains: While the paper claims computational benefits from replacing the MAE model with a CNN-based data augmentation strategy, it lacks specific measurements or comparisons to substantiate these gains. A quantitative analysis—such as GPU hours, memory usage, or training time—would provide stronger evidence of the efficiency improvements in DQ V2.\n\n2. Missing Baselines: I noticed that some recent coreset selection baselines for deep learning are missing: D2 Pruning[1], CCS[2], Moderate[3]. Those baselines seem to have a stronger performance than the proposed methods.\n\n3. Missing evaluation on ImageNet-1k: the paper argues that DQ-V2 is more efficient than DQ, but the method is only evaluated on the ImageNet subset. Previous methods including DQ all conducted evaluation on ImageNet-1k. It will be good to include an ImageNet-1k evaluation to demonstrate the scalability of the proposed methods.\n\n4. The data augmentation part is confusing: the goal of data quantization and coreset selection is to reduce the size of the training dataset, but the data augmentation method proposed in the paper expands the datasets -- the final expanded training dataset can be even larger, which is contradicted to the goal of coreset selection.\n\n5. Ablation study on data augmentation: The paper would benefit from a more detailed ablation study to assess the effectiveness of the data augmentation method used in DQ V2. Testing different data augmentation configurations (e.g., no augmentation, alternate augmentation techniques) would clarify its impact and help refine the methodology.\n\n[1] Maharana, Adyasha, Prateek Yadav, and Mohit Bansal. \"D2 pruning: Message passing for balancing diversity and difficulty in data pruning.\" ICLR 2024\n\n[2] Zheng, Haizhong, Rui Liu, Fan Lai, and Atul Prakash. \"Coverage-centric coreset selection for high pruning rates.\" ICLR 2023\n\n[3] Xia, Xiaobo, Jiale Liu, Jun Yu, Xu Shen, Bo Han, and Tongliang Liu. \"Moderate coreset: A universal method of data selection for real-world data-efficient deep learning.\" ICLR 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": 2 }, "primary_area": null, "questions": { "value": "The biggest question is what specific negative effects MAE actually introduces, as the authors' experiments and analysis do not clearly convey any significant drawbacks to using MAE." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The method proposed by the authors does indeed achieve comparable or even higher results without using MAE.\n\n2. The authors conducted extensive ablation studies on the parameters of the method itself, including experiments on patch size and data selection methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper examines the limitations of the DQ method and proposes corresponding improvements. The authors believe that using a pretrained MAE in DQ may cause issues, so they conducted experiments to see the impact on DQ when MAE is removed. The experiments, in a way, demonstrate the importance of MAE. The authors suggest using Tobias data augmentation as a substitute for MAE. According to their results, it is possible to achieve accuracy comparable to or even better than the previous DQ without using MAE." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation of this paper is somewhat unclear. From my understanding, the main value of DQ lies in reducing dataset size and storage requirements. However, as shown in Table 1, this method actually increases the storage usage of DQ. The problem it addresses is the need for a pretrained MAE in the original DQ, yet the authors' experiments do not highlight any obvious issues caused by using MAE. In my view, the authors have optimized a relatively minor aspect while losing sight of one of DQ’s key contributions. It would be beneficial for the authors to further elaborate on the advantages of this method.\n\n2. The logic of the proposed method is unclear. The authors first apply Tobias data augmentation, followed by dataset selection—what is the advantage of this sequence? What would the outcome be if Tobias data augmentation were added directly at the end based on DQ?\n\n3. The conclusions regarding line 210 may have some bias, as MAE was pretrained on ImageNet, which likely results in better reconstruction performance on ImageNette. The variables here are not limited to dataset size, so the effectiveness may not necessarily be due to the dataset size alone. It could also be influenced by the effectiveness of MAE 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": 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": 2 }, "primary_area": null, "questions": { "value": "Please see weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Using semantical-aware data augmentation to remove the pre-trained MAE model in DQ is interesting.\n2. The paper is well-organized.\n3. Experimental results show that the proposed DQ_v2 eliminates the drawbacks of DQ's dependence on pre-trained.\n4. The proposed method achieves performance improvement on multiple datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes DQ_v2, a corset selection method. To remove the pre-trained MAE in DQ, the authors investigate a data augmentation scheme, which can simulate the steps of pixel compression and reconstruction in DQ. Finally, the authors show the performance on several benchmark datasets, including CUB-200, Food-101, and ImageNet. The idea of using data augmentation to replace pre-trained MAE in DQ is somewhat novel to me. However, some critical concerns remain, please see weakness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In line 278, the authors say that the corset contains both original and augmented images. However, as far as I know, most existing corset selections only select original images from the datasets, meaning that there are no augmented images in corsets. So is this a fair comparison between DQ_v2 and other corset selection methods?\n2. The literature review section lacks comprehensiveness. Numerous recent studies closely related to the topic have not been studied, such as [1-5], which may affect the context and clarity of the proposed approach.\n[1] Tan, Haoru, et al. \"Data pruning via moving-one-sample-out.\" Advances in Neural Information Processing Systems 36 (2024).\n[2] Xia, Xiaobo, et al. \"Moderate coreset: A universal method of data selection for real-world data-efficient deep learning.\" The Eleventh International Conference on Learning Representations. 2022.\n[3] Yang, Shuo, et al. \"Dataset pruning: Reducing training data by examining generalization influence.\" arXiv preprint arXiv:2205.09329 (2022).\n[4] Maharana, Adyasha, Prateek Yadav, and Mohit Bansal. \"D2 pruning: Message passing for balancing diversity and difficulty in data pruning.\" arXiv preprint arXiv:2310.07931 (2023).\n[5] Yang, Suorong, et al. \"Not All Data Matters: An End-to-End Adaptive Dataset Pruning Framework for Enhancing Model Performance and Efficiency.\" arXiv preprint arXiv:2312.05599 (2023).\n3. In the semantic data augmentation section, the authors enhance diversity by replacing image backgrounds. However, it’s unclear if the potential for semantic ambiguity was considered—for instance, whether the new backgrounds might inadvertently introduce other objects, which could affect the intended semantics.\n4. The authors report only storage costs, but I recommend adding a comparison of training costs as well. This would provide a more comprehensive assessment of the method’s efficiency and practical applicability.\n5. The practical significance of the proposed method is unconvincing due to limited experimental validation. In the experimental section, all benchmark comparisons are with methods published before 2021. The compared baselines are outdated. While authors claim the comparison with state-of-the-art, many existing SOTA methods [1-5] are not compared. This weakens the method’s practical performance and 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": 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 goal of DQ is to reduce training data volume and improve data efficiency. Since the proposed method uses data augmentation, does it significantly increase the dataset size, potentially resulting in similar training costs as regular training?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The overall writing of the paper is smooth and easy to understand.\n- DQ V2 replaces MAE-based quantization with a simple augmentation strategy, achieving better performance without pre-trained models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Dataset Quantization V2 (DQ V2), an enhanced version of the original Dataset Quantization (DQ) method, focusing on efficient coreset selection without relying on large pre-trained models like MAE. Instead, DQ V2 integrates a new data augmentation strategy called Tobias, which uses randomly initialized CNNs to preserve the semantic regions of images while replacing background areas, mimicking the effect of pixel quantization. Extensive experiments demonstrate that DQ V2 achieves improved performance and training stability across multiple datasets, while also reducing computational complexity. The results suggest that DQ V2 provides a practical solution for data compression and coreset selection, paving the way for further enhancements in semantic-aware data augmentation and broader applications in complex visual tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper claims good scalability for the proposed method, but the experiments are still focused on smaller datasets and do not include evaluations on mainstream large-scale datasets like ImageNet-1k.\n- The coreset selection methods chosen for comparison, such as GraNd, Grad-Match, and GC, are from 2021. The paper should include comparisons with more recent coreset selection and dataset quantization methods." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper proposes an efficient core set selection method based on semantically-aware data augmentation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024rethinking,\ntitle={Rethinking Dataset Quantization: Efficient Core Set Selection via Semantically-Aware Data Augmentation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xajif1l65R},\nnote={under review}\n}" }, "abstract": { "value": "Dataset quantization (DQ) is an innovative coreset selection method to choose representative subsets from large-scale datasets, such as ImageNet. Although DQ has made significant progress, it heavily relies on large pre-trained models (like MAEs), leading to substantial additional computational overhead. We first identify that removing this pre-trained MAE model degrades DQ’s performance and increases the variance in model training. Where MAE plays a crucial role in introducing prior knowledge and implicit regularization into the training process. Second, we investigate a data augmentation scheme that can simulate the steps of pixel compression and reconstruction in DQ by simply using a randomly initialized ResNet model. This randomly initialized ResNet model can take advantage of the inductive bias of CNNs to locate the semantic object region and then replace the other region with other images. Therefore, we can use a random model or trained model in the early training stage to enhance semantic diversity while selecting important samples. We remove the module that contains the pre-trained MAE model and integrate the data augmentation scheme into the DQ pipeline, which formulates a new simple but efficient method, called DQ v2. Our method achieves performance improvements across multiple datasets, such as ImageNette, CUB-200, and Food-101." }, "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": [ "Coreset Selection", "Dataset Quantization", "Data Augmentation", "Efficient Deep Learning", "Semantically-Aware Augmentation" ] }, "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/6f309d87fdd9e243320102cf97700445d1aecbbe.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": "Rethinking Dataset Quantization: Efficient Core Set Selection via Semantically-Aware Data Augmentation" }, "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": "I think the logic in lines 181–184 is incorrect, though the result is ultimately correct. The logic appears to use the implication $ A \\implies B $ to conclude $ \\neg A \\implies \\neg B $, which is not valid. The result holds, however, because $ P = \\sum_{i=1}^k w_i P_i^{\\leq n} $ and $ Q = \\sum_{i=1}^k v_i Q_i^{\\leq n} $, so if $ P \\neq Q $, then $ \\sum_{i=1}^k w_i P_i^{\\leq j} \\neq \\sum_{i=1}^k v_i Q_i^{\\leq j} $ for at least $ j = n $.\n\nFor line 451, it should be $ \\delta x_0 x_1 \\to \\delta x_0 x_k $. Additionally, some of the calculations in lines 441–458 are straightforward, so a few lines could be removed to streamline the presentation.\n\nIn the proof of Proposition 7, when setting parameters, it would be helpful to specify how small the parameter $ \\eta_0 $ ( and $h_0,\\delta$) needs to be to ensure that the relative error of the marginal probability remains smaller than $ O(\\epsilon / n) $. Could you clarify why machine precision is considered here? Would $ \\mathrm{poly}(N / \\epsilon) $ bits suffice to represent $ \\eta_0 $, where $ N $ denotes the input size?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The main contribution of this paper is the first result: deciding whether two mixtures of product distributions are the same. Suppose we have $k$ product distributions $P_1, P_2, \\ldots, P_k$, where each $P_i$ is an $n$-dimensional product distribution, i.e., $X \\sim P_i$ is a vector $(X_1, X_2, \\ldots, X_n)$. The paper takes the prefix of $X$, namely $X^{\\leq j} = (X_1, X_2, \\ldots, X_j)$ for $j \\leq n$. This distribution is denoted by $P^{\\leq j}_i$. Then, they consider the mixture of $P^{\\leq j}_1, P^{\\leq j}_2, \\ldots, P^{\\leq j}_k$, denoted by $P^{\\leq j}$. The algorithm decides whether $P^{\\leq j}$ and $Q^{\\leq j}$ are the same for all $j$. The algorithm is based on induction from $j = 1$ to $j = n$. The base case is trivial. The difficult part is that for $P^{\\leq j}$ and $Q^{\\leq j}$, the support of the distribution can be as large as $\\exp(\\Omega(j))$. To reduce the computational cost, the algorithm finds a \"sketch\" of the two distributions. One needs to check whether $P^{\\leq j}(x) = Q^{\\leq j}(x)$ for exponentially many $x \\in \\Sigma^{j}$. For each $x$, the algorithm views $P^{\\leq j}(x) = Q^{\\leq j}(x)$ as a linear equation. Instead of checking an exponential number of linear equations, the algorithm finds a basis of the linear system, and the size of the basis is $\\text{poly}(n)$. Then the algorithm only need to check the equations in the basis.\n\nOverall, the algorithm and the definition of $P^{\\leq j},Q^{\\leq j}$ are simple and clever, and I think deciding whether two mixtures of product distributions are the same is a basic problem in statistics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of computing the total variation distance (TV-distance) between high-dimensional distributions from a computational standpoint. For distributions with compact descriptions, calculating TV-distance poses significant challenges, as the direct approach incurs exponential time complexity relative to input size. The authors present two main contributions:\n\n(1) They provide a polynomial-time algorithm for determining whether two mixtures of product distributions are identical. \n\n(2) They establish the computational hardness of approximating the TV-distance between two arbitrary Ising models.\n\nThe first result is based on a simple and clever algorithm, the second result comes from the standard hardness result of approximating partition functions of Ising models. \n\nDue to the limited time for reviewing, I went through all the proofs and understood the main ideas; however, I did not verify every detailed calculation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The hardness result follows from standard results. Proposition 6 provides a self-reduction for the Ising model, which is used in the standard counting-to-sampling reduction. Therefore, the proof of Proposition 6 could be omitted. Proposition 8 essentially states that one can fix the value of a vertex $v$ by adjusting the function $h(v)$, allowing the TV distance to encode the marginal distribution.\n\nThe relationship between Theorem 1 and Theorem 2 is not very strong, as they pertain to different 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": 3 }, "primary_area": null, "questions": { "value": "Can the equivalence checking algorithm be extended to other type of mixtures, say Bayesian networks with bounded treewidth?\n\nIs there any hope to soften the hardness assumption in estimating TV distance between Ising models by considering restricted classes of models?\n\nDo the authors have some recommendations for practical algorithms that could approximate the TV distance between Ising models despite the hardness result?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper addresses important computational questions about the total variation distance, which is fundamental in probability and statistics.\n\nThe algorithm for equivalence checking of mixtures of product distributions is new and provides a practical solution to a non-trivial problem.This hardness result bridges complexity theory and statistical measures and provides insight into why certain computational tasks are hard.\n\nThe proofs are well written, the results are accessible." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper discusses computational aspects of the total variation (TV) distance between probability distributions and has two contributions:\n\nIt provides a deterministic polynomial-time algorithm for testing whether two mixtures of product distributions are equivalent, which means the algorithm can decide whether the TV distance between them is zero.\n\nIt demonstrates that, unless NP is contained in RP, there cannot exist an efficient algorithm to estimate the TV distance between arbitrary Ising models. This result points out the computational hardness of this problem." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It would be nicer if the paper could elaborate more on the practical applications of the equivalence checking algorithm with regard to performance on real-world data.\n\nThe hardness result could also be pushed further by thinking about the possibility of approximate algorithms with different complexity assumptions.\n\nIt would be even more applicable and helpful with more examples or case studies." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Was a poly-time randomized algorithm for checking equivalence of mixtures of product distributions known previous to this work?\n- To the best of your knowledge, what was known from previous work about approximating TV distance between Ising models?\n- Do your hardness result tell us anything about hardness of computing an additive approximation for the TV distance between Ising models?\n- Does the problem of estimating the TV distance between Ising models remain to be hard if one is given the values of the partition function $Z_1$ and $Z_2$ for each of the Ising models?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- Mixtures of products are a fundamental family of probability distributions and checking their equivalence is one of the most basic questions about them. \n- The algorithm uses an interesting novel idea of keeping track of bases for the solution spaces of certain equations. This idea might find applications for testing equivalence of other classes of distributions.\n- The problem of estimating the total variation distance between two Ising model distributions is quite natural, as Ising models are a very well-studied class of probabilistic models.\n- The hardness result only relies on the assumption that NP is not in RP, which is a very mild complexity assumption." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper has two main contributions: (1) An algorithm for checking equivalence for mixtures of product distributions (2) A hardness reduction for approximating the total variation distance between two Ising models.\n\nThe problem of checking equivalence of product distributions is takes the following as input: we are given $w_1,\\cdots, w_k, P_1,\\cdots,P_k$ and $v_1,\\cdots, v_k, Q_1,\\cdots,Q_k$ where:\n1) Each of $P_1,\\cdots,P_k$ and $Q_1,\\cdots,Q_k$ is a product distribution over $\\Sigma^n$.\n2) $w_1,\\cdots, w_k$ are the weights for $P_1,\\cdots,P_k$ and $v_1,\\cdots, v_k$ are the weights for $Q_1,\\cdots,Q_k$.\nSo, the first distribution $P$ is the mixture of $P_1,\\cdots,P_k$ each weighted by $w_1,\\cdots, w_k$ and the second distribution $Q$ is the mixture of $Q_1,\\cdots,Q_k$ each weighted by $v_1,\\cdots, v_k$. The task is to determine wether the two mixture distributions are the same.\n\nHere is an example (from page 2) an equal mixture of $Bern(1/3)$ and $Bern(2/3)$ equals to the distribution $Bern(1/2)$. This illustrates how a mixture of two product distributions can equal a different product distribution.\n\nThe paper gives a deterministic algorithm whose run-time is $O(nk^4 |\\Sigma|^4)$ which is polynomial in the input size which equals $kn|\\Sigma|$. One idea the algorithm uses is that if the two mixtures $w_1P_1+,\\cdots, +P_kw_k$ and $v_1Q_1+,\\cdots, +v_k Q_k$ are equal, then (after a normalization) the mixtures $w_1P_1+,\\cdots, +P_jw_j$ and $v_1Q_1+,\\cdots, +v_j Q_j$ are also equal for every $j$. The algorithm utilizes the idea by iteratively establishing equivalence for $j=1,2..., k$ at every step $j$ establishing equivalence for $j+1$ assuming equivalence for $j$ (or to find some $x$ for which the probabilities of two distributions differ). As the paper shows, such steps can be achieved by keeping track of bases for solutions of certain types of equations.\n\nIsing models are a fundamental family of probability distributions over $\\{\\pm 1\\}^n$. Probability of each $x$ in $\\{\\pm 1\\}^n$ is proportional to $\\exp(P(x))$ where $P$ is a degree-2 multilinear polynomial. Ising models are a very well-studied family of distributions modeling systems for which random features have only pairwise interactions (this is because each term in $P$ has at most two variables). \n\nA Fully Polynomial-time Randomized Approximation scheme FPRAS is a randomized algorithm that gives a multiplicative $(1+\\epsilon)$-approximation to a desired quantity. Here, this means that one is given a pair of degree-2 polynomials $P_1$ and $P_2$ describing a pair of Ising models, and the goal of the algorithm is to output a multiplicative $(1+\\epsilon)$-approximation to the TV distance between the pair of Ising models. The paper shows that no poly-time algorithm can achieve this task (under a basic complexity-theoretic assumption).\n\nThe hardness proof proceeds by developing an approximation-preserving reduction to the problem of approximating the partition function of an Ising model, which is known from previous work to be hard to approximate. As an intermediate step, the reduction goes through the problem of approximating the marginal of an Ising-model distribution on one of the coordinates $x_i$ (this is referred to as the problem of approximating the atomic marginal)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The algorithm for mixtures of product distributions can only check whether P=Q exactly. The paper would be stronger if it gave an algorithm for approximating the distance between P and Q.\n- The paper rules out FPRAS for TV distance between a pair of Ising models, but it seems that there could still be a constant-factor approximation algorithm, and the paper would be stronger if this question was also addressed (i.e. it was shown that this is also hard, or an algorithm was given)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "The paper mentions some open questions but is there any chance that the techniques used, e.g., in the first theorem can be extended to other types of more general distributions? (I personally doubt it)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper give two new results about computing TVD. \nThe polynomial time algorithm for TVD of mixtures of product distribution has its main strength in the simplicity and clarity of the approach.\nThe second result extends the studies on the complexity of dealing with Ising models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies some properties about the complexity of computing the Total Variation Distance between distributions. The authors consider the case of the mixure of product distributions and the case of Ising models. In the first case they show a polytime algo that has access to the marginals and checks the equivalnece of two mixtures. For the second case they show hardness of approximating TVD by an FPRAS under the hypothesis NP not included in RP.\nThe two results are sound and well written. \nHowever, the contribution is limited. \nI think neither of the two results are very significant alone and together the overall contribution is not very much more." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper appears to be a gluing of two minor results with little connection between them. \nThe second result builds upoon the Jerrum and Sincler previous analogous study. The first result bears more novelty, although it would not be, in my opinion, sufficient for a paper at ICLR.\nI think the main issue is with the specificity (very particular cases) of the two problems solved." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "see above" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies two fundamental questions about high-dimensional probability distributions.\n\n1. Is there a polynomial-time algorithm which can decide whether two *mixtures of product distributions on the hypercube* are equal to each other?\n2. Is there a polynomial-time algorithm which can accurately approximate the total variation distance between two Ising model distributions?\n\nThe paper answers (1) affirmatively, giving a very nice linear algebraic algorithm which can test equivalence between mixtures of product distributions on the cube, given access to the parameters of the distributions.\nThe paper answers (2) negatively, showing that unless NP is contained in RP (that is, unless NP-complete problems have randomized polynomial time algorithms), the total variation distance between Ising models is inapproximable.\n\nStrengths:\nThe paper studies truly fundamental problems which will be of broad appeal to the ICLR audience.\nThe results are convincing, and the equivalence-checking algorithm is very nice.\n\nWeaknesses:\nThese problems are so fundamental that it is a little hard to believe they are not already well studied.\nIt would be nice if the related work section were expanded.\nIt would be nice if the algorithm also meant you can test equivalence given samples.\n\nOverall, I recommend accepting the paper to ICLR, on the basis that the results are about a very fundamental set of problems, and the equivalence testing algorithm is very clean.\n\n\nQuestions:\n- Does result (1) mean you can also do equivalence testing from samples?\n- Is the linear algebraic method you describe similar to any existing equivalence tester in the literature?" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "see above" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024computational,\ntitle={Computational Explorations of Total Variation Distance},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xak8c9l1nu},\nnote={under review}\n}" }, "abstract": { "value": "We investigate some previously unexplored (or under-explored) computational aspects of total variation (TV) distance.\nFirst, we give a simple deterministic polynomial-time algorithm for checking equivalence between mixtures of product distributions, over arbitrary alphabets.\nThis corresponds to a special case, whereby the TV distance between the two distributions is zero.\nSecond, we prove that unless $\\mathsf{NP} \\subseteq \\mathsf{RP}$ it is impossible to efficiently estimate the TV distance between arbitrary Ising models, even in a bounded-error randomized setting." }, "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": [ "total variation distance", "TV distance", "mixtures of products", "equivalence checking", "Ising models", "computational complexity", "FPRAS" ] }, "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/19c037430de9aa16c97d445fac7859a894c9b5ad.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Computational Explorations of Total Variation Distance" }, "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 description of the platform and recruitment process for human annotations is unclear. Who are the annotators (e.g., crowdworkers), and what are the recruitment criteria?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The provided analysis regarding the LLM judges' sensitivity to prompts, error types, a lack of robustness, and the leniency bias are interesting and valuable to future studies.\n\n2. The paper is well-written and the findings are clearly presented." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work provides an examination of LLM judges regarding their performance and vulnerabilities in a reference-based evaluation setting for QA tasks. Using human annotation as the gold standard, a series of judge models are evaluated. For evaluation metrics, the manuscript proposes using Scott’s $\\pi$ instead of accuracy, highlighting it as a main finding. It also shows that while less capable judges perform poorly at the instance level, i.e., giving the same decision as the human annotators, they achieve higher correlation with humans at the system level, i.e., producing a ranking of evaluated models by aggregating the instance-level decisions. Further analyses are conducted on changes in recall and precision scores, sensitivity to prompts, robustness, and leniency bias." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It appears that some of the main findings in this work are either not well-supported, may lack generalizability, or have been discussed in previous work.\n\n1. Lack of generalizability: The task setting of the LLM judges selected in this work is reference-based evaluation of QA, which differs from the common application scenario where LLM judges evaluate various tasks without a gold reference (e.g., AlpacaEval, Arena Hard). Access to gold references makes the evaluation task significantly easier. Therefore, the findings in this work may not generalize well to more open-ended, general evaluation settings. While it is stated that this task setting was chosen to reduce human disagreement, there exists a related dataset, LLMBar [1], which can be used to perform a more general evaluation of LLM judges, achieving over 90% human agreement (evaluation accuracy).\n\n2. The finding that automatic evaluation metrics (specifically LLM judges) have a higher correlation with human evaluations at the system level than at the instance level has already been identified and well-discussed in related work on evaluating automatic evaluation of natural language generation tasks [2][3][4][5]. For example, [5] shows that automatic metrics can achieve higher system-level correlation with humans when they evaluate more instances. Therefore, this finding itself is not a novel contribution.\n\n3. The manuscript proposes using Scott’s $\\pi$ instead of accuracy as the evaluation metric for LLM judges, claiming that it \"appears to be better able to discriminate among various judge models.\" However, this claim is not well-supported, as the only evidence provided is that Scott’s $\\pi$ yields scores with a wider numerical range than accuracy, which could potentially be achieved by trivially rescaling the range. Further examination is needed to verify this claim, such as by demonstrating that Scott’s $\\pi$ offers greater statistical power, with tighter confidence intervals or a lower p-value in significance tests. Additionally, the notion of separability defined in Arena Hard [6] would be useful for comparing evaluation metrics.\n\n4. The finding that the true negative rate (resulting in a lower recall score) falls quickly with less capable judges does not hold when the two lexical-similarity-based metrics, exact match (EM) and Contain, are excluded. In fact, all the small LLM-based judges achieve higher recall scores than precision scores. The observed low true negative rate/recall score of EM and Contain is expected, as these metrics rely on lexical similarity and are likely to mark an answer that is correct but lexically different from the reference answers as incorrect.\n\n\nReferences\n\n[1] Zeng, Zhiyuan \"Evaluating large language models at evaluating instruction following.\" ICLR 2024\n\n[2] The price of debiasing automatic metrics in natural language evalaution (Chaganty et al., ACL 2018)\n\n[3] Re-evaluating Evaluation in Text Summarization (Bhandari et al., EMNLP 2020)\n\n[4] A Statistical Analysis of Summarization Evaluation Metrics Using Resampling Methods (Deutsch et al., TACL 2021)\n\n[5] Re-Examining System-Level Correlations of Automatic Summarization Evaluation Metrics (Deutsch et al., NAACL 2022)\n\n[6] Li, Tianle, et al. \"From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline.\" arXiv preprint arXiv:2406.11939 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The error analysis in Table 2 shows judges struggle with under-specified answers. Could you provide examples of or qualitatively explain the under-specified answers that fooled even the best judges?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- They focus on a specific scenario with high inter-human agreement which is an attempt to isolate the judge model behavior from task ambiguity. \n- Several dimensions are explored: 1) model sizes and families, 2) Multiple metrics, 3) Error analysis provided. \n- Insights such as ``smaller models can rank exam-takers as effectively as larger ones'', and the attempted explanation that \"chat models may \"unlearn\" some knowledge during alignment\"; \n- The work also provides some recommendations for practitioners using LLM as judges, e.g. using Scott's $\\pi$ along with accuracy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper offers a study of LLMs-as-judges. The authors investigate 13 models (2B to 70B), evaluating 9 different \"exam-taker\" models on the TriviaQA benchmark. They found 1) Only the largest models achieve reasonable alignment with humans, though still falling short of inter-human agreement, 2) Scott's π provides better discrimination between judges than percent agreement, and 3) Even models with lower alignment scores can effectively rank exam-taker models. Through detailed analysis, the paper uncovers several vulnerabilities in judge models, including sensitivity to prompt complexity and a tendency toward leniency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The scope remains limited to TriviaQA. For \"short, factual\" answers, consider adding the \"LLMBar\" datasets, which have high human agreement rates > 90%. Sufficient examples can be used according to your dataset selection criteria [1]. Without the inclusion of additional datasets [1], it remains unclear how well the ranking ability would transfer. \n- The original claim (line 316-318) about judge performance being worse at identifying correct answers could be an artifact of including metrics that are overly strict about exact wording matches rather than semantic meaning. The finding does not appear to be surprising or novel. \n- Further analysis would be beneficial: show example outputs from each judge model and identify common errors; In Appendix I, where the authors justify the sample size, adding a power analysis would be ideal.\n\nReferences:\n\n[1] [Evaluating Large Language Models at Evaluating Instruction Following](https://arxiv.org/pdf/2310.07641) \n[2] [The NarrativeQA Reading Comprehension Challenge](https://arxiv.org/pdf/1712.07040)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "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 explanation of why Scott's Pi should be used instead of Kappa in judging-the-judges scenarios is a significant contribution that will benefit future researchers.\n* The comprehensive analysis across multiple dimensions (alignment metrics, ranking correlation, error analysis) provides valuable insights into the strengths and limitations of different judge models.\n* The comparison between LLM judges and lexical judges (EM, Contains) offers a novel and important perspective. This insight becomes increasingly critical as NLP tasks grow more complex, helping inform efficient evaluation strategies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper evaluates the LLM-as-a-judge paradigm by comparing 13 judge models against human evaluations on TriviaQA, focusing on a scenario with high human agreement to isolate judge performance issues. \n\nThe results show that while the largest models achieve reasonable human alignment, they still fall notably short of human-human agreement levels and exhibit various vulnerabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The evaluation relies solely on TriviaQA, making it difficult to deconfound the root cause: whether the best model's performance stems from better alignment, knowledge of TriviaQA content, or simply being favored by other LLMs. Other unusual findings may also be specific to TriviaQA: in Figure 1.a, EM's instability compared to Contains likely results from references providing multiple correct answers.\n* The paper lacks sufficient content for an ICLR long paper. I suggest expanding the scope by:\n * Including more evaluation datasets covering other types of tasks, such as objective long answers (e.g., code writing), using LLM judges to rank exam takers, etc.\n * Moving Appendix B (issues with Kappa) to the main paper and adding more experiments and analysis. This lesser-known fact would make the paper more informative and impactful." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* Did the authors consider other datasets, or non-binary notions of answer quality?\n* Did the authors consider evaluating the alignment across models to understand how they might be ensembled to mimic a better judge?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "* The paper is clear and nicely visualizes the relevant findings\n* The authors explore a dozen models as judges\n* The authors use manual annotation to carefully unpack the judge behavior, especially the observation about judge leniency" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors examine models as judges on TriviaQA questions, comparing how often humans agree with their answers. While they find alignment is high, they can use Scott Pi metrics to distinguish the quality of the judges. They also find models are frequently lenient judges in a binary setting of “correct” vs “incorrect”." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* As a reader I'm having difficulties understanding the overarching goals of the paper. Typically researchers use LLMs as a judge for longer form, more subjective questions, where answer coherence, style, and correctness are all part of the judgement. But for TriviaQA, the chosen dataset, the questions have clear, short answers with reference documents, meaning Exact Match is already a strong metric. Here, humans are simply reporting the binary value “correct” vs “incorrect” on the model answers, which seems to have little to do with “human alignment” and more to do with which model got the right answer? Could the authors provide more information on how they think these insights may or may not generalize to more more complex judging tasks, as well as discuss the limitations of their findings?\n* The choice of TriviaQA is extremely relevant to the reported results. Could the authors justify this choice? And have they considered comparing their results on other types of datasets?\n“well-aligned models tend to produce more false negatives than false positives.” This does not seem to be supported by Figure 3b, where the most correct model’s errors are mostly false positive? Could the authors please provide details to explain this—in case I am misunderstanding?\n* Could the authors add more details discussing the extent to which their contributions are novel and provide the community with actionable? A discussion section would be helpful here." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Do you think the results generalize to different types of datasets/tasks?\n\nWhat do you think it's an acceptable agreement percentage?\n\nWhat are the takeways from this study? What are the best practices that can be adopted when using LLM-as-a-judge?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "* Thorough and timely study\n* Several interesting experiments" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a study on the performance of LLM-as-a-judge expressed as agreement with human assessment. The LLMs that are judged are run in an exam-taker context, using TriviaQA dataset. The study shows that only the largest and more recent Lllama-3.1 models approach human judgement. The rest observe widely different scores, some as low as 60% agreement. Different prompting styles seem to make a difference, with single digit performance improvements for the performant models.\n\nI think it would be nice to summarize all the takeaways in a section and best practices for doing assessment using LLM-as-a-judge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* I would have liked to see more datasets; in fact, I would suggest reducing the number of exam-taker (e.g., I find the base models less interesting) and use different datasets" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We conduct a comprehensive study of the LLM-as-a-judge paradigm in a relatively controlled setup and report many interesting findings about its strengths and weaknesses" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024judging,\ntitle={Judging the Judges: Evaluating Alignment and Vulnerabilities in {LLM}s-as-Judges},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xam3sR3ffY},\nnote={under review}\n}" }, "abstract": { "value": "Offering a promising solution to the scalability challenges associated with human evaluation, the LLM-as-a-judge paradigm is rapidly gaining traction as an approach to evaluating large language models (LLMs). However, there are still many open questions about the strengths and weaknesses of this paradigm, and what potential biases it may hold. In this paper, we present a comprehensive study of the performance of various LLMs acting as judges, focusing on a clean scenario in which inter-human agreement is high. Investigating thirteen judge models of different model sizes and families, judging answers of nine different ‘examtaker models’ – both base and instruction-tuned – we find that only the best (and largest) models achieve reasonable alignment with humans. However, they are still quite far behind inter-human agreement and their assigned scores may still differ with up to 5 points from human-assigned scores. In terms of their ranking of the nine exam-taker models, instead, also smaller models and even the lexical metric contains may provide a reasonable signal. Through error analysis and other studies, we identify vulnerabilities in judge models, such as their sensitivity to prompt complexity and length, and a tendency toward leniency. The fact that even the best judges differ from humans in this comparatively simple setup suggest that caution may be wise when using judges in more complex setups. Lastly, our research rediscovers the importance of using alignment metrics beyond simple percent alignment, showing that judges with high percent agreement can still assign vastly different scores." }, "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": [ "LLMs", "NLP", "LLM Evaluation", "LLM-as-a-Judge", "Benchmarks" ] }, "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/90088059d25efa77ec389180e2148d22173ae320.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/1ee7874cf919a1d4a4559eedaf35d180c35a2bf3.zip" }, "title": { "value": "Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges" }, "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": { "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": "- What do the abbreviations ‘MC’ and ‘SA’ mean in Table 1?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Proposing a multi-modality VLM benchmark is meaningful and has basic value, which can greatly help the communities of chip design and LLMs.\n- The benchmark is notably challenging due to the diverse types of visual content, allowing significant room for research and improvement.\n- This paper offers five distinct data categories in data collection and introduce their details, enhancing the diversity of the benchmark." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a benchmark in terms of chip design. Such benchmark is characterized as VQA tasks, parallel to existing text-based tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- ChipVQA contains only a total of 142 samples, which limits its scalability and utility. Specifically, there is respectively only one sample each for certain categories (flow, equations, neural nets), reducing the benchmark's effectiveness for evaluating specific content types.\n- As a benchmark, this paper lacks a further dev/test splitting, which restricts the flexibility for developers to conduct training and fine-tuning.\n- While the paper describes the benchmark as multi-modal, it only incorporates text and images. Although there are various types of visual samples, such as diagrams and graphs, all are treated as images in the experiments.\n- This paper lacks the discussion of an alternative aspect of VLMs, such as CLIP [1], which emphasize visual capabilities over language components.\n- Some typos: A missing space after \"ChipVQA\" in line 175, and a labeling error where \"Figure 1\" should read \"Figure 3.\"\n- The authors did not follow the citation instruction, as citations should be in parenthesis when the authors or the publication are not included in the sentence.\n\n[1] Radford, Alec, et al. \"Learning transferable visual models from natural language supervision.\" International conference on machine learning. PMLR, 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": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I have some concerns about this benchmark.\n1. The number of QAs in this benchmark is only 146, which is too small to represent the five major class problems in the chip design. The small scale will make the evaluation results difficult to reflect the ability of VLM. The rare QAs belonging to specific small problems may lead to results that are not robust.\n2. The small benchmark could only be used for fast evaluation, which cannot be used to improve the chip design ability of VLM. Although the authors’ claim “demonstrates promising potential to enhance LLM/VLM problem-solving capabilities with minimal training overhead” in the conclusion, the small number of QAs in the benchmark makes this work not promising enough.\n3. The experiments cannot sufficiently support the effectiveness of this benchmark. The authors said “unlike existing benchmark efforts targeting at most undergraduate level engineering question Yue et al. (2024),”. However, the authors never compare the performance of VLM on the other benchmarks, such as Yue et al. (2024), with ChipVQA to justify its superiority.\n4. Experiment 4.1 seems to verify a common conclusion, that more knowledge will help VLM understand and reason. However, the low pass rate in Table 3 actually highlights the hardness of ChipVQA. Experiment 4.2 also seems to verify a common conclusion, that the higher image resolution will help improve the answer quality of VLM.\n5. The quotation marks are used in mistake in multiple places, such as “”Derive” on page 5, etc.\n\nIn conclusion, I think a large-scale high-quality VLM benchmark for chip design will be more attractive to the researchers. The authors could enlarge the limited number of QAs in the benchmark and provide more sufficient experiments and insights of applying VLM in chip design problems." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors proposed a new benchmark, ChipVQA, to evaluate the existing VLMs ability to understand and reason in chip design, which is a specific and important research area.\n2. ChipVQA is considerably challenging even for the most advanced VLM model, GPT-4o.\n3. The collected QAs span various chip design areas from abstract architecture design to realistic semiconductor manufacturing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors proposed a new benchmark, ChipVQA, to evaluate the existing VLMs ability to understand and reason in chip design, which is a specific and important area. ChipVQA is considerably challenging even for the most advanced VLM model, GPT-4o. Meanwhile, the collected QAs span various chip design areas from abstract architecture design to realistic semiconductor manufacturing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The limited number of QAs in this benchmark fails to adequately represent the chip design field and does not provide sufficient potential to support VLM development.\n2. The experiments are insufficient to demonstrate the superiority of this benchmark." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 4 }, "primary_area": null, "questions": { "value": "1.Whether the benchmark has some questions, which are not included by the anonymous repo, with higher quality?\n2.Whether the authors try to redraw the manuscripts of analog circuits by software like Visio to test the performance of VLM?\n3.Whether the authors test senior undergraduate students in majors related to Integrated Circuit on this benchmark as a comparison? I think only after this, they can say that majority of some previous benchmark primarily cover content up to the level of undergraduate engineering courses and their benchmark is an exception." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "ChipVQA is the first work which constructs a multiple fields and multi-modal benchmark in chip design and tests the performance of mainstream VLMs on the benchmark. Results reveal that this benchmark is challenging for current capabilities of VLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a comprehensive benchmark including 142 VQA questions covering five chip design disciplines: Digital Design, Analog Design, Architecture, Physical Design and Semiconductor Manufacturing. It is the first benchmark suite in the field of multi-modal chip design knowledge. Moreover, some experiments are implemented to test this benchmark, and it is proved that the benchmark is challenging for the capabilities of current VLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the authors emphasize the quality of their benchmark and say that these questions are developed by “seasoned chip design experts, each with over ten years of industry experience” and majority of some previous benchmark primarily cover content up to the level of undergraduate engineering courses, as a reviewer who majored in Integrated Circuit and System during undergraduate studies, I think the most questions of their benchmark showed in the anonymous repositories only have the level of undergraduate courses in majors related to integrated circuit. Moreover, most of the questions are common for students in majors related to Integrated Circuit. The authors only collect them together from textbooks, course exams, manuscripts and so on, and some of the images of the questions are rough, hastily written manuscripts which make it even difficult for VLMs to recognize the content.\n\nApart from the benchmark quality, there’s not much novelty in this work, because authors mainly test some mainstream VLMs on this benchmark. In the part where the authors demonstrate their discovery, the third, the forth, and the fifth are obvious. And the first and the second also doesn’t have much value, considering that some image paired with questions are just hastily written manuscripts." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present a benchmark suite for VLM on chip design and manufacturing knowledge" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024chipvqa,\ntitle={Chip{VQA}: Benchmarking Visual Language Models for Chip Design},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xao3fIJC6M},\nnote={under review}\n}" }, "abstract": { "value": "Large-language models (LLMs) have exhibited great potential to as-\nsist chip designs and analysis. Recent research and efforts are mainly\nfocusing on text-based tasks including general QA, debugging, design\ntool scripting, and so on. However, chip design and implementa-\ntion workflow usually require a visual understanding of diagrams,\nflow charts, graphs, schematics, waveforms, etc, which demands\nthe development of multi-modality foundation models. In this paper, we propose ChipVQA, a benchmark designed to evaluate the\ncapability of visual language models for chip design. ChipVQA includes 142 carefully designed and collected VQA questions covering five chip design disciplines: Digital Design, Analog Design, Architecture, Physical Design and Semiconductor Manufacturing. Un-\nlike existing VQA benchmarks, ChipVQA questions are carefully\ndesigned by chip design experts and require in-depth domain knowledge and reasoning to solve. We conduct comprehensive evaluations\non both open-source and proprietary multi-modal models that are\ngreatly challenged by the benchmark suit. ChipVQA examples are available at\nhttps://anonymous.4open.science/r/chipvqa-2079/." }, "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 LLM; Chip Design and Manufacturing; VQA" ] }, "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/00763d656f22547ec6b04f886275a34687498fde.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": "ChipVQA: Benchmarking Visual Language Models for Chip Design" }, "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": 2 }, "primary_area": null, "questions": { "value": "- In Line 55-58 in the introduction- 'To address these concerns, this study proposes evaluating LLMs using a multiple-choice question test based on entirely fictional medical knowledge. By doing so, we aim to determine whether traditional evaluations are sufficient for assessing the clinical knowledge and reasoning abilities of LLMs for the medical domain, free from the influence of pre-existing data.' This seems to indicate the motivation of the study. Can the authors say more? What were the hypotheses they wanted to study? Did they expect the model to perform poorly? and why? By creating a similar QA study on are they establishing anything different?\n\n- Is it possible to provide more information on the development of the 'Glianorex' textbook? How were the models steered to generate the content? \n\n- There is very limited information on how the models are set up experimentally- i.e. what was the input uniform across all 14 models given the differences in context windows? \n\n- How long were the paragraphs used for generating the questions/options? Is there a chance that not all the questions that were generated were about the fictional gland? \n\n- Can the authors comment more on why the generated questions were difficult for medical professionals? Is it because they were unfamiliar with the working of the fictional gland? or because the answers were linguistically hard to disambiguate? Can they provide examples?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The study generates a 'fictional' textbook on a fictional gland and uses the synthetic data to evaluate model performance on QA.\n- I found it fascinating that the doctors found no major flaws in the generated QA. \n- The study evaluates a variety of LLMs (in zero-shot setting)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript aims to assess the QA performance of LLMs on 'fictional' data. Authors generate a textbook about an inexistent gland, generate QA from paragraph fragments from this generated textbook and subsequently evaluate the QA using LLMs in a zero-shot setting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It is known that evaluations for using LLMs for healthcare are insufficient and incomplete. Newer, better ways of benchmarking them are necessary, for e.g., using real life patient data. LLMs, due to their size and scale are known to memorize their training data. LLMs performing well on QA benchmarks indicate they encode the knowledge (complete/incomplete). With this case study, the aforementioned evidence is only revalidated and makes a limited case for novelty. As mentioned in the future work section, may be interactive, scenario-based assessments are an interesting direction!\n\n- Line 216- 'By generating entirely fictional content, we ensured that no pre-existing data could influence the models’ - this seems hard to be convinced of- even if the gland is fictional, the model is exposed to data about the other organs/biological aspects of the human anatomy that play a role in relation to this gland. The LLM may also be exposed to information/vocabulary on how a gland is supposed to function and what their possible medical conditions that occur to other glands. This could also be a possible explanation as to why models show better performance than a random baseline.\n\n- Line 488 in conclusion: 'This study demonstrates that LLMs can achieve high scores on multiple-choice questions based on entirely fictional medical knowledge, even without prior exposure to the content.' I disagree that this knowledge is completely fictional. Although the gland is fictional, the information of how 'a gland' is supposed to functional is not fictional medical knowledge along with other information about the human anatomy.\n\n- Yes, the models was able to do well on ~80-90 questions. What about others? Could the authors discuss errors? Where did the models go wrong? Were the errors the similar to the ones highlighted in previous work in MedQA? Why were they hard? Did the medical experts find them easier? \n\n- I find the quality of the generated QA may not be completely equivalent/ comparable to existing work- \n Minor corrections\n- Line 457: missing text\n- Line 485: incomplete sentence" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "LLM Generation Chain:\n- How do you control for the fact that both content and questions are LLM-generated?\n- Have you considered comparing performance between LLM-generated questions and human expert-generated questions about the same fictional content?\n\nRelation to Known Limitations:\n- How does this work advance our understanding beyond existing studies showing LLMs lack reasoning capabilities (GSM8K, Physics of Language Models, etc.)?\n- Could you compare your findings about medical MCQ performance to similar pattern-matching behaviors documented in math, physics, and symbolic reasoning tasks?\n- What makes medical MCQs different from other domains where LLM pattern-matching has been studied?\n\nBenchmark Comparisons:\n- How does model performance on your fictional medical questions compare to performance on real medical benchmarks like USMLE?\n- Could you analyze if models exploit similar patterns in both real and fictional medical questions? e.g. Have you considered creating paired real/fictional questions with matched reasoning requirements to isolate the effect of domain knowledge vs. pattern matching?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The study evaluates a comprehensive range of models - from proprietary systems like GPT-4 to open-source models like Mistral and domain-specific medical models. The multilingual evaluation in both English and French adds valuable cross-linguistic insight. Testing across model scales (from 7B to 110B parameters) and comparing both base and medically fine-tuned versions provides a thorough performance landscape.\nThe core concept of testing reasoning gaps through fictional medical content is important given the high-stakes nature of healthcare applications. With the increasing deployment of LLMs in medical settings, understanding their limitations in medical reasoning versus pattern matching becomes critical for patient safety.\nThe experimental setup allowed for controlled model performance testing without the confounding variable of pre-existing medical knowledge. The methodological approach of creating a fictional medical domain to isolate reasoning capabilities represents creative thinking about AI evaluation challenges.\nThe statistical analysis is solid, including Cohen's d effect sizes and significance testing, provides some quantitative backing for the findings. \nCode is openly available and use of harness is good practice.\nThe authors are also transparent about the limitations of their approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a study examining the effectiveness of multiple-choice questions (MCQs) in evaluating Large Language Models' (LLMs) medical knowledge and reasoning capabilities. The authors devise an experiment where they use GPT-4 and Claude 3.5 Sonnet to generate textbooks about a fictional gland called \"Glianorex,\" along with corresponding multiple-choice questions in both English and French. They then evaluate various LLMs, including proprietary, open-source, and domain-specific models, on these questions in a zero-shot setting. The models achieve surprisingly high average scores of around 64%, despite having no prior knowledge of this fictional medical content. \n\nThis points to a significant methodological concern: since both the content and questions were generated by LLMs, the high performance likely demonstrates LLMs' ability to recognize and reconstruct patterns in LLM-generated text rather than any genuine medical reasoning capabilities. While the authors interpret their results as evidence that MCQ-based evaluations may not adequately assess medical knowledge, their experimental design inadvertently reveals more about LLM-to-LLM pattern recognition than about the limitations of multiple-choice testing in medical AI evaluation. This highlights a broader issue in AI evaluation methodology, where the tools used to test AI systems may be inherently biased towards the systems' pattern-matching capabilities rather than their actual understanding or reasoning abilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. LLM Chain of Generation:\nThe experimental pipeline (textbook → questions → answers) is LLM-generated, creating a closed loop that primarily tests LLM-to-LLM pattern recognition. This begs the question if this tests more than other approaches that have shown models fragility to lexical substitution or paraphrasing methods.\n\n2. Well-Established Lack of Reasoning:\nMultiple studies have already definitively shown LLMs don't perform actual reasoning:\n- GSM8K and other math reasoning benchmarks show LLMs struggle with novel mathematical reasoning\n- \"Physics of Language Models\" work shows LLMs operate through pattern matching rather than causal understanding\nThis study doesn't meaningfully advance beyond these existing findings. \n\n3. Missing Comparisons and related work:\n- No comparison between performance on these fictional medical questions versus real medical benchmarks (like USMLE)\n- No comparison to human-generated alternative versions e.g. drug name substitution benchmarks \n- Related work should include work on lexical substitution https://aclanthology.org/D14-1066.pdf , dataset contamination https://arxiv.org/abs/2404.18824 and robustness e.g. https://arxiv.org/abs/2406.06573, https://arxiv.org/abs/2406.12066\n\n4. Missed Opportunity for Medical Evaluation:\nInstead of showing that LLMs don't reason (which we know), they could have investigated what specific patterns in medical MCQs make them vulnerable to exploitation by LLMs, which would be more useful for improving medical testing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Is there existing work on benchmarks for specific medical domains, such as oncology? Including one in the related work section could enhance context.\n\n2. In the discussion's training paragraph, you mention “the improved performance of finetuned models Internist.ai and Meerkat over their base versions underscores the impact of domain-specific training on enhancing LLM capabilities.” Could you clarify what you mean by \"LLM capabilities\" in this context?\n\n3. Is the generated textbook open-sourced? If so, please provide a link to the anonymous version of the textbook." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The introduction of a fictional gland is a novel way to control for LLMs’ memorized real-world knowledge. The knowledge about the fictional gland is generated in a comprehensive and detailed way. \n\n2. Human experts were involved to ensure the quality of the generated medical knowledge and questions, and to provide a human performance baseline. Expert involvement makes the experiments more accountable and trustworthy. This is important for applications in healthcare and medicine because these are mission-critical scenarios.\n\n3. The study uses statistical analysis to provide solid evidence supporting its claims. \n\n4. The inadequacy of MCQ evaluation method is a long-standing and important issue, especially in clinical scenarios. The investigation is well-motivated. In addition, the authors clearly outline the motivation for the study and the implications of their findings, making it easy for readers to grasp the significance of the research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses an key problem: the inadequacy of current multiple-choice question-answering evaluation methods for Large Language Models (LLMs) in medical contexts. It highlights that LLMs' test-taking abilities, i.e., their capability to rely on statistical patterns rather than understanding, impact their performance on medical benchmarks. This phenomenon is particularly relevant in settings where LLMs may select correct answers based on learned statistical associations rather than true comprehension. To demonstrate this, the authors devised a test based on fictional medical knowledge, effectively removing the influence of prior knowledge for the models. The results indicate that LLMs still perform well on questions about fictional medical knowledge. The authors claim that their finding suggest that LLMs have good test-taking ability that prevent current evaluation methods from effectively assessing the true clinical reasoning capabilities of these models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The claim that \"entirely fictional content\" removes pre-existing data influence is overstated. While the fictional setup minimizes real-world knowledge's impact, it does not fully eliminate it. The generated content of the fictional gland using LLMs stills adhere to the knowledge framework of medicine. Thus, there is pre-existing foundational medical knowledge from LLMs embedded in the content and questions even though no pre-existing data specific for the fictional gland exists. LLMs’ reasoning ability with prior knowledge is not entirely ruled out. This claim, central to the paper’s message, is thus an overstatement. This is a limitation of the method in isolating LLMs' test-taking abilities.\n\n2. The statistical analysis is appreciated, but the authors do not clearly explain the apparent disparity between the small and negligible performance differences derived statistically and the visible differences shown in Figure 1. This may confuse readers unfamiliar with advanced statistical analysis.\n\n3. A known issue with multiple-choice evaluation is selection bias/position bias for option order and IDs. Studies show LLMs may have positional preferences (e.g., favoring option A), making it necessary to control for option order, as positional bias could affect model performance significantly. The authors do not address this issue.\n\n4. The discussion on the distribution of correct answers and models' \"inferential strategies\" lacks clarity. Terms like \"inferential strategies\" are uncommon in this field and require further elaboration. Additionally, the paragraph addressing language in the discussion section could benefit from better clarity.\n\n5. While the paper highlights issues with current multiple-choice evaluation, it does not propose and develop solutions. A method that can mitigate the issue and suggest potential paths forward would make the contribution more impactful, moving beyond identifying a problem to offering ways to address it.\n\n6. Presentation flaws are present, such as incomplete sentences in the limitations section, which disrupt readability. For instance, the \"Knowledge coherence\" paragraph in the limitations section is unfinished, and the \"Model selection\" paragraph in the same section ends abruptly with \"This selection.\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 was not entirely clear whether the exam-takers (humans and language models) had access to the textbook prior to answering the questions or was the textbook only used for generating QA pairs. Additionally, did each question provide any context at all?\n* It would be useful if the authors could provide more information on the statistical analysis carried out including:\n * How was the standard error of the mean computed? The reported results are for accuracy so what was the mean taken over? \n * Additionally, how was this computed for the human evaluation? The error bars were smaller for the human results which is surprising given that these were based on a smaller sample of questions\n * How was the 2-way ANOVA carried out? What were the 2 factors used?\n* The authors provide some proxy metrics for alignment between model’s answers, such as performance differences, however why did the authors not test answer alignment more directly, using a metric such as Cohen’s kappa? This would allow significantly more insight into how similar performance was between models.\n* I’m not sure how informative figure 2 is. If performance of most models was >0.6 accuracy it’s somewhat given that the correct answers would be skewed towards a majority correct. Using metrics such as Cohen’s kappa or a similar approach would be more informative to understanding model agreement. Additionally, this figure would be easier to understand if the x axes were the same scale.\n* Why did the authors provide no comparison of human answers with model answers? Given that they only used 100 questions as their sample, it’s possible the sample used to test human performance was a harder subset of questions.\n* It would be helpful to see more in-depth analysis of the QA pairs the models got correct and incorrect. For example, were correct answers more likely for specific sub-topics? Were questions based on particular textbook sections more likely to be answered correctly?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors take a novel and interesting approach to assessing the extent to which language models rely on internal knowledge in benchmark evaluations, including expert evaluation of their synthetic data to ensure the validity of the generated questions. The results may have important implications for the field of language model evaluation and certainly warrant further investigation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper the authors generate a multiple-choice QA dataset from a synthetic textbook based on fictional medical knowledge. The authors find that despite performance in humans being at expected chance levels, all language models tested display performance significantly above-chance. This implies that language-model performance on various benchmarks might be the result of general exam-taking ability, rather than testing actual medical knowledge. If true, this could be an important finding to the field of language model evaluations and would warrant further investigation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The results presented here are only for questions in a single domain, related to questions for a synthetic textbook on only a single fictional topic. It is not clear whether these results would generalise to other domains. Additionally, it is possible that the language models used to generate the synthetic text may have imputed this data with real knowledge, which could partially explain the strong performance for every language model tested." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Evaluating LLMs with fictional medical benchmarks reveals traditional MCQs assess pattern recognition over clinical knowledge, indicating a need for improved evaluation methods." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024multiple,\ntitle={Multiple Choice Questions and Large Languages Models: A Case Study with Fictional Medical Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xawA8X5dHq},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) like ChatGPT demonstrate significant potential in the medical field, often evaluated using multiple-choice questions (MCQs) similar to those found on the USMLE. Despite their prevalence in medical education, MCQs have limitations that might be exacerbated when assessing LLMs. To evaluate the effectiveness of MCQs in assessing the performance of LLMs, we developed a fictional medical benchmark focused on a non-existent gland, the Glianorex. This approach allowed us to isolate the knowledge of the LLM from its test-taking abilities. We used GPT-4-Turbo and Claude 3.5 Sonnet to generate two comprehensive textbooks on the Glianorex in both English and French and developed corresponding multiple-choice questions in both languages. We evaluated various open-source, proprietary, and domain-specific LLMs using these questions in a zero-shot setting. The models achieved average scores around 64%, with minor performance differences between larger and smaller models. Performance was slightly higher in English than in French. Fine-tuned medical models showed some improvement over their base versions in English but not in French. The high performance across models suggests that traditional MCQ-based benchmarks may not accurately measure LLMs' clinical knowledge and reasoning abilities, instead highlighting their pattern recognition skills. This study underscores the need for more robust evaluation methods to better assess the true capabilities of LLMs in medical 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": [ "large language models", "medicine", "benchmark", "evaluation", "clinical knowledge", "multiple choice questions" ] }, "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/7f4069de689d52997213b9d8ffd02b4934f64f94.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/95a8ef51676e40de7866efa441e87c94e9606049.zip" }, "title": { "value": "Multiple Choice Questions and Large Languages Models: A Case Study with Fictional Medical 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": 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. Existing equivariant neural networks with geometric inductive biases have been outperformed, both in terms of performance and efficiency, by Transformer-style architectures (ie, something as simple as torch.nn.TransformerEncoder). Was any ablation done that compares Deep Signature with such architectures that \"tokenize\" the trajectory intervals? \n\n2. I'm curious what the learned coarse-grained beads/maps look like for GPCR proteins – is it creating CG beads only based on atoms within a locality? How does this compare to naively taking atoms within some radial ball and considering them a bead? More often than not, these naive CG choices work well in practice, without the need to overcomplicate it with a learnable method. I'd like to see whether this \"learned GCN-based coarse-graining\" is really necessary.\n\n3. How sensitive is the path signature transform to very dynamic conformational changes in a short duration of time? For instance, as I mentioned above, fast-folding proteins undergo significant changes in 3D shape in just short simulations. Can this method capture this dynamics well enough?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Incorporates appropriate symmetries for (temporal) 3D point cloud learning. In particular, the invariance to time reparameterization is key because the underlying MD simulations might be prone to random restarts and miscellaneous artefacts preventing them from being a smooth \"video\" (ie, the MD itself might be erratic with the same state sampled repeatedly).\n\n2. Deep Signature _learns_ the ideal CG beads relevant to the task, bypassing the need to manually remove degrees of freedom (eg: CA-level coarse-graining, as done in existing methods). The use of the signature transform allows for local and global temporal interactions to be captured well, as opposed to learning representations on separate frames. \n\n3. I'm confident this method can be used for good representation learning of trajectory information for tasks beyond the ones mentioned in the paper (eg: interpolation in the latent space of protein conformations, perhaps for ensemble generation).\n\n4. Demonstrates relatively good performance compared to weak and strong temporal graph learning baselines (like GraphLSTM)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Deep Signature, a framework for characterizing dynamics on graphs. Deep Signature involves three key components: (1) graph coarsening using a deep spectral clustering module, (2) computation of path signatures to capture global inter-node interactions using iterated integrals over time, and (3) a two-layer MLP for property prediction.\nThe authors tested Deep Signature on three datasets: (1) gene expression dynamics, classifying into degradation or dimerization types, (2) GPCR dynamics, distinguishing between active and inactive states, and (3) EGFR mutant dynamics, predicting drug sensitivity. They conducted an ablation study validating the contribution of each loss component in the model and performed limited comparisons with methods based on static structures, dihedral angles, and GraphLSTM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper does not clarify the level of coarsening necessary to make path signature computations feasible. For instance, in gene regulatory dynamics, the graph was reduced from 100 nodes to 30, while the EGFR dataset was reduced from approximately 5000 atoms to 50 nodes, but the level of coarsening for the GPCR dataset is unspecified. This raises questions about whether 50 nodes is a computational limit for the method. Additionally, while protein structure can be intuitively coarsened into backbones, sidechains, and motifs, the interpretation for coarsened graphs in gene expression dynamics is unclear. It is uncertain if the coarsened nodes correspond to gene hubs or other biologically meaningful groupings.\n\nThe equation used to model GRN dynamics (Eqn 13 in Section B.1) appears incorrect. Specifically, for dimerization (when f=2), the concentration should should be squared under Michaelis-Menten kinetics, rather than simply doubling the decay rate. This could lead to inaccurate modeling of gene expression dynamics and affect the results in this section.\n\nThe comparisons used in the experiments are limited and static, primarily involving the first and last frames (head, tail, and head & tail) which do not capture temporal dynamics. The authors do not benchmark against dynamic approaches that consider time-varying information, such as MDTraj [1], Timewarp [2], and DSR [3]. Including these comparisons would provide a more rigorous evaluation of Deep Signature's effectiveness relative to established tools for molecular trajectory analysis.\n\nThe authors claim that the coarsened dynamics in Fig 6c follow the same trend as the original dynamics, yet this similarity is not quantified. Providing a quantitative measure, such as correlation coefficients between the original and coarsened dynamics at various coarsening levels, would better support this claim. Additionally, the paper would benefit from comparisons of the authors' coarsening strategy against other deterministic and learnable methods for protein graph coarsening to demonstrate its effectiveness and fidelity in preserving dynamics.\n\nThe description of the cross-validation process and test set creation is confusing. The authors mention “for each running, we evaluate the prediction accuracy of our method on an independent unseen test set and report the averaged results,” but it is not clear how this set is constructed. If results are averaged, please report the standard deviation in all the tables.\n[1] MDTraj: A Modern Open Library for the Analysis of Molecular Dynamics Trajectories. Biophysical Journal, 2015.\n[2] Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics. NeurIPS 2023\n[3] DSR: Dynamical Surface Representation as Implicit Neural Networks for Protein. NeurIPS 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. I think the classification task using evolution trajectories is not a common setting. Could you explain why this problem setting is reasonable comparing to the one frame classification setting? Do we really need the complete trajectory to do classification?\n\n2. I do not understand your setting in the EGFR dynamics experiment since I'm not an expert on this domain. Could you please explain why the trajectory can be labeled according to its sensitivity towards the drug? I think the sensitivity should be a number rather than a 0/1 label." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The presentation of the method is very clear and easy to understand.\n2. The motivation of applying signature transform is reasonable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In the paper, the authors introduce the Deep Signature framework to capture complex dynamics using the evolution trajectories. The Deep Signature framework includes spectral clustering, signature transform and a classifier. Additionally, the authors show that the framework satisfies the desired symmetry constraints. Experiments on gene regulatory dynamics, EGFR mutation dynamics and GPCR dynamics exhibit the empirical performance of the framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of experiments on large dataset. As the paper claims, the Deep Signature framework can capture large-scale complex dynamics. So I think experiments on datasets with large amount of data and system size are necessary. But the paper only includes the experiments on datasets with large system size.\n\n2. I think the baseline in this paper is too weak. For example, the author should compare the strong baseline with graph transformer architecture[1] based on the first/last frame of the trajectory. Comparations between these strong baseline may strengthen the empirical performance of the framework.\n\n[1].Ying, Chengxuan, et al. \"Do transformers really perform badly for graph representation?.\" Advances in neural information processing systems 34 (2021): 28877-28888." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. How were the hyperparameters chosen, such as the loss coefficient parameters $\\lambda_i$, the number of nodes in deep spectral clustering model, and time interval $[r_i, r_{i+1}]$ in path signature transform?\n2. When visualizing critical pathways and interatomic interactions on the EGFR dynamics, why were three atoms identified specifically? Could more than three be selected?\n3. Have the authors considered comparing their approach with advanced time series classification algorithms？\n4. How does the computational efficiency of the proposed method compare to baseline methods?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Authors develop an end-to-end framework to characterize interatomic interactions and dynamics of large-scale molecules, it shows improvement on three benchmarks and provides interpretability. \n2. The size of the system is reduced by deep spectral clustering module without any expert knowledge. \n3. The framework's desirable properties are supported by theoretical analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a framework, Deep Signature, designed to analyze the dynamics and interatomic interactions of large-scale molecular system. The method uses a deep spectral clustering model to capture coarse grained dynamics, a path signature transform module to characterize interatomic interactions through iterated integrals, and a classifier for property prediction. Theoretically, Deep Signature is shown to maintain desirable symmetry properties. The method demonstrates improved accuracy across three benchmarks and demonstrates interpretability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors compare their approach to baseline methods, but a more comprehensive comparison with some SOTA baselines would provide a more robust evaluation. \n2. The manuscript would benefit from a comparison of deep spectral clustering module with existing coarse graining methods. \n3. An analysis of the model’s sensitivity to hyperparameters would provide insights into its robustness and reproducibility." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024deep,\ntitle={Deep Signature: Characterization of Large-Scale Molecular Dynamics},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xayT1nn8Mg},\nnote={under review}\n}" }, "abstract": { "value": "Understanding protein dynamics are essential for deciphering protein functional mechanisms and developing molecular therapies. However, the complex high-dimensional dynamics and interatomic interactions of biological processes pose significant challenge for existing computational techniques. In this paper, we approach this problem for the first time by introducing Deep Signature, a novel computationally tractable framework that characterizes complex dynamics and interatomic interactions based on their evolving trajectories. Specifically, our approach incorporates soft spectral clustering that locally aggregates cooperative dynamics to reduce the size of the system, as well as signature transform that collects iterated integrals to provide a global characterization of the non-smooth interactive dynamics. Theoretical analysis demonstrates that Deep Signature exhibits several desirable properties, including invariance to translation, near invariance to rotation, equivariance to permutation of atomic coordinates, and invariance under time reparameterization. Furthermore, experimental results on three benchmarks of biological processes verify that our approach can achieve superior performance compared to baseline 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": [ "Molecular dynamics; representation learning; graph neural network; path signature" ] }, "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/a94e9f9ca2b6d60c25dcea316010ae4721139a73.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": "Deep Signature: Characterization of Large-Scale Molecular Dynamics" }, "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": "No." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Tensor Operation Approximation is used to reduce the communication overhead of FL.\nVarious test results across small datasets are demonstrated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The edge applications leveraging FL are highly constrained by resources, in terms of computation and communication. Existing efficient FL solutions either compromise accuracy or ignore memory usage. They introduce the ordered-layer-freezing and Tensor Operation Approximation for reducing memory footprint & communication overhead." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "layerfeezing technique is not new. Novelty seems to be marginal" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 overall contributions seem marginal and further evaluations are necessary to uncover the true potential and limitations of the proposed technique. Besides addressing the above-raised comments in the weakness section, following questions also need particular attention.\n\n1.\tWhat general guidelines or thresholds can be provided for setting the scaling factor s in TOA, beyond using a grid search, to balance accuracy and communication costs effectively? or maybe is there a way to determine the optimal value of s without using grid search?\n2.\tThe TOA technique in FedOLF uses a fixed scaling factor across all clients, which might not fit well for devices with different processing power (different hardware). Therefore, it is important to explore a more flexible, client-specific scaling factor to improve FedOLF's efficiency on diverse IoT devices?\n3.\tTo emulate system heterogeneity, why only the case of uniform clusters is considered? \n4.\tIn the context of Section 3.5, how many clients with full network training capacity are necessary for this technique to offer reasonable (and fast) convergence? Also, can this technique be used when most of the devices are significantly memory constrained. If yes, what are the limits? Also, how does the proposed technique compare with the state-of-the-art techniques under such scenarios. \n5.\tHow well the technique performs in cases with a skewed distribution of classes (and general categories) across devices? Are there any data-level assumptions that have to be valid for the technique to offer reasonable results? \n6.\tWhat is the impact of the proposed technique in cases with significant computational power imbalance between devices? What is the impact of such cases on the overall training time?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tApplying TOA to frozen layers preserves model accuracy while reducing the communication cost.\n\n2.\tThe experimental results of FedOLF on different iid and non-iid datasets and architectures demonstrate its adaptability and advantage in both energy and accuracy over baseline methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed approach, FedOLF, combines the layer freezing method and Tensor Operation Approximation (TOA) to reduce the memory and computation requirements of the training process in federated learning settings for the memory-constrained IoT devices/systems. Experimental results show better overall accuracy with low memory, and energy usage compared to prior works, for different datasets like EMNIST, CIFAR-10, CIFAR-100, and CINIC-10." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe paper offers limited novelty, as FedOLF's approach combines previously explored techniques, partial neural network freezing and TOA. The combination of existing techniques, although seems useful, requires more thorough evaluation to distinguish the scenarios and system configurations in which it outperforms the state-of-the-art techniques and scenarios and systems configurations in which it offers sub-optimal results. Specifically, how well the technique performs compared to state-of-the-art for system configurations in which most of the clients are resource constrained (in terms of memory and compute both)? Also, the paper should demonstrate the performance of the proposed technique for non-uniform clusters of devices. Additionally, the proposed technique should be evaluated for cases with a skewed distribution of classes. All these analyses will provide a detailed performance comparison with the state-of-the-art techniques and a better understanding of the significance of the novel contributions presented in the paper.\n \n2.\tFedOLF assumes that \"low-level layers across various local models usually have higher degrees of Centered Kernel Alignment (CKA) similarity,\" allowing frozen layers to generalize across clients without retraining. While Table 1 shows strong accuracy compared to the benchmark models in non-IID settings, supporting FedOLF's effectiveness, it does not directly validate this assumption of universal feature similarity. In real-world scenarios with highly diverse data (e.g., medical vs. satellite images), this may not hold true, and the frozen layers will become less useful for some clients. To validate this assumption, the experimentation should include more types of non-IID characteristics, such as label distribution skew, feature distribution skew, and quantity skew across clients. Additionally, a theoretical analysis of the conditions under which CKA similarity holds or fails, as well as experiments for validating this similarity across different layers and datasets, would be useful. Moreover, it could be interesting to perform a more in-depth analysis of the performance of the individual layers when considering the heterogeneity of the data; it may offer more detailed insights into the practical limits of the proposed approach.\n \n3.\tThe number of frozen layers is determined just based on the memory footprint of different options, and the computational requirements are not considered. The authors should highlight how this is scalable to systems composed of devices with diverse set of computational capabilities. Moreover, is considering just the memory footprint for real systems sufficient? If not, what modifications are required in the proposed technique, and how incorporating computational requirements into the layer freezing decision can improve the performance of the proposed technique for practical systems?\n\n4.\tBy freezing specific low-level layers, FedOLF implicitly assumes that the representation error from these frozen layers will diminish as training progresses. While the paper claims that these errors vanish empirically, a detailed analysis of error propagation through frozen layers seems missing. To address this, a layer-by-layer analysis of error propagation throughout training, using visualizations like heat maps to track error dynamics, is required. Additionally, experiments that vary the number and position of frozen layers could provide crucial insights into how these changes impact error propagation and overall model performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "it is not clear to me how the communication energy consumption is captured in this experiment. How do the authors provide their relative weights? \n\nRegarding Table 1. There is no evidence that the final achieved accuracy would be better in the proposed system compared to the rest. Is the target number of iterations that give this result? How the performance of the above algorithms change with the iteration number T?\n\nSomeone would expect the benefits of the method to become more profound as the number of layers in the network increases. However, Table 1 shows that this is not the case and the gap between the proposed method and STL closes (even for T=500). Can the authors comment on that?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The idea (and proof) of bounding the approximation of the true gradient through layer freezing under certain condition is important and can be used for the design of topologies that operate under such regime. \n\nThe experimental results show significant performance gains in the specific benchmark tests." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes the FedOLF framework that aims to improve the efficiency of FL for resource constrained devices. The key observation behind the work is that by allowing the update of the layers of a network in an ordered manner, energy and memory costs can be reduced without significant sacrifices in the performance of the system. The authors move further and enhance the efficiency of their system by employing Tensor Operation Approximation demonstrating better results in their setting compared to traditional quantization approaches. \n\nOne of the main claimed contributions of the work, which is also the basis of the method, is the vanishing representation error. That is by imposing an ordering of the layers that are frozen in any update, the authors demonstrate a bound on the representation error of the active layers as a function of the last frozen layers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The idea of tackling the energy and memory consumption in FL by frozen a number of layers is not new and many works have already utilised this approach. The novel claimed contribution is the strategy of the ordering of the updating. Given that it is well known that more high level layers are more important for specific tasks compared to the early layers that have the job to extract more generic features, such approach is not so surprising.\n\nThe memory footprint can also be tackled by re-evaluating part of the results in an on-demand manner. As such, someone can trade off time for training vs memory footprint. Given that for many devices the computational cost (in terms of energy) is less than the communication cost, it is unclear how such approach would compare on the overall energy cost vs quality of training with the proposed method that chose to freeze some of the layers for addressing the memory requirements.\n\nA key observation/assumption that underpins the theoretical working of the method is that the upper bound B is likely smaller that one. However this is only based on few networks and datasets and I found difficult to justify the above statement \n\nIn the experimental section, it was not clear to me what is the assumption of the quality of the training of the original network. For example, I would expect ordered layer freezing to perform well if the original model was not very far (and especially the initial layers) from the true one, where if you give to the system a network with random parameters (extreme case) then the initial layers would not be very informative and any imposed freezing on them would not lead to meaningful training for the rest of the layers. The authors should comment on that." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose Federated Learning with Ordered Layer Freezing (FedOLF) as a solution to mitigate the energy consumption and memory footprint of Federated Learning in resource-constrained devices while maintaining accuracy." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024energy,\ntitle={Energy and Memory-Efficient Federated Learning with Ordered Layer Freezing and Tensor Operation Approximation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xbW6EGve6a},\nnote={under review}\n}" }, "abstract": { "value": "The effectiveness of Federated Learning (FL) in the context of the Internet of Things (IoT) is hindered by the resource constraints of IoT devices, such as limited computing capability, memory space and bandwidth support. These constraints create significant computation and communication bottlenecks for training and transmitting deep neural networks. Various FL frameworks have been proposed to reduce computation and communication overheads through dropout or layer freezing. However, these approaches often sacrifice accuracy or neglect memory constraints. In this work, we introduce Federated Learning with Ordered Layer Freezing (FedOLF) to improve energy efficiency and reduce memory footprint while maintaining accuracy. Additionally, we employ the Tensor Operation Approximation technique to reduce the communication (and accordingly energy) cost, which can better preserve accuracy compared to traditional quantization methods. Experimental results demonstrate that FedOLF achieves higher accuracy and energy efficiency as well as lower memory footprint across EMNIST, CIFAR-10, CIFAR-100, and CINIC-10 benchmarks compared to existing 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": [ "Federated Learning", "Resource-Constrained devices", "Computation and Communication Overheads", "Layer Freezing", "Tensor Operation Approximation" ] }, "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/d7b62618a52ea46c2a11f8da40d9ff3911353474.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/91d8c84d778fec97c540847bd236046b4a828899.zip" }, "title": { "value": "Energy and Memory-Efficient Federated Learning with Ordered Layer Freezing and Tensor Operation Approximation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Can the authors explain more about parameter count?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- CWPS demonstrates high performance on various tasks and scenarios.\n- The proposed framework is simple but effective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Current methods for knowledge transfer such as fine-tuning or weight sharing only offer coarse-grained sharing solutions and do not search for optimal parameters for sharing. This paper introduces Channel-Wise Parameter Sharing (CWPS) for efficient knowledge transfer. By refining the granularity of shared parameters from the layer level to the neuron level, they achieve fine-grained parameter sharing to address the coarse-grained problem. The authors also propose a simple method to search for suitable parameters for sharing. The proposed method achieves state-of-the-art results on several benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experiment is mainly done using ResNet-50, lacking the experiment using other models such as DenseNet and EfficientNet, or even transformer-based or MLP-based models. \n- Lack of experiments with different depths of the model" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Would it be possible to visualize which channels are masked at each layer and whether there is a discernible pattern?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.Innovative Approach: CWPS offers a new perspective on knowledge transfer by enabling fine-grained parameter sharing, which is a significant departure from traditional coarse-grained methods.\n\n2.Enhanced Efficiency: The method is designed to be efficient in terms of parameter sharing, which can potentially lead to better performance and faster training times.\n\n3.Comprehensive and Plug-and-Play: CWPS is presented as a comprehensive solution that can be easily integrated into existing networks without the need for extensive modifications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Channel-Wise Parameter Sharing (CWPS), a novel approach to knowledge transfer that enhances the efficiency and effectiveness of sharing parameters across different tasks or new data. Traditional methods like fine-tuning and layer-wise parameter sharing often provide coarse-grained solutions, which struggle to effectively identify shared parameters, thus limiting performance and efficiency. CWPS addresses these limitations by introducing fine-grained parameter sharing, achieved by refining the granularity from layers to neurons, allowing for the explicit composition of model neurons and utilization of knowledge from previous tasks. The paper also presents an effective search strategy to minimize computational costs and simplify the determination of shared weights. CWPS is designed to be comprehensive, plug-and-play, and has strong composability and generalization abilities, making it theoretically applicable to any network with linear and convolution layers. The method is evaluated across various datasets in incremental learning and multi-task learning scenarios, demonstrating superior precision-to-parameter ratio performance compared to existing methods, regardless of the backbone used." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.Since the supplementary materials of the paper include experiments with the Transformer model, it seems inconsistent that the main text focuses on convolution as the primary subject of discussion, rather than the Transformer. Moreover, the performance of Vision Transformer (ViT) in Table 5 of the appendix shows suboptimal results, which somewhat undermines the argument that the algorithm can generalize across different backbones effectively.\n\n2.Why were experiments not conducted on datasets of a scale comparable to ImageNet in Tables 1 and 2? Observing the right side of Figure 4, it appears that all datasets have some relation to ImageNet. Therefore, wouldn't conducting a downstream task on a dataset of similar scale to ImageNet, but not necessarily entirely relevant, further demonstrate that the method's effectiveness is not solely due to knowledge transfer from ImageNet, but rather that the algorithmic design itself contributes significantly to the 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": 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": "N/A" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper addresses an important problem in transfer learning. With more and more large foundation models available to the public, these models contribute a very large pool of knowledge learned through vast volume of data. The techniques investigated in this paper can be greatly beneficial to tapping into the potential of such models.\n\n2. The proposed method seems fairly universal. Even though the experiments in the paper are mostly conducted on CNNs, the method can be potentially applied to state-of-the-art transformer networks as well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a method for knowledge sharing to tackle the problem of transfer learning. The method uses existing models as a knowledge pool, from which a layer-wise composition is constructed tailored to the new task to be solved. This composition is referred to as the parent model. Together with the parent model, a child model of the same architecture is trained. The two models are combined channel-wise using learnable masks. Experiments on multiple benchmarks are conducted to demonstrate the effectiveness of the approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The training pipeline seems unnecessarily convoluted and impractical. If I understand this correctly, given a new task, a child model first needs to be trained for a short period of time. This child model is then used as reference to construct the parent model. Afterwards, the two models are combined with learnable masks to tackle the aforementioned new task. Compared to standard fine-tuning, the advantage of the proposed method appears to be its parameter efficiency. However, the parameter count itself is not the most faithful reflection of the complexity of the model. Stats such as memory consumption and training time are more important, and accurately reflect the Flops needed to train the model. Given that the performance of the proposed method is generally worse than fine-tuning, the contribution of the work is less than convincing.\n\n2. A major motivation of transfer learning is to exploit the knowledge learned on large volume of data and apply it to domain where data may be scarce. In this paper, however, it appears that all data of the target task is assumed to be available. This significantly reduces the practical value of the work. On the other hand, if the effectiveness of the proposed method can be demonstrated in a few-shot setting, where only a small number of labelled examples are available, I think this can make the paper much stronger. For a point of reference, I believe a recent work by Zhang et al. (2024) shows that a learnable linear combination of models in the weight space can achieve very strong performance in few-shot learning and also test-time adaptation. In the latter case, no labelled data is available. I would encourage the authors to compared against this method.\n\n3. The writing of the paper is not the best, with numerous typos. For instance, in L201, I think the authors meant \"existing\" instead of \"exiting\".\n\nRefs:\n- Zhang et al., Knowledge Composition using Task Vectors with Learned Anisotropic Scaling. NeurIPS 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": 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. There is a lack of explanation regarding some baselines, as well as the approach to parameter sharing. Additionally, there are several prompt-like methods, such as DualPrompt." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. CWPS has lower overhead for additional parameters compared to previous methods.\n2. By using CPMS to determine the relationships between different tasks, it can effectively quantify these connections." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on multi-task learning and multi-domain learning. Existing methods have not taken into account the characteristics of neurons, so a channel-wise parameter sharing method, CWPS, is proposed, achieving better results than the baseline." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper argues that sharing at the neuron or channel level is important, but I think this requires more justification, as I didn’t see sufficient evidence for it in the paper. So, it doesn’t seem very different from other parameter-sharing methods, and the motivation behind the approach doesn’t seem very clear to me.\n2. The overall design of the method appears relatively simple, and the comparison results lack novelty. More experiments are needed to uncover the core reasons behind the improvements this method brings.\n3. The writing is poor." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a novel fine-grained parameter-sharing method for efficient and comprehensive knowledge transfer, addressing issues with current coarse-grained sharing solutions." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024cwps,\ntitle={{CWPS}: Efficient Channel-Wise Parameter Sharing for Knowledge Transfer},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xbXydoejvY},\nnote={under review}\n}" }, "abstract": { "value": "Knowledge transfer aims to apply existing knowledge to different tasks or new data, and it has extensive applications in multi-domain and multi-task learning.\n The key to this task is quickly identifying a fine-grained object for knowledge sharing and efficiently transferring knowledge.\n Current methods, such as fine-tuning, layer-wise parameter sharing, and task-specific adapters, only offer coarse-grained sharing solutions and struggle to effectively search for shared parameters, thus hindering the performance and efficiency of knowledge transfer.\n To address these issues, we propose Channel-Wise Parameter Sharing (CWPS), a novel fine-grained parameter-sharing method for Knowledge Transfer, which is efficient for parameter sharing, comprehensive, and plug-and-play.\n For the coarse-grained problem, we first achieve fine-grained parameter sharing by refining the granularity of shared parameters from the level of layers to the level of neurons. The knowledge learned from previous tasks can be utilized through the explicit composition of the model neurons.\n Besides, we promote an effective search strategy to minimize computational costs, simplifying the process of determining shared weights.\n In addition, our CWPS has strong composability and generalization ability, which theoretically can be applied to any network consisting of linear and convolution layers.\n We introduce several datasets in both incremental learning and multi-task learning scenarios. Our method has achieved state-of-the-art precision-to-parameter ratio performance with various backbones, demonstrating its efficiency and versatility." }, "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": [ "Transfer Learning", "Multi-Domain Learning", "Multi-Task 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/ddaa2d469c90af9ae8aae975b2763fd5c724eed5.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/3fc0b70f1be0b45b1823ac00fa9d323759590456.zip" }, "title": { "value": "CWPS: Efficient Channel-Wise Parameter Sharing for Knowledge Transfer" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In the experiment, the motion of the real robotic hand is recorded and then replayed in the simulation to show the contact locations can accurately match. I am not clear the motivation here.\n2. In the texture feature extraction framework, how the conditions to classify points as texture features work?\n3. In pseudo-tactile data integration part, it is shown that the system is able to simulate tactile feedback without the need for real sensors. How the feedback works? Is the pseudo-tactile information used for some closed-loop controller?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "This work introduces an approach to extract pseudo-tactile information from vision and contact locations for robotic grasping tasks. The topic is interesting, the pipeline is presented with details, and the experiment results show the effectiveness of the approach with high accuracy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a framework to acquire point cloud representations of objects and simulate the contact locations when using a dexterous robotic hand for grasping tasks. The point cloud data are first processed to filter the background and texture feature points are determined afterwards. The real grasping data are derived from the hardware and can be reproduced in a simulation environment with the corresponding contact locations and texture features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The overall contribution is marginal. The vision-based \"pseudo tactile\" features are derived using some common tools for point cloud data processing. The simulation for finger tip localization is from replicating and transforming the real motion data into the simulated environment. It is also claimed that the combination of this vision-based information and the fingertip contact points can enhance tactile feedback reliability in robotic grasping. However, this is not clear. The experiment does not show how the robotic grasping is improved." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- What is the purpose of using simulation?\n- How is the tactile information from the point cloud used in contact localization?\n- How is ground truth obtained for contact localization? What is the precision/resolution of the ground truth measurement?\n- What does policy finetuning have to do with this paper?\n- How are the texture feature points used once they're determined by thresholding?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Open-sourced dataset of point cloud images is made available." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a dataset of point cloud images of a wide variety of everyday objects. A dexterous hand model is built in an Isaac Sim simulator for real-time contact localization by replicating the real-life set up in simulation and using measured joint angles." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Details are so unclear it is difficult to fully understand the paper. For instance, \n- Paper repeatedly talks about the \"Y component of the normal vector\" without clearly defining a coordinate frame. Authors mention employing \"policy fine-tuning techniques\" on page 7 without ever mentioning a policy up until this point.\n- The role of simulation in the paper is not clear. The paper mentions \"This real-time linkage of each joint’s degrees of freedom with actual dexterous hand movements and the simulation platform allowed us to record the spatial coordinates of each grasping contact point in the simulation accurately.\" This sounds like a simple forward kinematics problem that would only require a mathematical model of the robot – not a full-fledged simulation.\n- Paper claims that intel realsense has “sub-mm accuracy”. This is not supported by documentation from the manufacturer: https://dev.intelrealsense.com/docs/tuning-depth-cameras-for-best-performance?_ga=2.110331777.520332705.1730517789-101245430.1730517789#section-verify-performance-regularly-on-a-flat-wall-or-target\n- Section 4.2 claims to assess localization precision. It is unclear what the ground truth is, how this is being measured and what measurement is being compared against this ground truth. Referenced Table 4 is difficult to understand, has no mention of errors or comparative ground truth. \n- An RMSE error is reported with no clarity on what quantities are being compared.\n- It is also unclear how related work Section 2.3 is connected to this work. Pseudo-haptics is associated with giving humans touch sensory feedback, whereas “pseudo-tactile” in this paper is related to analyzing the surface tactile properties of objects.\n\nMinor comment:\nPaper is very poorly formatted. Main result tables are placed in the appendix and are difficult to understand. Results in Tables 4,5 and 6 have poor choice of units (meters when dealing with textures that are likely sub-cm scale), and numbers are represented with arbitrary precision with no regard for the precision and error rates of the depth measurement device, ie. Intel Realsense cameras. Bullet points in the results section seem to have headers that have the exact font formatting as the rest of the 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": 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": "There lacks of the definitions of the \"grayscale variance\" and \"Y-component\", the authors could define them properly when they first appear.\n\nTable 1 and equations (1)-(4) are describing the transformation from camera frame to the world frame to extract point cloud which seems to be fundamental computer vision techniques. I would suggest to move them to the Appendix since it is not novel in this work." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The work has clear writing and good structure which makes it easy to follow and understand. Figures are well-made and informative. It is well-motivated to address the hard-to-acquire tactile perception by using vision for dexterous robotic hands." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presented an approach to acquire object surface information including textures and geometries, and proposed a simulation-involved approach to locate fingertips of robotic hand for grasping. The results demonstrated the object grasping with different number of fingers and object surface feature extraction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I do not think this work has solid contribution or concrete experimental results. The proposed method simply combining existing techniques such as extracting point cloud from RGBD camera, and using simulator to simulate grasping. Instead of quasi-static grasping, I would encourage authors to extend it to more dynamic manipulation tasks and explore whether the extracted object features can be leveraged for these complicated tasks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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": "Apart from the weaknesses I have mentioned, I still have some questions for the authors to respond.\n1. I noticed that the author evaluated the localization results by \"comparing simulated 3D fingertip coordinates with real-world measurements\", how to acquire accurate real-world contact positions with 3D point cloud input or other approach?\n2. The authors use Intel RealSense Camera for capturing depth images and convert them into point clouds. As far as I know, depth images generated by D435 can be very noisy, thus the estimation of surface normal vectors can be very biased, how to deal with this problem?\n3. the authors mentioned that objects with glass material is included in the dataset, but the depth images generated from RealSense Camera on transparent objects are catastrophy. How can the author obtain accurate point cloud information under this circumstances?\n4. In Sec. 3.3, the author mentioned \"We also employ policy fine-tuning techniques, using a small amount of real-world grasping attempt data to fine-tune the model\", but since the method is not deep-learning based, how to conduct the fine-tuning stage in the pipeline?\n\nRight now, I'm inclined to reject this paper. However, if the authors are willing to answer my questions well, I'll consider raising my score." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper addresses the problem of obtaining tactile information during grasping based on only vision perception, and provides a clear method for object surface texture extraction with 3D point cloud input. The strengths are listed below.\n1. The authors provide abundunt and clear explanation for method presentation, and present a simple yet effective approach for point cloud preprocessing and feature extraction.\n2. The authors give a thorough representation on the expermental setup, and conduct real-world experiments on a dexterous hand for validation.\n3. The result of the experiments seems very ideal, indicating the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the challenges of acquiring haptic perception during grasping by a mechanical dexterous hand and proposes solutions. The main tasks of the research include the acquisition of “pseudo-tactile” information about everyday objects through vision and the construction of a dexterous hand (RH8D) model in Isaac Sim for real-time fingertip contact localization. The study establishes a scientific link between simulated 3D coordinates, actual 3D coordinates, and pseudo-tactile information derived from the point cloud, which is quantified by normal vector and grayscale ANOVA. Experimental results show that the method is able to clearly extract the surface texture of an object, accurately locate the fingertip contact point in real time, and provide haptic information at the contact point." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I recognize some of the article's contributions, I still have the following concerns.\n1. From my point of view, this paper is a bit lack of novelty, because some parts of the method chapter are biased towards engineering practice rather than innovation, such as point cloud preprocessing, camera coordinate system transformations, and other types of work are actually common in robotics work, and cannot be listed as points of innovation. Meanwhile, the texture extraction method in the section overlaps most with [1].\n2. While generating pseudo-haptic sensing is one of the important contributions of the article, I didn't see that the authors had measured how good the quality of the generated haptic signals were, both quantitatively and qualitatively.\n3. The article's experiments still seem inadequate to me and lack comparison with previous work. For the contact position localization part, are there any previous baselines that can realiza this? For example, 3D point cloud keypoint prediction baselines, etc. Meanwhile, the authors didn't conduct the ablation studies on the proposed method, such as the different effectiveness on selections of KDTree radius, normal threshold, etc.\n\n[1] Budiyanta, N. E., Yuniarno, E. M., & Purnomo, M. H. (2021, December). Human point cloud data segmentation based on normal vector estimation using pca-svd approaches for elderly activity daily living detection. In TENCON 2021-2021 IEEE Region 10 Conference (TENCON) (pp. 632-636). IEEE." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This study extracts pseudo-tactile information from everyday objects using vision, enabling real-time localization of fingertip contact points and their corresponding pseudo-tactile 3D point cloud information in robotic dexterous hand grasping." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024visionbased,\ntitle={Vision-Based Pseudo-Tactile Information Extraction and Localization for Dexterous Grasping},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xcHIiZr3DT},\nnote={under review}\n}" }, "abstract": { "value": "This study addresses the challenges of tactile perception in robotic dexterous hand grasping by focusing on two main tasks: 1) Acquiring tactile information from everyday objects using vision, termed \"pseudo-tactile\" information, and 2) Building a Dexterous Hand (RH8D) model in Isaac Sim for real-time fingertip contact localization. Utilizing Isaac Sim enables safe, cost-effective experimentation and high-precision simulations that facilitate data collection for model validation. The research establishes a scientific connection between simulated 3D coordinates, actual 3D coordinates, and pseudo-tactile information derived from point clouds, quantified through normal vectors and grayscale variance analysis. Results demonstrate the ability to extract clear object surface textures, accurately locate fingertip contact points in real-time (with precision up to $0.001 m$), and provide tactile information at contact points. This framework enhances robotic grasping capabilities and offers low-cost sensory data. The source code and dataset are publicly available now." }, "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": [ "Pseudo-Tactile Information", "Dexterous Grasping", "Vision-Based Perception", "Robotic Localization" ] }, "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/a45afd60172c0e0d8c40b02c5471fa532026c1ef.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/716d6fb35281c2cbea867edbdf570a8204fa4c45.zip" }, "title": { "value": "Vision-Based Pseudo-Tactile Information Extraction and Localization for Dexterous Grasping" }, "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 above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper is well-written, the ideas are easy to follow, and the architecture is well described.\n* The idea of using INRs for time series, and especially for imputation tasks, is a timely and important topic for the community.\n* The results of the experiments are convincing, especially when the missing rate is high." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This article proposes a new model for the problem of time-series imputation that uses Implicit Neural Representations (INR) for time-series modeling. The model employs a convolutional network to extract multi-scale features from the time series, followed by a transformer encoder. The time series are modeled by an INR decomposed into three sub-functions modeling the trend, seasonality, and residuals. The parameterization of the INR is performed based on the output of the transformer. To facilitate the learning of the residual function, clustering is performed on the different variables of the time series to dedicate a specific MLP to each cluster. Experiments on 7 datasets are conducted, comparing the proposed approach to various state-of-the-art algorithms. An ablation study is also carried out to demonstrate the usefulness of each component." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The greatest weakness of the paper, in my opinion, lies in the experimental section and the comparison to existing approaches. There are at least two other papers using INRs for time-series imputation: [1], which is also set in the context of extremely missing observed data (95%), and [2], which uses trend/seasonality/residual decomposition and is also set in the context of a high missing rate. The paper does not cite or compare to these two approaches. Given the similarity of ideas and the relatively few papers on the subject, it seems necessary for the proposed model to be compared to these two approaches.\n\nA minor weakness is the section on clustering, which seems questionable to me. The assumption is that when variables are related, they remain related over time – as far as I understand, the clustering is done on the entire series. This clustering is only used in the residual part to facilitate learning. I wonder if the same result could not be achieved with better regularization of the MLP network used.\n\n[1] Time Series Continuous Modeling for Imputation and Forecasting with Implicit Neural Representations, Le Naour et al., TMLR 2024\n[2] HyperTime: Implicit Neural Representation for Time Series, E. Fons, A. Sztrajman, Y. El-Laham, A. Iosifidis, S. Vyetrenko, NeurIPS 2022 SyntheticData4ML" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "As seen in weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well-written.\n2. The targeted problem is important." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a solution to the time series imputation problem, particularly for datasets with high proportions of missing observed values. The approach leverages Implicit Neural Representations (INR), which are known for their ability to model continuous functions, potentially enhancing the accuracy of missing data interpolation. The authors introduce three distinct functions to represent multivariate discrete datasets, corresponding to the trend, seasonal, and residual components. Additionally, a clustering module is incorporated to group channels with similar distributions, further refining the handling of inter-variable relationships and improving imputation quality. Experimental validation demonstrates the method's superior performance compared to existing techniques, particularly in extreme missing data scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The word \"novel\" appears multiple times throughout the paper; however, techniques such as INR, clustering, and time series decomposition (including trend, seasonal, and residual components) are well-established and widely utilized in existing research. Additionally, the approach of learning parameters to calculate imputation results through mathematical methods has been explored in prior studies[1]. Although the combination of these elements may provide certain technical contributions, the work lacks a distinctive level of innovation to set it apart from previous works.\n[1] Liu, Shuai, et al. \"Multivariate time-series imputation with disentangled temporal representations.\" The Eleventh international conference on learning representations. 2023.\n\n2. In Equations 9 and 10, the authors use fixed formulas to represent parts of the INR function (trend and seasonal components). While this choice enhances the interpretability of ImputeINR, it also reduces its ability to handle various complex datasets. Could the authors explain why they used these two fixed functions to represent the trend and seasonal components of time series data? Detailed insights will be preferred.\n\n3. The average experimental results obtained by ImputeINR are quite impressive; however, this performance is primarily evident in datasets such as BAQ and IAQ, while the improvements observed on widely used datasets like ETT and Weather are quite modest. Does this indicate that ImputeINR has stringent requirements concerning dataset distribution? The authors are encouraged to clarify which types of datasets are most appropriate for imputation with ImputeINR.\n\n4. The experimental settings require more detailed clarification. BAQ, IAQ, and Solar are datasets containing tens of thousands of time steps, while the experiments presented in this paper utilize only a few hundred or fewer. Furthermore, the training and testing set split ratios vary across datasets, such as Weather and Solar. Similar concerns are observed in other datasets as well.\n\n5. Incorporating SOTA baselines, such as CSDI[2], would enhance the credibility of the experimental results.\n[2] Tashiro, Yusuke, et al. \"Csdi: Conditional score-based diffusion models for probabilistic time series imputation.\" Advances in Neural Information Processing Systems 34 (2021): 24804-24816." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weak points listed above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "S1. Imputing time series data is an important problem.\n\nS2. Various baselines are considered in experiments.\n\nS3. The visual analysis is conducted, to improve the readability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an imputation approach to address the incomplete time series with high missing rates,. By leveraging implicit neural representations (INR) to learn continuous functions, ImputeINR can generate fine-grained imputations even when substantial values are missing. The method includes a multi-scale feature extraction module to enhance the imputation's fine-grained and global consistency. Additionally, ImputeINR uses a specific form of INR continuous function to separately learn trend, seasonal, and residual information, and an adaptive group-based framework to model complex residual information. Experiments on seven datasets show ImputeINR's superior performance in high absent ratios in time series." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. This paper focuses on imputing time series data with high missing rates, e.g., 70%, 90%. However, it is doubted whether this scenario is important and commonly observed in real applications, since there is no real scenario provided to support such a motivation. Please provide specific examples of real-world scenarios where such high missing rates occur, or to justify why addressing these extreme cases is important even if they are rare.\n\nW2. The core idea is to use the ability of INR to learn continuous functions and achieve interpolation. To adapt INR for time series imputation, trend, seasonal and residual items are considered, which are typical operations for modeling temporal data. It is suggested to further highlight the contribution and novelty of proposed framework. Please explicitly compare the proposed approach to existing methods that use trend, seasonal, and residual decomposition, and clearly state what specific innovations this method introduces beyond these typical operations.\n\nW3. As claimed in W1, in experiments, all the datasets are originally complete and only artificial missing values are considered in the evaluation. It is necessary to use real-world incomplete data with large missing rates in experiments to demonstrate the applicability of proposed methods as well as the motivation scenario.\n\nW4. It is also suggested to consider the application study using real-world incomplete datasets, to investigate the performance of using proposed techniques to serve real scenarios.\n\nW5. In Table 2, mean and median values are the same in different mask rates. Please explain why the mean and median values are the same across different mask rates, or to verify if this is correct.\n\nW6. In Table 2, Mask rate is represented inconsistently, e.g., 10% or 0.1." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In the Introduction, the authors state that ImputeINR is the first method designed specifically to handle extremely sparse observed data. However, to my knowledge, other recent studies have addressed this issue, including CSDI (Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation) and SSSD (Diffusion-based Time Series Imputation and Forecasting with Structured State Space Models). The author should include these methods in their experimental comparisons or a comparative discussion of these approaches in the literature review and highlight how ImputeINR are different and better from these approaches for extremely sparse data scenarios.\n\n2. Could the author provide recent advancements in INR for time series, limitations of current INR approaches, or how INR addresses missing values in existing work? For instance, to my knowledge, \"Time Series Continuous Modeling for Imputation and Forecasting with Implicit Neural Representations\" also provide an INR framework for data imputation, for which this is not being reviewed in this work.\n\n3. It is not so clear to me how the author generate the (missing) mask. Could the author clarify how their masks are generated in detail? In Section 4.1, the authors mention that missing values are generated by randomly masking values based on a specified mask rate. Does this imply the missing value are generated according to the missing rate? If so, have the authors also considered block missing scenarios, as proposed in SSSD, where entire segments are missing?\n\n4. Could the authors clarify why the last column in Table 2, showing imputation results with the mean/median baseline, are identical?\nIs this an error in reporting or is it a characteristic of the mean/median imputation method. If this is not an error, the authors could explain and discuss this pattern. \n\n5. Additional detail on the computational complexity of ImputeINR would be helpful. Specifically, how does its computational burden compare to that of other baseline methods?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The author introduces an innovative ImputeINR model, addressing the complex problem of time series imputation under high missing ratios. The method shows strong performance on benchmark datasets with diverse characteristics, achieving state-of-the-art imputation accuracy, especially under extreme mask rates.\n2. The paper includes extensive experiments, ablation studies, and robustness analyses to validate the contributions of each component, showing ImputeINR’s consistency across diverse settings.\n3. The paper provides clear explanations and visualizations, particularly for clustering results, aiding the reader's understanding of ImputeINR’s inner workings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel approach for imputing missing values in time series data, particularly focusing on cases with high proportions of missing values. The proposed method, ImputeINR, employs implicit neural representations (INR) to model time series as continuous functions, allowing for fine-grained interpolation even with sparse observations. ImputeINR incorporates a multi-scale feature extraction module to capture various temporal patterns and a novel adaptive group-based architecture that leverages clustering to group variables with similar distributions. Finally, numerical experiments across seven datasets and five different levels of missing data were conducted to demonstrate the performance of ImputeINR. Overall, the paper looks promising and organized, but I have some questions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please refer to the questions sections." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose ImputeINR, a time series imputation method based on adaptive group-based implicit neural representations." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024imputeinr,\ntitle={Impute{INR}: Enhancing Time Series Imputation with Adaptive Group-based Implicit Neural Representations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xcPN6Or88c},\nnote={under review}\n}" }, "abstract": { "value": "Time series data frequently exhibit the presence of missing values, rendering imputation a crucial process for downstream time series tasks and applications. However, existing imputation methods focus on discrete data points and are unable to effectively model sparse data, resulting in particularly poor performance for imputing substantial missing values. In this paper, we propose a novel approach, ImputeINR, for time series imputation by employing implicit neural representations (INR) to learn continuous functions for time series. ImputeINR leverages the merits of INR that the continuous functions are not coupled to sampling frequency and have infinite sampling frequency, allowing ImputeINR to generate fine-grained imputations even on extremely absent observed values. In addition, we introduce a multi-scale feature extraction module in ImputeINR architecture to capture patterns from different time scales, thereby effectively enhancing the fine-grained and global consistency of the imputation. To address the unique challenges of complex temporal patterns and multiple variables in time series, we design a specific form of INR continuous function that contains three additional components to learn trend, seasonal, and residual information separately. Furthermore, we innovatively propose an adaptive group-based framework to model complex residual information, where variables with similar distributions are modeled by the same group of multilayer perception layers to extract necessary correlation features. Since the number of groups and their output variables are determined by variable clustering, ImputeINR has the capacity of adapting to diverse datasets. Extensive experiments conducted on seven datasets with five ratios of missing values demonstrate the superior performance of ImputeINR, especially for high absent ratios in time series." }, "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": [ "time series imputation", "implicit neural representations" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/63000132df19e4b22261c740292d0cdf8109bc13.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/2e088e4499233ace74242e93d1329227f9af3261.pdf" }, "title": { "value": "ImputeINR: Enhancing Time Series Imputation with Adaptive Group-based Implicit Neural Representations" }, "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": "See Weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Pros:\n1. The paper presents a unified approach to detecting multiple types of LLM misbehavior, whereas previous work typically focused on individual issues.\n2. The authors provide detailed visualizations and analysis showing how different patterns in the causal maps correspond to different types of misbehavior, as demonstrated in Figures 2-5 and the extensive experimental results in Section 4.\n3. The method shows impressive performance, particularly for detecting lies, jailbreaks, and toxic content, with AUC scores consistently above 0.95 across different models and datasets. This is particularly notable given that the method works with the first generated token, allowing for early detection of potential misbehavior." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces LLMSCAN, a novel method for detecting various types of misbehavior in Large Language Models (LLMs) through causal analysis of the models' internal mechanics. The approach monitors both input tokens and transformer layers to create causal maps that capture how different components contribute to the model's outputs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Cons:\n1. The performance on bias detection is notably weaker than other tasks, with AUC scores ranging from 0.71 to 0.78. While the authors acknowledge this limitation and provide some analysis in Section 4.3, they could have explored more deeply why their approach struggles with this particular type of misbehavior and potential solutions.\n2. While the authors provide a public code repository, the reproducibility section (Appendix C) could be more detailed, particularly regarding the specific hyperparameters used for the detector training and the process for selecting attention heads for token-level analysis.\n3. The evaluated LLMs are relatively limited and small-scale. It is suggested that the authors also evaluate on models with a larger scale such as llama-70B. Otherwise, the effectiveness of the proposed method is relatively limited." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Can more sophisticated causal analysis be applied for LLM detection?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- A proactive solution by identifying and preventing the intent to generate misbehavior of LLM based on internal patterns before a token is generated is a good idea.\n- The idea of analyzing internal patterns is interesting.\n- The idea of constructing a causality distribution map for token and layer is interesting" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new method called LLMScan for detecting misbehavior in large language models (LLMs) usage. LLMScan uses causality analysis to identify the LLM's internal activation pattern to indicate potential misbehavior, such as generating untruthful, biased, harmful, or toxic responses. The authors demonstrate the effectiveness of LLMScan through experiments on various LLMs and tasks, achieving high accuracy in detecting different types of misbehavior." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "No baseline.\nThe work claims to provide many baselines, but they are not presented in the relevant experiment result tables.\n\n\nPoor presentation.\n- The terminology provided in this work is not precise, leading to confusion for the reader. E.g., \n - The name of causality seems misleading. The process described in ``Computing the causal effects of tokens'' is a measurement of sensitivity to token replacement by the attention scores. Since the target is an internal attention score, the term causality seems not to be an appropriate description. \n - The term causal map is also weird. The causal map provides a vector value and not a 2-dimensional matrix value. \n- Besides the example of {“Who developed Windows 95?”. The truthful response is “Microsoft Corporation”. The untruthful response is “Bill Gates”} (Line 263) is also very curious since Bill Gates is the CEO of Microsoft Corporation at that time.\n- Other typo/errors: \n - Table 2 duplicate first and second row" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "It is clear that the detector (binary classifier) can effectively identify whether a prompt can lead to misbehavior in LLM. I am curious whether the detector can be extended to precisely point out which misbehavior the LLM has given a malicious prompt." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The observation that the difference in causal maps between truthful and untruthful responses is significantly important for advancing research on the robustness and security of LLMs.\n\n2. The experiments are sufficient, and the results show the effectiveness of the LLMScan." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents LLMScan, a novel approach for detecting various types of misbehavior in large language models (LLMs). LLMScan is composed of two key components: a scanner and a detector. The scanner assesses the influence of each token through attention scores and evaluates the contribution of each layer by examining differences in output logits. These token influences and layer contributions are used to form a causal map and serve as features, which are then processed by a multi-layer perceptron (the detector) to identify misbehavior in the LLM. The paper demonstrates the effectiveness of LLMScan by detecting four types of misbehavior across 13 public datasets using three well-known LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper leverages the ATE (Average Treatment Effect) to evaluate the causal effect of each token and each layer ($ATE=E[Y|do(T=1)]-E[Y|do(T=0)]$).\n\nMy concern is replacing a token with a placeholder \"-\" and calculating the difference in attention scores may not fully align with the rigorous causal inference. Here are some reasons: \n\n- In the context of causal inference, when we conduct the intervention on a token, we should expect the downstream effects on the intervention, leading to the change of other tokens. That said, this paper implicitly makes a strong assumption that all tokens are independent of each other. However, this assumption may not be held in NLP.\n\n- From the perspective of estimating causal effect, the method only measures a difference in attention scores given a single instance, which is actually more like deriving a counterfactual sample if we assume tokens are independent. \n\nNote that I was not questioning the effectiveness of this method. I just don't think the technique described constitutes a rigorous causal analysis. The proposed approach provides useful insights, but labeling it as causal analysis without following the established principles can be misleading." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a novel method for scanning LLM's \"brain\" and detecting LLM misbehavior using causal analysis on input tokens and transformer layers, enabling early detection of lies, harmful and outputs." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024llmscan,\ntitle={{LLMS}can: Causal Scan for {LLM} Misbehavior Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xdGsiYNfje},\nnote={under review}\n}" }, "abstract": { "value": "Despite the success of Large Language Models (LLMs) across various fields, their potential to generate untruthful, biased and harmful responses poses significant risks, particularly in critical applications. This highlights the urgent need for systematic methods to detect and prevent such misbehavior. While existing approaches target specific issues such as harmful responses, this work introduces LLMScan, an innovative LLM monitoring technique based on causality analysis, offering a comprehensive solution. LLMScan systematically monitors the inner workings of an LLM through the lens of causal inference, operating on the premise that the LLM's `brain' behaves differently when misbehaving. By analyzing the causal contributions of the LLM's input tokens and transformer layers, LLMScan effectively detects misbehavior. Extensive experiments across various tasks and models reveal clear distinctions in the causal distributions between normal behavior and misbehavior, enabling the development of accurate, lightweight detectors for a variety of misbehavior detection 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": [ "Large Language Model", "LLM Safety", "LLM Misbehavior Detection", "Causality Analysis", "Model Scan" ] }, "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/8937373704c26e897d2dd360872d85f33c449cc1.pdf" }, "presentation": null, "primary_area": { "value": "causal reasoning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "LLMScan: Causal Scan for LLM Misbehavior Detection" }, "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": "N/A" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper introduces a new uncertainty estimation metric based on NLL that avoids the need for multiple sequence generations, which is a common bottleneck in existing methods.\n- By eliminating the need to generate multiple output sequences, the proposed method significantly reduces computational overhead, making it more practical for large-scale applications.\n- The method achieves or surpasses the performance of existing state-of-the-art uncertainty estimation methods across different models and tasks.\n- The approach shows strong performance across various model architectures, sizes, and training stages, demonstrating its broad applicability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel approach to uncertainty estimation in natural language generation (NLG) models. The authors propose using the negative log-likelihood (NLL) of the generated sequence as a surrogate for uncertainty estimation. By leveraging the theoretical framework of proper scoring rules, they demonstrate that NLL can serve as an effective uncertainty metric. This approach simplifies the estimation process because it only requires the likelihood of the generated sequence under the model, avoiding the need for multiple samples. The theoretical foundation is well-established within the framework of proper scoring rules, and the empirical results demonstrate the method's superiority over existing metrics across various models and tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed metric focuses on statistical uncertainty derived from model probabilities but does not explicitly account for the semantic aspects of generated text. Incorporating semantic uncertainty would provide a more holistic estimation, capturing discrepancies between the generated content and the underlying meaning or intent. While the authors briefly discuss this limitation in the conclusion, it remains a significant issue that warrants deeper exploration, possibly through additional methods or combined metrics.\n- While the experiments are extensive, they focus primarily on free-form question-answering tasks. Additional experiments on other types of NLG tasks (e.g., dialogue generation, story generation) would strengthen the claims.\n- The paper spans 7 pages, whereas the conference allows submissions up to 10 pages. This unused space represents an opportunity to expand on key areas such as additional experiments, detailed analyses, or discussions that could further strengthen the paper's contributions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 Weaknesses)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Computational Efficiency: The proposed uncertainty measure requires only a single output sequence, significantly reducing computational costs compared to methods that generate multiple sequences, making it highly scalable for real-world applications.\n- Theoretical Soundness: Using MAP as the metric of uncertainty is grounded in established principles of proper scoring rules, ensuring theoretical robustness while simplifying the complexity of uncertainty estimation for natural language generation models​." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a MAP-based approach to estimating uncertainty in large language models (LLMs) to improve the reliability of generated text. Traditional Monte-Carlo uncertainty estimation methods rely on generating multiple output sequences, a process that is computationally intensive and inefficient at scale. This study introduces a streamlined method that estimates uncertainty using only the negative log-likelihood of the most probable output sequence, eliminating the need for multiple sequences. The proposed approach maintains theoretical rigor and outperforms or matches existing methods across a range of tasks and models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Questionable Efficiency: It seems to me that obtaining the MAP sequence (argmax) is non-trivial. While seemingly at the end of the day we only get one sequence, taking efforts to approximate it to be the MAP could be no computationally cheaper than sampling a lot of candidates, which is exactly what the paper is claiming to avoid. It would make the paper more compelling, if the authors can briefly study how well the argmax sequence is approximated, and if the approximation of obtaining argmax is not quite good, what is the worst-case performance 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": 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’s $\\mathcal{D}$ in Line 115?\n\n2. Could you provide more justifications on why NLL may be better than other sampling-based baselines?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed metric alleviates the need for sampling to estimate the uncertainty in natural language generation. \n\n2. The derivation of the different uncertainty terms and defining aleatoric and epistemic uncertainty is helpful.\n\n3. The presented experiments cover a few backbone models and representative tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Traditional uncertainty estimation relies on sampling-based methods, which inevitably incurs additional computation cost. This work addresses this limitation and proposes to measure the uncertainty solely based on the negative log-likelihood of the most likely sequence. Empirical results demonstrate the performance of the proposed metric in distinguishing between correct and incorrect answers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The only metric proposed by this work is the zero-one score, which is one minus the predictive distribution for the most likely output sequence. Therefore, I find this is actually equivalent to propose $p(y=y’|x)$ as the confidence estimation, which has been widely applied in the machine learning community, whereas uncertainty is simply derived by 1-confidence. Consequently, this metric lacks technical novelty.\n\n2. Though the proposed NLL metric seems to be superior to baselines, this work lacks justification and insights on why the NLL is a better metric than the variants using sampling. \n\n3. Verbal explanations have been widely implemented in estimating the confidence level of LLMs. The author includes relevant discussion in Line 249. However, there is no empirical comparison to this type of baseline." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "083: What is \\mathcal{V} here? You need to introduce the vocabulary. \n\n094: “since \\mathcal{Y} scales exponentially with the sequence length T.” Here you defined \\mathcal{Y} as all possible sequences, which is an infinite set so it shouldn’t be growing. If you want to make this claim, you can define \\mathcal{Y}_t as the subset of all sequences with length <t. \n\n098-100: “We consider uncertainty for a given LMs, … a valid assumption” I am a bit confused, what is the assumption here?\n\n111: Here if you are sampling y’ from p(\\cdot |x, \\cdot) it is better to write it explicitly “y’ \\sim …” It can be a bit confusing here, and I am not sure how the discussion on “Proper Scoring Rules for Uncertainty Measures.” advances your main claims - if this is only about evaluation, you can defer this to later sections.\n\n127: Again, I am not sure how “Aleatoric and Epistemic Uncertainty” relates to your proposed method. The purpose of “related works” or “preliminaries” is to make people ground your work to existing literature, if you believe that your proposed method is connected with this literature, make it more explicit.\n\n898: “The reference answer sampled using beam search with a size of 20 is considered for assessing overall correctness, as it represents the most likely answer generated by the language model” - why is beam search of 20 = most probable answer? Do you have any guarantees that a beam size of x makes the generated sequence log-prob close (difference bounded) to the most likely sequence?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The paper studies an important topic, which is crucial for many applications (e.g. improving trustworthiness of LMs).\n\n2. The experiment results are good, which is surprising given the simplicity of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to quantify the uncertainty of a language model for a specific prompt using the log-likelihood of the most probable sequence. Empirical results show that this new measure is effective in quantifying the uncertainty of the model without having to generate multiple times." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The idea to use a single generation to “approximate the most likely output sequence (line 223)” is concerning - the motivation is to avoid generating multiple sentences, and yet in order for beam search to find the most probable sentence (even in a toy setting), it requires multiple samples (Appendix A, Figure 2). Practically, I don’t know how close a greedy sampled / top-k sampled sequence is close to the most likely sequence even of the same length. \n\n2. The contribution (using log-likelihood of one generation) is somewhat limited to empirical findings without any theoretical guarantees that one generation is able to find a sequence that is close to the most probable sequence. \n\n2. The paper is not very well written, making it difficult to understand what the authors want to convey. See the questions section." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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. Are eq (1-6) developed by you or prior work? \n2. what is the exact step to calculate Eq(8).\n3. Do you have real groundtruth measurement for generation correctness?\n4. How many generations do you need to estimate uncertainty for one sequence?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The uncertainty quantification for LLM is an important topic. \n2. The derivation of aleatoric and epistemic entropy is valid. \n3. The introduction of zero-one score is valid." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the uncertainty quantification for LLM. The paper first defines uncertainty as the expected score (which needs to be designed) of LLM prediction with respect to all possible parameters fitting the data. It then defines the scoring function as zero-one indicator of whether the generated sequence reaches maximal likelihood. The paper claims the estimation of the final uncertainty quantity only requires max-decoding (such as beam search). The paper evaluates the proposed method against prior baselines on 3 tasks for 6 LLMs. The paper used AUROC to measure accuracy and claims the proposed method achieves the best." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing of the paper is very blurred. It is not clear which part is from prior paper, which part is the original contribution in this paper. Eq(7-8) are proposed in this paper. But Eq(1-6) are unclear. \n2. The definition of uncertainty in Eq (1) seems to suggest there is a groundtruth y generation. The definition in Eq(2) is questionable. It is unclear what is the posterior distribution of parameter w. Does it need to have a distribution of parameter? What if the parameter is fixed. \n3. The actual estimation algorithm is not described. In particular, Eq(8) needs to estimate an expectation term. It is unclear how to estimate this part. There is no description in the paper. \n4. The evaluation approach and the metric used are quite questionable. It is unclear why this particular F1 threshold-based correctness is used. But the details of estimation such correctness is also not described well. The use of LLaMA 70B as the evaluator is also quite questionable." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a theoretically grounded uncertainty measure for LLMs that significantly reduces computational costs while maintaining state-of-the-art performance." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024rethinking,\ntitle={Rethinking Uncertainty Estimation in Natural Language Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xeP03R58RH},\nnote={under review}\n}" }, "abstract": { "value": "Large language models (LLMs) are increasingly employed in real-world applications, driving a need to determine when their generated text can be trusted or should be questioned. To assess the trustworthiness of the generated text, reliable uncertainty estimation is essential. Current LLMs generate text through a stochastic process that can lead to different output sequences for the same prompt. Consequently, leading uncertainty measures require generating multiple output sequences to estimate the LLM’s uncertainty. However, generating additional output sequences is computationally expensive, making these uncertainty estimates impractical at scale. In this work, we challenge the theoretical foundations of the leading measures and derive an alternative measure that eliminates the need for generating multiple output sequences. Our new measure is based solely on the negative log-likelihood of the most likely output sequence. This vastly simplifies uncertainty estimation while maintaining theoretical rigor. Empirical results demonstrate that our new measure achieves state-of-the-art performance across various models and tasks. Our work lays the foundation for reliable and efficient uncertainty estimation in LLMs, challenging the necessity of the more complicated methods currently leading the 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": [ "llm", "nlg", "uncertainty estimation", "uncertainty measures", "proper scoring rules" ] }, "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/950207a67d554f17fc2ad806037fcbb4f76f3f8e.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": "Rethinking Uncertainty Estimation in Natural Language Generation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weaknesses. Will consider adjusting my initial rating upon authors' responses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper presents a new way to distill AR language models into diffusion language models to bridge their performance gap, where the TCS distillation objective effectively connects different types of models. The paper's evaluation on a range of tasks demonstrates its improved performance on complex tasks like in-filling and arithmetic. Moreover, the potential for faster parallel generation is also a advantage over autoregressive counterparts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript introduce Distilled Diffusion Language Models (DDLM), a framework to distill pre-trained autoregressive (AR) language models into denoising diffusion language models. \nA key contribution is the Target Concrete Score (TCS) distillation objective, aiming to bridge the gap between AR and diffusion models. Specially, top-K and gradient-informed estimation are proposed to efficiently estimate the TCS. \nDDLM is evaluated for both discrete and continuous Diffusion LMs, on several language modeling and reasoning tasks, showing its effectiveness in improved performance of Diffusion LMs with faster parallel generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I think one of the major weaknesses lies in evaluation on more and widely-used standard LLM benchmark. \nThe authors only evaluate the models against PPL where PPL/likelihood of Diffusion LMs are often not exact and cannot serve as a standalone indicator for real language capabilities. Therefore, the paper should provide more detailed comparisons of DDLM with existing AR-LMs (e.g, LLAMA-3 8B) on downstream language tasks beyond GSM8K-Aug, such as BBH, MMLU, multilingual tasks like translation, etc. Plus, case study of samples generated by DDLM is needed to assess the behaviors of the model, especially for reasoning tasks.\nALL of these are important to convincingly demonstrate the proposed framework's ability to generalize across a wider range of language tasks and datasets. \n\nMoreover, despite the great promise of self-correction and bidirectional context in Diffusion LMs, AR-LMs can achieve similar results through reflection or complicated chain-of-thought reasoning, as demonstrated by O1. Additionally, open-ended reasoning is particularly challenging for Diffusion LMs because they require pre-specified sequence lengths. Faster parallel generation is good, but AR-LLMs enjoy many MLSYS optimizations thanks to exactly their autoregressive/recursive decomposition especially at inference time. \nAt the end of the days, what is the real potential of developing Diffusion LMs for natural language tasks, as an alternative for AR-LLMs? And to reach this goal, what major challenges need to be 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": 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": "see weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Knowledge distillation is a potential direction to enhance diffusion models. \n- The results are good. \n- This paper is well writen." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explore the possibility of distilling a pre-trained autoregressive (AR) language model (teacher) into a non-autoregressive diffusion (non-AR) language model (student), combining the best of both worlds. The authors propose TCS distillation, a theoretically grounded framework that bridges autoregressive and diffusion paradigms, which can be broadly applicable to both discrete and continuous diffusion models, with any pre-trained autoregressive teacher model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I am not sure whether several numbers in Table 1 is missing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How does the proposed TCS distillation method compare to other state-of-the-art distillation techniques, especially in terms of efficiency and performance?\n2. Could the authors provide more detailed experiments that systematically vary data size and model complexity to demonstrate the scalability of the proposed method?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed method shows the potential to improve learning efficiency and perplexity, which is demonstrated through experiments on language modeling tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel framework for distilling knowledge from pre-trained autoregressive (AR) language models into non-autoregressive diffusion models. The core contribution is the Target Concrete Score (TCS) distillation, a method designed to bridge the gap between autoregressive and diffusion paradigms. It can apply to both discrete and continuous diffusion models and leverages any pre-trained autoregressive teacher model. Experiments on language modeling tasks show improvements in pre-trained diffusion language models and the ability to train new models from scratch." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **The method's introduction lacks systematic clarity**: the paper does not provide a comprehensive introduction to the process of distilling from autoregressive (AR) models to diffusion models. It is challenging to discern the specific difficulties involved in this distillation process and how the current work addresses these challenges. It would benefit from a more structured explanation that highlights the novel contributions and breakthroughs of the proposed method.\n\n2. **The experimental comparisons are insufficient**: the experimental section lacks a thorough comparison with existing diffusion models, particularly in terms of perplexity (PPL). While the paper presents baseline comparisons, it fails to include benchmarks against state-of-the-art (SOTA) diffusion models, which is crucial for validating the effectiveness of the proposed method. For instance, Table 3 follows the experimental setup of Ye et al., 2024, but does not include a comparative analysis of their results, limiting the ability to assess the method's performance.\n\n3. **The validation of the method does not scale up in terms of model size and capability**: the paper does not sufficiently demonstrate the method's scalability, particularly in terms of distilling larger AR models. The ability to effectively distill knowledge from more complex AR models is crucial for validating the motivation behind transferring knowledge to diffusion models. However, the manuscript lacks discussions on whether the proposed method can scale up to handle larger models, which is a key aspect of assessing the practical viability of the 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": 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": "Do diffusion language models learned from scratch and learned with TCS distillation show similar patterns in intermediate generation steps?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors propose distillation from autoregressive models as an effective way to enhance the performance of diffusion language models.\n2. The proposed method is theoretically grounded.\n3. Empirical results show the effectiveness of the method in terms of perplexity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Target Concrete Score (TCS) to distill autoregressive language models to diffusion language models in order to enhance the latter one. The TCS method is applicable to a wide range of diffusion language models, both continuous and discrete ones. Comprehensive experiment supports the effectiveness of TCS." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors emphasize the error-correction ability of diffusion language models but do not show evidence to support it. Additionally, autoregressive models also have the potential to correct previous errors with chains of thought.\n2. Although this paper narrows the performance gap between autoregressive and diffusion language models, diffusion language models still underperform autoregressive models in most tasks without unique advantages. \n3. Insufficient experiments to study how the scales of teachers and students affect learning efficiency. It remains unclear whether the proposed methods help scale a diffusion language model." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Distilling a pre-trained autoregressive language model into a diffusion-based language model with proposed Target Concrete Score objective." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024distilled,\ntitle={Distilled Diffusion Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xfw92pDy2u},\nnote={under review}\n}" }, "abstract": { "value": "Transformer-based Large Language Models (LLMs) have demonstrated remarkable capa-\nbilities, yet their autoregressive nature forces sequential token-by-token decoding, leading\nto inefficiencies during inference. Furthermore, autoregressive language models lack in-\nherent self-correction abilities, which hinders their capacity to refine and improve gener-\nated content without relying on external prompting or retraining techniques. In contrast,\ndiffusion-based models offer the advantage of fast parallel generation through iterative\nrefinement, while leveraging bi-directional attention to utilize full context at once. How-\never, diffusion models are unable to match their autoregressive counterparts. This moti-\nvates us to explore the possibility of distilling a pre-trained autoregressive (AR) language\nmodel (teacher) into a non-autoregressive diffusion (non-AR) language model (student),\ncombining the best of both worlds. In this work, we present Target Concrete Score (TCS)\ndistillation, a theoretically grounded framework that bridges autoregressive and diffusion\nparadigms. TCS distillation is broadly applicable to both discrete and continuous diffu-\nsion models, with any pre-trained autoregressive teacher model. We propose techniques\nto make TCS distillation scalable and efficient for transformer-based models, and show\nhow it can both improve pre-trained diffusion language models and also train new mod-\nels from scratch. Through comprehensive experiments on language modeling tasks, we\ndemonstrate the effectiveness of our proposed 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": [ "diffusion language models", "discrete diffusion", "distillation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f8c8ace4bf4fef1915735d20ce50bd5e071ad7a1.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": "Distilled Diffusion Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- For the comparison of the average number of tokens in the generated rationale and the retrieved documents in Figure ```2```: What is the number of retrieved documents? Is it before or after subsetting to drafters?\n\n- Is the Multi-Perspective Sampling stage also considered in the latency analysis in Section ```4.5```?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The conceptual extension of the speculative farmwork proposed by the author is both novel and appealing, and supported by strong empirical results.\n\n- The paper targets a timely challenge in RAG and offers promising techniques to improve the efficiency for the system.\n\n- The experiments are comprehensive and thorough, in particular, the extensive ablation study and analysis provide great insights for a better understanding of the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a retrieval-augmented generation framework termed SpeculativeRAG, leveraging high-level concepts analogical to speculative decoding.\nThe framework exhibits better performance with lower latency via lunching a set of smaller model instances in parallel, each processes a subset of retrieved documents, to produce answer drafts and rationales.\nThe answer drafts and rationales are subsequently verified by a larger, strong base LLM to select the final answer." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The instruction-tuning of the small drafter LMs requires synthesis rationales generated by Gemini Ultra from an additional 40k instances. This presents an unfair comparison for baselines methods that do not have access to these external resources in the experiments.\n\n- It is unclear whether multiple instances of large verifier LLMs are also lunched in parallel. According to Line ```11~14``` in Algorithm ```1```, this seems to be the case. If so, the large memory overhead might significantly offset the latency gain in practical scenarios.\n\n- A relevant work [1] (first released ~4.5 month ago) is missing in the baseline and not mentioned in the related work Section. To the best of my understanding, [1] also proposes a very similar synthesis rationale-augmented generation approach for RAG, thus, might be a suitable method for baseline comparison.\n\n- The contribution and result of the paper could be further strengthened if newer generation of models (e.g., gemma-2, llama-3) are adopted in the experiments.\n\n[1] Wei et al, *InstructRAG: Instructing Retrieval-Augmented Generation via Self-Synthesized Rationales*. 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In Line 207-209, is the answer A obtained by human labeling or generated by LLMs given documents D? What if the D does not provide evidence for generating A?\n2. Except for the inference latency, what are the consuming token numbers of different methods?" }, "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-written and easy to follow. The authors present detailed descriptions of their methods, including prompts and experimental setups.\n2. The proposed method achieves obvious improvements among the baselines, and the ablation studies can verify the effectiveness of their method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces a new speculative decoding method to enhance the retrieval augmented generation. Specifically, a smaller distilled specialist LM is used to generate answer drafts based on randomly selected document subsets. After that, a larger generalist LM is used to verify the answer confidences according to the answer and rationale generation probabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The conducted experiments should include more recent RAG baselines, especially the speculative decoding methods (e.g., [1], [2]). \n2. It is a trivial idea and makes limited contributions compared to previous studies. As mentioned in this paper, many recent studies investigate the two-stage RAG, where a drafting stage produces answer candidates and the assembling stage generates final answers based on the drafts. There are only a few differences between them.\n\n\n[1] Ground every sentence: Improving retrieval-augmented llms with interleaved reference claim generation.\n\n[2] RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning in Long-Horizon Generation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 this is a paper target for the RAG framework, testing the proposed framework on some RAG benchmarks may be more reliable. Because the QA benchmark is mainly designed for reasoning, not for the RAG framework." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper trained an LLM (Drafter) to generate multiple answers paired with rationale (reference + explains). Then, they use another LLM (Verifier)to confirm the answers, which enhances the robustness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Speculative RAG – a framework that leverages a larger generalist LM to efficiently verify multiple RAG drafts produced in parallel by a smaller, distilled specialist LM. Each draft is generated from a distinct subset of retrieved documents, offering diverse perspectives on the evidence while reducing input token counts per draft." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main contribution of this paper is insufficient. Because the whole framework just involves clustering the documents and fine-tuning small LLM to generate answers with rationale. Then prompt Verifier model confirms the answer based on rationale." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. While the author has shown that the proposed method is effective, it is important to note that the generalist LM has not undergone fine-tuning. RAG serves to supply additional information to the LLM to mitigate knowledge gaps. However, if the generalist LM itself lacks relevant knowledge, can it reliably evaluate the rationale and drafts generated by the specialist LM, particularly when the specialist’s output is inaccurate or includes hallucinations? To better understand this limitation, could the author provide boundary cases or conduct an error analysis?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The designed specialist LM is lightweight and allows to analyze multiple documents and produce RAG drafts in parallel, which effectively reduces the computation cost and increases the inference efficiency.\n2. The distinction between specialist and generalist LMs clarifies their functional roles, reducing the risk of generalization degradation that may result from supervised fine-tuning (SFT) of a generalist LM. Specialists, optimized through targeted fine-tuning, focus on improving draft generation capabilities, thereby enhancing result accuracy.\n3. Applying a clustering algorithm, i.e. K-means, to pre-group retrieved documents and then uniformly sampling from each group, i.e., multi-perspective sampling, can mitigate analysis inaccuracies caused by incomplete information, enhancing robustness in practical applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper leverages the idea of speculative decoding to design a speculative RAG framework, which offloads computational burden to a smaller, specialist LM that serves as the RAG drafter for generalist LMs. Extensive experiments and ablation studies demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The quality of the SFT data constructed by the author remains unclear due to insufficient analysis. Additionally, further ablation studies are needed to evaluate the drafter’s performance across varying data volumes. Acquiring substantial SFT data through a proprietary model is resource-intensive. Therefore, it is essential to investigate whether a smaller dataset could yield satisfactory results or if a large volume of SFT data is indeed required to achieve optimal outcomes.\n2. P(Yes) or SR is commonly applied in hallucination detection. However, [1] highlights that LLMs without target SFT or aligned, while LLMs may differentiate between correct and incorrect results, the probability itself, i.e., P(Yes), is either uncalibrated or demonstrates poor calibration performance. Although Table 3 shows that incorporating SR enhances overall performance, there is insufficient evidence to validate the intrinsic effectiveness of P(Yes). This limitation weakens the robustness of the proposed method. Could the author also add an ablation study to independently assess the effectiveness of P(Yes)?\n3. It seems that results for key baseline models, such as CRAG and Self-CRAG, are missing for the three free-form datasets. This omission is concerning, as it is standard research practice to replicate baseline results when they are not available from prior studies, especially when the authors are comparing their approach with only a few previous works. Could the authors also include the performance metrics of CRAG and Self-CRAG on the free-form datasets for a more comprehensive comparison?\n4. According to the results in Table 1, Mixtral-Instruct-8x7B has already achieved high scores on TriviaQA and ARC-C (73.91% and 78.41%), and the improvement brought by Speculative RAG is limited (74.24% and 80.55%). These results may diminish the contribution of the proposed method, as the performance improvement brought by instruction tuning is more pronounced and stable (Mistral-7B vs. Mistral-Instruct-7B and Mixtral-8x7B vs. Mixtral-Instruct-8x7B). Could the authors provide more explanation and analysis to clarify this result?\n5. Although the datasets used in this paper include a variety of question-answering formats, such as free-form (short-form and multi-hop) and closed-set, they are all based on wiki-style or specific domains. To more convincingly demonstrate the method's effectiveness, the authors should further extend their evaluation to more realistic and domain-diverse RAG datasets, like FreshQA and BRIGHT. Could the author also discuss the potential challenges in applying their method to these datasets and propose specific experiments to show the performance?\n\n[1] Kadavath et al., Language Models (Mostly) Know What They Know, 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024speculative,\ntitle={Speculative {RAG}: Enhancing Retrieval Augmented Generation through Drafting},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xgQfWbV6Ey},\nnote={under review}\n}" }, "abstract": { "value": "Retrieval augmented generation (RAG) combines the generative abilities of large language models (LLMs) with external knowledge sources to provide more accurate and up-to-date responses. Recent RAG advancements focus on improving retrieval outcomes through iterative LLM refinement or self-critique capabilities acquired through additional instruction tuning of LLMs. In this work, we introduce Speculative RAG - a framework that leverages a larger generalist LM to efficiently verify multiple RAG drafts produced in parallel by a smaller, distilled specialist LM. Each draft is generated from a distinct subset of retrieved documents, offering diverse perspectives on the evidence while reducing input token counts per draft. This approach enhances comprehension of each subset and mitigates potential position bias over long context. Our method accelerates RAG by delegating drafting to the smaller specialist LM, with the larger generalist LM performing a single verification pass over the drafts. Extensive experiments demonstrate that Speculative RAG achieves state-of-the-art performance with reduced latency on TriviaQA, MuSiQue, PopQA, PubHealth, and ARC-Challenge benchmarks. It notably enhances accuracy by up to 12.97% while reducing latency by 50.83% compared to conventional RAG systems on PubHealth." }, "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": [ "generative model", "retrieval augmented generation" ] }, "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/5404b2c6d52a09993b93dd65a3797328da13ba75.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": "Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1)\tThe discussion in line 315 is a bit confusing. The authors say “MindSearch does not yield better performance in terms of facticity, but as per Figure 4 in the paper, factuality of MindSearch is preferred 70% of the time.\n\n2)\tPlease consider showing the example in Figure 5 in English." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1)\tThe paper demonstrates considerably better output responses for MindSearch, compared to proprietary AI-Search engines like Perplexity Pro and ChatGPT-Web.\n\n2)\tMindSearch also works considerably better than the closed-book and ReACT baselines on a variety of multi-hop question-answering datasets. \n\n3)\tExtensive analysis and evaluation provided in terms of the prompting strategy for WebPlanner along with using a graph-based methodology vs JSON-based and code-based." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes MindSearch, an LLM-based multi-agent information-seeking framework for complex multi-step information-seeking questions. MindSearch includes a Web Planner which decomposes the user query into atomic sub-questions as nodes in a dynamic graph and progressively extends the graph based on results from the WebSearcher. MindSearch considerably improves in response quality in terms of depth and breadth and also improves over the baseline react-based iterative search system." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1)\tWhile the paper only evaluates for final response quality, it does not consider the attribution quality of the generated response. Popular AI search engines like Perplexity.AI and ChatGPT-web also provide citations as part of the generated output. The authors do not discuss whether MindSearch provides any kind of attribution, and if yes, what does the citation quality look like (based on automatic evaluations like ALCE [1])\n\n2)\tNo analysis was provided with regard to the dynamic graph constructed by the WebPlanner. Does the number of hops in the question match the depth of the tree? How often is an incomplete graph created? Also, it would be interesting to see a cost analysis in terms of the number of search queries that MindSearch generates, in comparison to the baselines (ReACT specifically)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 happens when Webplanner and code style interaction is not employed ? Is query decomposition required for all queries in web-searcher ? There is also a lack of qualitative analysis of failure scenarios. What happens when response at one node of the chain is wrong ? Does it result in cascading failures. Is there ayn mechanism for the Webplanner to detect such mistakes with feedback from websearcher ? \n\n2. Was there any qualitative evaluation on the benchmark where several human subjects were involved in performing the task with corresponding measurement of time taken ? to compare to mindsearch ?\n\n\n3. how do you respond to the first point in the weakness 1." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The problem is both interesting and important. Multi-agent systems for complex QA tasks that are robust and effective \n\n2. Easy to follow and the methods are simple and well explained.\n\n3. Experiments that include inference cost analysis is well considered." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents MindSearch, a multi-agent system for complex tasks that uses large language models (LLMs) and search engines for complex web information-seeking tasks. MindSearch addresses complex queries by decomposing them and retrieving information hierarchically, modeling the process as an iterative graph construction to enhance precision and recall. By distributing tasks across specialized agents, the framework manages complex and extended contexts effectively. The authors show that experimental results using GPT-4o and InternLM2.5-7B MindSearch outperforms benchmarks like ChatGPT-Web and Perplexity.ai, with human evaluators preferring its responses." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The work fails to cite and compare to other relevant baselines. For complex QA tasks like HotpotQA or MusiqueQA self-ask[1] with search is a relevant baseline. Similarly Searchain[2] is particularly relevant as it also forms a global reasoning chain or graph where the query is decomposed into subquestions that comprise the nodes of the chain and this planning is similar in philosophy to Mindsearch. I think Assistantbench[3] released in July 2024 is also very relevant and useful to evaluate on. The method SeeplanAct proposed in the paper would serve as a strong baseline. SeeAct[4] is also a relevant baseline. While the authors have cited the same they have not compared to this approach. Other RAG baselines in AssistantBench are also relevant.\n\n2) Some claims are unsupported. For instance the claim made in abstract and section 2.3 regarding the utility of Mindsearch : “Mindsearch performs in 3 minutes tasks worth 3 hours of human effort” has no related evidence cited in the paper. Was there any qualitative evaluation on the benchmark where several human subjects were involved in performing the task with corresponding measurement of time taken ? to compare to mindsearch ?.\n\n3) The work also misses on some important ablations. What happens when Webplanner and code style interaction is not employed ? Is query decomposition required for all queries in web-searcher ? There is also a lack of qualitative analysis of failure scenarios. What happens when response at one node of the chain is wrong ? Does it result in cascading failures. Is there ayn mechanism for the Webplanner to detect such mistakes with feedback from websearcher ? The current approach is a simple tool use based approach which has been well explored in existing WebAgent based works. The additional analysis and error handling mentioned above may help strengthen and understand the core contributions of MindSearch\n\n[1] Measuring and Narrowing the Compositionality Gap in Language Models, Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, Mike Lewis\n\n[2] Search-in-the-Chain: Interactively Enhancing Large Language Models with Search for Knowledge-intensive Tasks, Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, Tat-Seng Chua [3] AssistantBench: Can Web Agents Solve Realistic and Time-Consuming Tasks?, Ori Yoran, Samuel Joseph Amouyal, Chaitanya Malaviya, Ben Bogin, Ofir Press, Jonathan Berant [4] GPT-4V(ision) is a Generalist Web Agent, if Grounded, Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, Yu Su" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "See 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 writing and framework of this paper are clear and easy to follow.\n\nS2: The method is novel, utilizing the agents WebPlanner and WebSearcher to perform web search tasks.\n\nS3: Extensive experiments are conducted, demonstrating both the effectiveness and efficiency of this approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel tool agent, MindSearch, which decomposes user queries into atomic sub-questions represented as graph nodes and progressively extends the graph based on the search results from WebSearcher. For each sub-question, WebSearcher performs hierarchical information retrieval using search engines to gather relevant information for WebPlanner. Extensive experiments are conducted, including both open-set and closed-set datasets, and using open-source models alongside close-sourced LLMs, demonstrating the its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1: In Figure 5, the words should also be accompanied by English translations.\n\nW2: For WebSearcher, how does the LLM select the most valuable pages from all the retrieved web content? More details should be provided. Additionally, regarding answer generation, the statement, \"After reading these results, the LLM generates a response to answer the original question based on the search results,\" requires further elaboration, such as information on input design or specific prompt construction.\n\nW3: For the open-set evaluation, five experts are chosen. The author should provide more details, including whether these experts had prior exposure to the answers generated by MindSearch. Furthermore, examples should be included to intuitively demonstrate the differences between the responses generated by MindSearch, ChatGPT-Web, and Perplexity.ai.\n\nW4: The author could provide information on token consumption to help the community manage the budget when using MindSearch in their projects." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- When constructing the DAG, how does MindSearch automatically create graph nodes? What are some tips for structuring question as graph nodes?\n- When large amounts of content are retrieved, how does WebSearcher reduce noise? And, as the rapid growth of web content can easily exceed the maximum context length of the LLM, how does WebSearcher effectively limit content length?\n- Additionally, could the experiment include more closed-source and open-source LLMs to further validate the effectiveness of the method?" }, "rating": { "value": 6 }, "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 logical approach to the problem, with a well-organized visual format that is easy to understand and read. \n- This method provides a novel question-answering retrieval method based on directed acyclic graphs, which makes the RAG more reasonable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a system called MindSearch, designed to emulate human cognitive processes to enhance web information retrieval and integration tasks. By combining large language models (LLMs) with search engines, the system addresses limitations in handling complex queries, fragmented information, and lengthy content through an LLM-based multi-agent framework.\n* WebPlanner: Simulates the cognitive process of multi-step information seeking by breaking down user queries into atomic subproblems, represented as nodes in a graph. The graph is then progressively expanded based on search results from WebSearcher.\n* WebSearcher: Conducts hierarchical information retrieval for each subproblem, using a search engine to gather valuable information for WebPlanner.\n\nThis multi-agent design enables MindSearch to search and integrate information from vast web sources within three minutes, equivalent to saving three hours of manual effort. MindSearch demonstrates significant response quality improvements in both closed-set and open-set QA tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The method part is not detailed enough to show the technical details. For instance, the design of DAG and the use of DAG is not clear.\n- Few baseline methods from the same category are included, and many RAG-based question-answering approaches are left unexamined, such as ChatKBQA, AutoReAct, etc. \n- The backbone was only tested on GPT-4 (close sourced) and InternLM2.5 (open sourced). Under this setting, it is hard to tell if the MindSearch will work for all (at least most) LLMs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024mindsearch,\ntitle={MindSearch: Mimicking Human Minds Elicits Deep {AI} Searcher},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xgtXkyqw1f},\nnote={under review}\n}" }, "abstract": { "value": "Information seeking and integration is a complex cognitive task that consumes enormous time and effort. Inspired by the remarkable progress of Large Language Models, recent works attempt to solve this task by combining LLMs and search engines. However, these methods still obtain unsatisfying performance due to three challenges: (1) complex requests often cannot be accurately and completely retrieved by the search engine once (2) corresponding information to be integrated is spread over multiple web pages along with massive noise, and (3) a large number of web pages with long contents may quickly exceed the maximum context length of LLMs. Inspired by the cognitive process when humans solve these problems, we introduce MindSearch to mimic the human minds in web information seeking and integration, which can be instantiated by a simple yet effective LLM-based multi-agent framework. The WebPlanner models the human mind of multi-step information seeking as a dynamic graph construction process: it decomposes the user query into atomic sub-questions as nodes in the graph and progressively extends the graph based on the search result from WebSearcher. Tasked with each sub-question, WebSearcher performs hierarchical information retrieval with search engines and collects valuable information for WebPlanner. The multi-agent design of MindSearch enables the whole framework to seek and integrate information parallelly from larger-scale (e.g., more than 300) web pages in 3 minutes, which is worth 3 hours of human effort. MindSearch demonstrates significant improvement in the response quality in terms of depth and breadth, on both close-set and open-set QA problems. Besides, responses from MindSearch based on InternLM2.5-7B are preferable by humans to ChatGPT-Web and Perplexity.ai applications, which implies that MindSearch can already deliver a competitive solution to the proprietary AI search engine." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "language model", "search engine", "multi-agent 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/c119427897b9714f0418071eb00cad13c8849ee5.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": "MindSearch: Mimicking Human Minds Elicits Deep AI Searcher" }, "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.", "Yes, Research integrity issues (e.g., plagiarism, dual submission)" ] }, "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 numerous existing token merging approaches that extend beyond their application in diffusion models and generative tasks. The proposed method appears to function as a plug-and-play token merging technique. How does it perform when integrated with baseline models and discriminative tasks? Are the improvements consistently observed across these models and tasks?\n\n- Could the authors provide more detailed information on the implementation of the tile-shaped regions?\n\n- The submodular-based destination selection appears analogous to Farthest Point Sampling (FPS). To my understanding, in most 3D applications, the FPS algorithm is implemented with CUDA to achieve acceptable speed. This step seems to contribute significantly to the computational overhead of the proposed method. Could the authors clarify the distinctions between the submodular approach and FPS, particularly in terms of efficiency?\n\n- In the original ToMeSD paper (applied in SD 1.5), the results indicate a reduction in inference time (s/img). However, in Table 1, even at higher compression rates (0.5 and 0.75), this reduction is not evident. Could the authors provide an explanation for this discrepancy?" }, "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-crafted and effectively articulates both the proposed methodology and the corresponding experimental outcomes.\n- The implementation of the approach is methodical and straightforward, which supports practical applicability.\n- The comprehensive implementation details, supplemented by the provided code, significantly bolster the reproducibility of the research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents three significant advancements aimed at enhancing the token merging mechanism in diffusion models for generative tasks: identifying representative merge destinations, optimizing the merging and unmerging processes, and reducing computational complexity. Specifically, the proposed method employs a greedy-based algorithm to determine a representative subset that serves as merge destinations. This is followed by an additional cross-attention operation and matrix multiplication to effectively execute the merging process. During the unmerging phase, the authors leverage the inverse (or transpose) matrix from the merging step, thereby improving the overall efficiency of the unmerging procedure. Moreover, the authors introduce strategies to merge only tokens located within the same local region and to share destination and merge matrices across iterations and layers, further mitigating computational costs. When compared to an existing approach (i.e., ToMeSD), the proposed method achieves notable improvements in text-to-image generation tasks across two datasets (GEMRec and ImageNet-1k) evaluated using three metrics (CLIP, DINO, and FID), highlighting its efficacy and substantial contribution to the field." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The title and scope of the paper may lead to potential misunderstandings. While diffusion models have applications beyond generative tasks, the experiments in this work are solely focused on generation. It would be advisable to revise the title to more accurately reflect the scope of the contributions.\n- The experimental evaluation is restricted to text-to-image tasks, which limits the generalizability and perceived practical impact of the proposed approach.\n- The discussion and comparative analysis do not sufficiently engage with related work on token merging, such as CrossGET [1] and TRIPS [2], which diminishes the thoroughness of the literature review.\n- The comparative evaluation is limited to ToMeSD, and there are notable inconsistencies when compared to the results reported in the original paper.\n\n\n\n[1] CrossGET: Cross-Guided Ensemble of Tokens for Accelerating Vision-Language Transformers, ICML 2024\n\n[2] TRIPS: Efficient Vision-and-Language Pre-training with Text-Relevant Image Patch Selection, EMNLP 2022" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses part." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tThe method uses a submodular function to identify a representative subset of tokens for merging and applies a GPU-efficient vectorized optimization algorithm.\n2.\tThe design of ToMA carefully considers the advantages and limitations of GPU computations.\n3.\tToMA achieves 30%-50% speedups without noticeable sacrifice in image quality." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To address the two main challenges in token merging, this paper introduces the TOMA method. TOMA first uses a submodular-based approach to select diverse tokens for merging. It then leverages efficient attention implementations to minimize merge overhead. By abstracting (un-)merging as (inverse) linear transformations, TOMA enables shared computation across layers and further accelerates processing by operating on tokens within local blocks to exploit image locality." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This work seems like an enhanced version of ToMeSD, focusing on updated merge rules and additional locality optimizations, but the contributions may not be substantial enough.\n2. Regarding experimental results:\n- The paper only tests on the SDXL architecture, limiting generalization claims. As noted in line 372, this method could be extended to SD2 and SD3, so more results on these structures are needed. Actually, using token merging in the DiT structure could theoretically offer greater speedups.\n- The results in Table 1 for ToMeSD are strange, as its inference time is longer than the baseline. Were torch and xformer versions verified to match the official implementation during testing? Without a correct ToMeSD implementation, comparisons may lose significance.\n- FID scores in Figure 5 exceed 25, unusually high for ImageNet.\n- The speedup achieved by ToMA is limited. At a ratio of 0.25, the improvement is just 10%, and while a ratio of 0.75 yields a 20% speedup, it comes with a significant decline in quality metrics.\n- The comparison methods are limited; it would be beneficial to include approaches such as “Token Downsampling for Efficient Generation of High-Resolution Images.”\n3. Some figures and explanations are unclear, e.g., the X-axis in Figure 5." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Refer to Weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The use of submodular optimization for token selection effectively reduces information loss during merging, improving quality retention compared to previous approaches.\n\n2. The paper's experiments, which utilize metrics such as CLIP, DINO, and FID on high-quality datasets, demonstrate ToMA's balance between efficiency and image quality." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Token Merge with Attention (ToMA) to optimize transformer-based diffusion models, addressing inefficiencies in existing token merging methods. By utilizing submodular optimization for token selection, efficient attention mechanisms, and leveraging token locality, ToMA achieves substantial computational speedups with minimal impact on image quality, making it compatible with modern GPU architectures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The use of locality to limit the scope of attention for computational efficiency, as implemented in ToMA, is not sufficiently novel. Similar approaches have already been explored in methods such as Sparse Transformer[1], DiffRate[2], ToFu (Token Fusion)[3], making it difficult to assess the unique contribution of ToMA.\n\n2. The experimental comparisons are primarily limited to ToMeSD, without benchmarking against other prevalent methods such as Token Pruning, Flash Attention, DiffRate[2], ToFu[3], and FRDiff[4].\n\n3. The paper is lack of qualitative visual analysis. Without sufficient visual examples, it is challenging to assess ToMA's performance meaningfully, especially in comparison to other acceleration methods.\n\n[1] Child, R., Gray, S., Radford, A., & Sutskever, I. (2019). Generating Long Sequences with Sparse Transformers. arXiv preprint arXiv:1904.10509.\n\n[2] Chen M, Shao W, Xu P, et al. Diffrate: Differentiable compression rate for efficient vision transformers[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 17164-17174.\n\n[3] Kim, M., Gao, S., Hsu, Y.-C., Shen, Y., & Jin, H. (2023). Token Fusion: Bridging the Gap between Token Pruning and Token Merging. arXiv preprint arXiv:2312.01026.\n\n[4] So J, Lee J, Park E. FRDiff: Feature Reuse for Universal Training-free Acceleration of Diffusion Models[J]. arXiv preprint arXiv:2312.03517, 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": { "value": "Nan" }, "flag_for_ethics_review": { "value": [ "No 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.\tI noticed that in the smaller steps of Figure 4, the average token intersections across different layers are significantly different. In these steps, could the sharing of both destinations and attention weights between layers lead to a notable loss in performance?\n\n2.\tHow is the scale of the set of destinations determined? Specifically, how does the size of 𝐷.\n\n3.\tThe terms \"Dino\" and \"Clip\" mentioned in line 475 should be aligned with the entries in Table. 3." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper exhibits several strengths:\n\n1.\tThe motivation and methodology are both reasonable and intuitive.\n\n2.\tThe generation model (SDXL) is significantly accelerated by merging and unmerging tokens before and after attention, along with additional speed-up settings, all without any loss in quantitative performance indicators." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Token Merge with Attention (ToMA) to tackle the issues of limited image quality due to the loss of important tokens and the inefficiency of attention mechanisms. The authors establish ToMA through three major components: a submodular-based token selection method, an efficient attention implementation, and (un-)merging as (inverse-)linear transformations. Based on this design, the paper significantly reduces the inference time for text-to-image generation model (SDXL)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper exhibits several Weaknesses:\n\n1.\tLack of qualitative comparison with ToMeSD.\n\n2.\tThe visual effects of ToMA are underrepresented, and quantitative indicators only partially reflect the quality of generation. More samples are needed to substantiate claims about \"the best trade-off between image quality and speed.\"\n\n3.\tCurrent Text-to-Image models (such as Flux and SD3) based on diffusion transformers have achieved new state-of-the-art results. While the paper states that ToMA can be applied to any attention-based T2I model, it is recommended that the authors verify ToMA's performance on the latest T2I models to enhance persuasiveness.\n\n4.\tIn the bottom of Figure 6, ToMA introduces considerable noise compared to the original result. Does this imply that, despite ToMA showing less performance loss in quantitative evaluations, it incurs greater performance loss in terms of visual perception?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "My primary concern is the lack of discussion regarding the DiT model. Could the authors provide additional results or discussions specifically related to DiT image generation models, such as PixelArt-Alpha?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The use of attention-based operations for token merging is well-designed and makes sense. Additionally, the authors' choice to make it an invertible function is highly meaningful, and they thoughtfully consider GPU implementations.\n\n- The authors also discuss sharing destination selections across steps, which indeed reduces computational costs and enhances the practicality of the overall approach.\n\n- The experiments are comprehensive and well-executed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Previous token selection methods have overlooked the relationships between tokens and have not utilized the latest attention implementations, limiting actual speedup.\nThis paper proposes a submodular function-based token selection mechanism and introduces an attention-based approach for merging and unmerging tokens. This design leverages the benefits of modern attention acceleration libraries and is reversible in nature.\nAs a result, the authors' method achieves an optimal trade-off between performance and efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The manuscript lacks discussion on the DiT model, focusing only on Stable Diffusion. In the \"Local Region\" section, it would be beneficial to include insights on how this technology could be adapted for DiT-like models. Without convolution layers, it is unclear if the locality is still evident enough to support the use of this method. Can you provide a brief analysis or discussion on how the locality assumptions might change for DiT models, and whether any modifications to the proposed method would be needed to accommodate those differences.\n\n- Some figures could be improved for better visualization.e.g., it is a little difficult to differentiate different methods in Figure 5.\n\n- The manuscript contains redundant content, specifically in lines L142-L150 and L151-L155, where identical information is repeated." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose an improved token merging algorithm to speed up diffusion." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024toma,\ntitle={To{MA}: Token Merging with Attention For Diffusion Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xhtqgW5b93},\nnote={under review}\n}" }, "abstract": { "value": "Diffusion models have emerged as leading models for image generation. \nPlug-and-play token merging techniques have recently been introduced to mitigate the heavy computation cost of transformer blocks in diffusion models. \nHowever, existing methods overlook two key factors: 1. they fail to incorporate modern efficient implementation of attention, so that, the overhead backfires the achieved algorithmic efficiency 2. the selection of token to merge ignores the relation among tokens, limiting the image quality. \nIn this paper, we propose Token Merging with Attention(ToMA) with three major improvements. Firstly, we utilize submodular-based token selection method to identify diverse tokens as merge destination, representative of the entire token set. Secondly, we propose attention merge, utilizing the efficient attention implementation, to perform the merge with negligible overhead. Also we abstract the (un-)merging as (inverse-)linear transformations which also allows shareable transformation across layers/iterations. Finally, we utilize the image locality to further accelerate the computation by performing all the operations on tokens in local tiles." }, "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", "Token Merge", "Attention" ] }, "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/731cdda030b454cf7258c80592cbcd9a846466d2.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/fcdac60247f71d978b4212fd423cabccddba78a2.zip" }, "title": { "value": "ToMA: Token Merging with Attention 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": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why were methods like LoRA and other recent parameter-efficient fine-tuning techniques not included in the comparison? Would the results hold up against these widely-used benchmarks?\n\n2. How does the method perform on more diverse and real-world tasks outside the provided datasets? Are there specific domains where L-MSA is particularly effective or less so?\n\n3. Given the still significant computational demands, how does L-MSA compare in terms of efficiency gains relative to other lightweight fine-tuning techniques? Is the trade-off between computational cost and performance improvement justified?\n\n4. The method currently selects one layer at a time for fine-tuning. Could this approach be extended to simultaneously fine-tune multiple layers, and if so, how would it impact performance and computational cost?\n\n5. Have the authors considered dynamically reselecting layers during the training process? Would this adapt better to changing training dynamics and potentially improve generalization?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Novel Layer Selection Metric:** The paper introduces a new metric for layer selection based on the Method of Successive Approximations (MSA), which is theoretically grounded and offers a fresh perspective on fine-tuning strategies.\n\n2. **Theoretical Foundations:** The authors provide a comprehensive theoretical analysis within the context of deep linear networks, which strengthens the credibility of the proposed approach.\n\n3. **Empirical Validation:** The paper demonstrates the effectiveness of L-MSA across multiple datasets and tasks, showing consistent improvement over several existing layer-wise fine-tuning methods.\n\n4. **Parameter-Efficiency Focus:** By targeting specific layers for fine-tuning, the method aims to reduce computational costs, addressing a key challenge in training large-scale models.\n\n5. **Clear Contributions to Layer-Wise Fine-Tuning:** The research highlights the potential of selectively fine-tuning layers to achieve better performance, contributing to the growing field of parameter-efficient fine-tuning methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel approach for efficient fine-tuning of large-scale models. It addresses the challenge of substantial memory consumption associated with such models by introducing L-MSA, a method that fine-tunes only selected layers based on a new metric derived from the Method of Successive Approximations (MSA). This metric guides layer selection and optimizes their fine-tuning, resulting in better model performance with reduced computational costs.\n\nThe paper provides a theoretical analysis within deep linear networks and evaluates L-MSA across various datasets and tasks. It compares the proposed method with other layer-wise fine-tuning techniques, demonstrating its superior performance. The experimental results show that L-MSA consistently identifies the most impactful layers for fine-tuning and optimizes them effectively, outperforming baseline methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Lack of Comparison with State-of-the-Art Methods:** The paper does not compare L-MSA with prominent methods like **LoRA**[Hu et, al], which limits the evaluation of its relative effectiveness and practical impact.\n\n2. **Disjointed Presentation:** The flow of the paper is disrupted by referencing equations and concepts out of order, requiring readers to frequently navigate back and forth, which hampers comprehension and readability.\n\n3. **Inadequate Contextualization of Contributions:** The novelty of the proposed method is not sufficiently contextualized against a broader range of parameter-efficient fine-tuning techniques, making it harder to assess its uniqueness and value.\n\n4. **Generalization Concerns:** The paper acknowledges that the approximated updated loss may not always guarantee strong generalization to test data, which raises questions about the robustness of the approach in diverse real-world scenarios.\n\n5. **Computational Demands:** Despite aiming for parameter efficiency, the method still involves substantial computational overhead due to both forward and backward propagation, which could limit its practicality for very large-scale models.\n\n6. **Limited Scope of Empirical Comparisons:** While the paper evaluates L-MSA across several tasks, the range of comparative baselines is not exhaustive, potentially missing out on broader insights.\n\n[Hu et. al, ICLR 2022, LoRA: Low-Rank Adaptation of 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": 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": "1. Why the Method of Successive Approximations is good for layer selection and layer fine-tuning?\n2. What is the goal of the proposed method? Which layer should be actually selected for fine-tuning?\n3. Compared with the baseline fine-tuning methods, what is the main advantage of the proposed method?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The task of parameter-efficient fine-tuning is crucial for practical applications of pretrained models.\n\n2. The authors performed an in-depth analysis on the effectiveness of fine-tuning at the layer level." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores layer-wise fine-tuning strategies for adapting large pretrained models. The authors found that fine-tuning specific layers can lead to varied performance outcomes, and selectively fine-tuning certain layers may enhance results. Building on these insights, they propose a novel layer-wise fine-tuning approach that leverages Pontryagin’s Maximum Principle to guide layer selection for fine-tuning. The effectiveness of this method is demonstrated through transfer learning from ImageNet to CIFAR100." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper's novelty is somewhat limited, as the idea that fine-tuning different layers can lead to varying performance outcomes has already been widely explored in previous research. It is not clear what is the main contribution of the paper. \n\n2. The proposed method, which relies on the Method of Successive Approximation, lacks clear motivation. It’s unclear why this approach would lead to improved layer selection or how it compares favorably to other existing methods. In other words, what are the benefits of the proposed approach?\n\n3. For the theoretical analysis, it is not clear what is the main contribution. The authors should also connect the theoretical analysis to the proposed method and discuss why the proposed method can lead to better layer selection and fine-tuning. \n\n4. The explanation based on PMP is interesting, but the main technical contribution is not clear. It would be helpful if the authors could clarify where the primary novelty lies.\n\nOverall, while the paper presents a reasonable idea, it suffers from poor writing, insufficient motivation and unclear 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": 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": "- [Suggestion] Some of the contributions listed in the paper need to be merged and toned down. The paper currently lists four contributions. not contributions. The first contribution is closely related with the fourth one. Hence my suggestion to merge them, Moreover, this fourth contribution is related to the empirical validation of the proposed method. In this regard, the proper validation of a proposed method is not a plus but a must for a decent piece of scientific work. Therefore, I would suggest toning down this claim.\n \n- [Suggestion] In Fig. 6, I suggest reporting the loss curves on a per-dataset basis. It might also be informative to discuss any trend observed during the training iterations.\n \n- [Question] In l.379, it is indicated that “…he results illustrate that our L-MSA metric consistently identifies layers associated with improved training loss, effectively pinpointing those that contribute to better training outcomes...“ May you indicate how do you observe this in that figure? From what I can interpret, the results from Fig. 4 are point estimates at a given training iteration. Therefore, it is hard for me to assess how this figure allows to observe a consistent behavior.\n \n- [Question] In l.408, it is stated “…Notably, using L-MSA for layer-wise fine-tuning results in performance improvements of up to 20% compared to full fine-tuning” May you indicate the source of this observation?\n \n- [Question] In Sec. 4.3 a “No Adaptation” baseline is included. In l.432 it is indicated that this baseline “…provides a reference point for model performance without fine-tuning”, May you motivate how good of a reference this baseline is for datasets like CIFAR-Flip and ImageNet-C considering the number of output classes is significantly different (CIFAR-Flip) and/or reduce (ImageNet-C)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- A theoretical analysis for the proposed method has been provided (Sec. 3 & Sec. A.1).\n \n- The empirical evaluation of the proposed method includes a good variety of datasets and competing methods. This allows to assess the applicability of the proposed method in different contexts.\n \n- The validation of the proposed method includes an ablation analysis this is effective towards obtaining insights on how the different components of the proposed method contribute to its performance and how it compares w.r.t. full fine-tuning.\n \n- The paper adequately highlights the limitations of the proposed method (Sec. 5)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method to do a more targeted fine-tuning process where instead of updating every single layer/parameter in the architecture, the layer that has the strongest effect in the total loss is selected and updated on each iteration of the fine-tuning process.\n\nTowards this goal the paper proposes a metric to select the layer to be updated and an algorithm for the fine-tuning of the selected layers.\n\nExperiments on several datasets (CIFAR-100, CIFAR-C, CIFAR-Flip, Living-17 and ImageNet-C) including several related methods (full fine-tuning, LIFT, LISA, surgical Fine-tuning and Auto-RGN) provide evidence on the performance of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The originality of the paper is somewhat reduced. While the method put forward by the paper outperforms (in some cases) the considered baselines, as admitted by the paper (Sec. 4.1)  there are already methods in the literature that aim at a more targeted fine-tuning process.\n \n- In its current form the paper does not feel self-contained, there are several aspects of the paper that are delegated to the appendix. For instance, the models considered in experiments of Sec. 4.2 are not explicitly indicated. In other cases, the protocols and motivations behind the conducted experiments are not clear. Here a proper balance must be achieved to ensure the paper provides sufficient details as to allow the reader to critically analyze the proposed method and its conducted validation.\n \n- While pairing a method with its corresponding theoretical analysis is very desirable, the extend to which such analysis is currently provided (Sec. 3) is just too shallow as to be informative.  Perhaps it could be completely moved to the appendix in order to allocate space to further elaborate on other parts of the paper that are currently not detailed enough.\n \n- Some statements seem to lack supporting evidence. For instance, in several parts of the paper (l.535) statements are made regarding to the reduced computational costs of the proposed method. The evaluation section is missing however a proper experiment addressing that aspect.\n \n- The improvement put forward by the proposed method is not that outspoken. For instance, in Table 1 it is observed that the proposed method is better in only 2/4 considered settings.\n \n- In Fig. 6, averaged results over different datasets are reported. This not only hinders the variations in performance across datasets, but also behavior observed for the different methods across the considered datasets.\n \n\n- Weak positioning; Good part of the related work is centered around other not directly related aspects. For instance, Sec. 6.1 discusses large architectures, which has close no link w.r.t. the proposed method.  Similarly l.497-502 from Sec. 6.2 is related to prompt-based methods. Thus having a relatively weak link with the proposed method. This weak positioning w.r.t. related efforts becomes more evident when we consider that almost none of the related methods considered in the experimental section (Sec. 4) are covered in the related work section. In addition, I would suggest looking at the two references below which seem to be very related to the proposed method.\n \n - Youngmin Ro, Jin Young Choi, \"AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks\", AAAI 2021\n - Basel Barakat; Qiang Huang, \"Enhancing Transfer Learning Reliability via Block-Wise Fine-Tuning\", ICMLA 2023" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Q1.The other works compared in this paper are all about NLP and LLM. Why did the authors only apply these works to image datasets and test them on image datasets instead of NLP datasets used in other papers?\n- Q2.Does L-MSA introduce additional memory overhead compared to randomly selecting other papers? I have not been involved in PEFT, but in theory it seems to require the same memory spikes as Full Fine-turn, which runs counter to the motivation of the authors.The authors should provide a detailed analysis of the memory usage of L-MSA compared to full fine-tuning and other PEFT methods, including peak memory usage and average memory consumption?\n- Q3.Does performing MSA on the loss of the network introduce more training time? Especially when the network size increases, it is necessary to calculate and compare the state and co-state variables layer by layer.\n- Q4.In Figure 5, in the experiments of ImageNet to CIFAR-100, it can be clearly observed that Full-Finetuning and Auto-RGN continue to converge, while L-MSA seems to have completed convergence. Can you show the results for more epochs?\n\nIf the author resolves my questions, I will consider raising my rating." }, "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 easy to follow.\n- The experimental results seem to be promising, surpassing Full Fine-tuning on part of the dataset" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The machine learning community has made remarkable progress with the advent of large-scale models, but these models become a significant obstacle by consuming large amounts of memory during training. In this paper, we propose a fine-tuning Method using the Method of Successive Approximations, L-MSA. Experiments on different data sets verify the effectiveness of L-MSA" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- This paper emphasizes the advantages on large-scale models, but the datasets used seem to focus on extremely small datasets such as CIFAR, which show the worst performance on experimental fine-tuning on Imagenet-C. The only LIFT that seems to outperform is a paper that has not yet been published through peer review. This makes me worry about its practical prospects.\n- Not enough experiments, this paper is motivated by the fact that large-scale models consume memory, but there is no comparison of memory overhead in the paper." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose L-MSA, a novel layer-wise fine-tuning approach, which encompasses both the criterion for layer selection and the algorithm for fine- tuning the targeted layer." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024lmsa,\ntitle={L-{MSA}: Layer-wise Fine-tuning using the Method of Successive Approximations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xi3sDtf8A0},\nnote={under review}\n}" }, "abstract": { "value": "With the emergence of large-scale models, the machine learning community has witnessed remarkable advancements. However, the substantial memory consumption associated with these models has emerged as a significant obstacle to large-scale training. To mitigate this challenge, an increasing emphasis has been placed on parameter-efficient fine-tuning methodologies, which adapt pre-trained models by fine-tuning only a subset of parameters. We observe that in various scenarios, fine-tuning different layers could lead to varying performance outcomes, and selectively fine-tuning certain layers has the potential to yield favorable performance results. Drawing upon this insight, we propose L-MSA, a novel layer-wise fine-tuning approach that integrates two key components: a metric for layer selection and an algorithm for optimizing the fine-tuning of the selected layers. By leveraging the principles of the Method of Successive Approximations, our method enhances model performance by targeting specific layers based on their unique characteristics and fine-tuning them efficiently. We also provide a theoretical analysis within deep linear networks, establishing a strong foundation for our layer selection criterion. Empirical evaluations across various datasets demonstrate that L-MSA identifies layers that yield superior training outcomes and fine-tunes them efficiently, consistently outperforming existing layer-wise fine-tuning 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": [ "layer-wise finetuning", "parameter-efficient fine-tuning", "method of successive approximations" ] }, "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/1e54a5d336fb0477b769b6c3fdbbedece45f7f28.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/a867fef986d3459ce644226439a031bbd29c1b76.zip" }, "title": { "value": "L-MSA: Layer-wise Fine-tuning using the Method of Successive Approximations" }, "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": "Refer to the weakness part." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThe method is novel which allows the client download prompts from others.\n2.\tThe experiments are extensive and show good performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces pFedMoAP that enables effective federated prompt learning for vision-language models like CLIP. The key innovation is allowing clients to download multiple pre-aggregated prompts as fixed non-local experts rather than being restricted to a single globally aggregated model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tIt is very similar to the prompt learning framework. How does your method differ from the classic prompt learning framework if you can directly access prompts from other clients?\n2.\tI noticed that similar work, like pFedPrompt, conducted experiments on large-scale datasets like UCF-101 or ImageNet. Would the scale of the dataset influence the performance?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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. In Section 4.1 **Implementation details.**, for CIFAR10&CIFAR100, why the participation rate is 10%? Does it mean only one client will be selected to join the training process in each communication round?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Implementing prompt learning into the distributed environment is an important topic for FL applications due to its efficiency in computation and communication.\n\n2. The proposed learning framework is simple and effective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a lightweight federated prompt learning framework to personalize the prompt learning process through the lens of mixture of experts (MoE). An attention-based gating network is also introduced to achieve efficient cross-client knowledge sharing. Experiments indicate that the proposed **pFedMoAP** performs better than state-of-the-art methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The quality of the writing is poor. The expression of the paper is somehow obscure, and not concise enough. Besides, there are many typos in the paper, e.g., lacking space between two adjacent words in Line 27, Line 198, Lin 415, and Line 469. \n\n2. The novelty is limited. The **pFedMoAP** seems like a naive combination of PromptFL and MoE. Although the selected non-local experts will be updated in each communication round, the intuition behind it is also not explained well, leading to weak convincingness. \n\n3. it is unclear why the attention-based gate network does not need to be uploaded to the server for aggregation. Please clarify.\n\n4. I think the ablation study of the paper is not enough. \n- To study the effectiveness of MoE-based textual feature enhancement mechanism, the authors should add experiments under different values of $K$ to confirm performance changes.\n\n- The author implements a dimension reduction operation in Section 3.3, please add experiments to show whether the performance will be influenced. \n\n- How about a larger number of clients, e.g., 100 or more?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Please refer to the weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The integration of the MoE method into prompt-based federated learning is an innovative concept.\n\n2. The paper is well-organized and effectively communicates its main contributions. The methodology is clearly explained, particularly regarding the incorporation of the MoE approach in the federated learning context." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a Mixture of Experts (MoE) approach for prompt-based federated learning, employing a local attention-based gating network to refine text features for better alignment with local image data. It also highlights the lightweight design of the prompt, exploring a configuration where the prompt is updated directly by other clients rather than relying on an aggregated prompt from the server." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While this approach utilizes inter-client sharing, it could expose prompt updates directly to other clients, which may lead to privacy concerns as prompt updates could be tracked, making the model susceptible to certain attack algorithms. \n\n2. In the related work section in the appendix (Line 783), the authors mentioned that previous work utilized inter-client sharing prior to aggregation. However, these works allow prompt aggregation in the server and do not share local prompts with other clients. This sentence is somewhat unclear; please provide a clearer version to avoid misunderstandings. \n\n3. The paper claims that the MoE-based method can leverage prompts from other clients. Please provide experimental evidence to show whether your method’s effectiveness is sensitive to the total number of clients and the value of $K$. \n\n4. Some related works are encouraged to be properly cited in Related work. For example, personalized federated learning [a,b] and prompt-based federated Learning [c]. \n\n[a] Li, et al, FedTP: Federated Learning by Transformer Personalization, TNNLS 2023. \n\n[b] Cai, et al, Fed-CO2: Cooperation of Online and Offline Models for Severe Data Heterogeneity in Federated Learning, NeurIPS 2023. \n\n[c] Pan et al, Federated Learning from Vision-Language Foundation Models: Theoretical Analysis and Method, NeurIPS 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please 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. pFedMoAP proposes a novel framework that personalizes the prompt learning process by allowing clients to download multiple pre-aggregated prompts as fixed non-local experts. This leverages the concept of Mixture of Experts, where a local attention-based gating network is implemented to generate enhanced text features that better align with the local image data on the client side.\n\n2. By facilitating a many-expert scenario (e.g., 10 experts) per client with negligible communication overhead, pFedMoAP achieves efficient and effective personalization. This approach overcomes the limitations of having too many or too few experts, which can lead to high communication costs or suboptimal performance, respectively.\n\n3. The algorithm consistently outperforms existing work across various heterogeneous federated settings, as demonstrated through extensive experiments on 9 datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces a new paradigm in federated prompt learning for pre-trained Vision-Language Models (VLMs) such as CLIP, challenging the traditional restriction that clients are limited to downloading a single globally aggregated model. This shift is particularly beneficial for lightweight prompts, enabling more flexible and effective knowledge sharing among clients." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Determining which prompt experts should be included in the pool and when to update or replace them requires careful management. If the selection process is not optimal, it could lead to suboptimal performance. Additionally, maintaining a diverse and up-to-date pool of experts is crucial but can be complex.\n\n2. The local attention-based gating network, while designed to be efficient, may still introduce optimization challenges. Ensuring that this network converges to a solution that effectively combines local and non-local prompts is not trivial, especially when dealing with heterogeneous data distributions.\n\n2. As the number of clients increases, managing a dynamic pool of prompt experts and maintaining efficient communication between the server and all clients can become more challenging. The complexity of coordinating updates and ensuring that each client receives the most relevant prompts could grow significantly." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose pFedMoAP, a novel personalized federated prompt learning framework for CLIP-like VLMs under data heterogeneity from a Mixture of Experts perspective." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024mixture,\ntitle={Mixture of Experts Made Personalized: Federated Prompt Learning for Vision-Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xiDJaTim3P},\nnote={under review}\n}" }, "abstract": { "value": "Prompt learning for pre-trained Vision-Language Models (VLMs) like CLIP has demonstrated potent applicability across diverse downstream tasks. This lightweight approach has quickly gained traction from federated learning (FL) researchers who seek to efficiently adapt VLMs to heterogeneous scenarios. However, current federated prompt learning methods are habitually restricted to the traditional FL paradigm, where the participating clients are generally only allowed to download a single globally aggregated model from the server. While justifiable for training full-sized models under federated settings, in this work, we argue that this paradigm is ill-suited for lightweight prompts. By facilitating the clients to download multiple pre-aggregated prompts as fixed non-local experts, we propose Personalized Federated Mixture of Adaptive Prompts (pFedMoAP), a novel FL framework that personalizes the prompt learning process through the lens of Mixture of Experts (MoE). pFedMoAP implements a local attention-based gating network that learns to generate enhanced text features for better alignment with local image data on the client, benefiting from both local and downloaded non-local adaptive prompt experts. The non-local experts are sparsely selected from a server-maintained pool, fostering collaborative learning across clients. To evaluate the proposed algorithm, we conduct extensive experiments across 9 datasets under various heterogeneous federated settings. The results show that pFedMoAP consistently outperforms the state-of-the-art alternatives, underscoring its efficacy in personalizing prompt learning for CLIP within the federated learning paradigm." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Federated learning", "prompt learning", "vision-language model", "mixture of experts" ] }, "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/4b93ce3fbfa4257d51a27424369e4287a2f6ab04.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/6b3cb5025f97bbcdcbdebfc02bd05abc18700b62.zip" }, "title": { "value": "Mixture of Experts Made Personalized: Federated Prompt Learning for Vision-Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "The main desideratum is a substantial overhaul of the paper's writing, especially as it pertains to the methodology Section 3. \n\nThere are other minor concerns, such as the exact conformal methods that the current proposed method is tested against. For instance, APS has been superseded in the literature by RAPS, so that should be included in the comparison. Also, importantly, at least a couple of the many recently appearing class-conditional and local conformal methods should also be included in the comparison for the sake of fairness; indeed, these methods are designed to be able to empirically attain partial versions of conditional coverage without receiving explicit constraints (such as fairness ones) as input, so they should provide a stronger baseline than methods like APS." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper's main asset is a reasonably broad evaluation of the proposed methods on a variety of datasets where fairness concerns may necessitate using one of the tabulated conformal fairness metrics. Compared to prior works the experimental section appears more extensive, both in terms of metrics and datasets. Moreover, conformal fairness guarantees were displayed on graph data in addition to the standard supervised tasks. The results indicate that explicitly enforcing fairness according to any of the tabulated fairness criteria is in fact often necessary, as evidenced by vanilla methods not satisfying non-enforced fairness coverage constraints. Furthermore, no drastic deterioration in efficiency guarantees is observed, indicating that the price to pay for fairness conditional coverage may often be acceptable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript introduces and studies a general-purpose conformal fairness framework. It consists of an algorithm that can enforce various notions of fairness (such as demographic parity, equalized odds, predictive parity, and several others contained in the ML fairness literature) on the coverage of prediction sets, in the sense of satisfying appropriate conformal guarantees on expected coverage conditional on sensitive labels/attributes. To demonstrate practical performance of the framework given the various fairness notions, experimental evaluation is provided on a variety of supervised learning tasks; besides the standard setting, graph conformal prediction with fairness constraints is also demonstrated in experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "--- The main weakness and bottleneck at this point is the writing of the manuscript. In particular, the writing of Section 3, which introduces the objectives, the general algorithm, and the analysis, is currently not acceptable. Indeed, lots of key notions and terms, both on the fairness side and on the conformal side, are not properly and unambiguously introduced and/or are discussed in arbitrary order. \n(Fairness: groups/group collections are never formally defined, nor are requirements on them specified (e.g. non-intersectionality etc); \"filters\" and \"filter functions\" are simply thrown into the presentation. No background on the fairness measures, no elaboration on how they are transformed into conformal analogues compared to original notions, except for an incoherent sentence starting with \"Essentially achieved...\" in line 130. Conformal prediction: auxiliary notions like \"interval widths\", \"label miscoverages\" appear right away in the pseudocode, which also uses clunky notation and for which the textual explanation is as confusing as the pseudocode notation.) This is not to say that one cannot fill in many of the details with enough imagination and enough expertise, but at the moment the writing is hardly structured and accessible enough.\n\n--- The theoretical/methodological side does not offer strong novel contributions; for the most part it simply wraps class-conditional and group-conditional conformal methodology into fairness-specific terminology --- and is currently doing it in a somewhat confusing way, given the currently unsatisfactory presentation as stated above.\n\n--- In addition, the literature background review, where it's claimed that there are \"very few prior efforts\" on fairness and conformal prediction, misses an established line of work on group-conditional fairness guarantees; these works study the enforcement of coverage guarantees (usually of both upper and lower bounds) on rich classes of subpopulations given by possibly arbitrarily overlapping groups.\n\nR. Barber et al: The limits of distribution-free conditional predictive inference [Journal of IMA 2020]\nC. Jung et al: Batch Multivalid Conformal Prediction [ICLR 2023]\nZ. Deng et al: Happymap: A generalized multi-calibration method [ITCS 2023]\nO. Bastani et al: Practical adversarial multivalid conformal prediction [NeurIPS 2022]\nI. Gibbs et al: Conformal Prediction with Conditional Guarantees [2023]" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "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": 2 }, "primary_area": null, "questions": { "value": "- How does the framework handle scalability with multiple sensitive attributes, especially computationally?\n\n- Why were existing fairness-aware methods not used as baselines for comparison?\n\n- How practical is the exchangeability assumption for real-world graph data? Any empirical evidence?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Originality: Introduces \"Conformal Fairness,\" extending Conformal Prediction (CP) to address fairness in uncertainty quantification, especially for non-IID data.\n- Quality: Provides strong theoretical backing with rigorous proofs and effective validation through experiments on graph and tabular datasets.\n- Clarity: Clearly defined theoretical concepts and a stepwise algorithm description make the methodology accessible to those with relevant background knowledge. Providing additional background information on key concepts would help readers who are less familiar with the topics to follow more easily\n- Significance: Tackles an important gap by combining fairness with uncertainty quantification. Adaptable to multiple fairness metrics and data types, making it broadly applicable in ethical AI contexts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces \"Conformal Fairness,\" a framework that extends Conformal Prediction (CP) to ensure fairness across sensitive groups. The authors develop a theoretically grounded algorithm to control gaps in coverage between these groups, leveraging the exchangeability assumption instead of the typical IID assumption. This allows the framework to be applied to non-IID data types, including graph data. Experiments on both graph and tabular datasets show that the framework can effectively control fairness gaps while maintaining coverage guarantees." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper lacks a detailed discussion of the fairness-efficiency trade-off. Quantifying acceptable efficiency losses when fairness is improved would make the results more actionable for practitioners balancing both aspects.\n\n- The extension of the exchangeability assumption to real-world data may not always hold. Adding empirical evidence or discussion on when this assumption is valid in practice would make the claims more robust.\n\n-Lack of comparison with all existing fairness-aware methods limits benchmarking. Adding baselines would clarify the framework's effectiveness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "I think it would be very helpful for the authors to explain more about their design choices. For example, why choose inverse quantile for miscoverage level? Explain a bit more about the intuitions behind lemma 3.1 and 3.2 would really help the readers." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Conformal prediction is a very interesting area and there are many interesting works in this area. It is a natural question to ask how to achieve fairness for models in this setting. The authors provide a very general framework that can achieve many different definitions of algorithmic fairness. The paper also did many interesting experiments in graph datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework to check and find the suitable threshold for conformal learning. Further, the framework can be easily extended to some constraints, like fairness constraints. The authors specifically studied achieving fairness properties for graph models. Further, the authors studied how to use their framework to check the fairness property of a model. The authors' algorithm first find the right use inverse quantile to check whether a threshold is good." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I personally think the author does not elaborate enough on why they chose this algorithm design. The writing of this paper could be improved." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Please see the weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper is well-written with a clear structure.\n\n2. The notion of conformal prediction is common, while its application on fairness is not widely studied to the best of my knowledge. The novelty of this notion is well justified.\n\n3. The conformal fairness metrics in Table 1 are intuitive and well-founded." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a notion of fairness, named $\\textit{conformal fairness}$. In contrast with standard fairness notions (e.g. demographic parity), this notion significantly applies conformal mapping, which the authors clearly defined. The authors developed an algorithm to achieve this type of fairness on graph datas and confirmed with experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the conformal fairness metrics are well-defined in Table 1, they were not carefully discussed in later results (Theorem 3.4). Specifically, Theorem 3.4 states that \"the coverage difference for Predictive Parity is satisfied\". In Table 1, the definition of Predictive Parity relies on a conformal prediction $\\mathcal{C}$, but the specific mapping/quantification is not discussed in the theorem statement.\n\n2. This is a continuation of the previous point. Suppose there is a conformal prediction in Theorem 3.4, how good is it, i.e. what shall the value of $\\alpha$ be, as in Equation 1?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Generic Framework for Conformal Fairness},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xiQNfYl33p},\nnote={under review}\n}" }, "abstract": { "value": "Conformal Prediction (CP) is a popular method for uncertainty quantification with machine learning models. While the method provides probabilistic guarantees regarding the coverage of the true label, these guarantees are agnostic to the presence of sensitive attributes within the dataset. In this work, we formalize \\textit{Conformal Fairness}, a notion of fairness using conformal predictors, and provide a theoretically well-founded algorithm and associated framework to control for the gaps in coverage between different sensitive groups. Our framework leverages the exchangeability assumption (implicit to CP) rather than the typical IID assumption, allowing us to apply the notion of Conformal Fairness to data types and tasks that are not IID, such as graph data. Experiments were conducted on graph and tabular datasets to demonstrate that the algorithm can control fairness-related gaps in addition to coverage aligned with theoretical expectations." }, "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": [ "Fairness", "Conformal Prediction", "Graph Neural 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/4df4abf5fa6f9e591d00bdee1717967002a28b2b.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/808fe5049890b5f355b94384745dcbdfec4f2b00.zip" }, "title": { "value": "A Generic Framework for Conformal Fairness" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How do you see vector-ICL applied in practical scenarios, considering decreasing inference cost and increasing context length of LLMs? Given your experiments show that without task-specific fine-tuning, vector-ICL consistently outperforms tokens-based ICL?\n\n- Did you explore instruction tuning after pre-training? Authors should discuss the tradeoffs between task-specific finetuning and more general approaches like instruction tuning/ human preference training." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The concept of using vectors for in-context learning is a new exploration, and the proposed methods for vector-ICL and evaluations show promising results.\n\n- Using light-weight trainable projectors with simple pre-training on general task is also not very expensive and can be integrated as a part of general LLM pre-training. \n\n- Experiments cover a wide range of tasks and modalities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the feasibility of vector-ICL, that extends the in-context learning capabilities of LLMs to continuous vectors. Authors use light-weight projectors to align embedding space of input text embedding with LLM. To train the projectors, they use general language modeling objective followed by task-specific objectives. Authors experimented with multiple tasks and modalities, e.g., text classification, summarization, time-series classification, fMRI decoding, etc." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Method lacks depth. Specifically, replacing any length text for any complexity task with a single embedding may not be sufficient. Including an ablation where text is replaced with a series of vectors (one vector per sentence/ chunk) would be helpful.\n\n- Currently, the method requires task-specific fine-tuning to outperform token-ICL. I think authors should explore RLHF/ instruction finetuning datasets/ objectives to avoid task-specific finetuning." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* Figure 4a: Why does the correlation vary so much between datasets and models of similar size?\n* Figure 4b: What do the axes represent?\n* lines 477-479: Can you expand on the block patterns and how they are explained?\n* How does the finetuning work precisely? Do you finetune with multiple shots already? If so, how many?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The proposed approach is simple to understand and implement.\n* The method is evaluated on a wide range of tasks and datasets, including multiple modalities.\n* Parts of the results suggests that the models perform better with more shots.\n* The paper is relevant to the ICLR audience." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new method for in-context learning. By encoding examples into single vector representations via a separate, frozen encoder, and training a simple projection on top of the vector, LLMs are trained to perform ICL from continuous vector representations alone (Vector-ICL). The projection is trained a) via pretraining on unlabeled data and can optionally be b) finetuned on a labeled corpus.\nThe method is evaluated on a range of tasks, including text only tasks like text classification and summarization, as well as multi-modal tasks that utilize encoders for brain scans, graph data, or time-series." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper claims that with finetuning the vector-ICL method outperforms standard ICL. However, The method's baseline is standard-ICL without any finetuning, i.e. the method compares a supervised method (albeit with a weak adapter projection P) with an unsupervised one. A baseline that is obviously missing is when finetuning the base model with standard ICL, perhaps using an adapter that is equally weak. For text classification and summarization, the soft prompting baseline may be adequate, but the reason for its inclusion is never discussed.\n* The method is not well-motivated. In lines 34-38 it states that many data modalities cannot be well represented in natural language, making continuous vector representations necessary for in-context learning. While this is true, many of the experiments on non-textual either don't benefit from multiple shots or need to be trained on downstream data to work (e.g. time-series classification and graph classification in Figure 3, bottom). What's the value of Vector-ICL here? If you need to finetune a model to perform a task, can it still be called in-context learning?\n* The paper structure could be better. The information pertaining to a particular task is scattered all over the paper, such that I found myself scrolling back and forth a lot. For example, I did not find Figure 3 particularly helpful because the information necessary to understand it fully is located far away from it (e.g. what the horizontal bar represents). I think it would be better to focus on fewer tasks in the main body of the paper, but explain these really well. The remainder could go into the appendix." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "q1: How does the proposed approach differ from soft prompt tuning (Li and Liang, 2021)?\n\nq2: In 3.3 it is mentioned that the LLM would be trained with ``conditional generation loss''. However, in Figure 2b that is referred to, the LLM is supposed to be frozen. Could you clarify this point and explain the training process in more detail, including which components are frozen and which are updated during a) pre-training and b) fine-tuning?\n\nq3: For Time Series (l 346), it is mentioned that you use the output of the last time step from Chronos-base. Doesn't this mean that the model is not able to use the full context of the time series but just the final state? What are the trade-offs between using the full context of the time series and just the final state?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "To the best of my knowledge, the proposed method is original (although some connections with soft prompt tuning should be discussed).\nThe paper is for the most part clear and the approach is interesting in the sense that it shows that LLMs can operate in-context on projected vectors. The results seem promising, as the learned projectors allow LLMs to tackle new tasks (graphs, fMRI etc) that are (assumed to be) not possible to tackle without projectors. The experimental part covers a wide range of tasks with different input types (8 tasks, with only 3 being classic text tasks and the proposed method beats its baselines in most cases. Clarity is generally good with some exceptions (see questions)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper is about projecting input data into an LLM embedding space, such that the LLM can perform in-context learning on the projections. The paper claims that this approach can improve the performance of the LLM on downstream tasks and opens up new application domains with Graphs, fMRI, and Time Series." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. A weakness is that the projectors require pre-training with a language modeling objective and task-specific fine-tuning. It feels this defeats the purpose of in-context learning (i.e. not needing any training data to tackle a new task) to some extent. Authors can reflect if a change of the proposed methods name would be needed here.\n2. Although the paper uses soft prompt tuning as a baseline, the relationship of the proposed approach with soft prompt tuning (Li and Liang, 2021; and follow-up work) is not explicitly discussed, i.e. as both methods learn some artificial token representations that are placed into the prompt.\n3. The paper assumes that regular ICL is not applicable to Graphs, fMRI, and Time Series. While arguably not a natural fit, this assumption should be supported by some experiments, e.g. just using the numerical PCA vectors as text, or some flattened representation of the graph edges. As people have even thrown the parameters of another LM at an LM -- just in text format, I could imagine that even those non-text data types could be successfully encoded as text in a more straightforward manner without the need for training artificial tokens.\n4. The experiments on time-series, graph classification and two fMRI tasks do not have a baseline, but only compare the pretrained-only projectors with fine-tuned projectors. It would be interesting to see how the proposed approach compares to some standard baseline of the respective domain, i.e. graph neural networks for graphs.\n\n\nMinor point: In terms of presentation, the results description could directly link to the results, even if repetitive, e.g., Fig 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": 4 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper identifies an interesting research problem and presents a well thought-out exploration. The idea of using embeddings for conditional generation has been explored n various indirect forms (e.g., [1]), but the paper's focus on ICL seems to be a useful contribution.\n\n[1] Text Embeddings Reveal (Almost) As Much As Text (Morris et al., 2023)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores using embeddings for examples in in-context learning (ICL), which may be more suited in non-textual tasks. The paper proposes a light-weight pretraining and finetuning scheme to obtain the encoder. The paper shows improvements over textual ICL on various non-textual tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed framework is back to encoder-decoder. In this context, previous works like Flan-T5 have thoroughly explored the model's ICL capabilities [2]. While the paper differs since it focuses on a single embedding and non-textual tasks, it's a bit odd to omit this context.\n\n2. Intuitively vector-ICL should not work as well as textual ICL for a lot of tasks. Clearly if the task requires a lot of symbolic data ICL is strictly more powerful than the proposed approach. But the paper seems to mostly show positive results. It'd be good to do some exploration of the cases vector-ICL is worse than textual ICL and by how much. \n\n[2] Scaling Instruction-Finetuned Language Models (Chung et al., 2022)" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We discover that large language models can effectively process and in-context learn from continuous representations from various domains, often outperforming regular ICL and domain-specific models across diverse tasks and modalities." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024vectoricl,\ntitle={Vector-{ICL}: In-context Learning with Continuous Vector Representations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xing7dDGh3},\nnote={under review}\n}" }, "abstract": { "value": "Large language models (LLMs) have shown remarkable in-context learning (ICL) capabilities on textual data. We explore whether these capabilities can be extended to continuous vectors from diverse domains, obtained from black-box pretrained encoders. By aligning input data with an LLM's embedding space through lightweight projectors, we observe that LLMs can effectively process and learn from these projected vectors, which we term Vector-ICL. In particular, we find that pretraining projectors with general language modeling objectives enables Vector-ICL, while task-specific finetuning further enhances performance. In our experiments across various tasks and modalities, including text reconstruction, numerical function regression, text classification, summarization, molecule captioning, time-series classification, graph classification, and fMRI decoding, Vector-ICL often surpasses both few-shot ICL and domain-specific model or tuning. We further conduct analyses and case studies, indicating the potential of LLMs to process vector representations beyond traditional token-based paradigms." }, "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", "in-context 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/85fc1ddf9c80f4dbb8482bf8a871242ea3d563d8.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": "Vector-ICL: In-context Learning with Continuous Vector Representations" }, "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": "- ***Challenges in Finding Nash Equilibrium***: Given that finding a Nash Equilibrium is PPAD-complete and challenging even with state-of-the-art solvers, could the authors provide more insight and detailed discussion on optimizing the equilibrium for protein applications, especially in high-dimensional cases? The paper could better illustrate why the benefits of introducing the equilibrium outweigh the additional challenges rather than implying that solvers are guaranteed to find the NE. This clarification would prevent confusion among readers not familiar with the literature, especially since multiple NE solvers are discussed while claiming guarantees of finding local optima.\n\n- ***Price of Anarchy Discussion***: The discussion of the price of anarchy briefly addresses the motivation for the problem formulation but is not sufficiently convincing. Providing more details and insights specific to protein design tasks could improve the presentation and strengthen the argument.\n\np.s. Due to my limited domain expertise, I am not confident in assessing the novelty of the problem formulation in this specific application." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- ***Innovative Formulation***: The paper introduces an interesting formulation by shifting the focus from kernel design for protein structures to an objective that reflects the insight that multiple separate contributing factors exist within protein design tasks.\n\n- ***Sample Efficiency Guarantee***: The algorithm comes with a sample efficiency guarantee, which is valuable for practical scientific experimental design tasks where data collection can be costly and demand a principled optimization solution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel method that frames the protein design task as an optimization problem of finding a Nash Equilibrium (NE) with unknown utility functions. It provides sample efficiency guarantees, and empirical results using multiple NE solvers suggest promising improvements over existing baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- ***Limited Novelty in BO Guarantees***: The Bayesian Optimization (BO) component naturally inherits the guarantees of the Upper Confidence Bound (UCB) algorithm. There doesn't appear to be a significant difference or improvement, especially considering that the method does not address specific challenges unique to protein design.\n- ***Lack of Comparison with High-Dimensional BO Methods***: BO methods tackling protein design typically involve specific treatments for high-dimensional spaces. Including comparisons against these methods would strengthen the paper, especially when ESM embedding allows the direct application of such 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": 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.\tDoes this algorithm work without assumption $|\\mathcal{X}^{(i)}|=d$? It means that in this case the numbers of possible decision variables are different.\n2.\tWhat is the v function in Algorithm 2? Is that the r function defined in Definition 3.1?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tCombinatorial Bayesian optimization (CBO) is a very important subfield in Bayesian optimization. While Bayesian optimization has been successfully applied to many areas, studying CBO helps us solve even more practical application problems, like protein design.\n2.\tSample complexity guarantees are provided to show the algorithm’s ability to achieve approximate Nash equilibrium.\n3.\tExperiments on protein design are comprehensive and ablation studies are provided." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies how to solve the combinatorial Bayesian optimization problem, which has many important potential applications like protein design. The key idea of this work is introducing the Nash equilibrium to Bayesian optimization, where a game is carefully designed for domain variables to play and the decision making is driven by equilibrium finding algorithms. It’s good to see substantial experiments in protein design in the end." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tTo be honest, as a ML researcher, I have limited background in Nash equilibrium and I think it could be beneficial for this paper to provide more information on introducing Nash equilibrium since this is an ICLR submission. Although locally optimal, why and how does Nah equilibrium help in CBO setting? Could you provide a brief comparison between the local optimality of Nash equilibria and global optimality in the context of CBO. Can we design some globally optimal CBO algorithms or are the authors aware of such algorithms? Also, it might be hard to theoretically investigate the performance difference between GameOPT and the global optimal solution, it is feasible to investigate it in experiments where a small finite decision space is used. That would greatly provide valuable insights into the method's effectiveness.\n2.\tI have concerns on the computational efficiency of Step 5 of Algorithm 1. Because the GP is assumed on function f which is now defined on the combinatorial space, how do you find the top B equilibria according to UCB? Could you provide more details on how to implement this step efficiently, discuss the computational complexity of this step compared to the overall algorithm, and explain any approximations or heuristics that may be used if an exact solution is computationally infeasible?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Is there any consideration of balance of exploration and exploitation? How is the data-efficiency guaranteed?\n2. What is the influence of batch size B? In what case may the algorithm perform worse?\n3. Why not compared baselines using tree-based surrogate models. As far as I know, the tree model is a suitable choice for protein design.\n4. How to prove that solving equation (2) is more tractable than solving other acquisition functions?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The discrete black-box optimization problems are important, and the proposed game-theoretical solution seems novel.\n2. The protein design is vital in drug discovery and healthcare.\n3. The presentation is clear and the idea is easy-to-follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studied the combinatorial Bayesian optimization in the discrete search space. Based on the theory of cooperative games and upper confidence bound of Gaussian processes (GPs), a new acquisition function that seeks to find the local equilibria of players was proposed. Based on batch and parallel computation, the proposed method achieved better scalability. The effectiveness and competitiveness were demonstrated in the application of computational protein design." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is no mechanism to avoid or mitigate stucking in local optimality. At the same time, as a Bayesian optimization method, the exploration-exploitation balance is not considered. As a preliminary solution, random restarts can be introduced into the search process, similar to what the trust-region Bayesian optimization algorithm does [1], to ensure exploration and global search ability. \n2. As local optima, the regret bound is not insightful enough to show the superiority. Currently, the regret is similar to what the original GP-UCB is, but only local optima are guarenteed. While the GP-UCB ensures global search and convergence. In my opinion, assuming convexity in the local region containing equilibria may help improve the theoretical results [2].\n3. The testing benchmark problems are not diverse enough. Some well-known synthetic problems are required for fair comparison, espectially that employed in the state-of-the-art work.\n4. The influence of hyper-parameters, especially the batch size B, is not clear. The baseline algorithms, especially the GP-UCB, are not tailored for batch optimization. It is recommended to consider smaller B, such as B=1, first. \n5. The literature review is insufficient. More baseline algorithms in combinatorial Bayesian optimization and general optimizers should be considered.\n\n[1] David Eriksson, et al.: Scalable Global Optimization via Local Bayesian Optimization. NeurIPS 2019.\n\n[2] Shuang Li, et al.: \"Why Not Looking backward?\" A Robust Two-Step Method to Automatically Terminate Bayesian Optimization. NeurIPS 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See the weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Large-scale combinatorial optimization is an important challenge, though Bayesian optimization (BO) may not be the most typical approach for tackling it.\n- While theoretical analysis is provided for the convergence rate to an $\\epsilon$-Nash equilibrium. However, it does not necessarily equate to finding the global optimum of the objective function, and the results are not significantly different from known results for existing game-theoretic BO literature." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles scalability challenges in combinatorial black-box optimization tasks, specifically when each dimension contains the same number of discrete points. While this assumption is not universally applicable in combinatorial Bayesian optimization, it is relevant in specific domains, such as protein engineering. The authors introduce GameOpt, a novel approach that leverages the uniform structure of the discrete search space. By reframing the combinatorial optimization problem as an $n$-player cooperative game, the method aims to identify an $\\epsilon$-Nash equilibrium at each iteration. This transformation enables the application of scalable solvers and polynomial-time equilibrium-finding techniques from game theory, offering a computationally efficient strategy for optimizing the acquisition function, particularly GP-UCB." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Clarity**\nThe weakest part of this paper is definitely the clarity, particularly in math. There are so many uncommon, unexplained, and undefined symbols, that make me frustrated. Detailed below:\n\n**Problem Statement**. What exactly is $x$? I initially understood it as a $d \\times n$-dimensional variable. However, in Line 107, the statement \"f(x) corresponds to --- the amino acid sequence $x$, and each $x$ can take $20^n$\". I've lost. Is $x$ meant to be a $d$-dimensional variable or a $d \\times n$-dimensional variable? I also can’t interpret the meaning of \"each $x$ can take $20^n$\". Shouldn't $d = 20$ in this case? Additionally, the definition of the fitness function $f(x)$ is unclear. Is $f(x)$ defined for each discrete variable, or is it applied to the entire variable set? For example, if $x$ is assumed to be $d$-dimensional, does our objective function take the form $\\prod_{i=1}^n f(x^{(i)})$? Or is $x$ actually $d \\times n$-dimensional, and thus $f(x)$ refers to something else? I’ll proceed under the assumption that $x$ is $d \\times n$-dimensional.\n\n**Section 3**. I understand the reward $r^{(i)}$ here is the UCB defined for $x$ as a $d \\times n$-dimensional input, and each player $i$ corresponds to one of the discrete variables. However, since UCB is defined on the entire $d \\times n$-dimensional space, how can we apply UCB as $r^{(i)}$ then? How do you select the other $n-1$ variables? Or does $r^{(i)}: \\mathcal{X} \\rightarrow \\mathbb{R}$ imply that each player takes a $d \\times n$-dimensional input and $r^{(i)}$ = UCB for all $i$? If that is the case, and all players share the same utility function, why do equilibria arise? Wouldn't this reduce to the same optimal points for all players? Additionally, in Definition 3.1, $r^{(i)}$ takes two arguments, but this is supposed to be UCB. Why does UCB take two arguments? What is arg eq? What exactly does Eq (2) represent? I'm also confused by the explanation on Lines 172-173: \"Intuitively, equilibria are computed by breaking down the complex decision space into individual decision sets.\" Could you precisely explain this process before jumping into the intuition? How is finding a Nash equilibrium equivalent to maximizing UCB? What is the payoff $v$ in Algorithms 2 and 3? I've lost again. \n\nI’ll stop pointing out specific issues here, but I want to note that the same level of confusion persists throughout the remaining sections (although the introduction was quite smooth).\n\n- **Lack of Critical Guarantee**\nIs finding a Nash equilibrium mathematically equivalent to maximizing the objective function $f$? If I understand correctly, this process is more akin to constructing a Pareto frontier, which contains the global maximum, making it more aligned with finding the set of local maxima. How, then, does this algorithm guarantee global convergence? If the goal of this work is local optimization or a heuristic approach, this should be clearly stated in the title or assumptions. Approximations can be valuable, but only if they provide some guarantee that they can recover the original problem under some conditions. Existing work, such as [1], provides such guarantees. I am not surprised that a local optimization algorithm finds local optima (as this work does) faster than a global optimizer (like [1]) in certain experiments, as they target different objectives. Even with these results, I wouldn’t choose this method for global optimization. When the function query is highly expensive, the computational overhead of the acquisition function becomes comparable. Moreover, [1] is not exactly state-of-the-art as it dates to 2022. There are likely more recent studies, such as [2], though performing a literature search is not my role here.\n- **Limited Evaluation to Assess Practicality**\nHeuristic approaches are completely acceptable if they prove useful in real-world applications. However, in that case, we expect the algorithm to be genuinely practical and perform best among existing heuristics. In the context of protein optimization, chemists would most likely start with deep learning-based approaches, especially diffusion-based methods, which are already well-established. There is a plethora of scalable and sample-efficient work in this area, such as [3, 4]. These methods seem like a more natural choice than using deep learning embeddings as features, as done in this work. Even with embeddings, the GP has no prior data on the objective function, meaning BO must start without a pretrained dataset. Alternatively, why not reduce the dimensionality of the embedding features to make them more tractable for the naïve GP-UCB? This could be achieved by adding just one additional layer to the transformer. In doing so, standard GP-UCB would likely perform just fine like popular latent BO doing. Since this paper does not compare its method against these popular alternatives or simpler alternatives, I cannot properly evaluate whether it represents a good heuristic. What I gather from this work is that it performs better than outdated methods on the limited tasks provided, but that alone does not demonstrate its superiority or practicality compared to modern approaches.\n- **Limited Novelty**\nThe equation on line 295 appears nearly identical to the existing optimistic game-theoretic approach presented in [5], but it is not cited. Additionally, the idea of treating each discrete variable as a player is not particularly novel. For instance, Shapley value GP [6] and additive kernels [7] can be seen as variants of this concept (although they treat dimensions as players). While the specific combination of game theory, combinatorial optimization, and Bayesian optimization may be novel, it ultimately feels like a straightforward combination of existing approaches.\n- **Missing Limitation**\nI understand that this method leverages the problem structure where each dimension contains the same number of discrete variables, $n$. However, in typical combinatorial Bayesian Optimization, the goal is hyperparameter optimization, where each dimension may have a different number of categorical values, and the space can sometimes be a mixture of continuous and discrete variables. This method is not applicable to such cases. In mixed-variable scenarios, the existence of a Nash equilibrium cannot be guaranteed.\n\n- **Citation**\n- [1] Daulton, Samuel, et al. \"Bayesian optimization over discrete and mixed spaces via probabilistic reparameterization.\"NeurIPS 2022\n- [2] Papenmeier, Leonard, et al., \"Bounce: reliable high-dimensional Bayesian optimization for combinatorial and mixed spaces.\" NeurIPS 2023\n- [3] Gruver, Nate, et al. \"Protein design with guided discrete diffusion.\" NeurIPS 2023\n- [4] Campbell, Andrew, et al. \"Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design.\" arXiv preprint arXiv:2402.04997 (2024).\n- [5] Han, Minbiao, et al., \"No-Regret Learning of Nash Equilibrium for Black-Box Games via Gaussian Processes.\" UAI 2024.\n- [6] Chau, Siu Lun, et al., \"Explaining the uncertain: Stochastic shapley values for gaussian process models.\" NeurIPS 2023\n- [7] Kandasamy, Kirthevasan, et al., \"High dimensional Bayesian optimisation and bandits via additive models.\" ICML 2015." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024optimistic,\ntitle={Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xiyzCfXTS6},\nnote={under review}\n}" }, "abstract": { "value": "Bayesian optimization (BO) is a powerful framework to optimize black-box expensive-to-evaluate functions via sequential interactions. In several important problems (e.g. drug discovery, circuit design, neural architecture search, etc.), though, such functions are defined over large $\\textit{combinatorial and unstructured}$ spaces. This makes existing BO algorithms not feasible due to the intractable maximization of the acquisition function over these domains. To address this issue, we propose $\\textbf{GameOpt}$, a novel game-theoretical approach to combinatorial BO. $\\textbf{GameOpt}$ establishes a cooperative game between the different optimization variables, and selects points that are game $\\textit{equilibria}$ of an upper confidence bound acquisition function. These are stable configurations from which no variable has an incentive to deviate$-$ analog to local optima in continuous domains. Crucially, this allows us to efficiently break down the complexity of the combinatorial domain into individual decision sets, making $\\textbf{GameOpt}$ scalable to large combinatorial spaces. We demonstrate the application of $\\textbf{GameOpt}$ to the challenging $\\textit{protein design}$ problem and validate its performance on four real-world protein datasets. Each protein can take up to $20^{X}$ possible configurations, where $X$ is the length of a protein, making standard BO methods infeasible. Instead, our approach iteratively selects informative protein configurations and very quickly discovers highly active protein variants compared to other 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": [ "Combinatorial Bayesian Optimization", "Game Theory", "Gaussian Processes", "Protein 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/3ae89ed0db94e1278d950a381ce4c10078e49a6b.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": "Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design" }, "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 refer to the Weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Attempts to explain the inner workings of ICL, based on reasonable assumptions and investigative tools.\n- The findings align with previous work, encouraging readers to accept the claims presented in the paper.\n- Visualized results help readers quickly grasp the core concepts and findings of the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to explain the mechanisms behind in-context learning (ICL) using the inference circuit framework.\n\nAccording to the authors, the ICL process consists of three internal steps:\n1. Summarize: Large language models (LLMs) encode each demonstration within its corresponding forerunner token, $s_i$ .\n2. Semantics Merge: The semantics of each demonstration and its label are combined into the representation of the label $y_i$ .\n3. Feature Retrieval and Copy: LLMs rely on the accumulated labels $y_{1:k}$ to respond accurately to the query $s_q$ , yielding the most appropriate answer.\n\nEach step is empirically validated using methods such as kernel alignment and embedding comparisons. The authors also seek to align their findings with those of prior research, reinforcing the credibility of the arguments presented in this work." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the proposed framework is logical and reasonable, it remains challenging to argue definitively that the core mechanism of ICL follows the assumptions presented in the paper. As noted in Section 5.2, there are exceptions that do not align well with the proposed framework, raising concerns that the explanations may be superficial and fail to capture the essence of ICL. This is understandable, as fully explaining the inner workings of neural networks is inherently difficult, if not nearly impossible.\n- I am somewhat unclear about the core novelty of this paper. As I understand it, the primary contribution seems to be the attempt to apply the existing inference circuit framework to the ICL of specific LLMs, including LLaMA 3. In Section 2.1, I did not find explanations that clarify why the procedure conducted in this paper is particularly innovative, compelling, or novel. More comprehensive comparisons with prior work employing induction or inference circuits to illustrate the inner workings of ICL would be helpful to underscore the merits and uniqueness of this study.\n- In Section 3.1, sentence embeddings generated by an external encoder (BGE M3) are compared to hidden representations computed by an LLM. Since these two representations come from different models, without any modification or fine-tuning, there is a risk that their vector spaces are not aligned or compatible. This raises concerns about whether this experiment is sufficiently reasonable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors proposed to use the mutual nearest-neighbor kernel alignment of the intermediate representations of LLMs and sentence embeddings produced by another pre-training model to assess the quality of these representations. This method is novel.\n\n2. Extensive analysis has been performed on all three steps of the proposed framework. Possible explanations have also been provided for many phenomena.\n\n3. The experiments are performed with real-world LLMs and datasets, which makes the insights more likely to be useful in practice.\n\n4. The paper is well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a three-step inference circuit to capture the in-context learning (ICL) process in large language models (LLMs):\n\n1. Summarize: Each input (both demonstration and query) is encoded into linear representations in the model's hidden states.\n\n2. Semantics Merge: The encoded representation of each demonstration is merged with its label, creating a joint representation for the label and demonstration.\n\n3. Feature Retrieval and Copy: The model retrieves and copies the label representation most similar to the query's representation, using this merged representation to predict the query's label.\n\nThis circuit explains various ICL phenomena, such as position bias, robustness to noisy labels, and demonstration saturation. Ablation studies show that removing steps in this process significantly reduces performance, supporting the dominance of this circuit in ICL. The paper also identifies some bypass mechanisms that operate in parallel, assisting the inference circuit through residual connections." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The majority of the analysis is based on associations without verifying their strength and whether those effects are causal. For example, Figure 2 right does not look significant enough for me. The peaks highlighted in Figure 5 also look pretty noisy to me.\n\n2. The causal evidence that the authors provided in the ablation study only shows the effect of deleting the hypothesized important components in ICL. What if unimportant components are deleted? Would they have a similar effect? Only if the unimportant components have a significantly weaker effect on ICL performance, can we draw a causal conclusion that the proposed three-step process dominates ICL." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N.A." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please 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 findings in this paper are clearly explained. Experimental results and visualizations enhance readability and help the audience follow the study's progression easily.\n\n2. This work is well-connected to existing ICL research, with discussions that compare its findings to prior studies on ICL explainability and demonstration selection.\n\n3. The study uses multiple LLMs, strengthening the generalizability of the findings across different model architectures.\n\n4. The insights provided are thought-provoking and have potential practical implications for ICL applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the mechanisms within large language models (LLMs) that enable in-context learning (ICL) tasks, breaking down the process into three distinct stages: summarization, semantic merging, and feature retrieval/copying. The study employs a variety of experiments across multiple LLMs to validate its findings. Overall, this paper presents valuable insights that can contribute to the field of LLM research, particularly within the ICL community." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. [Section 3.1] The authors used mutual nearest-neighbor kernel alignment to evaluate LLMs' summarization abilities. However, the term “summarize” lacks clarity. Does it refer to encoding capabilities similar to those in BGE?\n\n2. [Section 3.1] Additionally, the kernel alignment metric may not be sufficiently robust, as alignment scores in Figure 2 range only from 0.25 to 0.35, which is not significant enough. Consequently, the finding on “summarization” may hold only to a limited extent.\n\n3. [Section 4.1 – Copying from Text Feature to Label Token] It is unclear whether the copying mechanism is applied solely to label tokens or if it extends to other tokens within the input. Using results from other tokens as a baseline could provide a more nuanced understanding of the copying process.\n\n4. [Figure 5, Right] After layer 40, the classification accuracy drops significantly. The authors did not investigate potential reasons for this decline. Could it be due to the gradual degradation of copied information?\n\n5. The experimental setup in Section 5.1 is insufficiently detailed. For instance, how many attention heads are disconnected at each layer? Additionally, the experiments lack certain baselines, such as randomly disconnecting some attention heads to observe the impact on model performance.\n\nDespite these questions and weaknesses, I believe this paper still offers meaningful insights." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "The three-stage framework proposed in the paper is quite interesting. The questions I have raised are mentioned in the weaknesses section. Here, I would like to know what inspired you to propose this framework. Each part of the framework consists of very specific ideas—were they derived from repeated trial and error, or were they inspired by something else? Alternatively, are they improvements on a significant prior work?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- **Originality:** To my knowledge, the three-stage circuit proposed by the authors is a novel contribution.\n- **Quality:** The hypothesis put forward is reasonable, and the experiments are thorough with a well-crafted methodology.\n- **Clarity:** The arguments and evidence presented in the paper are clear, and the experimental descriptions are appropriately detailed.\n- **Significance:** Currently, ICL is one of the most important applications in the LLM field, and understanding the mechanisms behind ICL will greatly aid in enhancing its performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a three-stage ICL circuit hypothesis and provides thorough empirical examinations of the existence and significance of these stages. Within this circuit framework, many phenomena are explained, such as how Forerunner Tokens encode input text representations and the bias in input text encoding towards position. These findings present intriguing insights." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The three-stage ICL framework appears to have implicit applicability conditions, which I believe should be clarified.\n\nFor example, in Fig. 1 on page 2, a few-shot scenario with $ k=2 $ is presented, which indeed fits the three-stage ICL circuit framework. However, in a zero-shot scenario ($ k=0 $), step 1 may still exist, but steps 2 and 3 would not be applicable. In a few-shot scenario with $ k=1 $, steps 1 and 2 might still apply, but step 3 cannot exist.\n\nTherefore, the framework proposed in this paper should be limited to discussions of scenarios where $ k \\geq 2 $. A related question arises: if the focus is restricted to this scenario, what potential issues might emerge?\n\nFurthermore, if we condition on $ k \\geq C $ (where $ C $ is a fixed value), could this value vary depending on the problem type? For instance, in tasks like SST-2 and SST-5, which have different label set sizes, might the value of $ C $ differ across these scenarios?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We decompose In-context Learning into 3 operations and measure their operating dynamics to catch many inference phenomenon of ICL in Large Langauge Models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024revisiting,\ntitle={Revisiting In-context Learning Inference Circuit in Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xizpnYNvQq},\nnote={under review}\n}" }, "abstract": { "value": "In-context Learning (ICL) is an emerging few-shot learning paradigm on Language Models (LMs) with inner mechanisms un-explored. There are already existing works describing the inner processing of ICL, while they struggle to capture all the inference phenomena in large language models. Therefore, this paper proposes a comprehensive circuit to model the inference dynamics and try to explain the observed phenomena of ICL. In detail, we divide ICL inference into 3 major operations: (1) Summarize: LMs encode every input text (demonstrations and queries) into linear representation in the hidden states with sufficient information to solve ICL tasks. (2) Semantics Merge: LMs merge the encoded representations of demonstrations with their corresponding label tokens to produce joint representations of labels and demonstrations. (3) Feature Retrieval and Copy: LMs search the joint representations similar to the query representation on a task subspace, and copy the searched representations into the query. Then, language model heads capture these copied label representations to a certain extent and decode them into predicted labels. The proposed inference circuit successfully captured many phenomena observed during the ICL process, making it a comprehensive and practical explanation of the ICL inference process. Moreover, ablation analysis by disabling the proposed steps seriously damages the ICL performance, suggesting the proposed inference circuit is a dominating mechanism. Additionally, we confirm and list some bypass mechanisms that solve ICL tasks in parallel with the proposed circuit." }, "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": [ "In-context Learning; Induction Circuit; Mechanistic 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/3f39e829a743291c790f552ca58b45508f9ace60.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Revisiting In-context Learning Inference Circuit in 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": { "value": "The proposed model for video guardrailing, along with the benchmark for evaluating safety policies in existing MLLMs, addresses sensitive content types. Given the nature of the videos and policies (safety categories) evaluated, it is essential to examine the work for potential biases, privacy concerns, potential harms, and legal compliance." }, "flag_for_ethics_review": { "value": [ "Yes, Discrimination / bias / fairness concerns", "Yes, Privacy, security and safety", "Yes, Legal compliance (e.g., GDPR, copyright, terms of use)", "Yes, Potentially harmful insights, methodologies and applications" ] }, "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": "- SFT Baseline: Could the authors provide additional context for the \"SFT baseline\" mentioned in Figure 5?\n- Inference Cost: What accounts for the increase in inference cost with additional few-shot examples, as illustrated in Figure 5?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper addresses a critical topic by proposing guardrails for video MLLMs based on defined safety policies, which is timely and important with the rise of MLLMs.\n- It introduces a baseline model built upon the InternVL2-8B backbone and leverages two plug-in modules to (1) improve latency during training and inference, and (2) reduce positional biases related to the policy order.\n- The benchmark provides a comprehensive evaluation of existing MLLMs on video guardrailing tasks, demonstrating the model’s effectiveness across six safety policy categories, covering 30 subtopics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an MLLM-based video guardrail model that takes into account safety policies to provide a multi-label video content output including explanation, considering both the safety policies and the video content. The proposed model comprises two plug-and-play modules to improve latency of the guard rail model and mitigate positional biases by breaking down the safety guidelines. This work also introduces a benchmark for video guardrailing using multi-agent consensus and comparison across existing MLLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Details about the training and testing splits within the benchmark are insufficient, leaving questions about data partitioning.\n- The authors should clarify if any videos were discarded during dataset curation due to multi-agent discussion pipelines not reaching a consensus or human verification disagreements on final explanations. This clarification could shed light on the multi-agent approach's effectiveness in generating explanations that align with human perspectives, especially given video content's subjective nature." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "The example in the upper left corner of Figure 3 may have 2 issues.\n\n1. This type of data likely comes from real copyrighted videos, which may involve copyright infringement.\n2. The authors did not blur or mask faces in their examples, which could raise privacy concerns." }, "flag_for_ethics_review": { "value": [ "Yes, Legal compliance (e.g., GDPR, copyright, terms of use)" ] }, "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 design corresponding experiments and proofs to demonstrate that the mechanism producing the algorithm's effects aligns with their claims that the PEPE algorithm can \"allow each policy to be encoded independently and in parallel\" and that \"equivalent positional embedding ensures that different policies are treated without bias\"?\n2. Could the authors provide some quantitative analysis of the annotation quality? Can this pipeline approach human-level annotation quality? Compared to annotation by a single LLM/VLM, what advantages does incorporating Multi-agent Discussion bring?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The authors contribute a very large-scale benchmark for video security\n2. The authors propose the PEPE algorithm, which can mitigate positional bias in the input\n3. The authors propose the PAP algorithm, which maintains high recognition accuracy while reducing inference costs" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "SAFEWATCH is a new video moderation system that efficiently identifies harmful content and explains its decisions. It features two main innovations: PEPE (for parallel policy encoding) and PAP (for selective video analysis), both designed to make the system faster and more accurate. The researchers also developed SAFEWATCH-BENCH, a large dataset containing 2 million videos across six categories of harmful content, which they used to train and test their system." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The working mechanism of the PEPE algorithm lacks detailed theoretical explanation or experimental validation. The authors conduct ablation experiments to prove the effectiveness of the PEPE algorithm, but they don’t provide sufficient proof of its underlying principles. In lines 293-297, the authors claim that the PEPE algorithm can provide independent representations for each policy, which can alleviate the position bias problem in MLLM mentioned in lines 266-269. **However, regarding this claim, there are neither experimental designs nor mathematical proofs to support it. I have doubts about whether the mechanism behind the algorithm truly aligns with the authors' claims.**\n \n \n2. There is a lack of explanation regarding the effectiveness of the multi-agent propose-discuss pipeline mentioned in line 105. The authors mention in lines 105-106 that they use a novel pipeline for data annotation, but there is limited discussion about this pipeline. In the pipeline-related content, the authors do not cite any references, and based on my knowledge of related work, this pipeline has not been used in any previous work, making this the first work to employ this pipeline. Given this, **I am uncertain whether this pipeline can provide sufficiently high-quality annotation results**, and the authors have not provided any quality analysis of the annotation 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": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "The authors claim data contributions, but there is unsafe content in SafeWatch-Bench." }, "flag_for_ethics_review": { "value": [ "Yes, Discrimination / bias / fairness concerns", "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": "1. Could you provide some evaluation results for Safety-aware Event Sampling? Currently, its effectiveness or limitations are unclear.\n2. Is SafeWatch-Bench truly a video benchmark? Specifically, does it truly require reasoning across multiple frames for a model to achieve high performance?\n3. It appears that humans do not directly provide annotations for SafeWatch-Bench; instead, annotations are model-generated, with human reviewers checking if re-annotation is necessary. In the caption for Figure 2, you mention that the 1K test set has high-quality annotations. Were these test set annotations created directly by humans, or were they produced via the multi-agent propose-discuss-consensus pipeline?\n4. Could you describe how the preference pairs were curated for the preference post-tuning stage? Additionally, how were the challenging benign examples—those easily identified by humans but likely to mislead guardrail models—selected? A detailed explanation of these curation processes would be helpful.\n5. What is the quality of the synthetic videos generated by the GenAI models? Are they accurately aligned with the unsafe prompts? Was any video filtering applied to filter out misaligned videos?\n6. Could you clarify your model training data recipe? Are the unsafe videos used in stage-1 training and the guardrail tasks in stage-2 training drawn from the instruction-tuning dataset within SafeWatch-Bench? Additionally, how many samples are used for each stage, task, and dataset?\n7. Which layers are tuned in the preference post-tuning stage?\n8. In Table 3 of Appendix A.1, there is a column named \"Temporal location\". What does it mean?\n9. What is the SFT Baseline mentioned in Figure 5 and Table 5?\n10. It was claimed that prior datasets lack detailed descriptions of the videos, suggesting that SafeWatch-Bench offers a detailed description for each video. Is that correct?\n11. There is unsafe content in SafeWatch-Bench. Will SafeWatch and SafeWatch-Bench be released? If so, how do you plan to ensure their proper use?\n\nMinor comments:\nLine 749 has placeholder text." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Their model, SafeWatch, outperforms SOTA video guardrails on SafeWatch-Bench by 19.6% and on existing benchmarks by 15.4%, while reducing inference costs by an average of 25%. SafeWatch also demonstrates strong policy-following abilities, outperforming baselines by 20% in zero-shot adaptability to new policies. Additionally, both LLM-as-a-judge and human evaluators confirm the high quality of the explanations provided by SafeWatch.\n2. The design choices are well-founded, following best practices for efficient MLLM construction.\n3. This is an important area of study, with meaningful contributions (if these contributions are reproducible)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors of the paper propose a multimodal large language model (MLLM) called SafeWatch, designed to follow customized safety policies and provide multi-label video guardrail categorical outputs with answer explanations in a zero-shot manner. They also introduce SafeWatch-Bench, a large-scale video guardrail benchmark containing over 2 million videos spanning 6 safety broad categories and covering over 30 finer-grained risk categories to ensure comprehensive coverage of potential safety scenarios.\n\nThe technical contributions include:\n- Model Design: The authors introduce two key plug-and-play modules: Parallel Equivalent Policy Encoding (PEPE) and Policy-Aware Adaptive Pruning (PAP). \n - PEPE mitigates high latency from extensive input contexts and policy positional bias by dividing lengthy safety guidelines into independent chunks encoded in parallel with equal importance. \n - PAP, on the other hand, reduces latency by selecting the most relevant visual tokens for each policy while discarding those with low relevance.\n\n- Data: Each instance in SafeWatch-Bench is annotated with multi-label guardrail categories and detailed explanations. The dataset includes 2 million videos—both real-world and generative from various SOTA models—comprising an instruction-tuning set and a test set of 1K hand-selected, high-quality annotated videos across subcategories.\n\n- Training strategy: The authors fine-tune InternVL2-8B with their modeling changes on this new data via three stages, i.e., multi-task training, adaptive-pruning training, and preference post-tuning. \n - Stage 1: Only PEPE is trained during this stage on a large corpus of unsafe videos, as well as traditional VQA and captioning tasks on normal videos. \n - Stage 2: Both PEPE and PAP are fine-tuned on guardrail tasks. \n - Stage 3: Preference pairs are curated to enable the preference post-tuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Missing Evaluation: The evaluation of the Safety-aware Event Sampling step is absent. This step should be crucial for model performance, as the authors used TransnetV2 to segment videos into safety-aware events, sampling a single frame per event for further MLLM processing.\n2. Dataset Clarity: The dataset’s specifics and its exact use in model training remain unclear. For instance, there is no information on the average video length or the typical length of an explanation in SafeWatch-Bench. Additionally, the quality of the SafeWatch-Bench test set is not fully addressed, which is particularly important for an evaluation dataset.\n3. Reproducibility Concerns: Reproducibility in data collection and model training is questionable. For instance, Section 4.2 on \"multi-agent consensus video annotation\" provides a basic idea of the processes but lacks sufficient detail for replication (e.g., missing the prompts used, configurations such as the number of frames used for each model, etc.). Additional issues are noted in the \"Questions\" section." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. The author mentioned that Human Verification is used for curating the 2M video benchmark. Can the authors make a detailed description of the Human Verification procedure?\n\n2. More procedure should be enclosed in the appendix." }, "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 dataset is novel and the data curation procedure is well-organized.\n2. The proposed Parallel Equivalent Policy Encoding and Policy-Aware Adaptive Pruning can effectively encode the Safety Policy Guidelines and reduce the redundancy of video tokens.\n3. The results are good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a safety aware video understanding benchmark, including 2M human verified videos. The unsafe scenarios are separated into 6 classes. The authors also design a pipeline for automatically data generation. For the video understanding model, the authors propose the Parallel Equivalent Policy Encoding and Policy-Aware Adaptive Pruning to encode the Safety Policy Guidelines and reduce the redundancy. The result of the trained model is good comparing to both close- and open- source models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Dataset Construction: \n (1) The release and separation of the dataset is a concern. The author can only provide links to publicly available sources and annotations. But it is common that the link may fail.I understand that this is something unavoidable, but it will undoubtedly reduce the frequency of use and impact of this dataset.\n (2) 2M videos for human verification is a huge effort, the authors don't provide any details of the procedure.\n\n2. Model training：\n (1) Some of the training procedure is ambiguous, there should be more details about Preference Post-tuning procedure." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose an efficient MLLM-based video guardrail model and a large-scale instruction-tuning benchmark dataset to achieve customized safety policy and transparent explanations." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024safewatch,\ntitle={SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xjKz6IxgCX},\nnote={under review}\n}" }, "abstract": { "value": "With the wide adoption of generative AI and rapid growth of high-quality video generation, video guardrails have become more crucial than ever to ensure safety and security across platforms. Current video guardrails, however, are either overly simplistic, relying on pure classification models trained on simple policies with limited number of unsafe categories, which lack detailed explanations, or prompting multimodal large language models (MLLMs) with long safety guidelines, resulting in inefficient and impractical guardrails for real-world content. To bridge this gap, we propose SAFEWATCH, an efficient MLLM-based video guardrail model designed to follow customized safety policies and provide multi-label video guardrail outputs with content-specific explanations in a zero-shot manner. In particular, unlike traditional guardrails that encode entire policies autoregressive, causing inefficiency and bias, SAFEWATCH uniquely encodes each policy trunk in parallel and eliminates their position bias such that all policies are attended simultaneously with equal importance. In addition, to improve efficiency and accuracy, SafeWatch incorporates a policy-aware visual token pruning algorithm that adaptively selects the most relevant video tokens for each policy, discarding noisy or irrelevant information. This allows for more focused, policy-compliant guardrail with significantly reduced computational overhead. Considering the limitations of existing video guardrail benchmarks, we propose SafeWatch-Bench, a large-scale video guardrail benchmark comprising over 2M videos spanning six safety categories which covers over 30 tasks to ensure a comprehensive coverage of all potential safety scenarios. We have conducted extensive experiments, showing that SafeWatch outperforms all SOTA video guardrails on SafeWatch-Bench by 19.6% and 15.4% on existing benchmarks, while reducing inference cost by 25% on average. SafeWatch also demonstrates strong policy-following abilities and outperforms baselines by 20% in zero-shot adaptability to new policies. Additionally, both LLM-as-a-judge and human evaluators\nconfirm the high quality of the explanations provided by SafeWatch." }, "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 Guardrail Model", "Safe Foundation Models", "Efficient LLMs Inference", "LLM Safety", "Multimodal Foundation 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/a931c7c7320b297943eff52b2db77c595912603b.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": "SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The paper's theme is closely related to Safe RL, as mentioned. However, many Safe RL algorithms are not included as baselines for comparison. What is the rationale behind this omission?\n2. While PPO+Replacement has slower learning efficiency, it strictly ensures the satisfaction of constraints. This property is valuable in some online environments, but AM lacks this capability. Are there any proposed methods to address this limitation?\n3. In Figure 4(c), the performance of AM-SAC suddenly increases at 7500 steps. How can this phenomenon be explained?" }, "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 this work is clear: it addresses the SCMDP problem by utilizing a model to learn a feasible action space, effectively converting SCMDP into an MDP and improving the algorithm’s sample efficiency.\n2. The paper provides a step-by-step explanation of related concepts, making it very accessible to readers who may not be familiar with the field.\n3. The description of the action and state spaces for the tasks in the experiments is clear, and the importance of the AM algorithm is effectively illustrated through visualizations at the end of the experimental section." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce a novel approach known as action mapping (AM) within the context of deep reinforcement learning (DRL), showcasing its effectiveness, particularly when utilizing approximate feasibility models. Their results suggest that the integration of approximate model knowledge can improve training performance and enable agents to represent multi-modal action distributions, thus enhancing exploration strategies. By applying AM to both PPO and SAC, which represent on-policy and off-policy RL algorithms respectively, the authors provide comparative experimental results. These findings demonstrate that the AM method can transform state-wise constrained Markov Decision Processes (SCMDP) into Markov Decision Processes (MDP), thereby enhancing the sample efficiency of the original algorithms." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The absence of accompanying code makes it difficult to replicate the experimental results.\n2. The experimental section appears somewhat limited, as it only tests the method in two environments. This reduces the persuasive power and credibility of the results.\n3. The action mapping approach is not end-to-end, requiring pre-training with trajectory data before its application. This introduces additional costs, which are not adequately discussed in the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The most significant contribution is how action mapping allows agents to express multi-modal action distributions through a simple Gaussian in latent space, improving exploration. Ability to plan with approximate feasibility models is a notable advantage since perfect models are rarely available in practical applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the problem of how to efficiently train agents in environments with constraints (like a robotic arm avoiding obstacles or an aircraft maintaining non-holonomic constraints). Traditional DRL approaches struggle with poor sample efficiency and slow convergence in such constrained environments. The authors propose \"action mapping,\" which decouples the learning process into two parts: first training a feasibility policy that learns to generate all feasible actions for any state, then training an objective policy to select optimal actions from this feasible set, effectively transforming a state-wise constrained Markov Decision Process (SCMDP) into an unconstrained MDP. They validate their approach through two experiments - a robotic arm end-effector pose task with perfect feasibility models and a path planning problem with approximate feasibility models - demonstrating superior performance compared to common approaches like action replacement, resampling, and projection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper omits constraint violation plots for the path planning task, making it impossible to verify claims about performance with approximate feasibility models.\n \n- The feasibility model is a critical component of the proposed architecture, yet the paper lacks essential analysis and ablations of this component. Key questions remain unanswered: How is state space sampling performed during training? What metrics determine sufficient training of the feasibility model? How does the quality/approximation level of the feasibility model impact overall performance? Without these analyses, it's difficult to understand the method's robustness and its applicability to scenarios where perfect or near-perfect feasibility models aren't available.\n\nMinor Comments:\n\n- Typo on line 406, “actiosn”" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 ethics review 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": 1 }, "primary_area": null, "questions": { "value": "1. **[About Weakness 1]** Isn’t the required setup for the proposed method too strict? Other constrained RL methods estimate the cost function without prior knowledge, obtaining cost information similarly to rewards. In such cases, would the proposed action mapping approach still be applicable?\n\n2. **[About Weakness 2]** Could you provide experimental results in a wider variety of environments?\n\n3. **[About Weakness 3]** Could you also show performance comparisons with other constrained RL methods?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The method for training the feasibility policy is novel, allowing for fewer constraint violations and higher returns compared to other methods.\n2. Experiments were conducted in environments requiring constraints, such as a robotic arm task and a spline-based path planning.\n3. The approach is straightforward and can be combined with any RL algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an action mapping method that distinguishes between feasibility and objective policies during training. By pretraining the feasibility policy first and then training the objective policy, the approach enables more efficient learning within a reduced action set. Experimental results demonstrate that the proposed method results in fewer constraint violations and achieves higher returns compared to previous action replacement, resampling, and projection methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The assumption that the feasible policy can be pretrained seems overly strict. Pretraining requires prior knowledge of the cost function $C^\\tau(s;\\pi)$ and the feasibility model $G(s,a)$, which may be difficult to assume in general.\n2. The experimental environments appear limited. It would be beneficial to include comparisons in environments like Safety Gym or other constrained RL environments.\n3. There is a lack of baseline algorithms. Currently, the comparisons are limited to variants of action mapping, such as action resampling and projection. Direct comparisons with a wider range of methods, including Lagrangian approaches, would strengthen the evaluation. Even if these methods are primarily designed for standard CMDPs rather than SCMDPs, adjusting the constraint thresholds to be more strict would allow for a fair comparison under the SCMDP constraints used in the original 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": 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": "1. For the training in Section 4.1, does it require a ground-truth $g(s,a)$?\n2. What is the bound of the output of $\\pi_f$ and $\\pi_o$?\n3. What are the implementation details of the proposed method, e.g., network structure, hyperparameters?\n4. For Figure 4, why is SAC+Replacement not included in Figure 4c? And where is the constraint violation plot for SAC?\n5. Throughout the paper, it seems that the definitions in Equation 5-7 are not necessary, as in Section 4 and the experiments only the function $g$ is assumed. Can the authors provide more explanation on this point?\n6. How many seeds/trials are used in Figure 4?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The target problem to solve in this work is of great significance to many real-world applications.\n- The related works are introduced and discussed satisfactorily." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a strategy called action mapping for RL in continuous environments with state-wise constraints. The idea is to learn a latent action space and an action mapping model, with which the policy samples a latent action and the latent action is further mapped to a feasible action. The proposed method is evaluated in a robotic arm end-effector pose positioning task and a path planning environment, showing better performance than several existing methods in terms of higher returns and lower constraint violations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although the authors mentioned the reference [Theile et al., 2024], I found the content in Section 4 overlaps largely with the content in Section 3 and Section 4 of [Theile et al., 2024]. For example, Equation 16 in this paper is almost the same as the JS loss in Table 2 of [Theile et al., 2024]. Therefore, the novelty and contribution of this paper are questionable.\n- The whole training process is not clear. Adding a pseudocode of the proposed algorithm will help. Especially, the training of the feasibility model is not clear enough. In addition, many technical details are missing. Please see my questions below.\n- The idea of action mapping is closely related to the research on action representation learning [1-4]. The illustration in Figure 1 is very similar to the concept presented by Figure 1 in [1]. These related works should be included in the related work section for a detailed discussion.\n\n### Minors\n\n- The symbol $J$ in Equation 1,2 and Equation 3,4 are inconsistent.\n- The legends in Figure 4 are too small.\n\n---\n### Reference\n\n[1] Yash Chandak, Georgios Theocharous, James E. Kostas, Scott M. Jordan, Philip S. Thomas. Learning Action Representations for Reinforcement Learning. ICML 2019: 941-950\n\n[2] Boyan Li, Hongyao Tang, Yan Zheng, Jianye Hao, Pengyi Li, Zhen Wang, Zhaopeng Meng, Li Wang. HyAR: Addressing Discrete-Continuous Action Reinforcement Learning via Hybrid Action Representation. ICLR 2022\n\n[3] Pengjie Gu, Mengchen Zhao, Chen Chen, Dong Li, Jianye Hao, Bo An. Learning Pseudometric-based Action Representations for Offline Reinforcement Learning. ICML 2022: 7902-7918\n\n[4] William F. Whitney, Rajat Agarwal, Kyunghyun Cho, Abhinav Gupta. Dynamics-Aware Embeddings. ICLR 2020" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose action mapping, a method that integrates feasibility models into deep reinforcement learning, significantly improving training efficiency in constrained environments with continuous action spaces." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024action,\ntitle={Action Mapping for Reinforcement Learning in Continuous Environments with Constraints},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xjornbs7aT},\nnote={under review}\n}" }, "abstract": { "value": "Deep reinforcement learning (DRL) has had success across various domains, but applying it to environments with constraints remains challenging due to poor sample efficiency and slow convergence. Recent literature explored incorporating model knowledge to mitigate these problems, particularly through the use of models that assess the feasibility of proposed actions. However, integrating feasibility models efficiently into DRL pipelines in environments with continuous action spaces is non-trivial. We propose a novel strategy, termed action mapping, that leverages feasibility models to streamline the learning process. By decoupling the learning of feasible actions from policy optimization, action mapping allows DRL agents to focus on selecting the optimal action from a reduced feasible action set. We demonstrate through experiments that action mapping significantly improves training performance in constrained environments with continuous action spaces, especially with imperfect feasibility 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": [ "Constrained MDPs", "continuous action space", "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/66001385e4c21c2dc795d322e5db8aa8747b0762.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": "Action Mapping for Reinforcement Learning in Continuous Environments with Constraints" }, "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": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Their findings reveal that a minimal amount of labeled data, sometimes as few as 250 images, suffices for the effective recovery of these intrinsic images.\n2. The research claims a positive correlation between the generative quality of a model and the accuracy of its recovered intrinsic properties, which is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on investigating the intrinsic knowledge encoded within various generative models, such as GANs, autoregressive models, and diffusion models. The study aims to uncover whether these models inherently capture fundamental scene properties, including Depth, Surface Normals, Albedo, and Shading. Through the use of Low-Rank Adaptation (LoRA), a lightweight technique that introduces minimal learnable parameters, the authors propose a model-agnostic approach to recover these intrinsic features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Table 1, why do different generative models exhibit varying abilities to capture scene intrinsics without altering the generator head? For instance, even within the same model, such as SG-XL, all scene intrinsics (e.g., Normal, Depth, Albedo, and Shading) are recoverable for the FFHQ dataset, while none of these intrinsics can be captured for ImageNet. Are there any hypotheses regarding which properties of the models or datasets might influence their differing abilities to capture intrinsic features?\n\n2. While the paper asserts that “enhancements in image generation quality correlate positively with intrinsic recovery capabilities”, this claim seems not convincing enough due to the following reasons:\n\na) Figure 2 attempts to validate the claim by showing a relationship between generated image quality (measured by FID) and recovery errors. However, it includes only three generative models (SG-XL, SGv2, and VQGAN), which represent GAN-based and autoregressive models but exclude diffusion models. This limited selection makes it challenging to empirically confirm the claim. It would be better to provide more generative models in Figure 2, including diffusion models as well. Moreover, discussing the technical reasons behind such correlations would strengthen the argument.\n\nb) The claim lacks rigor due to inconsistencies. For instance, Figure 2 suggests that SG-XL outperforms SGv2 and VQVAE in generative quality, yet Table 2 shows SG-XL occasionally underperforming them, such as in Shading (on FFHQ). Moreover, factors beyond the generative model itself, like the dataset, also impact performance. For example, SGv2 performs worse in Depth but better in Shading when switching from FFHQ to LSUN Bedroom. Please provide a more in-depth statistical analysis of the correlation between generative quality and intrinsic recovery capabilities across all models and datasets.\n\n3. One of the paper’s objectives is to demonstrate that “tiny new parameters and data are enough for intrinsic recovery”. However, as mentioned in this paper, several existing works (e.g., [1]) have already shown that parameter-efficient adaptation methods, like linear probes, can effectively extract intrinsic features such as depth. This work merely replaces linear probes with LoRA and demonstrates its effectiveness, which feels somewhat incremental. Besides, it would be better to discuss more, e.g., it would be helpful to discuss further -- for example, whether there are particular scenarios where LoRA outperforms linear probes or if LoRA provides any advantages beyond performance.\n\n[1] Yida Chen, Fernanda Viegas, and MartinWattenberg. Beyond surface statistics: Scene representations in a latent diffusion model. 2023.\n\n4. The paper notes a lack of ground truth data for certain maps (like albedo), which raises the question of whether using a light physics simulator or tools like Unreal Engine could provide high-quality ground truth labels. This approach could also enable the creation of more complex scenes and lighting conditions for more rigorous model testing. I am curious about the feasibility of incorporating such simulators or tools into this work, and whether this approach could substantially enhance the reliability of the 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": 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 address the comments in the weakness section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- [S1: Originality] Though the proposed method is very standard, the paper presented a generic approach and demonstrated that the approach recovers intrinsic information across diverse generative models, including GANs, autoregressive models, and diffusion models. \n\n- [S2: Clarity] The motivation, methodology, and results are clearly articulated, enabling a good understanding of the research contributions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the intrinsic information (e.g., normal, depth, albedo, and shading) encoded in various generative models, including GANs, autoregressive models, and diffusion models. Key findings include (1) Intrinsic information is a byproduct of training generative image models; (2) Low-Rank Adaptation (LoRA) is a generic technique to study the intrinsic information encoded, better than other approaches such as linear probing and fine-tuning; (3) The quality of a generative model and accuracy of its recovered intrinsics (e.g., depth prediction accuracy) are positively correlated. To recover each intrinsic property, a structured prediction module (e.g., depth prediction) has been used to provide pseudo groundtruth during the LoRA optimization. Experiments have been conducted mainly on depth and normal prediction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- [W1: Deeper understanding] While the paper establishes what intrinsic knowledge is encoded, it doesn't delve into how generative models utilize this knowledge during image generation. \n - [W1.1] For example, the reviewer would like to understand if intrinsic property emerges with a half-trained generative model (or at different training epochs).\n\n- [W2] The reviewer find that most intrinsic property do not contain high-frequency details (e.g., sharp edges in the depth prediction). This could be the shortcoming of using LoRA fine-tuning in the design. Although multi-step diffusion inference has been added in Section 5, the question still remains if a single-step approach is sufficient to recover high-fidelity intrinsic information.\n - [W2.1] How does multi-step inference apply to other generative models such as StyleGAN?\n - [W2.2] What’s the motivation of applying LoRA to attention layers of a diffusion model?\n - [W2.3] The ablation studies on applying LoRA to different modules (Appendix B) is interesting. It seems to suggest that LoRA is not successful when applied to up or down blocks. What’s the insight behind the discovery?\n\n- [W3] The reviewer does not fully understand the claim that StyleGAN-XL trained on ImageNet is an exception (Line 365). For example, StyleGAN-XL achieved a much lower FID (Line 452) compared to other generative models, which seems to contradict with the claim that StyleGAN-XL’s limited ability to generate realistic images (Line 366). Please clarify this point in the rebuttal.\n\n- [W4] Specific to the comparison between SD-1.X and SD-2.1, how much of the performance difference can be attributed to the improved encoder-decoder? Is it possible to recover the intrinsic property by applying LoRA fine-tuning on encoder-decoder alone?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Related to the last weakness, Section 4.4 identifies LoRA as the superior method for intrinsic image recovery, but the underlying reasons remain unclear. Could the authors provide insights into its advantages over alternative methods?\n- The authors demonstrate a positive correlation between image intrinsics and generation quality. Could they elaborate on the nature of this relationship? Specifically, does better intrinsic understanding lead to improved generation quality, vice versa, or is there a bidirectional relationship between these aspects?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The work raises fundamental questions about the nature of internal representations in generative models and their relationship to generation quality, contributing to our understanding of these architectures.\n- The research demonstrates that world representations in powerful generative models are more concrete and accessible than previously assumed, establishing important connections to explainable AI.\n- The authors present compelling insights about the minimal data requirements for intrinsics recovery, suggesting that the underlying knowledge is already embedded in the model's parameters.\n- The methodology is straightforward yet effective, making the findings easily reproducible and applicable to various generative models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates whether current generative models possess a genuine understanding of the world they're generating, rather than just statistical pattern matching. The authors propose using the generation of image intrinsics (such as depth maps, normal maps, etc.) as a proxy for evaluating the models' physical understanding and explainability. They demonstrate that a simple adaptation method (LoRA) can effectively recover these intrinsic properties and conduct extensive analyses to support their hypothesis about models' world understanding." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper appears to have originated as a study specifically focused on diffusion models, with other generative architectures added later to broaden its scope. While Section 4's experiments and Section 5's generation quality refinements primarily use Stable Diffusion, validating these findings across different architectures would strengthen the paper's claim of developing a general approach.\n- The experiments in Section 4.2 focus solely on normal map recovery when determining optimal rank and dataset size. The authors should clarify whether these optimal settings generalize to other intrinsics or if each intrinsic property requires specific configurations.\n- The paper would benefit from a deeper analysis of why LoRA is particularly effective for intrinsic recovery. While Line 102 suggests that \"intrinsic information is distributed throughout the network,\" this insight isn't fully developed in Section 4.4 or elsewhere. Understanding this mechanism could also inform the development of future recovery methods.\n\nThe paper presents strong empirical findings, though deeper analysis is needed to enhance its impact on interpretability and explainable AI research." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "To AC: I can clearly understand the task, the proposed approach and the experiments in the paper. Overall, I find the paper supports the claims with properly designed experiments, but I cannot evaluate the importance and novelty of the paper at the moment since I don't follow literature on this intrinsic estimation topic. I would evaluate these two aspects later by referring to opinions from other reviewers. So please use this review with proper weight when make decisions." }, "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-written and easy-to-follow\n\n2. The approach is straightforward and effective by using LoRA to adapt pretrained generative models for downstream tasks beyond image generation. The predicted intrinsic results look decent.\n\n3. By using LoRA instead of retraining or finetuning the whole model, the approach requires less data and parameters to perform the image intrinsic estimation task. As an experimental work, the paper mostly supports the claims therein with extensive experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to use pretrained generative models extended with trainable LoRA layers as image intrinsic predictors. The proposed approach aims to learn effective intrinsic extractors with as few LoRA parameters and training samples as possible. Extensive experiments are conducted to primarily show that (1) with less LoRA parameters and data samples than state-of-the-art approaches, extracting intrinsic images is still possible (2) it is the prior knowledge in pretrained generative models that helps extract intrinsic image, or in other words, generative models do encode useful intrinsic knowledge though it is not clear how such intrinsic knowledge are used during generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Fig 8 and Fig 9 show that there is a performance peak as the number of LoRA params or training samples vary. Why there is a peak? For the case of fixed number of training samples, with more LoRA params than Rank 8, the performance degrades. Is this due to underfitting that there is no enough data to train larger LoRA well enough? Similarly, for the case of fixed number of LoRA params, is this due to overfitting that more data overfits a relative small LoRA?\n\n2. A relevant question to 1. If you increase the number of LoRA and the number of training data at the same time avoiding overfitting or underfitting, can you outperform existing approaches eventually as shown in Table 3? Furthermore, the paper shows qualitative results for all four intrinsics Surface Normal, Depth, Albedo and Shading throughout the paper, but shows quantitative results on only Surface Normal, Depth in Table 3. Is there is any reason for this practice?\n\n3. The paper overclaims a bit about \"minimal\" requirements on parameter updates and data. Why do you say minimal and in what sense it is minimal? Thought the approach is able to work with fewer params and data , the quantitative performance doesn't outperform existing approaches as shown in Table 3.\n\n4. Typo Table 3 caption, \"intrinsicsacross\"->\"intrinsics across\"" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Intrinsic images are encoded across generative models (GAN, Autoregressive and Diffusion). We can recover them using LoRA while using the same decoder head that generates the image." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024generative,\ntitle={Generative Models: What Do They Know? Do They Know Things? Let's Find Out!},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xkR3bcswuC},\nnote={under review}\n}" }, "abstract": { "value": "Generative models excel at mimicking real scenes, suggesting they might inherently encode important intrinsic scene properties. In this paper, we aim to explore the following key questions: (1) What intrinsic knowledge do generative models like Autoregressive models, GANs and Diffusion models encode? (2) Can we establish a general framework to recover intrinsic representations from these models, regardless of their architecture or model type? (3) How minimal can the required learnable parameters and labeled data be to successfully recover this knowledge? (4) Is there a direct link between the quality of a generative model and the accuracy of the recovered scene intrinsics?\n\nOur findings indicate that a small Low-Rank Adaptation (LoRA) can recover intrinsic images---depth, normals, albedo, and shading---across different generators (GAN, Autoregressive, and Diffusion) while using the same decoder head that generates the image. As LoRA is lightweight, we introduce very few learnable parameters (as few as 0.04% of Stable Diffusion model weights for a rank of 2), and we find that as few as 250 labeled images are enough to generate intrinsic images with these LoRA modules. Finally, we also show a positive correlation between the generative model's quality and the accuracy of the recovered intrinsics through control experiments." }, "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": [ "Visual knowledge", "Generative models", "Intrinsic Images" ] }, "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/df84f7acd2b4bdab7db09523ee3d18d81698d448.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/49ecba92fa656bcdcf4b388713cdfef7335ae689.zip" }, "title": { "value": "Generative Models: What Do They Know? Do They Know Things? Let's Find Out!" }, "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": "* The paper focuses on the visual areas of the brain. How would the results look if the entire brain was used? A small discussion would suffice. \n* How many parameters are ViT-H and CLIP? Are they comparable?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Novelty with instruction-tuned MLLMs: Overall, I haven’t seen too much work on brain encoding models with instruction-tuned MLLMs. This work is highly timely.\n* Well-designed and controlled experiment: The paper uses a controlled and well designed experiment for exploring brain fits.\n* Instruction breakdown: I really appreciated Figure 3 and 4. I thought it was pretty interesting to see how different instruction types fit responses in the visual cortex. This was a fairly novel result. \n* I also really liked Figure 6 which focused on shared variance. This was an interesting result that showed which instruction types had a higher degree of shared features which resulted in a higher degree of brain similarity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies brain-encoding fits of instructions-tuned MLLMs in comparison with traditional methods for designing multimodal embeddings such as via CLIP. To study this question, the paper feeds different types of instructions to the MLLMs in comparison to other baselines like CLIP or vision encoders. The paper finds that instruction-tuned MLLMs have much better fits with visual areas of the brain than vision embedding models and similar performance to CLIP. The authors break down the instructions by type to identify which type of instructions have stronger correlation with visual areas of the brain and do the same with visual concept groups. The authors also include a shared variance experiment that aims to characterize whether shared features were used for each instruction group that were relevant for encoding model performance. \n\nOverall, I liked this paper and would vote for acceptance outside of a few concerns. It answered a relevant question about instruction-tuned MLLMs that haven’t really been explored before." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Comparisons: Overall, I think it would be better to include a baseline with a non-instruction-tuned MLLM that has the same architecture. For example, maybe BLIP-2 instead of InstructBLIP? This would have really explored the role of instruction tuning more thoroughly in comparison. BLIP-2 should be available off the shelf and I believe this would be a salient comparison. I would also be curious about how a language model of similar size would do. \n* Another concern here is that improvement over ViT-H could be due to an increase in parameters. See questions. \n* Nit: Could Figure 3 be sharpened? The text is quite blurry and difficult to read. \n\nIncorporation of some of these comparisons would help raise my score/confidence." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": [ "Yes, Discrimination / bias / fairness concerns", "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": 3 }, "primary_area": null, "questions": { "value": "None" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper introduces a novel approach to evaluating the brain alignment of instruction-tuned MLLMs, providing valuable insights into how these models process multimodal information in relation to human brain activity.\n- The findings have implications for understanding how brain activity corresponds to the processing of multimodal information, which could be valuable for cognitive neuroscience and AI research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates the alignment of instruction-tuned multimodal Large Language Models (MLLMs) with brain activity, particularly focusing on how these models process visual information when prompted with natural language instructions. The study explores the brain alignment of MLLMs compared to unimodal and non-instruction-tuned multimodal models and assesses the effectiveness of various task-specific instructions on brain alignment. The paper presents a comprehensive analysis of how different instructions from MLLMs capture visual concepts and their correlation with brain activity, using fMRI data from the Natural Scenes Dataset (NSD)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The study relies on the NSD dataset, where subjects passively view images, which may not fully capture brain activity aligned with task-specific instructions. Active task engagement during fMRI scans could provide a more comprehensive evaluation.\n- How do the authors address ethical considerations regarding the use of fMRI data, especially in relation to participant privacy and data security?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "-\tIn Figure 2 it seems like CLIP is far outperforming the randomly initialized network. Why is there no asterisks?\n-\tIs Figure 6 summarizing the results across NSD-General? Why show only for one representative subject rather than the average of all?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "-\tInstruction-tuned multimodal models are an interesting way to investigate task tuning, as they have higher performance than prior fine-tuned models like the taskonomy set\n-\tThe comparison of retinotopic versus category-selective tuning is interesting, and may yield novel insights into high-level vision.\n-\tVariance partitioning allows a more fine-grained look into the contribution of different task tuning" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors test the ability of instruction tuned multimodal LLMs to match human visual cortex responses to static scenes in the NSD. They compare instruction tuned models to one standard multimodal model (CLIP) and a unimodal vision model.\n\nThe findings reveal that all multimodal models match visual cortex responses better in both low- and high-level visual regions. There does not seem to be a difference between CLIP and instruction-tuned models in any region. Within the instruction tuned models, there seem to be differences between tasks that best explain different parts of visual cortex, with lower-level tasks (e.g., color) better describing retinotopic areas and higher-level tasks (e.g., image captioning) better explaining higher-level tasks. Variance partitioning reveals that there is shared variance between many pairs of tasks, but some tasks like food labeling and scene recognition are more unique.\n\nIt is difficult to draw strong comparisons about these models compared to unimodal vision models though, because of the many differences (architecture, training data, task). As others have pointed out in prior work, advantages of multimodal models in visual cortex significantly decrease when you consider more balanced pairs (e.g., SLIP family models in Wang et al, and Conwell et al) and it seems likely that many of these differences would also diminish in more controlled modeled comparisons. If there are matched vision transformers trained on the same dataset, including them would strengthen the claims of the paper. \n\nThe task comparisons are interesting, but it would help for the authors to spell out what is at stake in these investigations. Does better fit of one type of instruction mean tell us what tasks to train neural networks on for improving human-aligned AI, or does it tell us something about the tuning properties of visual cortex. It would help to clarify this, particularly in the paper introduction and discussion." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-\tWhile the instruction model-brain comparisons are very interesting, it is not entirely clear what is at stake. Is the central question spelled out in the intro (lines 83-85) important for better human-alignment of AI, or in order to reveal tuning properites of the human brain? If the latter, the authors should flesh this out, and state some limits of this model comparison approach (see below). The primary goal of the study should be clarified in the introduction.\n-\tThe authors make a distinction between the advantage of multimodal models in high- versus low-level visual cortex. The results look very qualitatively similar in early versus late ROIs in Figures 2 (and also compared to other category selective regions in Figure 10) so if this is a major point, it should be backed up with a statistical analysis, (e.g., non-parametric anova, or permutation-based comparison of the vision-multimodal difference in both regions).\n-\tPrior work has shown that when architecture and training set are matched (e.g., the SLIP family models) advantages of multimodal models largely go away (e.g., Wang et al 2023, Conwell et al., 2022). It seems likely that the advantage of the multimodal models over unimodal vision models in this paper is similarly due to different/richer training data, rather than multimodality itself.\n-\tThe distinction between the different instructions/tasks could be made clearer. It seems like the tasks vary from low- high-level and many of the papers claims suggest this distinction. It would help to explicitly order the instructions and make this distinction clear early in the paper (e.g., Table 1). Without a priori labels for low- versus high-level tasks, some of the conclusions seems somewhat circular (eg a task that best explains retinotopic cortex is low-level). This ordering could also help make the results in Figure 3 more clear.\n-\tIt is difficult to see trends that are constant across the two different instruction-tuned models in Figure 3 and 11. InstructBLIP seems to show the trends highlighted in the paper, mPLUG-Owl looks quite random. Perhaps the re-ordering of the tasks/color bar suggested above will help. Alternatively, perhaps small differences across the models are being magnified in the winner-take-all plot. Either way, the authors should address these discrepancies.\n-\tAs a small point, the ROI labels in Figure 2 could be more intuitive. I believe “whole brain” refers to the NSDGeneral mask? If so, this should clarify that it is only visual cortex, not whole brain. pRF-Visual and FLOC-PLACES would be more clear as “retinotopic early visual cortex” and “scene-selective” or something similar\n-\tIt is unclear what Figure 5 is adding to the paper. Layerwise comparisons in different transformer networks are somewhat confusing and also seem dependent on the structure of encoders/decoders. The authors should consider moving this to the supplement. If it is important to the main findings, the author should clarify how the layerwise comparisons relate to the main text, and how they should be interpreted in light of any architectural differences across models (eg encoder vs decoder layers)\n-\tThe variance partitioning is a major strength of the paper, but it seems to only investigate shared variance between pairs of tasks. The more important question seems to be how much (if any) unique variance is explained by models tuned on any one task? And what these results can tell us about the brain." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How does the CLIP baseline work?\n - L235 states “we input both image and ground truth caption pairs” — are the CLIP image and CLIP text embeddings simply concatenated? Are you using the final pooled output from CLIP, or some intermediate layer? If you are using an intermediate layer, how do you pool across image patches / tokens?\n- In L29, what does “effectively capture” mean?\n - Does this mean that the MLLM embedding correlates with the “expected” brain region that is known to do a certain type of processing? This is not obvious from the discussion in L365.\n- Below are a few minor comments.\n - L280: typo; “random **initialization** of the 3 MLLMs”\n- Below is a comment on the limitations section, although this note is not important for my rating and I leave this to the discretion of the authors.\n - I wish the paper would also briefly discuss the limitations of fMRI itself, and how it is not exactly synonymous with “processing in the brain.” Specifically, fMRI is imprecise, as it “measures a surrogate signal” where the “spatial specificity and temporal response” is constrained [1].\n - Nevertheless, fMRI is the most available / accessible brain signal to study, and it is still interesting to study the relation between machine learning models and fMRI.\n\nReferences\n\n[1] Logothetis, N. What we can do and what we cannot do with fMRI. *Nature* 453, 869–878 (2008)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The central question the paper studies (the alignment of MLLMs and fMRI) is novel, as MLLMs have not been studied in prior work. The paper is also interesting because it presents some insights into the internals of MLLMs, for example, the fact that image captioning overlaps highly with other instructions.\n- The presentation of the paper is good. The experiments are well-motivated and interpreted with precise language. The paper is also thorough in its setup and appropriately references prior work.\n - The ablations relating specific instructions or model layers to brain regions (Sec. 6.2) were interesting, even if for some instructions the association was inconclusive.\n - Ablations, such as the baseline “cross-subject prediction accuracy” (L174) and “variance partitioning approach” (L262), are motivated by prior work. The usage of a ridge regression based model (L242) and Pearson Correlation as a metric (L252) for brain alignment is also standard practice." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the correlation between multimodal LLMs (MLLMs) and fMRI from the Natural Scenes Dataset (NSD). Specifically, they feed images and instructions to the MLLMs, cache the embeddings, and fit a linear model to map from MLLM embeddings to fMRI. They find that MLLMs exhibit higher brain alignment than vision-only models and CLIP. They also analyze the correlation of specific instructions and specific MLLM layers with brain regions. Finally, they do a variance partitioning analysis to quantify the overlap between different pairs of instructions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experiments could include additional ablations for the input to the MLLM.\n - For example, similar to the setup in Sec. 6.1, one could ablate feeding only the image or only the instruction to the MLLM, which reduce to “vision-only” or “LLM-only” baselines. This could help control for model size / other model statistics, i.e., ViT-H might also perform worse because it is a smaller model." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024correlating,\ntitle={Correlating instruction-tuning (in multimodal models) with vision-language processing (in the brain)},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xkgfLXZ4e0},\nnote={under review}\n}" }, "abstract": { "value": "Transformer-based language models, though not explicitly trained to mimic brain recordings, have demonstrated surprising alignment with brain activity. Progress in these models—through increased size, instruction-tuning, and multimodality—has led to better representational alignment with neural data. Recently, a new class of instruction-tuned multimodal LLMs (MLLMs) have emerged, showing remarkable zero-shot capabilities in open-ended multimodal vision tasks. However, it is unknown whether MLLMs, when prompted with natural instructions, lead to better brain alignment and effectively capture instruction-specific representations. To address this, we first investigate the brain alignment, i.e., measuring the degree of predictivity of neural visual activity using text output response embeddings from MLLMs as participants engage in watching natural scenes. Experiments with 10 different instructions (like image captioning, visual question answering, etc.) show that MLLMs exhibit significantly better brain alignment than vision-only models and perform comparably to non-instruction-tuned multimodal models like CLIP. We also find that while these MLLMs are effective at generating high-quality responses suitable to the task-specific instructions, not all instructions are relevant for brain alignment. Further, by varying instructions, we make the MLLMs encode instruction-specific visual concepts related to the input image. This analysis shows that MLLMs effectively capture count-related and recognition-related concepts, demonstrating strong alignment with brain activity. Notably, the majority of the explained variance of the brain encoding models is shared between MLLM embeddings of image captioning and other instructions. These results indicate that enhancing MLLMs' ability to capture more task-specific information could allow for better differentiation between various types of instructions, and hence improve their precision in predicting brain responses." }, "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": [ "brain encoding", "fMRI", "visual processing", "multimodal instruction-tuned models", "language decoder", "LLMs", "MLLMs" ] }, "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/5ab0da392e0f68af6635e8ac452c88cdf4adccbc.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": null, "title": { "value": "Correlating instruction-tuning (in multimodal models) with vision-language processing (in the brain)" }, "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": "See weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Introducing the concept of maxcut into attributed graph pooling is novel and meaningful, and it is worth noting the proposed method is feature-aware and differentiable.\n- The background and related work are introduced in detail and accompanied by illustrations, which greatly aid in understanding the proposed method.\n- It is encouraging that the proposed method was tested on three tasks, including maxcut partition, graph classification, and node classification." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an innovative method for computing the MAXCUT in attributed graphs which is fully differentiable, enabling joint optimization of the MAXCUT with other objectives. Leveraging the resulting MAXCUT partition, the authors develop a hierarchical graph pooling layer tailored for Graph Neural Networks. The authors claim that the pooling layer is sparse, differentiable, and effective for downstream tasks on heterophilic graphs, addressing a key challenge in graph learning by providing a versatile and adaptable pooling strategy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The authors are advised to theoretically and empirically analyze the complexity of MaxCutPool, showing whether it introduces additional computational overhead to GNNs.\n- It is also recommended that the authors provide the code to facilitate a better understanding of the proposed method and to ensure the reproducibility of the experiments.\n- The authors claim that MaxCutPool is particularly effective for heterophilic graphs, but no detailed analysis is provided. What is the underlying interaction between MaxCutPool and heterophilic graphs? Additionally, considering that traditional message passing (including GIN) is designed for homophilic graphs and is regarded as a low-pass filter, could this conflict with MaxCutPool and potentially lead to poorer results?\n- My primary concern is with the experimental section. In graph classification tasks, the proposed approach does not demonstrate satisfactory performance and even falls below the no-pooling baseline on nearly half of the datasets. In Table 3, the authors mark the performance of the no-pooling baseline in gray, even when it ranks highest, which is confusing.\n- Is MaxCutPool-NL learning $\\mathbf{s}$ solely with task loss after removing $\\mathcal{L}_{cut}$? Does this imply that the core concept of MaxCut has been removed? If so, it still appears to be comparable.\n- Although the work emphasizes its advantages on heterophilic graphs, it does not perform better than other pooling methods on the GCB-H and MUTAG datasets, which have the highest levels of heterophily.\n- Compared to graph classification, the performance on node classification is satisfactory. However, it seems that datasets are not the commonly adopted ones for node classification (such as Planetoid or OGB) and exhibit very low levels of heterophily.\n\nThe method and idea are very interesting, but the experiment does not sufficiently demonstrate their effectiveness and necessity. I will consider increasing the score once this issue is 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": 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": "* The method uses HetMP, which can be interpreted as a Laplacian sharpening kernel that enhances differentiation between signals across nodes, thereby reducing smoothness. This approach performs well on heterophilic graphs. However, in Figures 2(b-c), the steps seem to be based on a homophily assumption—where neighboring nodes tend to belong to the same cluster. Could you provide further clarification on this?\n* Beyond Equation (4), how does the proposed method perform when using other graph kernels or filters, such as high-pass filters? Additionally, how does the method perform on homophilic graphs?\n* Can the proposed method scale to graphs of arbitrary size, or are there practical limitations on the scalability threshold?" }, "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 generally reader-friendly, featuring detailed illustrations such as figures and hyperparameters. Theoretical analyses are also provided.\n- The proposed method is easy to understand yet remains effective, performing well in experiments.\n- This work includes comprehensive experiments covering MaxCut partition computation, graph classification, and node classification, which lend credibility to the conclusions presented in the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel GNN-based approach for solving the MaxCut problem in attributed graphs. The proposed method exhibits good performance across multiple experiments, particularly when evaluated on the newly introduced heterophilic graph benchmark dataset for graph classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please refer to the questions below." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In experiment 4.1, what were the features used for GNN? Were these learnable vertex embeddings?\n2. How does maxcutpool scale with the number of vertices in a graph?\n3. What is the intuition as to why learning the max cut operator provides a good pooling operator? Is it just that this is a high-frequency projection?\n4. In experiment 4.3, could you provide the tuned baseline for a GNN with no pooling?\n5. With $\\delta$ set to 2 in your node classification experiments, your propagation operator is already tuned towards heterophilic settings, and the stated intuition for why maxcutpool works is because it is projecting out high frequency components. Could you provide a node-classification results as a function of delta? Given that these datasets are all heterophilic, it's hard to tease out whether the improvements are due to the pooling or the propagation operator.\n6. How was hyperparameter tuning performed?\n7. Can this work be extended to link prediction in any meaningful way? Would it be possible to view super-node membership as a labeling trick?\n8. Does the unpooling operator involve any normalization?" }, "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 clearly presented\n2. The experiments are well defined and clearly motivated\n3. The authors make significant efforts to contextualize their work within the rest of the literature.\n4. The algorithm is clearly outlined in enough detail to reimplement" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Graph neural networks are constructed out of a few relatively simple building blocks -- A propagation operator, a message function, and a pooling function. These three components have received significant attention over the last few years, with both heuristic and graph theoretic approaches to improve the performance of GNNs on a variety of graph ML tasks. In this work, the authors present a novel pooling operator termed MaxCutPool, that is inspired by the classic graph theory problem of finding the max cut. They provide justification for the value of this pooling operator through robust experiments to find significant performance improvements." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper is well written, but not without weaknesses. In particular:\n\n1. The authors do not discuss the computational cost of running their pooling operators in comparison to others. While potentially less relevant for graph classification tasks, it is highly relevant for node classification tasks. Please include some discussion, at least in the appendix, of this.\n2. The authors do not justify _why_ it is expected that the max cut should provide a better pooling operator.\n3. The results in experiment 4.2 indicate only mild performance gains, and the presented method frequently underperforms the case with no pooling. It's was hard for me to understand from the text what the intuition for this is.\n4. The experiments in section 4.3 are hard to understand without a no-pooling option." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* l.289: Is CON always defined using the assignment matrix $\\mathbf{S}$ based on nearest neighbor aggregation?\n* How well does the differentiable MaxCut approximate MaxCut compared to more other algorithms in terms of approximation and time complexity? \n* Is the differentiable MaxCut required or does this method work with other algorithms for MaxCut?\n* How important is the nearest neighbor aggregation?\n* How do the other methods perform with the same level of hyperparameter tuning?\n* Why is there no reproducible implementation provided?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper proposes an intuitive pooling method. It is very nicely written and nicely describes the current state-of-the-art for graph pooling. I understood almost all of the details." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel graph pooling method based on the maxcut algorithm. As MaxCut assigns high scores to disconnected nodes, this paper motivates its connection to one-every-K pooling methods and score-based methods. They formulate a differentiable form of MaxCut which they integrate into a pooling layer that utilizes heterophilic message-passing layers. Several experiments comparing MaxCutPool to other pooling methods are conducted." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* W1: Auxiliarly loss: As with other top k pooling approaches, the selction is a discrete action and thus not differentiable. Optimizing $\\mathcal{L}_{cut}$ may not lead to an s that is well-suited for the downstream task. Claiming that the proposed pooling operator is fully differentiable is thus misleading.\n* W2: This paper proposes multiple parts without thoroughly analysing each of them:\n * i) Gradient descent for combinatorial problems are typically not preferred as the solution can arbitrarily bad due to local minima. As this paper proposes to use gradient descent for MAXCUT, there needs to be a more detailed comparison with existing methods, ideally in terms of time/space complexity, expected performance and guarantees. \n * ii) It also remains unclear whether this gradient based approach is better for downstream-tasks. The only experiment that compares the proposed MaxCutPool with existing MAXCUT algorithms is in Table 2, which shows slight improvements. To me it is not convincing to use this gradient based approach vs. traditional algorithms within MaxCutPool.\n * iii) While the nearest neighbor aggregation is proposed as a algorithms for many pooling methods, it is only evaluated in combination with MaxCutPool. As the difference is rather small it is unclear whether this step is needed. If the authors decide to propose such an algorithm, a more detailed evaluation would be needed. \n* W3: The overall contribution is quite limited. As pointed out by the authors, many algorithms for graph pooling exist and the theoretical and empirical benefits of MaxCutPool do not convince me.\n* W4 Experiments: \n * i) There is no reproducible implementation provided.\n * ii) Graph classification: There seems to be a large hyperparameter optimization ofr MaxCutPool, while \"all [other] pooling layers were used with the default hyperparameters\". This does not seem to be a fair comparison. Especially when only spliiting into train and validation, the search space needs to be of similar size.\n * iii) Heterophilic tasks are only defined for node classification tasks as heterophily is typically defined as having mostly different labels between adjacent nodes. As node labels do not exist for graph-level tasks, it is unclear what heterophilic graph classification means. The authors also did not define what heterophilic graphs are.\n \nAdditional minor points:\nl.156: Sharpening propagation operators cannot learn any kind of gradients, but sharp gradients.\nl.425: \"This is the first known instance of a non-expressive pooler passing the expressiveness test provided by this dataset, serving as a counterexample to Theorem 1 in Bianchi & Lachi (2023)\". To my understanding, Theorem 1 provides a sufficient condition not a necessary one." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A GNN-based approach for computing MAXCUT on attributed graphs, used to implement graph pooling." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024maxcutpool,\ntitle={MaxCutPool: differentiable feature-aware Maxcut for pooling in graph neural networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xlbXRJ2XCP},\nnote={under review}\n}" }, "abstract": { "value": "We propose a novel approach to compute the MAXCUT in attributed graphs, i.e., graphs with features associated with nodes and edges. Our approach is robust to the underlying graph topology and is fully differentiable, making it possible to find solutions that jointly optimize the MAXCUT along with other objectives.\nBased on the obtained MAXCUT partition, we implement a hierarchical graph pooling layer for Graph Neural Networks, which is sparse, differentiable, and particularly suitable for downstream tasks on heterophilic graphs." }, "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": [ "Graph neural networks", "graph pooling", "graph coarsening", "maxcut" ] }, "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/5632c3e0767e0bcf16ebe248b6f8a188dd180538.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "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": "MaxCutPool: differentiable feature-aware Maxcut for pooling in graph neural networks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "-/-" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper builds on a solid theoretical foundation of edge-dependent random graph models, extending existing work to improve clustering and variability.\n- The supplementary materials provide sufficient details for understanding and reproducing the experiments.\n- The paper demonstrates a clustering and variability across by fitting different real-world datasets.\n- The paper offers a _potentially_ useful tool for generating random graphs with enhanced clustering and variability by fitting model parameters to specific datasets.\n- The authors provide a thorough reproducibility package, including all experimental workflows, parameters, and scripts, though this package could benefit from more comprehensive documentation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a generative random graph model that aims to balance realism of the generated graphs with computational efficiency of the generation process, seeking to increase clustering and variability in graphs generated by traditional models like Erdős-Rényi and stochastic block models. To achieve this, the paper uses an edge-dependent graph model (EPGM), which they use to create controlled dependencies among edges, and proposes a tractable algorithm for graph generation. The study includes a theoretical framework, practical algorithms, and experimental validation to demonstrate the generation of graphs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Validation**: The term \"desirable\" graph properties---namely, realistic structures, patterns, and variability---is vague, highly domain-dependent, and lacks formalization in the paper, making it challenging to assess whether the proposed model achieves these goals. A formal definition and visual examples earlier in the text could maybe help clarify these aims.\n- **Limited Exploration of Desirable Properties**: The paper does not fully explore the binding approach’s effects on other important network properties, such as modularity, conductance, and resilience, which are also highly desirable in certain domains.\n- **Dependence on Parameter Fitting**: The model generates graphs after fitting model parameters to real-world networks, rather than directly from scratch. It is unclear whether this model could lend itself to straightforward manual parameterization to generate synthetic graphs independently, a feature offered by models like the Lancichinetti-Fortunato-Radicchi (LFR) benchmark graph generator. This suggest that the graph generator has a narrow practical use.\n- **Scalability**: It is unclear if the model can efficiently generate realistic extremely large graphs, given that the experimental results focus on relatively small networks.\n- **Unclear Tractability**: The explanation of the binding schemes' tractability could be more detailed and intuitive, especially as it seems that binding is computationally slower than edge-independent methods, which could hinder its scalability to larger graphs.\n- **Comparison with Established Benchmarks**: The paper does not compare with the Lancichinetti–Fortunato–Radicchi (LFR) benchmark data generator, which also generates graphs with **power-law degree distributions, variability, and high clustering**. Given that the LFR model shares similar motivations and goals with the proposition, a comparison here would add necessary context. Additionally, faster and more efficient follow-up models to LFR align well with the goals of this study, particularly in terms of tractability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "(a) Similar effort can be found in the exchangeable network models, in view of relaxing conditional independence among edges. See for example \n\nHarry Crane and Walter Dempsey. Edge exchangeable models for interaction networks.\nJournal of the American Statistical Association, 113(523):1311–1326, 2018. PMID:\n30467447.\n\nWeichi Wu, Sofia Olhede, and Patrick Wolfe. Tractably modeling dependence in networks\nbeyond exchangeability. Bernoulli, 31(1):584 – 608, 2025\n\nPlease discuss more on this in literature.\n\t\t\n(b) The organization of the paper can be improved. For example, Definition 3.1 and Theorem 3.3 seem to take up too much space without demonstrating any clear importance. By contrast, the key definition of binding is placed in the appendix. \n\n(c) The authors focus a specific subclass ( good subsets ) of EPGM in section 4, however, why are the subsets considered are important? The paper can be improved if more motivation besides the feasibility are given. \n\n(d) The authors should compare and discuss more general models, such as the graphon model and the exponential random graph model as mentioned in their Section 3.2 in addition to the four specific models in Section 5.4. These models exhibit edge dependence, and is able to produce many triangles (especially the exponential random graph). It would be beneficial if the authors could provide a discussion on these models. \n\n(e) Theorem 5.7 provides the probability result for 3-motifs only, and the proof uses an enumeration method that could be difficult to extend to higher order motifs, which can be restrictive sometime in practice. A way to alleviate this is to provide codes in package for motives 4,5,6 as discussed in page 21. Also, please discuss how the closed-form varies with changes in \\( p \\), \\( g \\), and \\( R \\) for readers to better understand the results. Is it possible to consider several motives together? \n\n(f) Besides triangle, other network motives such as transitivity, can be important. Could the binding methods extend to 'transitivity' and 'triangle' simultaneously? Also please add some reference for the importance of triangle since the paper focuses on this quantity. \n\n(g)The authors conduct fittings in their simulations, however, the discussion of fitting in the main text is quite limited. If possible, the authors could provide some discussion on the properties of fitting, such as consistency.\n\n(h) The results in Table 1 is a little confusing. In the caption it writes ``The statistics are averaged over 100 random trails\". But the number of triangles are solved by tractability results. The data set is given, so what does the `random trial mean'? For edge independent models, densities of triangles can be computed tractably, too. On the other hand, the results in Table 1 might not be very meaningful. First, the results show that a parameter of a model is close to the corresponding real world data. However, what really matter is that the proposed method can produce a graph with similar property in high probability, which can be evaluated using simulated mean squared of $($ground true $\\Delta$-generated $\\Delta$$)^2$. Second, the comparison is not meaningful. For instance, in the case of the ER model, the authors estimate the connection probability matrix using the ER graph, then optimize their model parameters (number of triangles) based on a loss function that relates to the triangles, and subsequently compare the number of triangles from their model with those from the ER graph. Then the results are surely better since the authors are optimizing the number of triangles (with respect to the observed number of triangles) in a large class of model that include the ER graph. On the other hand, ER model, or other independent edge model is not necessarily good for the real data set in addition to the discrepancy of the numbers of triangles. Therefore, improving in triangle numbers comparing with those possibly inappropriate model could be less significant in practice. The authors could consider comparing with other models that take into account triangles, such as exponential random graphs and others.\n\t\t\n(i) Please discuss in detail the differences between LOCLBDG and PARABDG, as these two methods are distinct (as evidenced by the results in Figure 3). It would be beneficial to explain under which scenario one should choose LOCLBDG (or PARABDG)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors propose a general binding algorithm, as well as local and parallel binding algorithms, to generate networks with high variability and clustering. They also present closed-form results concerning triangles and discuss time complexities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the concept of EPGM, which is wider than the edge-independent graph models. The authors propose a general binding algorithm, as well as local and parallel binding algorithms, to generate networks with high variability and clustering. They\nalso present closed-form results concerning triangles and discuss time complexities. The authors conduct simulation studies to evaluate their models. Overall, the idea of binding is interesting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Only provides results for triangles\n2. Focus on a specific subclass of EPGM\n3. Simulation results are not very convincing" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Two kinds of questions are asked above regarding 1) the modeling side of binding, and 2) possible guarantees that binding alleviates the dense subgraph count vs overlap issue that motivates the work.\n\n### Typos / minor\n- Lines 3-5 in Algorithm 2: Is this simply a for-loop from $1$ to $R$? That might be slightly clearer if so. \n- Line 371 \"EGPMs\" -> \"EPGMs\"" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The central problem that is tackled, alleviating limitations of edge-independent / inhomogeneous ER graphs, is well-motivated in the text.\n- The proposed concept of binding can be applied to augment several different edge-independent random graph models.\n- Experiments demonstrate the effective of binding at matching triangle on four well-known kinds of RGMs and several well-known graph datasets.\n- There is theoretical work guaranteeing the tractability of the sampling algorithms.\n- The organization of the paper is generally logical and clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes the concept of binding to upgrade edge-independent random graph models, which are RGMs in which each edge is sampled independently. Binding allows for sampling a graph with given marginal probabilities for each edge, but with some dependence between the edges, allowing for more realistic graph distributions in terms certain statistics, like high clustering coefficient. The idea is to partition all possible edges, then to sample a single uniform random variable for each group in the partition, and add edges to the sampled graph if the random variable is less than edge's marginal probability. This adds edge dependence in the sense that if $e$ and $e'$ are in the same partition, and the marginal probability of the former edge is higher, then the sampling of the former guarantees the sampling of the latter. The paper then proposes a scheme for finding a partition at sampling time, as well as a parallelized scheme which relaxes the edge partition to independent edge groupings. Finally, the paper presents theoretical validation of the tractability of their methods, showing that the samples are generated in quadratic time; and empirical validation of the increased quality of samples when using binding, showing that upon fitting parameters of their model to match the input graph's triangle count, samples indeed match the triangle count closely, while not harming how well the node degree and node pair distance distributions are matched." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some writing could be clearer. For example, at the end of page 1, the authors say they will explore EPGMs, but it is not clear at that point what it is, or whether it is a previously-proposed concept versus something being proposed in this paper. More natural language descriptions of math, e.g., Property 4.4, would facilitate reading.\n- Given that binding is perhaps the main concept of this work, its formal definition should be in the main paper, and there should probably be some natural language description of the idea of general binding along with Algorithm 1 (e.g., as given in the summary above). There could also be more description of the modeling implications of binding: Does binding reflect some natural process that gives rise to real-world graphs? On the other hand, could the model assumption in binding be limiting in some way?\n- At least in the main paper, the theoretical work gives guarantees about the tractability of the sampling algorithms, but there is no proof that binding can alleviate modeling issues beyond Proposition 5.2, which is fairly vague - \"binding produces higher or\nequal subgraph densities\", but how much higher? E.g., does binding solve the motivating issue in Theorem 3.3 that edge-independent models cannot generate many triangles with low overlap?\n- There could be more discussion of binding" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "None" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "My questions are mostly about how \"realistic\" this random process is:\n\n1. What are some types of real-world network properties you believe *cannot* be captured by the binding processes you outlined in the paper?\nCan it capture realistic community structure faithfully?\n\n2. One thing missing from the paper, as outlined in the weaknesses part, is an experimental part showing the this specific model is useful for downstream applications (and not just as proxy for estimating quantities that are practically useful in downstream applications, such as triangle density). I would imagine that one way to instantiate it is by pre-training some graph neural network on this kind of graphs and show that this pretraining helps improve the performance on some applied benchmark. Do you think it will be possible/realistic to design such an experiment? In what context do you think this random model could come up useful?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- An elegant, novel and flexible method to generate random graphs. Although the generated graphs from, say, the local binding model do not yet seem fully realistic to me, this is a good step in the right direction and the basic binding primitive is natural and analyzable.\n\n- The paper is well written in general. The mathematical statements are clear, intuitive, interesting, and believable (I have not fully checked the proofs).\n\n- The topic of \"realistic\" random graph models is very important and this paper certainly advances the literature (even if by a small step) on this topic." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new configurable way to generate random graphs with dependencies between the edges, in a fine-grained way that make it possible to control parameters such as the number of triangles, clustering coefficients, etc. \n\nThe simplest and most famous random graph models, such as Erdos-Renyi, have edges appear independently of each other (in Erdos-Renyi for example the probability of each edge to appear is $p$, independently of all others). This kind of behavior makes analysis easier, but on the other hand it provably cannot capture properties of realistic graphs, such as a large number of triangles/motifs, and the tendancy of realistic graphs to be relatively clustered.\n\nMore modern random graph models, that are better at capturing real world characteristics, tend to have dependencies between the edges. One example is random geometric graphs, where points are randomly located on a high-dimensional sphere. But still, most of these models are not very flexible, they are controlled by some global parameters but it may be difficult/impossible for the designer to build a graph that matches specific desired properties.\n\nThis paper provides several variants of a new and simple method, called binding, to generate random graphs with high variability (i.e., that do not match a single specific pattern) but also being realistic (having high triangle density, high clusterability, etc). The core idea is simple, and stems from the following question: what is the random graph model with all edges appearing with probability $p$ (possibly dependently), and, say, the largest expected number of triangles? It turns out that this is the distribution which is equal to a complete graph with probability $p$, and empty graph otherwise. Binding is a similar process, picking the same \"threshold\" to groups of random nodes in order to create small cliques / dense graphs. The proposed model allows to work with predetermined edge probabilities, not necessarily uniform.\n\nThe authors propose a couple of different binding methods, one that is sequential and another one that is parallelizable and easier to analyze. They prove several theoretical results: for example, that it is possible to efficiently match a target expected number of triangles by a suitable optimization process. The authors also run some experiments demonstrating the verstality of this method in practice, showing ability to generate diverse and realistic degree distributions or obtain a large triangle coefficient. Another experiment shows that the running time is decently fast, especially for the parallelizable version." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experimental section does not provide a lot of information beyond the theoretical part of the paper. It is relatively clear to me, without the experiments or even looking at the proofs, that this model will be able to capture a target triangle density, for example. I would have hoped for the experimental section to show other directions in which this model is realistic. (See questions below.)" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We explore edge probability graph models (EPGMs) beyond edge independency, and show that realization beyond edge independence can produce more realistic structures while maintaining high tractability and variability." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024exploring,\ntitle={Exploring Edge Probability Graph Models Beyond Edge Independency: Concepts, Analyses, and Algorithms},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xljPZuprBA},\nnote={under review}\n}" }, "abstract": { "value": "Desirable random graph models (RGMs) should \n*(i)* generate *realistic* structures such as high clustering (i.e., high subgraph densities),\n*(ii)* generate *variable* (i.e., not overly similar) graphs, and \n*(iii)* remain *tractable* to compute and control graph statistics.\nA common class of RGMs (e.g., Erd\\H{o}s-R\\'{e}nyi and stochastic Kronecker) outputs edge probabilities, and we need to realize (i.e., sample from) the edge probabilities to generate graphs.\nTypically, each edge's existence is assumed to be determined independently for simplicity and tractability.\nHowever, with edge independency, RGMs theoretically cannot produce high subgraph densities and high output variability simultaneously.\nIn this work, we explore realization beyond edge independence that can produce more realistic structures while maintaining high traceability and variability.\nTheoretically, we propose an edge-dependent realization framework called *binding* that provably preserves output variability, and derive *closed-form* tractability results on subgraph (e.g., triangle) densities in generated graphs.\nPractically, we propose algorithms for graph generation with binding and parameter fitting of binding.\nOur empirical results demonstrate that binding exhibits high tractability and generates realistic graphs with high clustering, significantly improving upon existing RGMs assuming edge independency." }, "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": [ "Random graph models", "edge dependency", "triangle density", "subgraph densities", "tractability", "variability" ] }, "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/103eb1afb82ca41bc9e79104d539f16a1b0ff899.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "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/6667cb63203da5a8e910bf5f6c37ae2d5135f98e.zip" }, "title": { "value": "Exploring Edge Probability Graph Models Beyond Edge Independency: Concepts, Analyses, and Algorithms" }, "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": "- Please elaborate on the limitation of the proposed method (e.g., computational cost).\n- Is the proposed method robust against noise?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- PE-GQNN combines quantile prediction and distribution recalibration in a single model (not two-stage models), enhancing both predictive accuracy and calibration efficiency.\n- By limiting GNN operations to specific features and introducing target values near the output layer, the model effectively prevents data leakage and improves computational efficiency.\n- The authors use pinball loss for quantile regression, which allows one to provide a regularization effect, improving prediction reliability across diverse quantile levels.\n- Extensive experiments on multiple real-world datasets demonstrate PE-GQNN’s superior performance in both predictive accuracy and uncertainty estimation compared to existing methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Positional Encoder Graph Quantile Neural Network (PE-GQNN), a novel model for spatial data prediction. PE-GQNN integrates Positional Encoder Graph Neural Networks (PE-GNN) with Quantile Neural Networks (QNN) to improve predictive accuracy and quantify uncertainty. Experiments on real-world datasets show that PE-GQNN consistently outperforms existing methods in predictive accuracy and uncertainty calibration." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed method is practically useful, but the way PE-GNN and QNN are combined is somewhat straightforward.\n- There are several innovations in the architecture, but they are all practical techniques, not theoretically sophisticated.\n- There is no discussion of computational cost for high-dimensional data.\n- There is no discussion of the shortcomings 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": 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 Figure 4(a) does not have legend. What does the 9 curves represent? It is unclear to me how does these 9 curves represent 10 samples.\n2. Instead of using MADECP, what is the performance when using popular adopt interval? For example, 95% confidence interval?\n3. The real-world geographic data usually collect in a temporal manner, so how does the data is separated into different train/val/test segments? It leads to data leakage issue for sure if the separation does not only use historical data for training.\n4. The sensor of the geographic data also will be removed/replaced/re-deployed/defected from time to time in real world applications. Does the proposed method also robust in these cases?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The framework is easy to understand and follow.\n2. The final experimental results indicate some effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduced a distribution-free uncertainty quantification framework by integrating PE-GNNs, Quantile Neural Networks for geographic data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The term \"fully nonparametric\" could be misleading. It would be more accurate to use \"distribution-free.\"\n\n2. The novelty of the proposed method is limited. Quantile regression is already proposed and widely used in distribution-free uncertainty quantification. For example:\n - \"Single-model uncertainties for deep learning.\" Advances in Neural Information Processing Systems 32 (2019).\n - \"Image-to-image regression with distribution-free uncertainty quantification and applications in imaging.\" International Conference on Machine Learning. PMLR, 2022.\n\n Additionally, using KNN-Graph with geographic coordinates has been proposed and widely applied in different domains:\n - \"Dynamic graph CNN for learning on point clouds.\" ACM Transactions on Graphics (TOG) 38.5 (2019): 1-12.\n - \"Spatiotemporal graph convolutional networks for earthquake source characterization.\" Journal of Geophysical Research: Solid Earth 127.11 (2022): e2022JB024401.\n\n3. PE-GNN does not claim the method as SOTA. So which baseline are considered as SOTA in this submission? PE-GNN is better than GNN but not the SOTA. At least there are more available models in applying GNN for earthquake data and traffic 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": 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": "What does ST stand for in the last sentence of the 1st paragraph of Section 2?\n\nIn Figure 4(a), why does the predicted density of PE-GQSAGE look like Gaussian distribution? The proposed method outputs predicted quantiles at the given tau. So, the prediction by the proposed method does not necessarily to be Gaussian shape." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Proposed some techniques for quantifying uncertainty in spatial regression.\nExperiments with three datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a graph neural network-based method for uncertainty quantification in spatial regression." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The technical contribution of this paper is limited since the proposed method is a combination of Positional Encoder Graph Neural Networks and quantile regression. The proposed method introduces some techniques, such as the use of response variables y and quantile parameters tau in neural networks. However, the novelty of these techniques is incremental.\n\nThe experimental results are not convincing. There have been proposed many quantile regression methods in neural networks; i.e., outputting variance in the last layer, the use of sparse Gaussian processes at the last layer. These methods can be easily combined with graph neural networks. The comparison with such existing methods can demonstrate the effectiveness 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": { "value": "The paper has not ethics concerns founded." }, "flag_for_ethics_review": { "value": [ "No 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. It is better to add some simple examples to illustrate the integration process and novelty in the process on how to design this new graph. \n\n2. The paper shall demonstrate the complexity change for this new PE-GQNN graph." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "There are several strengths demonstrated in the paper:\n1. The paper introduces the Positional Encoder Graph Quantile Neural Network (PE-GQNN), a new approach that integrates PE-GNNs, Quantile Neural Networks, and recalibration techniques in a fully nonparametric framework, requiring minimal assumptions about the predictive distributions.\n2. The paper has demonstrated the results on three datasets: California Housing, Air Temperature, and 3Droad with 6 different approaches including the proposed PE-GQNN." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors proposed the Positional Encoder Graph Quantile Neural Network (PE-GQNN) as a new framework to enhance predictive modeling for geographic data. The major contributions of this paper are listed as the following: The empirical results showed the capability of PEGQNN to achieve lower MSE, MAE, and MPE compared to traditional GNN and PE-GNN. Also, PE-GQNN demonstrated substantial improvements in predictive accuracy and uncertainty quantification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weaknesses of this paper are listed as the following:\n\n1.\tThe innovation of this paper seems incremental. Positional Encoder Graph Quantile Neural Network (PE-GQNN) is just a simple combination of PE-GNN with Quantile regression model.\n\nFirst, the paper shall illustrate in detail the challenges in the integration process. Normally a good integration will include some short cuts to reduce the total cost while comparing with the cost of simple addition of several algorithms together directly. Please try to add some \"novel points\" or new ideas to demonstrate your merits in integration.\n\nSecond, it is better to add some simple examples to illustrate the integration process and novelty in the process on how to design this new graph. Please illustrate which cost you saved compared the cost that you simply integrate several different phases from literature.\n\nThird, will the new integrated framework achieve higher performance compared simply by adding several phases together? What other advantages do you have for the new framework?\n\n\n2.\tThe paper shall demonstrate the complexity change for this new PE-GQNN graph. \n\nIs this new graph a simple integration? \nHow much is the increase of the total complexity? \nIs there any new approach to lower the total complexity?\n\nPlease provide a specific complexity analysis comparing PE-GQNN to PE-GNN, including time and space complexity.\n\n\n3.\tFor the experiments, the following should be addressed. \n\nFirst, the paper presented experimental results obtained from three datasets: California Housing, Air Temperature, and 3Droad. It seems that the paper lacks discussion about whether other kinds of datasets are suitable for this new approach. Also, how much the total cost changes to implement this new approach.\n\nSecond, please discuss the generalizability of their approach beyond the three datasets used. Also, please address what characteristics of a dataset make it suitable for PE-GQNN.\n\nThird, additionally, it would provide valuable practical insights if the authors can demonstrate a comparison of implementation costs between PE-GQNN and existing methods." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose PE-GQNN, which combines GNNs, quantile loss, and recalibration to improve uncertainty quantification and predictive accuracy in spatial data, outperforming current methods without increasing computational complexity." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024positional,\ntitle={Positional Encoder Graph Quantile Neural Networks for Geographic Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xlrpVyMIwz},\nnote={under review}\n}" }, "abstract": { "value": "Positional Encoder Graph Neural Networks (PE-GNNs) are a leading approach for modeling continuous spatial data. However, they often fail to produce calibrated predictive distributions, limiting their effectiveness for uncertainty quantification. We introduce the Positional Encoder Graph Quantile Neural Network (PE-GQNN), a novel method that integrates PE-GNNs, Quantile Neural Networks, and recalibration techniques in a fully nonparametric framework, requiring minimal assumptions about the predictive distributions. We propose a new network architecture that, when combined with a quantile-based loss function, yields accurate and reliable probabilistic models without increasing computational complexity. Our approach provides a flexible, robust framework for conditional density estimation, applicable beyond spatial data contexts. We further introduce a structured method for incorporating a KNN predictor into the model while avoiding data leakage through the GNN layer operation. Experiments on benchmark datasets demonstrate that PE-GQNN significantly outperforms existing state-of-the-art methods in both predictive accuracy and uncertainty quantification." }, "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": [ "Graph Neural Networks (GNNs); Quantile regression; Geospatial data; Uncertainty quantification; Calibration; Model recalibration." ] }, "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/6ebb1a286fcba9b1b6ca861b716ffb25fcb94659.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": "Positional Encoder Graph Quantile Neural Networks for Geographic 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": 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": "- Fig 5, is there a reason why $\\alpha=1$ performed best in all cases?\n- It would be interesting to have some theoretical justification for why the interventions, especially the power law sampling, lead to increased training speed." }, "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 overall clearly written.\n- The observations are indeed interesting and surprising especially since the interventions improve training without discriminating between datapoints." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors conduct a controlled set of experiments on three synthetic settings. They observe the loss plateau and find that by biasing the training distribution away from the test distribution to reduce data diversity at the start of training, one can accelerate training. Results from various simple and counterintuitive data interventions interventions suggest that simpler data filtering techniques can match or outperform complex optimization methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- From a loss landscape perspective, the initial speedup seen upon reducing data diversity suggests that the parameters are getting stuck in a local suboptimal minimum that may lie near the initialization. However, numerous methods have been developed to avoid getting stuck in such minima such as adding regularization, introducing learning rate schedulers, etc. Therefore, while reducing data diversity may seem to help speed up, it may not contribute to finding a better generalization solution at all. \n- The paper claims that a purely data-driven approach can have the same speedup as optimization methods; however, low similarity and high variance are also often associated with sharp minima in the loss landscapes. Therefore, the authors could consider comparing methods like sharpness-aware minimization, momentum, etc. in this context. Moreover, PCGrad is designed for training across multiple tasks, that are expected to have potential interference. However, I wonder how scalable it would be in a single-task setup for increasing model size, even though it does improve training speed. \n- While the authors mention that their synthetic settings are simple, I think finding the \"optimal\" amount of biasing needed for initial speedup can be quite challenging on real data, as it may require an extensive search over hyperparameters. I wonder if the authors could suggest any potential strategies for finding the optimal amount of biasing in real-world 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": 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": "# High-level questions\n1. Section 2.2. There is a notion of \"task\" in linear regression and fact memorization problems. What is a task in the sparse parity problem?\n2. Section 2.2. Why study 3-layer MLP here but transformers for the other two problems?\n3. Line 165. What is the $i^th$ token $d_i$? If these are the elements $s, r, n_1, ..., n_{k_noise}$, it doesn't make sense to have the transformer predict noise token $n_t$ given $s, r, n_1, ..., n_{t-1}$ since the noise tokens are drawn independently. If I'm reading this wrong, what is the tokenization procedure here?\n\n# Low-level questions\n1. Line 32: \"followed by abrupt learning to low loss\". What does this question mean? Are the authors pointing to the \"grokking\" line of work [1]?\n2. Line 93. Typo: $x$ is in $\\mathbb{R}^d$ but $(x, y)$ is not.\n3. Equation 3. What is $\\tau$?\n4. Line 132. What is $\\mathcal{D}$ here? Uniform distribution over the $d$-dimensional hypercube?\n5. Line 162. What is the ICL-LR setting? It is likely the 1st problem, but this abbreviation is introduced here for the first time.\n\n[1] Power, A., Burda, Y., Edwards, H., Babuschkin, I., & Misra, V. (2022). Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper presents experiments on different setups, involving both transformers and MLPs that are used in practice. There has been recent interest in using transformers/MLPs to learn sparse parities [1, 2] and in-context linear regression [3, 4], which the paper studies. The general problem in the paper: of accelerating training is of interest to the ML community. The paper takes approaches this problem from the lens of data curation goes back to Curriculum Learning [5].\n\n[1] Barak, B., Edelman, B., Goel, S., Kakade, S., Malach, E., & Zhang, C. (2022). Hidden progress in deep learning: Sgd learns parities near the computational limit. Advances in Neural Information Processing Systems, 35, 21750-21764.\n\n[2] Edelman, B. L., Goel, S., Kakade, S., & Zhang, C. (2022, June). Inductive biases and variable creation in self-attention mechanisms. In International Conference on Machine Learning (pp. 5793-5831). PMLR.\n\n[3] Garg, S., Tsipras, D., Liang, P. S., & Valiant, G. (2022). What can transformers learn in-context? a case study of simple function classes. Advances in Neural Information Processing Systems, 35, 30583-30598.\n\n[4] Zhang, R., Frei, S., & Bartlett, P. L. (2023). Trained transformers learn linear models in-context. arXiv preprint arXiv:2306.09927.\n\n[5] Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009, June). Curriculum learning. In Proceedings of the 26th annual international conference on machine learning (pp. 41-48)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper demonstrates with experiments that lowering training data diversity accelerates training. In 3 experiment setups with synthetic data---in-context linear regression in transformers, learning sparse parity functions with MLP, and fact memorization in transformers---the authors argue that a variety of interventions can accelerate training (e.g. interventions on the gradient updates or on the training dataset)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "# Motivation\nThe paper aims to accelerate training for the sake of optimization only, but that is not the central goal in learning. In fact, research in ML optimization techniques serve to fulfill the **central goal of generalizing to the test distribution**. Moreover, the paper frequently finds in their experiments that faster training (using the methods they present) often leads to worse generalization, so I doubt the benefits of the proposed methods.\n\n**In fact, the authors mention this exact severe limitation in line 52.** I'd find insights into generalization more helpful than just focusing on training optimization.\n\n## On reducing data diversity to optimize faster\nIt is not shown that accelerating training by reducing data diversity leads to increased generalization to the test distribution. I think that would be a good argument for \"reduce data diversity -> thus train faster -> thus generalize to test distribution better\". As an example, consider the setting where we want to learn the target function $f(x) = x^2$ given a finite number of training samples $x_1, ..., x_n$. We can forget about the target (test distribution) and instead sample from the function $g(x) = 1$, _just for the sake of low data diversity_. Optimizing on samples to learn $g$ is faster, but this is far from the goal of learning the target $f$.\n\nOn another note, decreasing the number of training tasks/samples to reduce training data diversity should intuitively lead to overfitting. Figure 4 shows this exact phenomena, so I'm not convinced that accelerated training with the paper's data curation methods is actually helping in learning.\n\n# Techniques\nIn Section 3, the paper argues that biasing batch gradients can accelerate training. Isn't this technique the same as SGD with momentum or an adaptive gradient method like AdaGrad or Adam? I'm not sure this is a novel contribution.\n\nFinally, the paper dos not conduct data-diversity experiments on real-world problems such as image classification (even on well-studied datasets like CIFAR or MNIST) and language generation. I appreciate that the authors mention this limitation in Section 6 and 7, but I believe the current version of the manuscript is severely lacking with just synthetic data experiments.\n\n# Writing\nThe paper needs numerous citations in the introduction to substantiate claims. Several typos and clarifications needed:\n1. Citations for lines 41-44.\n2. Line 96. It's verbose to call the task $f_w$ when we can simply write $w^T x + \\epsilon$ as used in Line 93. Lines 96-105 can be written in 3-5 lines at max.\n3. Line 297. The paper conflates \"faster optimization\" with \"learning\". The model trains faster, but does not generalize (or \"learn\") better or faster.\n\nMoreover, the statement \"low data diversity accelerates training\" is repeated throughout the paper (paraphrased in many ways), making the paper more verbose than necessary. I understand that it is the paper's main argument, but it is overly used in my view." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "1. There is a wide body of work on generalization through a theoretical lens, typically taking advantage of both model complexity and train/test distribution similarity. How does this work fit into this literature?\n\n2. Does this happen in more conventional language modeling (or non language modeling tasks)?\n\n3. Why does this effect happen at all? I don't see any justification of this phenomenon, only a few demonstrations of it." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The authors take a stance that, to my knowledge, has not been studied before in machine learning. If this were a well understood, real phenomenon, I think it could be impactful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper argues that convergence using SGD can be more efficiently obtained, in some cases, by using a training data distribution that has less in common with the test distribution. This argument is of course counterintuitive, and it disagrees with a wide body of fundamental research in the field around how an ideal training dataset is one that has more in common with the general distribution, not less." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall I found this paper’s argument unconvincing, as they experiment only on three synthetic settings and provide no arguments as to why this effect might occur. Of course, this effect cannot be shown forever, and eventually this skewing of the training distribution will have a negative effect on test performance. When and why does this happen, and what can be gleaned from this work about more conventional model training? \n\nAs the authors write themselves in the limitations section, they do not have an understanding of when or why this occurs, and do very limited evaluation and show preliminary results. I realize this section was written to get ahead of criticism they anticipate to this effect, but the work seems very incomplete with these questions left unanswered. The future work in the discussion section, for example, is very much in-scope, and should be part of the paper the authors are writing.\n\nAt this time, I feel that the presented investigation is not thorough enough to meet the bar for publication in ICLR." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Your approach and optimization algorithms can achieve the same goal, so how would you convince others to use your method instead of an optimization algorithm? \n\nWould combining your approach with other optimization algorithms lead to more significant performance improvements?" }, "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 experimental setup, demonstrating how intentional data biases can speed up learning in synthetic tasks.\n\nThe use of visuals makes its findings easier to understand.\n\nThis research is useful in situations where training efficiency is a priority." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper examines methods to accelerate training by reducing data diversity, focusing on synthetic tasks like in-context linear regression, sparse parity, and fact memorization. Traditional approaches typically improve training by adjusting optimization algorithms to mitigate plateaus; however, this study found that simple interventions in data sampling, such as reducing task diversity or sampling with non-uniform distributions, can achieve similar benefits. These findings provide insights into data filtering and curriculum learning approaches, suggesting that less diverse but strategically chosen training data could enhance model efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The application scope is narrow to meet real-world needs, and it lacks an explanation for the observed phenomena." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Reducing data diversity can speed up training in a variety of synthetic settings." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024lowering,\ntitle={Lowering Data Diversity can Accelerate Training: Case Studies in Synthetic Tasks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xlxDTVAbNM},\nnote={under review}\n}" }, "abstract": { "value": "We identify a loss plateau at the start of training in the three synthetic settings of in-context linear regression, sparse parity, and fact memorization. While careful tweaks to the optimization algorithm can mitigate these plateaus, we find that a simpler orthogonal approach of *lowering the data diversity*, and in doing so, biasing the training distribution *away* from the test distribution, counter-intuitively also speeds up training. This connection between data diversity and training speed holds for three different diversity-*reducing* interventions across our varied synthetic settings. Our findings offer a new perspective on data filtering and curriculum learning for training machine learning 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": [ "synthetic tasks", "data diversity", "curriculum learning", "data filtering", "learning plateaus", "batch gradients" ] }, "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/e87ef6894e9f1c1925402adf3cbd2645006bab84.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": "Lowering Data Diversity can Accelerate Training: Case Studies in Synthetic Tasks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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.It is recommended to expand both the U-MATH datasets size and the number of subjects.\n\n2.It is recommended to expand the µ-MATH datasets size.\n\n3.In Table 4, you only use accuracy to present the results. Since the study involves math problems, which are more complex than simple classification tasks, could you consider adding additional evaluation metrics like perplexity or WinoGrande ACC (to assess whether ambiguous problems are correctly identified)? This would give readers a clearer picture of how well the models truly understand and respond to university-level math questions. For more details, you might refer to examples in this paper: https://proceedings.mlr.press/v235/dao24a.html." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well-organized, providing a clear outline of the datasets, experimental setup, and evaluation metrics. The authors explain each component in a structured manner, making it accessible to readers.\n\n2. The datasets include a range of mathematical subjects and problem types, which reflects an effort to cover diverse aspects of mathematical reasoning, though the depth and breadth could still be improved.\n\n3. The introduction of U-MATH and µ-MATH provides additional benchmarks for evaluating LLMs in mathematical tasks, which may offer a reference point for similar studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the U-Math datasets, based on university-level mathematics, addressing the issues of insufficient thematic diversity and a lack of visual information question types in current datasets for evaluating the mathematical abilities of large language models. The U-Math datasets was tested on several large language models, revealing that the highest accuracy for text-based tasks was only 53%, while the highest accuracy for visual tasks was only 30%." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the U-MATH datasets consists of 1,125 samples and covers six subjects, the sample size is still too small. Evaluating the mathematical abilities of large models using a limited amount of data is not sufficiently convincing.\n\n2. Although the 340 samples in the µ-MATH datasets have been carefully selected to provide a challenging test, a larger sample size could enhance the representativeness of the evaluation, especially across different topics and problem types." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "What measures have been taken to mitigate potential biases introduced by using LLMs as judges for solution correctness?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1. The inclusion of university-level problems offers a significant advancement over existing datasets that mainly focus on elementary or high school-level tasks.\n\nS2: By integrating visual tasks alongside traditional textual ones, the dataset challenges LLMs to interpret and reason across multimodal formats.\n\nS3: µ-MATH introduces a novel approach to evaluate LLMs' ability to assess solutions, addressing biases and limitations in current evaluation practices." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces U-MATH, a comprehensive university-level mathematical benchmark designed to evaluate the performance of Large Language Models (LLMs) in solving advanced mathematical problems. The dataset consists of 1,125 problems sourced from university coursework, covering six core topics such as Algebra, Calculus (Differential and Integral), Multivariable Calculus, Sequences, and Series, with approximately 20% of the tasks involving visual components. To complement the U-MATH dataset, the authors also present µ-MATH, a meta-evaluation set for assessing the accuracy and reliability of LLM-based evaluators." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1: The reliance on LLMs as judges (e.g., GPT-4o) to evaluate free-form answers could introduce biases and inconsistencies, particularly since LLMs may struggle with complex derivations or nuanced interpretations of mathematical expressions.\n\nW2: The µ-MATH set includes LLM-generated solutions, which may limit the diversity and challenge of evaluation due to inherent model tendencies or training biases. This could result in less rigorous meta-evaluation as models may overfit to known patterns or heuristics." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "I don't have further questions." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.U-MATH Benchmark: This is a publicly available dataset of university-level math problems, covering six topics: Pre-Calculus, Algebra, Differential Calculus, Integral Calculus, Multivariable Calculus, and Sequences & Series. A unique aspect of this dataset is its inclusion of open-ended questions that require LLMs to perform multi-step reasoning.\n2.µ-MATH Meta-Evaluation Benchmark: This benchmark is specifically designed to test LLMs’ ability to assess the correctness of mathematical solutions. It contains 340 questions selected from U-MATH, accompanied by LLM-generated answers manually labeled as correct or incorrect, aimed at evaluating the capacity of LLMs to act as “judges.”\n3.Model Comparison: The paper compares the performance of various LLMs, including general-purpose models, specialized math models, and multimodal models, demonstrating the significant challenges LLMs still face in both text and visual tasks. For instance, the highest accuracy for text-based questions is 53%, while performance on visual questions is even lower, with an accuracy of only 30%.\n4.Challenges for LLMs as Math Judges: LLMs perform poorly when evaluating mathematical solutions, with the best-performing LLM judge achieving an F1 score of only 76% on µ-MATH, indicating that there is still room for improvement in this task." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduced a new benchmark dataset, U-MATH, designed to evaluate large language models (LLMs) on university-level math problems. The proposed U-MATH benchmark includes 1,125 college-level math problems collected from real educational materials, covering six core mathematical subjects, with 20% of the problems involving image understanding. Additionally, the paper introduces a meta-evaluation dataset named µ-MATH, aimed at assessing the ability of LLMs to judge the correctness of mathematical solutions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The U-MATH dataset introduced in the paper supplements the current math datasets by addressing college-level gaps, while the µ-MATH meta-evaluation dataset enables assessment of large models’ ability to evaluate college-level math solutions. However, aside from knowing that this training set focuses on university mathematics and includes six subjects, we lack information about the dataset’s question diversity, difficulty, reasoning steps required to solve the problems, and other aspects. Additionally, the dataset’s size may be insufficient.\n2.The paper mentions that the dataset has been released but does not provide an access link, so I have no direct way to review the dataset.\n3.The experiments in the paper provide valuable insights into the capabilities of current text-based and multimodal LLMs in solving university-level math problems.\n4.The paper states that U-MATH aims to promote further research and improve LLMs' ability to handle complex math problems. How is \"complex\" defined here? Does it refer to higher-grade, more challenging (for humans) knowledge, or does it mean problems requiring more and deeper reasoning steps?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 are there no examples of problems that require visual input?\n\nThe accuracy when using LLM as a judge is not provided, especially for higher mathematics problems where answers may be in different forms but are actually equivalent, indicating that it is easier to make mistakes compared to comparing a single form of answer." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper demonstrates a high-quality collection of problems that are well-balanced across six core mathematical subjects. This ensures a comprehensive evaluation of LLMs across different areas of mathematics.\nThe problems sourced from actual teaching materials add a layer of authenticity and practical relevance to the benchmark, ensuring that the skills assessed are applicable to real-world academic standards.\nThe creation of µ-MATH for meta-evaluation is an innovative approach to assessing the ability of LLMs to evaluate mathematical solutions. This adds another layer of complexity and originality to the benchmarking process, focusing not just on problem-solving but also on the assessment capabilities of the models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces U-MATH, a novel benchmark designed to evaluate the mathematical reasoning capabilities of Large Language Models (LLMs) at the university level. It comprises 1,125 unpublished, open-ended problems sourced from actual teaching materials, balanced across six core mathematical subjects, with 20% of the problems requiring image understanding. Additionally, the paper presents µ-MATH, a meta-evaluation dataset aimed at assessing the ability of LLMs to evaluate free-form mathematical solutions. The experiments conducted reveal significant challenges in advanced mathematical reasoning and visual problem-solving, with the best-performing models achieving only 53% accuracy on text-based tasks and 30% on visual problems. The paper also highlights the difficulty LLMs face in assessing solutions, with the highest µ-MATH F1-score being 76%, indicating room for improvement in LLMs’ evaluation capabilities. The datasets and evaluation code are open-sourced to facilitate further research." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the inclusion of visual elements in 20% of the problems is a step forward, the remaining 80% are text-based. The paper could benefit from expanding the visual problem set to better assess and train LLMs in multimodal mathematical reasoning, which is increasingly important for real-world applications.\n\nThe paper focuses on university-level mathematics, but it is unclear how well the findings generalize to other levels or types of mathematical reasoning. Future work could explore the transferability of the models trained on U-MATH to other mathematical domains." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "U-MATH, a challenging university-level math benchmark with both textual and visual problems, and additional μ-MATH benchmark to evaluate solution assessment capabilities." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024umath,\ntitle={U-{MATH}: A University-Level Benchmark for Evaluating Mathematical Skills in {LLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xlxGsX1pc7},\nnote={under review}\n}" }, "abstract": { "value": "The current evaluation of mathematical skills in LLMs is limited, as existing benchmarks are relatively small, primarily focus on elementary and high-school problems, or lack diversity in topics. Additionally, the inclusion of visual elements in tasks remains largely under-explored. \n \nTo address these gaps, we introduce **U-MATH**, a novel benchmark of 1,125 unpublished open-ended university-level problems sourced from teaching materials. It is balanced across six core subjects, with 20\\% of problems requiring image understanding. Given the open-ended nature of U-MATH problems, we employ an LLM to judge the correctness of generated solutions. To this end, we release **$\\boldsymbol\\mu$-MATH**, an additional dataset to evaluate the LLMs' capabilities in assessing solutions.\n\nThe evaluation of general domain, math-specific, and multimodal LLMs highlights the challenges presented by U-MATH. Our findings reveal that LLMs achieve a maximum accuracy of only 53\\% on text-based tasks, with even lower 30\\% on visual problems. The solution assessment proves challenging for LLMs, with the best LLM judge having an F1-score of 76\\% on $\\mu$-MATH.\n \nWe open-source U-MATH, $\\mu$-MATH, and evaluation code on GitHub." }, "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 (LLMs)", "Mathematical Reasoning", "Benchmarking", "University-Level Mathematics", "Multimodal", "Automatic Evaluation", "Solution Assessment" ] }, "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/5b892a9eaff28812a984c7732d30f2fb644b7533.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": "U-MATH: A University-Level Benchmark for Evaluating Mathematical Skills in LLMs" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. “We initialize the final projection layer of DiT to output zeros.” Do you mean zero-initialize all the weight and bias in the projection layer? Or do you use gated operation? More explanations are welcomed.\n2. Please consider adding the result of Make-an-Audio and AudioLDM in Table 2 and Table 3, as they are also very important baselines to compare with.\n3. In Table 6, the ETTA trained on AudioCaps shows 3.00 FAD on the audiocaps evaluation set. This is significantly lower than the current state-of-the-art. It would be helpful if the author could explain this result.\n4. Line 510: Please say more about the loss divergence issue? It seems contradictory as the evaluation metrics still look normal. More explanations would be helpful." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The author proposed a synthetic data generation pipeline and created the first million-size dataset with high audio-text correlation.\n2. The author has implemented a comprehensive list of objective evaluation metrics and shows interesting results during comparison.\n3. The author mentioned they will open-source their code for reproducibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper empirically studies the way to improve the current text-to-audio generation system, including creating a new synthetic audio-caption paired dataset, an improved architecture, and other related settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As there are a large number of evaluation metrics used. It would be helpful to explain what each evaluation metrics focus on. For example, there are three types of Frechet Distance - what are the differences?\n2. It is not clear how ETTA perform without pretraining on the large-scale datasets. For example, what would the result look like if the ETTA model is trained on AudioCaps from scratch (I assume this is a common setup)?\n3. Lack of subjective evaluation. As the author mentioned, there is not yet an effective objective evaluation metric for the TTA task. Out of the seven main conclusions in the paper, only one is backed by the subjective evaluation. This can make the conclusion less convincing. \n4. The architectural improvement seems a bit marginal. The improvement of ETTA seems to come from the new synthetic dataset mostly." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Please refer to Weaknesses#2." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The contribution of the dataset: Experiments demonstrate that AF-Synthetic significantly improves performance.\n2. It provides a practical guide for hyperparameter tuning in the field of TTA.\n3. Each experimental conclusion is highlighted with a purple-bordered box, which makes the paper very reader-friendly and pleasant to navigate." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The primary focus of this paper is not to explore novel model designs, but rather to provide a comprehensive understanding of the current paradigms in TTA models. It seeks to identify critical factors that contribute to performance improvements and to evaluate scalability concerning data and model size. Additionally, the paper introduces AF-Synthetic, the first million-size synthetic caption dataset with strong audio correlations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As the authors themselves mentioned, the paper doesn't propose a novel method. Moreover, as a paper based on extensive experiments, while it offers valuable conclusions, it lacks an innovative insight or discovery that stands out. Therefore, I believe this paper might be a better fit for the Dataset & Benchmark track. \n2. I would also suggest that the authors include the following three points: \n - Could the authors further analyze the \"potentially mode-collapsed mode\" part? In theory, the FD metric should be able to measure sample diversity. \n - Since the authors have already tried logit-normal *t-sampling*, could they also test the *Min-SNR Weighting Strategy*? \n - Regarding the Auto-guidance section, it might be worth experimenting with removing CFG (Classifier-Free Guidance) for the first 40% of steps and then adding CFG for the remaining 60% of steps to see if this improves the FD 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": 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 can be the reason behind mode-collapsed models consistently produce good scores across multiple metrics, while the backbone for these metrics are different?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The work aims to push the performance of diffusion-based text-to-audio models and show better results than open source models. The improvements come from a better dataset and an extensive evaluation of design/hyperparameter choices built on top of SOTA models. I appreciate the amount of work put into creating the dataset and try many different hyperparameters and design choices, and the commitment to open source the code that appears to push the field ahead." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on pushing the performance of diffusion-based text-to-audio models by introducing a new synthetic dataset and revisiting design choices. \n- New synthetic dataset (AF-Synthetic): The authors improve the pipeline from AF-AudioSet to generate 1.35M captions using audio samples from multiple datasets. The result is a large dataset with high CLAP score, while existing datasets are either small or have simple to no filtering method. The authors also point out that AF-Synthetic captions are different than existing datasets.\n- The model is built upon Audio-VAE (from stable-audio-tools), Diffusion Transfromer (Peebles & Xie, 2023) with some changes in the architecture.\n- The authors compare across datasets and metrics to discover the optimal choice for the number of function evaluations and classifier-free guidance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Since this is a purely empirical paper, my main concerns are mostly about evaluation and results.\n\n- Since the dataset is curated using CLAP scores, I find the CLAP score not a reliable indicator, while FAD and KL use old models to extract features.\n- There are no subjective scores in the results except for Table 9 which compares the final model with others.\n- The improvements in ETTA-DiT, OT-CFM and t-sampling are not consistent across metrics.\n- Using AF-AudioSet and AF-Synthetic yields similar results in AudioCaps and MusicCaps despite much larger size (Table 6-7)\n\nOverall, I find that a few conclusions in the paper are not very helpful (e.g. increasing model size improves the performance), especially when we have inconsistent metrics and lack of subjective scores. While it's good to see many hyperparameters being assessed, the contribution is lack of novelty since it does not propose new techniques or reveal surprising findings. The lack of reliable metrics, which authors also admit in the final section, also weaken claims and conclusions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": { "value": "No ethic concerns" }, "flag_for_ethics_review": { "value": [ "No 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. For the comparison of the proposed AF-Synthetic dataset, the paper indicates that both AF-AudioSet and Sound-VECaps have around 161K samples remaining after filtering with a CLAP score of 0.45. Since both datasets are primarily developed from AudioSet, is there a significant overlap in the subset that achieves a higher CLAP score? If so, could this raise concerns regarding the reliability of using the CLAP score as a filtering metric?\n2. What are the key differences between AF-AudioSet and the proposed AF-Synthetic, apart from the fact that the latter generates ten captions per sample and selects the one with the highest CLAP score?\n3. What distinguishes the AdaLN layer in the proposed ETTA-DiT structure from the one used in the original DiT model mentioned in a previous section?\n4. How does the ETTA model perform with improvements limited to the ETTA-DiT component (e.g., when trained solely on AudioCaps)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper introduces a novel filtering pipeline to select the best captions using the CLAP score, developing one of the largest high-quality audio-language datasets. The proposed ETTA generation model is trained across a wide range of diverse audio datasets and achieves competitive scores on both AudioCaps and MusicCaps, demonstrating state-of-the-art performance in both text-to-audio and text-to-music tasks.\n\nExternal ablation studies investigate the system's performance across different model sizes and training/sampling strategies. Subjective evaluations highlight significant improvements in the proposed ETTA system over baseline models, proving its enhanced ability to generate complex and imaginative captions.\n\nOverall, this paper presents interesting work in the field of audio generation. The authors first introduce a large-scale dataset and then present a state-of-the-art generation system trained using this dataset. Various experiments demonstrate the effectiveness of different methods or modules within the system, concluding with an analysis of the ETTA system's limitations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a 1.35M high-quality audio-caption dataset utilizing the state-of-the-art (SoTA) audio-language model, Audio Flamingo, referred to as AF-Synthetic. Additionally, the paper introduces a text-to-audio system based on the Latent Diffusion Model (LDM) with a Diffusion Transformer (DiT) backbone. Experimental results demonstrate that the proposed Elucidated Text-To-Audio (ETTA) system achieves SoTA performance across multiple metrics on both the AudioCaps and MusicCaps datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors claim the effectiveness of the proposed dataset by demonstrating the SoTA performance of the ETTA system trained on AF-Synthetic. However, the paper lacks sufficient experiments directly comparing the dataset's contribution, such as showing how models like Tango or AudioLDM perform when trained with AF-Synthetic.\n2. All the metrics used for evaluating the text-to-audio system are objective, which is generally sufficient but may not always reflect real-world performance. Therefore, including subjective evaluations like the Mean Opinion Score (MOS) would be beneficial, particularly for comparing models that achieve top performance across some metrics.\n3. The authors attempt to compare the performance between different baseline models. While ablation studies already demonstrate the effectiveness of the proposed ETTA strategies, the paper lacks experiments where the proposed model is trained on the same dataset (e.g., AudioCaps) to clearly illustrate the improvements contributed solely by the system itself.\n4. The ETTA model mainly builds on the backbone of the Stable Audio system. Beyond the training and sampling strategies (Flow Matching and ODE solvers, which are standard improvements in current LDM systems), the key enhancement appears to be the use of Adaptive Layer Normalization (AdaLN) within the DiT structure and the T5-base model for text embedding. These techniques are already implemented in the original DiT paper [1] and other baseline models. As a result, the model seems more like an engineering application with limited novel contributions.\n\nOverall, this is an interesting paper with great effort, I am willing to change the score if the author can solve the questions and fulfil the following experiments:\n1. Train baseline models like Tango or AudioLDM on AF-Synthetic and compare to their original performance.\n2. Train ETTA on other datasets like AudioCaps or TangoPromptBank and compare performance to when trained on AF-Synthetic. These experiments would help separate the contributions of the model architecture from the dataset.\n3. Develop human evaluations on a subset of generated samples, comparing ETTA to top baseline models, such as Mean Opinion Score (MOS) or other equivalent subjective ratings on audio quality, text relevance and so on.\n4. More clearly articulate the novel aspects of their model architecture compared to existing work like DiT and Stable Audio. \n\n[1]. Peebles, William, and Saining Xie. \"Scalable diffusion models with transformers.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We elucidate the design space of text-to-audio and present ETTA with state-of-the-art result and improved abilities to generate creative audio." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024elucidating,\ntitle={Elucidating the Design Space of Text-to-Audio Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xmgvF0sLIn},\nnote={under review}\n}" }, "abstract": { "value": "Recent years have seen significant progress in Text-To-Audio (TTA) synthesis, enabling users to enrich their creative workflows with synthetic audio generated from natural language prompts. Despite this progress, the effects of data, model architecture, training objective functions, and sampling strategies on target benchmarks are not well understood. With the purpose of providing a holistic understanding of the design space of TTA models, we setup a large-scale empirical experiment focused on diffusion and flow matching models. Our contributions include: 1) AF-Synthetic, a large dataset of high quality synthetic captions obtained from an audio understanding model; 2) a systematic comparison of different architectural, training, and inference design choices for TTA models; 3) an analysis of sampling methods and their Pareto curves with respect to generation quality and inference speed. We leverage the knowledge obtained from this extensive analysis to propose our best model dubbed Elucidated Text-To-Audio (ETTA). When evaluated on AudioCaps and MusicCaps, ETTA provides improvements over the baselines trained on publicly available data, while being competitive with models trained on proprietary data. Finally, we show ETTA's improved ability to generate creative audio following complex and imaginative captions — a task that is more challenging than current 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": [ "audio generation", "text-to-audio", "synthetic data", "diffusion", "flow matching" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/79f2fd64cae74742cebec0806882a9b061294cb8.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": "Elucidating the Design Space of Text-to-Audio Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]