Datasets:

babytreecc

/

ICLR2025review

like 7

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper is well-organized and easy to read. \n2. The proposed method presents a reasonable approach to improve the reasoning performance of LLMs by combining weak and strong LLMs. \n3. The approach is practical and has the potential for broad application.\n4. The experimental results reveal that the proposed method significantly enhances performance on various reasoning tasks compared to both the weak and strong LLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a collaborative framework that integrates a specialized weak model with a general strong model to enhance the reasoning performance of LLMs. In this framework, the weak model generates detailed initial drafts and background information tailored to specific domains, while the strong model refines and enhances these drafts utilizing its advanced reasoning capabilities. A feedback loop is implemented to fine-tune the weak model based on the preferences of the strong model, fostering an adaptive and synergistic relationship. Experimental results indicate that the proposed method outperforms both the basic weak and strong LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The technical innovations introduced in this paper appear to be somewhat limited, as the concept of leveraging both weak and strong LLMs has been extensively explored in prior research, including works such as “Your Weak LLM is Secretly a Strong Teacher for Alignment” and “Synthesizing Text-to-SQL Data from Weak and Strong LLMs.”\n2. A more comprehensive evaluation would enhance the study by comparing the proposed method against a more comprehensive array of advanced baseline models. Currently, the comparisons are limited to several basic baselines. Incorporating more sophisticated weak-strong collaboration methods and state-of-the-art techniques would provide stronger validation of the proposed method's effectiveness.\n3. To demonstrate the versatility of the proposed method, it would be advantageous to conduct experiments using different open-source LLMs of varying sizes." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. Why does the main experiment use the strong model GPT-3.5-Turbo for the ethical dataset, instead of maintaining consistency with other domains by using GPT-4?\n2. Why was the learning rate set to 1.41e-5? Intuitively, this seems like an uncommon number, was it determined by searching different learning rates?\n3. Typo: There is inconsistent formatting of the name 'Llama-3' throughout the paper. For example, it is written as \"LLama-3-8B\" in Table 1, \"LLaMA3-8B\" on line 481, and \"Llama3-8B\" on line 381.\n4. In the main experiment, were the results for Llama-3-8B obtained using a few-shot setting? The IfQA paper used two evaluation methods: a supervised setting and a few-shot setting. If the few-shot setting was not used, intuitively, the output form of the model might not be controllable. Similarly, when using Llama-3-70B and Llama-2-70B as strong models for evaluation, were few-shot settings adopted?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The research topic regarding the collaborative interaction between a specialized weak model and a general strong model is very important" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a weak-strong collaboration mode, in which a weak model fine-tuned on domain-specific datasets first generates drafts, while a strong model refines them. By utilizing feedback from the strong model to perform preference optimization, the performance of the weak model is further improved." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of novelty: The concept of weak-strong collaboration explored in the paper, essentially using feedback to correct large language models, is not a novel idea and has already been extensively researched [1]. The two collaboration strategies: standard refinement bears strong resemblances to prior works [2], and preference enhancement that leverages DPO for inconsistency alignment is also not new. It’s just old wine in a new bottle, wrapping up a story of the interaction between a specialized weak model and a general strong model.\n2. The datasets used in the experiments lack representativeness: (1) Domain selection: In addition to the three domains selected, more typical mathematical reasoning datasets should be included, such as GSM8k and MATH, which have been widely used in previous model collaboration work [3][4]. (2) Dataset selection: For the medical domain, the choice of MedMCQA, which is limited to a multiple-choice format, is too narrow. There should be more focus on broader and more practical long-form QA datasets like K-QA [5].\n3. Lack of baselines for model collaboration/ensemble: The main experiment mainly compares the proposed collaboration approach with only weak or strong model strategies, omitting critical baseline comparisons, such as self-refine [6], and other ensemble strategies such as multi-model debate [7], self-consistency.\n4. Some specific experimental settings were not clearly stated, for example, the retrieval knowledge base used by FLARE in three selected domains was not mentioned\n5. The Preference Enhancement Interaction lacks generalizability, as the acquisition of preference pairs is specific to a strong model. This specificity might limit the effectiveness and generalization when collaborating with different strong models.\n6. Questioning the experimental results: The results presented in Table 1 raise concerns about the necessity of weak-strong collaboration. In the Counterfactual and Medicine domains, weak models without SFT are much stronger than strong models, e.g., Llama-3-8b (68.57) vs. GPT-3.5-turbo (22.62). Similarly, in the Ethics domain, the performances were comparable. If weak models can perform on par with or better than strong models, is the use of weak-strong collaboration justified? Does the motivation for using a stronger model to assist weaker ones still stand?\n7. Concerns about the high costs for strong models compared to minor performance improvements in weak models: The proposed collaborative approach, compared to merely using a weak model for SFT, only brought minor improvements (shown in Table 1). However, this process requires the strong model to refine and evaluate the output of the weak model, which brings significant API costs.\n8. Lack of in-depth analysis of the improvements brought by the cooperation strategy, for example, the paper does not specify in which aspects the strong model has improved the weak model, nor does it detail the types and percentages of errors detected in the weak model by the strong model. Furthermore, the frequency with which the weak model adopts feedback from the strong model is not discussed. More comprehensive case studies are needed to understand these dynamics fully, rather than merely providing a superficial overview.\n\n[1] Automatically Correcting Large Language Models: Surveying the Landscape of Diverse Automated Correction Strategies. Pan et al. TACL 2024\n\n[2] Small Models are Valuable Plug-ins for Large Language Models. Xu et al. ACL 2024 Findings\n\n[3] Learning to Decode Collaboratively with Multiple Language Models. Shen et al. ACL 2024\n\n[4] Ensemble learning for heterogeneous large language models with deep parallel collaboration. Huang et al. NeurIPS 2024\n\n[5] K-QA: A Real-World Medical Q&A Benchmark. Manes et al. BioNLP 2024\n\n[6] Self-Refine: Iterative Refinement with Self-Feedback. Madaan et al. NeurIPS 2023\n\n[7] Improving Factuality and Reasoning in Language Models through Multiagent Debate. Du et al. arXiv 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": 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) Do the authors have a vision on how the proposed CoWEST is different from the LLM cascade methods such as CAT[1]?\n\nb) How the sampling will impact on the performance?\n\nc) How's the evaluator's quality? Have the author consider using logits or a trainable method (e.g., a MLP) to serve as the evaluator? Since self-critique sometimes may results LLM is always more confident with the content generated by itself, while logits or trainable methods can be more fair.\n\nd) In 127-129, using \\citep()\n\ne) In line 207, \"referred to as\"? there are more, please check the writing for readibility.\n\n**Reference**\n\n[1] Cascade-Aware Training of Language Models." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "a) The proposed CoWEST shows remarkable improvements over SOTA methods such as RAG-based methods.\n\nb) The interaction design between the weak LLM and the strong LLM is interesting compared to existing methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work focuses on the challenge that current large language models (LLMs) often struggle with specific domains or downstream tasks. To tackle this, we propose a collaborative framework, CoWEST, which integrates a weak LLM with a strong LLM. In CoWEST, the weak LLM is first fine-tuned for a specific domain or task, and then the strong LLM’s general capabilities are leveraged to enhance the fine-tuned weak LLM’s output. Additionally, a preference tuning paradigm is used to evaluate the collaborative output against that of independent models. Extensive experiments demonstrate the effectiveness of the proposed CoWEST framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "a) The sampling method for preference tuning is not clear, lack the the sampling statistics (e.g., sample distribution, average sample size etc.).\n\nb) The evaluator is like a self-critique and more evaluator quality details such as score criteria, comparisons with human evaluation etc. should be included.\n\nc) Minor writing issues." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 LLM abbreviation of L121 is repeatedly defined.\n2. The reference form of L127-L128.\n3. What if the weak model and the strong model are the same?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Considering the challenges of real-world scenarios, the issues that this paper focuses on are necessary. Strong-weak model collaboration is one of the promising directions.\n2. \"Using weak models for domain adaptation and then strong models for reasoning\" can be seen as a RAG method in which the weak model after domain adaptation generates evidence context, and then the strong model uses the evidence in this domain for reasoning. This may actually increase the amount of information for strong model reasoning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a paradigm of weak-strong model cooperation, in which the large model with stronger reasoning ability is responsible for reasoning with the background knowledge and drafts generated by the small model. Furthermore, the authors propose to fine-tune the weak model to adapt it to the preferences of the strong model to achieve so-called adaptive cooperation. The proposed method achieves the improvement in the F1 performance in three datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors explore the framework of weak-strong model cooperation, but I think it still needs to be better explained, that is, how the proposed feedback loop and interaction strategy go beyond the static cooperation method. I think the claims of L111-L115 are a bit far-fetched (considering that the weak model still reasons first during reasoning, and then the strong model uses the output of the weak model for reasoning). In addition, the writing needs to be improved, there are many small errors, and some claims are confusing to readers.\n2. The paper focuses on improvements such as performance scores (F1), but lacks qualitative analysis of how the models collaborate in real-world scenarios. In fact, I am still confused about the example in Figure 6, how to show the role of the strong model? There is also limited information about how the feedback loop between weak and strong models affects the interpretability or usability of the output in complex reasoning tasks, but it is one of the important contributions emphasized by the authors. I suggest that the authors add some qualitative examples that can show how collaboration improves responses (in terms of factual accuracy, reasoning chain, or coherence).\n3. The paper acknowledges the computational cost of fine-tuning large models, but the authors do not provide much insight into the scalability of COWEST when it is extended to larger weak models or more complex tasks, such as multi-hop questions that exploit the strong reasoning capabilities of large models. In addition, the resource impact of the feedback loop (e.g., computational overhead) is not discussed in depth, where the two inferences in the Inference stage increase the computational cost.\n4. The authors should conduct comparative experiments on transferring domain knowledge to strong models in the case of longer contexts." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a synergistic collaboration framework where a smaller, specialized model and a larger, general-purpose model work together, using preference finetuning to enhance problem-solving in specialized tasks." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024synergistic,\ntitle={Synergistic Weak-Strong Collaboration by Aligning Preferences},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3iJ7eSj2rE},\nnote={under review}\n}" }, "abstract": { "value": "Current Large Language Models (LLMs) demonstrate exceptional general reasoning and problem-solving abilities but often struggle with specialized tasks or domains requiring proprietary information due to their generalized training and size constraints. Fine-tuning large models for every specific domain is impractical because of inaccessibility to black-box model parameters and high computational costs. We explore a solution to this challenge: can a collaborative framework between a specialized weak model and a general strong model effectively extend LLMs' capabilities to niche but critical tasks? We propose a dynamic interaction where the weak model, tailored to specific domains, generates detailed initial drafts and background information, while the strong model refines and enhances these drafts using its advanced reasoning skills. To optimize this collaboration, we introduce a feedback loop by fine-tuning the weak model based on the strong model's preferences, fostering an adaptive and synergistic relationship. We validate our framework through experiments on three datasets. We find that the collaboration significantly outperforms each model alone by leveraging complementary strengths. Moreover, fine-tuning the weak model with strong model's preference further enhances overall performance.\nOur collaborative approach achieves an average F1 score improvement of 3.24\\% over the weak model alone and 12.17\\% over the strong model alone across all 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": [ "Weak-Strong Model Collaboration", "Preferences Tuning", "Large Language Model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/84abf7a1736735061ba70240d2871355b6ca1421.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": "Synergistic Weak-Strong Collaboration by Aligning Preferences" }, "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. Could you provide unadjusted(A) in Table 4?\n2. Besides average error, could you draw a histogram which can better validate the claims?\n3. How do you empirically decide when the weight difference is going to affect the conclusion? E.g., Adam shows larger weight difference, while ImageNet shows larger weight difference. It does not seem clear to me they all supports the claim that STE works the same as other gradient estimators on various settings." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The claim is strong that other gradient estimators works similar as STE in QAT.\n2. Experiments show that the weight difference is small to support the claim." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors theoretically analyze the weight difference in QAT when trained with different gradient estimators. Under certain conditions, the authors show that the weight difference is small which means that there's no need to try another gradient estimator other than STE. Empirical results show that the weight difference is small when adopting the proposed weight initialization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The mirror room story does not appear closely connected to the theoretical analysis.\n2. Assumption 5.1.1 violates Figure 1 where the gradient could be zero.\n3. From Table 4, Adam leads to larger weight difference.\n4. For more complicated task like ImageNet, the weight difference is much larger than MNIST." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In Eq. (5), why the $E^{(t)}$ is defined differently for SGD and Adam?\n\n2. Line 200, \"Most multi-bit gradient estimators proposed in the literature are cyclical.\" Can you specify exactly which estimators are cyclical?\n\n3. Typos in Eq. (9)-(12), $f_{STE}^{(t)}$ should be $\\nabla f_{STE}^{(t)}$.\n\n4. In assumption 5.1.1, should $| \\hat{Q}^\\prime (w) |$ be $\\hat{Q}^\\prime (w)$ without the absolute value?\n\n5. Does Table 4 suggest that more than 95% quantized weights from $\\hat{Q}$-net and STE-net are *identical*? This finding seems counterintuitive." }, "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 overall well presented.\n2. The concept of mirror effect is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies behavior of gradient estimators, including straight through\nestimator (STE), for weight quantization. It is shown that a large class of\nweight gradient estimators is approximately equivalent to the STE during training using SGD and Adam." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "A primary concern is that the key claims and several major concepts lack mathematical rigor. Additionally, the main theoretical results provided are too limited to substantiate the claims:\n\n1. Contribution 1 states that '... all nonzero weight gradient estimators lead to approximately equivalent weight movement for non-adaptive learning rate optimizers ...'. However, the term 'approximately equivalent weight movement' lacks a precise mathematical definition. It would be helpful to formalize this concept, perhaps by specifying the conditions under which these movements are considered 'approximately equivalent.' \n\n2. According to Section 6.2, 'approximately equivalent weight movement' appears to refer to high 'Quantized Weight Agreement' or a small 'Normalized Weight Alignment Error ($\\bar{E}$)'. Again, these metrics require explicit mathematical expressions for each. Additionally, this interpretation is not fully supported by the main theoretical results (e.g., Theorem 5.1), which only derive the increment in weight alignment error between two consecutive iterations, rather than a direct measure of agreement or alignment over the entire optimization trajectory. \n\n3. For the error bounds in Eq. (6) and (7), there is insufficient justification for why the gradient error terms should be small, nor any clear indication of how small these terms are. It is insufficient to merely claim that a term is 'small' and then disregard it. These errors could accumulate significantly over iterations, potentially undermining the main conclusions.\n\n4. The use of mathematical notation is poor, which possibly lead to incorrect derivations. For example:\n - The Euclidean norm should be denoted by $\\|| \\cdot \\||$ rather than $\\| \\cdot \\|$ as in Eq. (5)-(12) and other instances.\n - It should be explicitly stated that $Q$ and $M$ are applied *element-wise* to the weight vector. Additionally, it would be preferable to use bold letters to represent vectors and to distinguish them from scalars.\n - In Eq. (10), you have three vectors, $\\nabla f_{Q}^{(t)}$, $\\hat{Q}^\\prime$, $M^\\prime$, how are they multiplied together? The manner in which they are multiplied together is unclear. Furthermore, the residual term in Eq. (10) should not be a scalar $O(\\eta^2)$, but rather a vector. I believe that the second-order error term also depends on the *model size*, i.e., the dimension of $w$.\n\n5. The experiments is only conducted for one instance of $\\hat{Q}$ (HTGE Pei et al.), which is insufficient. Additionally, the expression of $\\hat{Q}$ from HTGE is missing.\n\nMinor comments:\n\n1. A key reference on the theoretical analysis of STE is missing, specifically: *Yin et al., Understanding Straight-Through Estimator in Training Activation Quantized Neural Networks, ICLR2019.*" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 weaknesses" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The theoretical insights are interesting, unexpected and (to my knowledge) novel. They offer better understanding and insight into how gradient estimators work, which appeals to me.\n- The paper is generally well written and easy to read. I appreciate how the authors lead their result with an intuitive explanation and illustrative graphic. This makes the following theory much easier to intuit.\n- The experiments shown in the paper provide good evidence for the theoretical results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a theoretical analysis of gradient estimators used in quantization-aware training. The authors show that in the case of quantized weights (but full precision activations) many extant gradient estimators for the quantization operation are approximately equivalent, if the learning rate and weight initialization are adjusted, and the learning rate is small. They then verify empirically their theoretical results, on image classification benchmarks, demonstrating that models trained with different gradient estimators indeed show high weight agreement and similar accuracies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major\n\n1. The claims relating to practical impact feel overstated (\"practitioners can now confidently choose the STE\"). The problem setting that the authors explore (full precision activations, quantized weights, uniform fixed point quantization, small learning rate) is rather specific, and practitioners may be interested in quantized activations, or low-precision floating point or larger learning rates etc. I would prefer if the authors tempered their claims.\n2. The experiments, although they demonstrate the theory well, are limited. They do not show finetuning from a pretrained full precision checkpoint, as is common for QAT in practice (I would expect this setting to match well with the theory since 1. QAT finetuning is done with a lower learning rate typically and 2. the gradient norm is likely to be low after initializing from a pretrained model). They do not show results other than 2-bit weight quantization even though they say results are similar. They do not show the practical limits of their theory, e.g. how weight alignment degrades/evolves over training or how much the learning rate needs to be increased for the error terms to start having a large impact.\n\nMinor\n1. Presentation could be improved in a number of ways. \n 1. Use of booktabs for tables. Place all table captions above the tables.\n 1. Tables are hard to parse when skimming -- would benefit from more descriptive captions/grouping table 3 with 4\n 2. All quotation marks are incorrectly rendered by LaTeX.\n 2. Fig. 3 would look better with the bins.\n2. The choice of training recipes are not explained -- it is unclear why the first 10 epochs are done using the same gradient estimator. Is it because the gradient norm is too high at the start of training resulting in the weights quickly diverging?\n3. I think the point made in line 305 should be made more prominent. I think it is quite important that the reader is made aware that the gradient error is small/zero since Q-net and STE net will quantize to similar weights." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Lines 400-404: Why was the initial 10% of training for the ImageNet-ResNet setup kept identical? What would the results be without this measure?\n\nLines 527-528: Regarding the statement that high learning rates might be the reason the equivalence is not observed in other studies, the authors write, \"we expect that this counter-argument will not stand the test of time, since by our main results, the higher learning rate masks the fact that models with novel $\\hat{Q}$ and the STE are still approximating the same process.\" Could you elaborate on what you mean by \"will not stand the test of time\" given that equations (6) and (7) indicate learning-rate-squared errors? It seems that higher learning rates would increase these errors. Why should these differences not enable advantages for novel $\\hat{Q}$?\n\n\nLines 530-539: Can the authors comment on the potential implications of their results for the studies mentioned in the last paragraph, which propose additional innovations alongside novel gradient estimators?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Equations (6) and (7) provide a rigorous theoretical contribution regarding the small difference in weight movement for different gradient estimators.\n\nThe presentation and writing style is very clear with helpful intuition such as the analogy of the \"funhouse mirror\". Additionally, the learning rate tweak in the experiment provides a practical comparison for the magnitude of the differences." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors examine how different gradient estimators affect quantization-aware training. They theoretically demonstrate that, in the limit of small learning rates and with minor adjustments to the initialization and the learning rate magnitude, most gradient estimators yield equivalent weight movements. Consequently, they suggest that the Straight-Through Estimator, treating the gradient as if no quantization occurred in the backward pass, performs as well as any other more sophisticated alternative. Their theoretical claims are complemented by an empirical investigation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors themselves acknowledge in Section 8 that many publications introduce more than just a new gradient estimator, raising questions about the broader practical impact of their findings. Novel gradient estimators are often proposed with other techniques to enable any benefits. As a result, the findings may have limited applicability beyond specific configurations.\n\nIt is somewhat difficult to draw clear conclusions about practical applications from the experiments. While the learning rate tweak provides a useful comparison, the differences in Tables 4 and 5 are challenging to interpret, particularly without comparisons across different initializations. A convincing experiment could be, to find a custom gradient estimator in the literature, which has been shown to improve validation accuracy over STE, replicate the results and then demonstrate that both perform equally with proper adjustments. The authors do not provide such a comparison, raising questions about whether the custom estimator was applied correctly in their experiments or if its potential advantages were overlooked.\n\n\nMinor comments:\n\nline 245: Typo: $Q(w_{\\hat{Q}}^{(t)})=Q(w_{\\hat{Q}}^{(t)})$\n\nlines 329-340: Equations (8) to (12): \"$\\nabla$\" missing before $f^{(t)}_{STE}$\n\nlines 502-504: Missing figure number for Figure 3" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We prove and empirically demonstrate that custom gradient estimators are equivalent to the straight-through estimator for quantized neural network optimization." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024custom,\ntitle={Custom Gradient Estimators are Straight-Through Estimators in Disguise},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3j72egd8q1},\nnote={under review}\n}" }, "abstract": { "value": "Quantization-aware training comes with a fundamental challenge: the derivative of quantization functions such as rounding are zero almost everywhere and nonexistent elsewhere. Various differentiable approximations of quantization functions have been proposed to address this issue. In this paper, we prove that a large class of weight gradient estimators is approximately equivalent with the straight through estimator (STE). Specifically, after swapping in the STE and adjusting both the weight initialization and the learning rate in SGD, the model will train in almost exactly the same way as it did with the original gradient estimator. Moreover, we show that for adaptive learning rate algorithms like Adam, the same result can be seen without any modifications to the weight initialization and learning rate. These results reduce the burden of hyperparameter tuning for practitioners of QAT, as they can now confidently choose the STE for gradient estimation and ignore more complex gradient estimators. We experimentally show that these results hold for both a small convolutional model trained on the MNIST dataset and for a ResNet50 model trained on ImageNet." }, "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": [ "quantization", "deep learning", "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/c745d2831deb070851b86757bc9435cdd4da1dfd.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": "Custom Gradient Estimators are Straight-Through Estimators in Disguise" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The model (entanglement entropy) is giving 100% test accuracy on the MC dataset in Table 1. Is there any explanation for this?\nIn table 1, in the QSANN model, when only CLS token was used the test accuracy dropped from 100 to 56%. What could have been the possible reason?\nWhy was the comparison only with QSANN model?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Novelty: The paper proposes entanglement based attention and novel methodology for computing attention. Three measures of entanglement(Entanglement entropy, SWAP test and concurrence) are used for measuring the entanglement between the queries and the value vectors. The proposed method is evaluated in both classical and quantum datasets. The proposed model was compared with scaled-dot-product attention and another quantum attention method QSANN model for various vision and NLP datasets. \nSoundness: The paper is well written. The paper claimed that the quantum entanglement based attention is having a better generalization gap across all datasets. The experimental results of the paper supports this claim. Experiments were conducted extensively on various NLP and vision datasets with clear figures and tables. \nSignificance: With limited number of works done on quantum computing w.r.t Transformers, this work has relevance and future applications\nRelation to prior works: The previous related works are discussed comprehensively in this paper. \nReproducibility: The authors have provided the source code of the experiments" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper incorporates quantum entanglement into the attention mechanism of a Transformer encoder by using a measure of entanglement to compute the attention matrix." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Clarity: In section 4.3, the authors have defined various methods of entanglement. But how these methods are applied in regard to self attention is not clear. The authors have stated the methods used for computing key, query quantum states and attentions, but exactly how it is done on mathematical terms is not defined. A mathematical expression of the measures of entanglement for computing the attention, would have made it clearer. In Figure 5, there is no explanation of how the parameterized quantum circuit is generated.\nThe transformer model consists of only two sequential attention layers. Eventhough the performance of the model with varying data sizes were studied, the performance of the model with varying model sizes have not been studied. Is the model underperforming on larger datasets because of smaller model size?\nA qualitative analysis of the behavior of attention maps and the interactions between various positions, if included, would have given a better understanding 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": 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 weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is easy to follow.\n- Introducing quantum in classical computers is meaningful and interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the potential of quantum entanglement for attention in Transformers. Authors use the entanglement of quantum states as a co-relation criterion for the Attention layer in the Transformer. The method is evaluated on both language tasks, vision tasks, and quantum datasets. Experiments show the potential of quantum entanglement attention to have better generalization ability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Dataset used for experiments is too small. The transformer is a large-scale model that requests large-scale data to learn meaningful features. Besides, quantum entanglement attention shows its outperformance from Figure 2 when the size of the dataset is less than 1000, which is not practical.\n- The paper claims that quantum entanglement attention has better generalization ability, which is the difference between train and test accuracy. However, as stated above, the transformer is a large-scale model that requests large-scale data, which means the transformer would easily overfit in small datasets. This has resulted in poor accuracy of transformers on small datasets. For example, in CV tasks, transformers generally require 200-300 epochs on ImageNet to match the accuracy of CNN (which also requires the corresponding number of epochs), while in CIFAR datasets, transformers require 7200 epochs to match the accuracy of CNN, which only requires 200 epochs.\n- It's vital to visualize or analyze the attention of quantum and classical. If the elements of quantum attention matrix is all same, the transformer treats all token equally, which means the transformer model is about to degenerate into an MLP which could generalize better than transformer when dataset is small. It's hard to conclude that the benefit is coming from quantum entanglement operation.\n\n\n- The details of the transformer model should have a description. What's the dimension of the transformer? How many blocks do you stack? \n- The details of datasets for experiments should have a more clear description, e.g., MC and RP datasets.\n- The details of training should have a description. How many epochs? How do you train the transformer?\n- The illustrations in the paper should be improved. The current illustration is confusing and the content is not clear, especially Fig 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": 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. Given that current quantum hardware is noise-prone, how do the authors envision the entanglement-based attention mechanism performing in noisy conditions? Are there plans to test this model in simulated noisy environments or on NISQ devices to verify its stability?\n\n2. Although this paper proposes using entanglement entropy for a quantum implementation of the attention mechanism, it lacks an in-depth analysis of the theoretical foundation and effectiveness of this approach. It is recommended that the authors enhance the theoretical exploration of the role of quantum entanglement in the attention mechanism, especially by explaining from a quantum information perspective why it performs exceptionally well on certain tasks. Additionally, a discussion on the theoretical basis and advantages of Hilbert space (related to the Quantum Feature Map, QFM) and the parameter efficiency of quantum models compared to classical models would be beneficial.\n\n3. Can the authors include a table comparing the parameters and architectures of the quantum and classical models to clarify any computational trade-offs? This would help readers understand the efficiency and scalability implications of the proposed approach.\n\n4. Have the authors considered evaluating the entanglement-based attention mechanism against other classical models, such as MLPs, to provide a broader baseline comparison? This could clarify whether the quantum approach offers unique benefits over simpler classical architectures.\n\n5. Several relevant works on quantum self-attention and quantum Transformer models are missing from the current paper. Could the authors consider adding the following references to provide additional context and background on prior work in this area?\n\nShi, Shangshang, et al. \"A natural NISQ model of quantum self-attention mechanism.\" *arXiv preprint arXiv:2305.15680* (2023).\nShi, Jinjing, et al. \"QSAN: A near-term achievable quantum self-attention network.\" *arXiv preprint arXiv:2207.07563* (2022).\nDi Sipio, Riccardo, et al. \"The dawn of quantum natural language processing.\" *ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)*. IEEE, 2022." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper presents an innovative approach that integrates quantum entanglement into the Transformer’s attention mechanism, using entanglement entropy to calculate attention coefficients. The method demonstrates improved generalization and reduced overfitting on small classical and quantum-generated datasets, providing a robust evaluation against classical and other quantum attention models. This work contributes to quantum-classical hybrid models, showing potential in data-limited applications and opening avenues for further exploration in quantum-enhanced machine learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an approach that integrates quantum entanglement into the attention mechanism of Transformer models, proposing an entanglement-based attention layer. By encoding query and key vectors as quantum states, entangling them through a parameterized quantum circuit (PQC), and using entanglement entropy to calculate attention coefficients, the method aims to enhance Transformer performance on specific tasks. Experimental results demonstrate that this quantum-based attention layer outperforms classical attention on smaller classical datasets and quantum-generated datasets, showing a superior generalization gap and reduced tendency to overfit. The work provides valuable insights into leveraging quantum properties within classical machine learning frameworks, especially for data-limited applications, and contributes to the emerging field of quantum-inspired hybrid models. This research lays the groundwork for further exploration of quantum resources as subroutines in machine learning models, particularly Transformers, offering new possibilities for performance improvements in specialized scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper does not address the impact of noise on the proposed quantum model, which is crucial given the current limitations of noisy intermediate-scale quantum (NISQ) hardware. Quantum systems are inherently sensitive to noise, and without examining how noise affects the model’s performance, it is unclear whether the proposed entanglement-based attention mechanism can be effectively implemented on real hardware. To improve the practical relevance, I recommend adding noise simulations or discussing how hardware noise might affect entanglement performance in attention mechanisms, which would make the work more applicable to real-world quantum devices.\n\n2. Although the authors introduce entanglement entropy for the attention mechanism, the paper lacks a rigorous theoretical foundation to explain why entanglement specifically improves generalization in small-data scenarios. There is little discussion on the advantages of Hilbert space representations (related to quantum feature mapping, QFM) or why quantum entanglement should provide performance benefits over classical models, especially from a quantum information perspective. I recommend that future work include a deeper theoretical exploration of the role of quantum entanglement in attention mechanisms. This could involve discussing Hilbert space properties, parameter efficiency, and the specific benefits of quantum versus classical models, to clarify the approach’s underlying strengths and limitations.\n\n3. The paper does not provide a detailed comparison of parameters between the quantum and classical models, which could help clarify the computational trade-offs of the proposed approach. Including a summary table of model configurations and hyperparameters would enhance transparency, allowing readers to better understand the computational costs associated with each method.\n\n4. The paper only compares its entanglement-based attention mechanism with a simplified Transformer model. It would be helpful to compare against other classical models, such as MLPs, to demonstrate the quantum model’s relative performance more comprehensively.\n\n5. The paper does not reference several recent works that are highly relevant to quantum self-attention and Transformer models. Key papers, such as Shi et al. (2023), Shi et al. (2022), and Di Sipio et al. (2022), explore similar mechanisms and should be cited for completeness. These references would provide additional context and underscore where this work contributes new insights to the existing literature." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "A few concerning points are listed as follows, and I hope the authors could clarify these before I change my mind in this paper's decision.\n\n1. How about the scalability/complexity of this model, or, what is the scaling with respect to the vector size?\n\n2. Can it show more concrete relations between quantum entanglement and enhancement, to evaluate whether stronger entanglement leads to stronger model performance?\n\n3. If the quantum circuit size becomes larger, will the quantum model keep its advantage on classical datasets as in Figure 2?\n\n4. There have been existing papers differently adapting attention mechanisms, such as Cherrat and Kerenidis (Quantum 2024), Ren-Xin Zhao (IEEE Trans. Pattern Anal. Mach. Intell. 2024), and Khatri (arXiv:2406.04305). What is your advantage or novelty compared to their works?\n\n5. Could you please provide a resource analysis including time complexity, qubit number, or number of measurements...\n\n6. Please clarify several concepts, including \"learning rate scheduler\", \"data reuploading layers\" In Appendix A. \n\n7. What is the number of trainable parameters in this model?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The attention mechanism is a cornerstone of modern machine learning, and the potential enhancements offered by quantum computing are compelling.\n2. Exploring the synergy between quantum computing capabilities and entanglement is valuable, and this paper provides promising numerical evidence.\n3. The quantum circuits are relatively simple and could likely be implemented in near-term quantum computers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an entanglement-based attention layer integrated into a classical Transformer, where the traditional dot product operation between query and key vector pairs is replaced by a quantum feature map circuit and an entanglement measurement. Leveraging quantum circuits introduces quantum entanglement into the attention mechanism. Numerical experiments indicate that entanglement entropy outperforms other entanglement metrics, and the entanglement-based layer demonstrates advantages over its classical counterpart in classification tasks within vision and NLP domains. For both quantum-generated and classical datasets, the model shows improvements in classification accuracy and a reduced generalization gap." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Circuit size: The model proposed in this paper involves a simulated quantum circuit with only 6 qubits; meanwhile, the quantum circuit in Figure 5 is too simple, introducing only local entanglement. Experiments with more qubits (such as 10~20) could significantly improve the soundness of the paper.\n\n2. Motivations: The introduction talks a lot about well-known concepts, but the motivations or insights to replace the dot product with entanglement in the attention mechanism are not sufficiently discussed.\n\n3. Efficiency: My understanding is that the entanglement measurement needs to be performed as many times as the number of attention coefficient matrix elements, which is quite inefficient. The algorithmic/time complexity should be explicitly discussed in this paper.\n\n4. Missing details: There is no specific explanation for the query state and key state; are they row or column vectors of Q and K? Detailed circuit implementations should be provided. Also, the complexities of the different entanglement measurements are not compared in Section 4.3.\n\n5. Concerns about model performance. The numerical results in Figure 2 suggest that the classical model performs better as the sample size increases, potentially diminishing the practical value of this model. I wonder if the small size of the quantum circuit limits performance when using large sample sizes. The training curves are not stable, which could be improved by adjusting hyperparameters. Is there any reason behind the instability of the training curve?\n\n6. Citation mistakes. The introduction refers to 'Systematic benchmarking of existing quantum approaches suggests that entanglement may not play a significant role' but cites no paper. In Section 4.1, QFM methods are reviewed but not proposed by Khan et al. (2024). In Section 4.3, Quantum State Tomography lacks citation, where a typo occurs (FST)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We study quantum entanglement entropy in attention models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024quantum,\ntitle={Quantum entanglement for attention models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3jRzJVf3OQ},\nnote={under review}\n}" }, "abstract": { "value": "Attention mechanisms in deep learning establish relationships between different positions within a sequence, enabling models like Transformers to generate effective outputs by focusing on relevant input segments and their relations. The performance of Transformers is highly dependent on the chosen attention mechanism, with various approaches balancing trade-offs between computational cost, memory efficiency, and generalization ability based on the task.\n\nQuantum machine learning models possess the potential to outperform their classical counterparts in specialized settings. This makes exploring the benefits of quantum resources within classical machine learning models a promising research direction. The role of entanglement in quantum machine learning, whether in fully quantum or as subroutines in classical-quantum hybrid models, remains poorly understood. In this work, we investigate whether quantum entanglement, when used as a resource, can improve the performance of the attention layer in Transformers.\nWe introduce an entanglement-based attention layer within a classical Transformer architecture and numerically identify scenarios where this hybrid approach proves advantageous. Our experiments on simple standard classification tasks in both vision and NLP domains reveal that the entanglement-based attention layer outperforms classical attention, showing superior generalization on quantum-generated datasets and in settings with limited training data for classical datasets. Additionally, it demonstrates a smaller generalization gap across all tested datasets. Our work contributes towards exploring the power of quantum resources as a subroutine in the classical-quantum hybrid setting to further enhance classical 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": [ "Attention models", "Quantum entanglement", "Transformers" ] }, "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/c06efeae808fb8191f327d232a0b3b4a195d9275.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": "Quantum entanglement for attention 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": "Please see above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. MathScape contains 1325 high-quality human-collected real multimodal mathematical problems. \n2. Authors conduct an analysis of the relationship between answer length and performance, which is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a new multimodal mathematical evaluation benchmark called MathScape, which consists of 1325 problems. MathScape combines both figures and mathematical text descriptions into images, which presents a challenge to multimodal large language models. This paper also introduced a two-stage evaluation method to evaluate long responses to math questions. They tested several MLLMs in different data-splitting methods to show results from different perspectives." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In the first challenge, the author said no existing benchmarks have both the mathematical description and figures being captured together in a single image. However, in MathVerse, one category of questions does provide descriptions and figures together. For the second challenge, the author claims existing method cannot assess long-form responses. But MathVerse proposes a method to assess the correctness of each step of a chain-of-thought response. Authors should conduct a more comprehensive literature review of the multimodal mathematical evaluation domain.\n2. This paper is not well organized and written. This means it is not easy to read and understand. For example, section 3.1 is oversimplified. The authors did not mention where they collected the mathematics question and what is the original format of the question documents. Besides, it’s not clear what kinds of annotations are done. What is “knowledge-based classification”? \n3. The proposed two-step evaluation method heavily relies on LLM’s ability to decompose and judge the answer. This may cause some errors in the progress. Did the authors examine how accurate LLMs are on each of the evaluation tasks?\n4. For the evaluation part:\n 1. The model “GLM4V” is without citation, and it is an open-sourced model from my knowledge. (https://huggingface.co/THUDM/glm-4v-9b). Besides, the open-source models in Line 278 are not cited properly. These kinds of format errors cause the paper to be hard to read.\n 2. Some reference performance is not provided: e.g., frequent choice, random choice, and human performance.\n 3. DeepSeekV2 is not in the evaluation setup models, did you mean DeepSeek-VL? \n 4. The performance on proof questions is higher than on choice and solution questions. This is uncommon and the reason given by the authors is not convincing. They said, “The structured format and clear information in proof questions make them easier”. However, when testing models on different kinds of questions. The format of questions is supposed to be similar unless the question format (structure or non-structure) is the primary research topic. \n 5. The authors provide limited insights of the performance on MathScape. Results such as “the closed-source models are more accurate than open-source ones” reveal little information.\n5. MathScape claims that it is the first to combine both figures and mathematical text descriptions in a single image. What unique challenge does this format of data bring to models? Did the authors dive deep into analyzing the different challenges present by MathScape and other multimodal mathematical benchmarks?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "* Problem Sources and Copyright: The problems are stated to be collected from school exams and homework, which raises questions about the original sources and copyright status of these data samples.\n* Fair Compensation: The dataset collection process involved human reviewers for quality control, but it is unclear whether these reviewers received fair compensation for their work." }, "flag_for_ethics_review": { "value": [ "Yes, Legal compliance (e.g., GDPR, copyright, terms of use)", "Yes, Other reasons (please specify below)" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Evaluation Method Validation: How does the proposed two-step evaluation method compare with traditional evaluation methods in terms of reliability and validity?\n* Token Limit Impact: How does the 2048-tokens generation limit affect the results, especially for verbose models? What percentage of responses are truncated by this limit?\n* JSON format output: The constraining model to output JSON format is known to decrease the quality of the generated content (e.g. https://arxiv.org/abs/2408.02442v1). Why the authors choose to stick to this method? What is the impact of such format constrains in current settings?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Benchmark Size and Coverage: The dataset covers a wide range of topics and difficulty levels; 1.2k samples allow for a statistically significant assessment of MLLMs in each subject (except for equations).\n* Data Quality Control: Post-photo quality control and classification is great addition, allowing reviewers to filter unreadable inputs.\n* Evaluation Approach: The two-step evaluation method with sub-task scoring might reduce judgment errors and allows for more fine-grained analysis of the evaluation results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces MathScape, a new benchmark that evaluates the mathematical capabilities of Multimodal Large Language Models using photo-based math problems. Unlike previous benchmarks, MathScape integrates problem statements and visual elements within a single image using a print-and-photo or display-and-photo approach. The authors collected 1,325 images of school-level mathematical problems in multiple choice, free-form, and proof formats (38%, 56%, and 5% respectively). They evaluated 11 closed and open-weight Large Language Models and provided a case study. The results demonstrate that MathScape is challenging even for state-of-the-art models, particularly in the stage of extracting problem statement from image input." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Insufficient Dataset Details: More comprehensive information about the dataset’s creation, sources, human annotators education level and potential biases would strengthen the paper.\n* Limited Language Scope: The focus on Chinese problems limits the applicability of the benchmark to other languages and educational contexts. (Please clearly state the language scope in the abstract and/or in the introduction).\n* Evaluation Method Reliance on LLMs: Using LLMs for scoring may introduce biases, as these models may share similar limitations with the models being evaluated. The judgment error is not addressed in the paper's results or case study.\n* Lack of Comparative Analysis: Given that all of the problems are available in textual format, the paper will benefit from including correlation analysis between original problems and photo-converted problems solve rate." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "More visualization results of the evaluation process can also help to understand the proposed evaluation strategy." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The data collection process is delicate and clearly stated with clear figures.\n\n2. The classification process of math problems are well defined and reasonable.\n\n3. The author provide detailed analysis of accuracy and answer length. This provides some insights to future math MLLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes MathScape, a new benchmark for multimodal math problems to evaluate the related capabilities of MLLMs. The collected datasets contain images with both math figures and questions. The author also uses a two-step evaluation method to first extract answer and then judge the correctness using LLMs. The author evaluate different MLLMs on this new benchmark with detailed analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The contribution of this paper is overclaimed. To the best knowledge, MathVerse contains six versions of a problem and the 'vision-only' one also contains both math figures and question in the image, similar to the contribution of this paper.\n\n2. The two-step evaluation cannot be viewed as an important contribution, since MathVista also uses an LLM (ChatGPT) to extract answers from the free-form response of models as the first evaluation stage.\n\n3. The evaluation of some math-domain MLLMs is missing on MathScape, for example: G-LLaVA and Math-LLaVA.\n\n4. Human performance is needed on MathSacpe for better reference." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "no" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Based on Weakness 1, please elaborate on the most significant contribution of this benchmark, compared to existing multimodal math reasoning benchmarks. You can ignore the parallel research, but I think the related work is not comprehensive yet.\n\n2. I think some of current MLLMs may suffer from different lingual contexts. Therefore, is it possible to expand your work to English problems, or explore the performance difference between Chinese and English.\n\n3. The evaluation part should include GPT-4o if possible. Besides, it should dive deeper into analysis of bad case category proportions and more bad case analysis.\n\n4. I wonder if geometric problems are the hardest type, as it also needs a more complex visual perception of specific components such as angles and lines. \n\n5. The performance tables need to include parameter size for each open-source models. Also, a scaling analysis is needed if possible." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.Originality: The paper presents MathScape, an innovative benchmark that combines real-world math problems captured in images with their correct answers, closely mirroring real-world scenarios and providing a more comprehensive assessment of MLLMS.\n\n2.Quality: The benchmark covers a wide range of difficulty levels, question types, and knowledge areas, which is commendable. \n\n3.Clarity: The paper is structured with clear explanations of the benchmark construction process, evaluation approach, and results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new benchmark termed MathScape for assessing the capabilities of Multimodal Large Language Models (MLLMS) in solving mathematical problems that involve both visual and textual information. MathScape addresses the gap in existing benchmarks by offering a more realistic testing environment with image-based math problems. The benchmark is designed to evaluate the theoretical understanding and application ability of MLLMS through a categorical hierarchical approach. Finally, the paper reports on a multi-dimensional evaluation of 11 advanced MLLMS, revealing the challenges posed by the benchmark and identifying current limitations of these models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I think authors should be aware that except for those previous works you mentioned, there are many other mathematical reasoning benchmarks this year [1,2,3,4,5], especially with a similar focus on multimodal reasoning. Hence, two of your contributions (New Persepective and New Benchmark) may lack novelty. Besides, New Method (i.e., how you construct and evaluate) is a fair but not strong contribution to the MLLM community. \n\n2. The paper indicates that the dataset primarily consists of Chinese problems. I think this will narrow the contribution as well. Besides, educational levels (i.e., primary/middle/high school) are highly different between China and Western countries. So it is better if you can address this limitation, such as including a comparison of educational standards or proposing how the benchmark could be adapted for different educational systems.\n\n3. The analysis is not sufficient for benchmark work. For example, we need to know the proportion of diverse reasons why the best model provides incorrect answers (e.g., failure to retrieve the visual information; misunderstanding of positioning; etc.) in both the whole dataset and each dimension. Furthermore, more bad cases are needed.\n\n4.The evaluation focuses on a set of state-of-the-art models, but it might be beneficial to include GPT-4o, which has been proven for its effectiveness for complex reasoning. Besides, math-specific MLLMs should be included as well, since you also mentioned them in your related works.\n\nReferences:\n\n[1] We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning?\n\n[2] IsoBench: Benchmarking Multimodal Foundation Models on Isomorphic Representations\n\n[3] CMM-Math: A Chinese Multimodal Math Dataset To Evaluate and Enhance the Mathematics Reasoning of Large Multimodal Models\n\n[4] CMMaTH: A Chinese Multi-modal Math Skill Evaluation Benchmark for Foundation Models\n\n[5] ErrorRadar: Benchmarking Complex Mathematical Reasoning of Multimodal Large Language Models Via Error Detection" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024mathscape,\ntitle={MathScape: Evaluating {MLLM}s in Multi-modal Math Scenarios through a Hierarchical Benchmark},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3jvgm61l9S},\nnote={under review}\n}" }, "abstract": { "value": "With the development of Multimodal Large Language Models (MLLMs), the evaluation of multimodal models in the context of mathematical problems has become a valuable research field. Multimodal visual-textual mathematical reasoning serves as a critical \nindicator for evaluating the comprehension and complex multi-step quantitative reasoning abilities of MLLMs. However, previous multimodal math benchmarks have not sufficiently integrated visual and textual information. To address this gap, we proposed MathScape, a new benchmark that emphasizes the understanding and application of combined visual and textual information. MathScape is designed to evaluate photo-based math problem scenarios, assessing the theoretical understanding and application ability of MLLMs through a categorical hierarchical approach. We conduct a multi-dimensional evaluation on 11 advanced MLLMs, revealing that our benchmark is challenging even for the most sophisticated models. By analyzing the evaluation results, we identify the limitations of MLLMs, offering valuable insights for enhancing model performance." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multimodal Large Language Models", "Math Ability", "Benchmark" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/66a6faa7becf9b5c1af605bc045e8424c2a65215.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": "MathScape: Evaluating MLLMs in Multi-modal Math Scenarios through a Hierarchical Benchmark" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could the authors provide the parameter counts for each layer of the SD model before and after sparse pruning?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The writing is very clear, and the main idea is highlighted effectively." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a pruning strategy for Diffusion models, using mask pruning to achieve progressive multi-step pruning. Ultimately, it realizes 1:2 pruning according to the Ampere architecture. During training, knowledge distillation is used to transfer knowledge from the full model to the pruned model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The pruning strategy is based on existing structures, with a relatively simple motivation. There are already other methods that achieve similar results, such as using linear attention or directly training a smaller model with distillation. \n2. Compared to directly using STE-based pruning, it does not further reduce the computational load.\n3. In Section 3.2, \"Transfer learn sparse diffusion models\" strategy is mentioned, but it does not explain the significant differences between this strategy and the progressive sparse training strategy discussed in Section 2.2. If the focus is solely on testing with perturbed datasets, it may not constitute a significant contribution.\n4. A generalized pruning strategy suitable for Transformer networks has not been proposed; simply relying on data perturbations is insufficient to demonstrate applicability to other datasets. Further testing on additional datasets, such as CelebA-HQ, LSUN Church, would be beneficial.\n5. Many of the latest comparative algorithms from 2024 are not mentioned, such as \"Pruning for Robust Concept Erasing in Diffusion Models\" and \"LD-Pruner: Efficient Pruning of Latent Diffusion Models using Task-Agnostic Insights.\"\n6. There is no comparison of the parameter counts for each layer of the SD model before and after sparse pruning. It is recommended to include a chart in the appendix to illustrate this.\n7. While Section 2.3 mentions applying perturbations to the dataset, it does not provide specific details on how the perturbations were implemented.\n8. The experiments only validate the FID score as a single metric; it is advisable to explore additional metrics, such as SSIM." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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.\tThe authors mentioned that “it does not mean that the greater the sparsity, the better the FID”. Please discuss the reason and why you choose 2:4 sparse.\n2.\tPlease discuss the reason ASP performs so worse in all experiments.\n3.\tPlease also clarify why your method and STE-based pruning fulfill the same MACs.\n4.\tPlease explain the reason that the FID of the proposed method in Fig. 3a obtain a lower FID in the first several steps.\n5.\tWhy the initial FID of 2:4 sparse in Fig.3b and Fig.3d is different?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThis paper is well-written.\n2.\tThe motivation is clear enough.\n3.\tThe organization of this paper is great." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to improve the efficiency of DM by sparse matrix for 2:4 sparse acceleration GPU. The authors improve the STE method and propose to gradually transfer knowledge from dense models to sparse models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThere is a typo in Eq5. Please also check all equations. Moreover, not all symbols have been explained.\n2.\tThe experiments are relatively limited. Specifically, only two U-ViT and DDPM are tested on the proposed pruning, which are proposed in 2022 and 2020 respectively. More recently proposed DiT or other methods should also be included.\n3.\tThe limitation and discussion are missing 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": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please address the weakness stated above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The 2:4 sparse model calculation offers practical values for practitioners using NVIDIA Ampere architecture GPUs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work aims to reduce the computation of Diffusion Models during inference. The authors suggest a method of straight-through estimation, which applies sparse masks to layers of a pretrained diffusion model and then employs transfer learning for training. Then, they use the same sparse mask during inference to improve compute efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While it may have some practical value for practitioners using NVIDIA Ampere architecture, the same technique may not benefit other practitioners or general researchers without access to Ampere architecture. \n\n- Besides, the straightforward idea of using masked training is neither interesting nor technically new. \n\n- More disappointingly, the speed acceleration due to this customized training for a particular architecture increases by x1.2 only. Studies related to reducing time steps for Diffusion inference or diffusion quantization/pruning methods may be more effective in achieving the same purpose." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* In Table 3, some variants (e.g., patch size = 2 and mlp_ratio = 2) are slower than the dense model, why do you think this is?\n* I think it would strengthen the effectiveness of SparseDM if the author show that it can also be applied to models like Stable Diffusion." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The paper introduces a simple fine-tuning method that converts existing diffusion models into sparse models, enabling them to be used in scenarios with limited computing power, such as on mobile devices.\n\n* The observations about fixed sparse training are interesting.\n\n* Experiments on various generation scenarios verify the effectiveness of SparseDM compared to baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces SparseDM, which converts existing diffusion models into sparse models that fit in a 2:4 sparse operator on the GPU. Specifically, the authors propose a Straight-Through Estimator (STE)-based fine-tuning framework that learns sparse masks. These sparse masks accelerate GPU inference speed up to 1.2. Comprehensive experiments validate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Weakness 1: More clarifications on Section 2.3.**\n\nIn Section 2.3, the authors claim that diffusion models only consider the distribution shift of the noisy data while sparse pruning methods only consider the model's weight change. Then, referring to RFR, the authors convert the model's weight changes resulting from sparse pruning methods into data changes for the diffusion model's training process. However, typical diffusion models have indicators for perturbed data (such as the noise schedule and timestep embedding), and it is unclear how these relate to perturbations caused by sparse training.\n\n**Weakness 2: Lack of analysis of fixed sparse training**\n\nI am not sure why fixed sparse training would be more effective than traditional progressive sparse training. Based on the experimental results, it seems that fixed sparsity applies a consistent distribution shift across all noise levels in diffusion training, whereas progressive sparse training gradually shifts the predefined noise levels, which may hinder the diffusion training process. However, this claim has not been theoretically verified, so the authors should provide theoretical proof to demonstrate the relationship between diffusion training and sparse training." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024sparsedm,\ntitle={Sparse{DM}: Toward Sparse Efficient Diffusion Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3kADTLbKmm},\nnote={under review}\n}" }, "abstract": { "value": "Diffusion models have been extensively used in data generation tasks and are recognized as one of the best generative models. However, their time-consuming deployment, long inference time, and requirements on large memory limit their application. In this paper, we propose a method based on the improved Straight-Through Estimator to improve the deployment efficiency of diffusion models. Specifically, we add sparse masks to the Convolution and Linear layers in a pre-trained diffusion model, then transfer learn the sparse model during the fine-tuning stage and turn on the sparse masks during inference. Experimental results on a Transformer and UNet-based diffusion models demonstrate that our method reduces MACs by 50% while increasing FID by only 0.44 on average. Sparse models are accelerated by approximately 1.2x on the GPU. Under other MACs conditions, the FID is also lower than 1 compared to other 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 models", "sparse pruning", "2:4 sparsity" ] }, "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/fc8524ef642e2c3611ca910a36dc747a9becf3a4.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/47c77c2cd47cc71808204e066cc809a38721c237.zip" }, "title": { "value": "SparseDM: Toward Sparse Efficient 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": 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. Please justify your assumption stated at line 161.\n2. For Line 1224 Figure 5, why do you only choose to conduct the ablation study on the ChChMiner dataset? Ablation studies on larger datasets are needed.\n3. Following your design, IGIB-ISE should effectively identify the core substructure of molecules, why did the model not improve the classification accuracy more? As it reduces redundant information, why does it occupy a larger space? More analysis is needed to identify factors that may limit the improvement. What are the potential enhancement may be introduced to address these limitations?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThis paper has good clarity. It is well-written with a clear structure. In a concise but informative style, readers would find it easy to understand the key concepts, backgrounds, and methods.\n2.\tTheir work also brings new insights into the MRL area. They noticed the inefficiency of current methods, where using the complete profile of an interacting molecule could not only be unnecessary but also comprises generalizability. And they proved the effectiveness of their method through experiments. \n3.\tIn general, they bring new ideas to the MRL area: Interactive Graph Information Bottleneck (IGIB). Bottleneck-based methods are widely used in many areas and receive satisfactory results. In this paper, they integrated it into the ISE framework for further optimization. It is also the method that leverages the model’s performance to outperform all baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To alleviate the problems in current methods of molecular relational learning: insufficient consideration of molecular interactions and failure to capture high-quality substructures, this paper introduces an IGIB (Interactive Graph Information Bottleneck)-ISE (Iterative Substructure Extraction) method. Their work achieves better performance than current SOTA models in terms of accuracy, generalizability, and interpretability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\t(General Assumption) Most molecule interactions may depend on each molecule’s substructures, but does this apply to all molecule interactions? If not, the assumption at line 161 is somewhat arbitrary, where some edge cases could be ignored by this model. This assumption needs to be further justified. \n2.\t(Time and Space Complexity) While the model outperforms all the baseline models, it spends much more time processing DDI Datasets. Compared to CMRL, with around 1% accuracy improvement, this model costs 5.8 ~ 7.1x more time and 6.4 ~ 9x more space. This may lead to expensive computation. The trade off between the performance and computing cost needs to be examined. \n3.\t (Ablation Experiment) Most experiments are designed well, but the experiment in line 1224 is less persuasive. Among all the datasets for the drug-drug interaction prediction task, ChChMiner has the fewest data points. Besides, since molecular interaction prediction tasks are different from DDI, a separate experiment would be good. \n4.\t (Improvement) While IGIB-ISE achieves good performance, ISE fails to outperform all Category II methods in Table 1 (line 324) and some Category II methods in Table II (line 378). Also, the improvement of IGIB-ISE is not that noticeable in the classification task." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In general, the paper is a solid contribution but the presentation should improve. Please answer to my questions above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(S1) The paper solves a timely problem and presents a sound solution that fully exploits the relationships among substructures.\n\n(S3) Due to its substructure alignment, IGIB-ISE outperforms previous techniques on several datasets.\n\n(S3) The method is well-motivated and builds on previous graph information bottlenecks, ELMO and expectation maximization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper describes a method to improve molecular relational learning using information theoretical loss functions on a subgraph of the molecules. The technical contribution lies in the coupling of graph information bottlenecks with expectation maximization. The results show the approach's superiority both in deductive and inductive scenarios. The method is well-motivated, and the experiments are solid." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(W1) Missing explicit objective function: The paper first explains the solution and then reaches the objective in Equation 8. I find this presentation counterintuitive. Why not present the objective first and then explain how to compute it?\n\n(W2) In the modelling of the graph there is no feature vector associated with nodes/edges. Are the graphs without attributes? Molecules should have information about the type of bonds among atoms.\n\n(W3) Notation without introduction: The paper uses notation without introducing it. Examples include:\n\n- $\\mathbf{Y}_\\mathcal{G}$\n- Line 216: the symbol *, is it a matrix multiplication?\n- $\\||$ in line 218\n\n(W4) If sim is symmetric cosine similarity, what is the need for computing both $sim(F_1, F_2)$ and $sim(F_2, F_1)$?\n\n(W5) It is not clear how Eq. 5 ensures that the two structures are aligned since $H_1$ and $H_2$ refer to two different embeddings spaces, or is the alignment enforced by the two matrices $I_{12}, I_{21}$? Please explain and motivate.\n\n(W6) What is the Gumbel sigmoid and how does it help in this case?\n\n(W7) It is not clear whether Eq. 16 is a lower bound on Eq. 8 or what is the relationship with Eq. 8? Is that an approximation or a heuristic? This aspect should be clarified in the text.\n\n(W8) In Figure 4, the focus of the network substantially changes over iteration. This seems to indicate that the method struggles with convergence. Is that expected or is it a sign of instability?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "**1. The discussion of the limitations of Category II methods is confusing.** \n- It is understandable that core substructures often play a crucial role in molecular interactions. But, Figure 1 (a) does not deliver a relevant message to support this argument. \n- In addition, from Figure 1 (a), it is unclear why integrating the complete profile of an interacting molecule into the substructure generation can be overwhelming. \n- It's unclear why Category II carries the risk of compromising generalizability. After reading the cited paper [1], it's still very confusing. There is no clear evidence from [1] to support this statement. \n- It's unclear why the authors mention \"Activity Cliffs\" here. \n\n**2. Limited Discussion of Method Robustness.**\nAs an interactive method, what happens if the EM algorithm finds optimal solutions during iteration? The lack of guidelines for selecting optimal iteration numbers based on dataset characteristics leaves important practical questions unanswered.\n\n**3. Technical Clarity Issues.** \n- Line 160, what is Y_G? Should it be Y?\n- In Tables 6-7, your method should be named ISE-IGIB or IGIB-ISE?\n\n**4. Computational Overhead.** \n- Tables 6 and 7 show IGIB-ISE takes more than 700% execution time and 1000% memory compared to one baseline DSN-DDI, with around 1.5% DDI performance improvement. I don't appreciate such results. The authors do not sufficiently address this limitation or propose potential optimizations. \n- The experiments focus on relatively small molecules. There is no discussion or analysis of how the method scales with molecular size, which is important for applications involving larger molecules. \nThe memory requirements (Table 6-7) suggest potential scaling issues.\n\n[1] Mechanisms of drug combinations: interaction and network perspectives" }, "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 novel approach to molecular interaction learning. Rather than handling entire molecular structures or extracting substructures independently, it introduces an iterative refinement process guided by molecular interactions. \n\n- Using EM algorithms for substructure extraction is creative, treating substructures as latent variables that get refined through iterations. This is a fresh perspective on the molecular interaction learning problem.\n\n- This work has a substantial potential impact on drug discovery and materials science. The ability to identify and understand interacting substructures between molecules is crucial for these fields." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the Iterative Substructure Extraction (ISE) framework for molecular relational learning, addressing how molecules interact through their core substructures. The framework combines an Expectation-Maximization algorithm for iterative refinement with a new Interactive Graph Information Bottleneck (GIB) theory to ensure extracted substructures are minimal yet influential. Through experiments on datasets covering both regression and classification tasks, the combined IGIB-ISE approach demonstrates improved accuracy and interpretability compared to existing methods for predicting molecular interactions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The discussion of the limitations of Category II methods is confusing.\n\nLimited Discussion of Method Robustness. \n\nTechnical Clarity Issues. \n\nComputational Overhead." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Can the authors validate the interactions between multi-molecule interactions?\n\n2. Why the interaction is computed as $H_1=F_1^{(1)}||F_1^{(2)}$?\n\n3. The way to extrapolate the core substructure is **very similar to [1]**. What's the difference between this paper and [1]?\n\n4. What's the complexity of the method? Can you compare the training and inference time with baselines?\n\n5. Can you validate your method on larger datasets?\n\n[1] Capturing substructure interactions by invariant Information Bottle Theory for Generalizable Property Prediction" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper introduces an innovative method for core substructure extraction using the EM algorithm, which effectively captures molecular interactions.\n\n2. IGIB theory ensures a precise and compact extraction of interactive substructures.\n\n3. The method is extensively validated across various molecular relational learning tasks, including drug-drug interaction and solvation energy prediction, showing clear improvements over state-of-the-art methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a framework called ISE to improve MRL by focusing on the interaction between core substructures of molecules. The model iteratively refines the core substructures using the EM algorithm. Additionally, the IGIB theory is proposed to capture minimal but most influential substructures, enhancing the efficiency and generalizability of the extraction process. Through extensive experiments, the IGIB-ISE framework demonstrates superior performance compared to existing methods in terms of accuracy, generalizability, and interpretability for molecular interaction prediction tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Some parts of this work is very similar to [1]**. The key idea and many formulas are similar. For example, they all utilize similar methods to extrapolate core substructures (Section 3.4 in this paper and Section 3.2 in [1]). The only difference here seems to be this paper extrapolates the core substructure from a pair of graphs while [1] extrapolates the core substructure from one graph.\n\n1. The framework is validated on interactions between two molecules. It does not extend to more complex scenarios like multi-molecule interactions, which are important in real-world biochemical environments.\n\n2. The method requires more iterations, increasing resource consumption and time. This may limit its scalability for very large datasets or complex molecular systems.\n\n\n\n[1] Capturing substructure interactions by invariant Information Bottle Theory for Generalizable Property Prediction" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024iterative,\ntitle={Iterative Substructure Extraction for Molecular Relational Learning with Interactive Graph Information Bottleneck},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3kiZ5S5WkY},\nnote={under review}\n}" }, "abstract": { "value": "Molecular relational learning (MRL) seeks to understand the interaction behaviors between molecules, a pivotal task in domains such as drug discovery and materials science. Recently, extracting core substructures and modeling their interactions have emerged as mainstream approaches within machine learning-assisted methods. However, these methods still exhibit some limitations, such as insufficient consideration of molecular interactions or capturing substructures that include excessive noise, which hampers precise core substructure extraction.\nTo address these challenges, we present an integrated dynamic framework called Iterative Substructure Extraction (ISE). ISE employs the Expectation-Maximization (EM) algorithm for MRL tasks, where the core substructures of interacting molecules are treated as latent variables and model parameters, respectively. Through iterative refinement, ISE gradually narrows the interactions from the entire molecular structures to just the core substructures.\nMoreover, to ensure the extracted substructures are concise and compact, we propose the Interactive Graph Information Bottleneck (IGIB) theory, which focuses on capturing the most influential yet minimal interactive substructures. In summary, our approach, guided by the IGIB theory, achieves precise substructure extraction within the ISE framework and is encapsulated in the IGIB-ISE}\nExtensive experiments validate the superiority of our model over state-of-the-art baselines across various tasks in terms of accuracy, generalizability, and interpretability." }, "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 Relational Learning", "EM Algorithm", "Substructure Extraction", "Interactive Graph Information Bottleneck" ] }, "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/ea31c56fe8c87654bf3f849f4c6226f7c7b4a910.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": "Iterative Substructure Extraction for Molecular Relational Learning with Interactive Graph Information Bottleneck" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1] The derivation of your DiLap operator appears to be flawed. In particular, in the second line of Equation (12) when you pull $s_i$ out of the sum, the term $s_i$ arises $d_i^{out}$ times and therefore, Line 2 should be $s_i - \\frac{1}{d_i^{out}} \\sum s_j.$ The correction of this error means that you should be working with the random walk Laplacian (see e.g. [1]), which would be far more intuitive. To me it seems that for it to be possible to accept this paper at ICLR, the derivation of your operator needs to be corrected and the subsequent experiments should be adjusted. \n\n2] I am unsure what adjacency matrix you use in Line 314 to sparsify the matrix $\\tilde{\\mathcal{C}}.$ Are you using the adjacency matrix corresponding to the graph in which you have added the node feature similarity edges? And if not, could you hypothesise how severe the impact may be of calculating the commute times on a rewired graph and to then message pass with the original graph. \n\n3] Your method appears to be relatively memory intensive. In particular, you seem to require the evaluation of the exponential function fo the dense matrix $\\tilde{\\mathcal{C}}$. Empirical evaluation of, not only the time, but also memory complexity of your method in comparison to your baseline methods would be very valuable. \n\n4] The ablation study in Table 2 is very interesting! I think it should be extended in scope to also extend to homophilic datasets and to also include other commonly used message passing operators, such as the symmetrically normalised adjacency matrix used in the GCN and the PageRank matrix used in the PPrGo model [2].\n\n5] It does not seem sensible to me to compare your sparisfied commute time based CGNN to the CGNN$_{ppr}$ using the dense PageRank matrix. It seems to be fairer to me to either compare dense versions of both matrices or sparse versions of both matrices. In particular, since you sparify your commute time matrix with the adjacency matrix, it would be interesting to compare your model to a PageRank-based scheme, where the PageRank matrix is also sparsified with the adjacency matrix. \n\n6] Minor comments:\n\n6.1] The abbreviation \"SPD\" is used in Line 45 before its definition. \n\n6.2] The contributions of your paper are not explicitly listed. \n\n\n[1] Von Luxburg, U., 2007. A tutorial on spectral clustering. Statistics and computing, 17, pp.395-416.\n\n[2] Bojchevski, A., Gasteiger, J., Perozzi, B., Kapoor, A., Blais, M., Rózemberczki, B., Lukasik, M. and Günnemann, S., 2020, August. Scaling graph neural networks with approximate pagerank. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 2464-2473)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The rewiring scheme that you propose is simple, but rather nice in my opinion. It would be interesting to see further study of its impact on the overall graph structure. \n\n- Your proposed CGNNs are compared to a comprehensive set of baseline models, which great to see. \n\n- The analysis of the application scope of your proposed model is very strong indeed and something that is generally not done enough in our literature." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In the submitted manuscript, the authors propose a novel digraph Laplacian, which is later used to more efficiently calculate the commute time of pairs of nodes. They furthermore propose a simple node feature based rewiring scheme, which allows them to ensure that the resulting graph gives rise to an aperiodic, irreducible Markov Chain, which has a unique steady state. The authors then propose to calculate commute times on this rewired graph, to transform these commute times by taking the exponential function of this matrix and subsequently sparsifying it with the adjacency matrix. This then allows the authors to propose a variant of the DirGNN, called CGNN, in which edges are reweighted by their transformed commute times. The authors finally evaluate the empirical performance of their CGNNs against a large variety of baseline models on a large number of datasets and find consistently good, although sometimes marginal performance improvements. They furthermore analyse these results and provide several insightful further experiments on runtimes and ablation studies of different model components." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please find further details on my listed weaknesses in my questions below.\n- The proposed model boils down a weighting of the DirGNN by an efficiently calculated function of the commute times, which is a rather trivial change. \n- The derivation of the DiLap matrix appears to be flawed. \n- Your ablation studies could be improved and extended." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "As outline in the weaknesses above, can the authors provide a principled reason for why features aggregated from a node's neighbors should be weighed by the commute distance of the node to its neighbors? Can the authors provide any theoretical justification for using commute distance for such weighing?\n\nThe authors provide one figure where they show for Chameleon and Squirrel data sets that an adjacency matrix constructed from nearest neighbors weighted by their commute distance more closely resembles an adjacency matrix constrained to connecting nodes within the same class. Can the authors include other empirical measures of how the weighting of neighbors can improve performance of GNNs? Can the authors look at properties of longer hops (going beyond one-hop neighbors) and how information is aggregated across nodes within or outside the same class? Would it be possible to construct a synthetic data set that would shed light on the mechanism behind why they see an improvement?\n\nHow does the proposed model's performance change with depth? The authors claim that the weighted neighbors avoids the problem of aggregation of irrelevant information as depth is increased? It would be informative to see an empirical demonstration of this. The authors should plot the performance of their model as a function of model depth and compare with existing models.\n\nIn addition, can the authors apply their approach a real world problem on directed graphs that requires long range information transmission, such as power grids or traffic flow data? This would bolster the empirical support for their method.\n\nAs noted in the weaknesses above, can the authors try alternative methods for graph rewiring that also produce sparse graphs, such as constructing a kNN graph using the node features? The authors should include an empirical comparison or provide theoretical justification for their method. Is the proposed similarity-based rewiring mode optimal?\n\nHow much of the improvement is coming from the fact that the Laplacian proposed by the authors is weighted by the stationary probability of random walks on the graph (or what the authors call the importance of a node)? Can the authors do an ablation study to disentangle this from the commute distance?\n\nThis reviewer was confused by the comment that general GNNs can outperform models tailored for directed graphs with hyper-parameter tuning. In Table 1, DirGNN, a model tailored for directed graphs, outperforms GCN. Can the authors clarify what they mean by this comment?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This is an interesting paper with a clever insight. It is sensible to say that not all nearest neighbors on a directed graph are created equal. Weighing the features of some neighbors more during the aggregation step of a GNN based on the shorter commute distance of those neighbors to the original node is an intriguing idea. Weighing based on the commute distance certainly sounds reasonable. The proposed method for rewiring a graph to make it irreducible and aperiodic while still retaining sparsity is also clever as is the weighted Laplacian that can be used to efficiently compute the commute distance leveraging its sparsity using methods such as randomized truncated singular value decomposition.\n\nThe state-of-the-art performance achieved using the author's method on some of the most commonly used directed graph data sets, such as Squirrel, Chameleon, etc. is impressive and provides reasonable empirical proof of the validity of the proposed approach. The authors also include solid empirical evidence on running times of their algorithm and convincing comparisons to PageRank for graph rewiring and calculation of commute distances." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a method for weighing the features of the neighbors of a node during aggregation step of GNN based on the commute distance of the neighbor to the node. The commute distance between nodes A and B is the average number of steps that a random walk takes traversing from node A to node B and back to node A. This distance is particularly relevant for directed graphs because although all nearest neighbors are one hop away from a node, their commute distances might vary because of the constraints imposed by the directions of the edges. One neighbor might require a circuitous path along many other nodes before returning to the original node (have longer commute distance) whereas another neighbor could be closer. The authors' key idea is to weight the importance of the features of the neighbors of a node during the aggregation step of GNN based on the commute distance of the node to those neighbors. Besides this weighing, the aggregation and update scheme that they use is based on that of Rossi et al. where the features of the incoming and outgoing neighbors are aggregated separately and used alongside the node's own features during the update step. The authors also propose an efficient way of computing commute distance. To do so, they introduce a weighted Laplacian for directed graphs that accounts both for the directional connectivity of the nodes and their importance (computed as the stationary probability at each node of a random walk on the graph). The authors also introduce a way to rewire a graph to ensure that it is irreducible and aperiodic while keeping the graph sparse (unlike alternative methods such as PageRank). The commute distance can then be efficiently computed using the sparse weighted Laplacian. Finally, the authors show empirically that their proposed approach improves on existing methods when applied to many standard directed graph data sets such as Squirrel and Chameleon." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "To this reviewer, the biggest weakness of the paper was that although weighing neighbors by commute time is sensible, it is not necessarily principled. Is there any reason to a priori expect that neighbors of a node that have shorter commute times to that node somehow contain more relevant features for learning on graphs? This seems to depend on the nature of the learning problem and the data set. Now, the authors can argue that their empirical evidence is sufficient to motivate their approach. However, more should be done here to support the author's proposal. Some evidence is provided in that an adjacency matrix constructed by weighing the neighbors by their commute distance more closely resembles an adjacency matrix constrained to edges that connect nodes within the same class. The authors should expand on this. What does this look like for other data sets? How does aggregation of information using these weighted neighbors look across multiple hops and longer distances across the graph? The authors should have come up with synthetic data sets that can elucidate the mechanism behind the improvement that they are seeing.\n\nIf empirical evidence is the main motivation behind the proposed schemes, the authors could have dome more to build a stronger case. An argument is made in the paper that with the weighing of the neighbors less irrelevant information is aggregated as the GNN models go deeper. The authors should empirically demonstrate this by showing how the performance of their model changes with depth and contrast with existing models. In general, it would have been very interesting to see the impact of the weights proposed in this paper on multi-hop GNN models, such MixHop, Shortest Path Networks, or DRew. In addition, it would have been more convincing if the authors had applied their approach to real world problems of directed graphs in addition to standard benchmark data sets used in Table 1 such as temporal web traffic data, power grids, traffic flow, etc.\n\nThe method proposed in the paper to rewire the graph to ensure irreducible and aperiodic graphs is also ad hoc and not very principled. The method certainly produces a sparse graph unlike PageRank, however, other approaches can also be used to generate irreducible graphs that are sparse such as generating a kNN graph based on the node features. It is not clear that the proposed method is optimal in any way other than that it outperforms the amended probabilities used in PageRank. See Questions below for suggestions on how the authors can evaluate this empirically." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "I have no further questions, though I would recommend that the authors address the previously noted weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper is engaging and well-written, with a thorough background review that enhances accessibility and readability. Key strengths I noted include:\n\n1. **Novel Approach with Significant Potential:** The proposed method, particularly the newly formulated digraph Laplacian, offers a fresh perspective with substantial potential for future research and applications.\n\n2. **Comprehensive Component Analysis (Section 5.3):** The inclusion of a component analysis strengthens the paper by providing an effective ablation study.\n\n3. **Clear Contribution and Baseline Comparison:** The authors clearly articulate their contributions, outlining the distinctions between their method and existing baselines. They explain where prior approaches fall short and demonstrate how their approach addresses these limitations.\n\n4. **Effective Visual Aids:** Figures 1 and 2 are well-designed and enhance understanding by clarifying details within the method.\n\n5. **Robust Experimental Validation:** he paper validates its approach across a wide variety of datasets and multiple baseline comparisons, highlighting the robustness and generalizability of the proposed method.\n\n6. **Reproducibility:** The authors provide code for reproducing their experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present a novel approach that integrates node-wise commute time into a message-passing framework, introducing the Commute Graph Neural Network (CGNN). The central contribution of CGNN is the development of a new directed graph Laplacian, designed to address path asymmetry in directed graphs. The authors demonstrate that CGNN outperforms existing baselines across most datasets and effectively motivates the significance of the problem they address.\n\nOverall, I found the paper well-executed and recommend it for acceptance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, this paper is strong in its methodology and results, though I have a few recommendations that could enhance its clarity and depth.\n\n1. **Graph Density in Rewiring Approach:** While I appreciate that the authors provided commute times before and after rewiring in Table 3, it would be beneficial to also examine how rewiring affects graph density. This additional metric could offer deeper insights into structural changes post-rewiring.\n\n2. **Unobserved Edges in Definition of $m_{i,in}^{(l)}$ and $m_{i,out}^{(l)}$:** Given that unobserved edges are introduced to the graph, I suggest adjusting the definitions of $m_{i,in}^{(l)}$ and $m_{i,out}^{(l)}$ to account for these edges, potentially assigning them a lower weight than observed edges. This adjustment could yield a more realistic representation of edge significance.\n\n3. **Model Complexity:** The model’s complexity is relatively high, even though it’s reported to be on par with other GNN models. This complexity, particularly in precomputation, might be a barrier in some cases. However, I do not consider this a critical issue, as future work could address and optimize this aspect.\n\n4. Inclusion of Synthetic Datasets: While the paper impressively covers a range of empirical datasets, the addition of synthetic datasets could improve interpretability. By embedding known patterns, synthetic data could highlight the model's strengths and limitations in detecting specific features.\n\n5. Reordering Related Work: Placing the Related Work section (currently Section 6) closer to the beginning would make the reading experience smoother, giving readers essential context before diving into the methodology and results.\n\nThese revisions would, in my opinion, strengthen the paper without diminishing its core 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": 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) What do you mean by the footnote on page 1?\n2) Why is the proposed Laplacian sparse? From Eq. (5), the matrix $P$ seems to be a complete matrix.\n3) What is the relationship between Eq. (5) and $D^{-2}L$ and why do you should choose Eq. (5)?\n4) Why does the rewiring procedure only minimally alters the overall semantics of the original graph?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1) The idea of considering commute time is novel and reasonable.\n2) The topic of directed graph neural networks is important.\n3) The source code is provided." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Commute Graph Neural Networks (CGNN) for directed graphs, which is based on a new digraph Laplacian matrix taking into the commute time on a (possibly rewired) strongly connected graph. Theoretical and empirical analysis is provided." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The paper is poorly written with unsupported claims. For example, the footnote on page 1 is not sensible, and it is not clear why the previous methods are undirectional during shortest path computation.\n2) It is unclear why the proposed Laplacian is sparse. From Eq. (5), the matrix $P$ seems to be a complete matrix.\n3) It is unclear what the relationship is between Eq. (5) and $D^{-2}L$ and why you should choose Eq. (5).\n4) Being strongly connected is too strong an assumption, and it is not clear why the rewiring procedure only minimally alters the overall semantics of the original graph.\n5) [1] mentions flow imbalance in directed graphs and is not discussed. It is also unclear whether the idea in [1] is considered undirectional by the authors.\n6) (minor) Grammar issues: e.g., line 115.5 \"notations. We\" should be \"notations, we\"\n\nReference:\n [1] He, Y., Reinert, G., & Cucuringu, M. (2022, December). DIGRAC: digraph clustering based on flow imbalance. In Learning on Graphs Conference (pp. 21-1). PMLR." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose an approach to integrate commute time into graph neural networks to enhance the analysis of directed graphs, effectively addressing the asymmetry and complex path interactions inherent in these structures." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024commute,\ntitle={Commute Graph Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3ktyyYGLxB},\nnote={under review}\n}" }, "abstract": { "value": "Graph Neural Networks (GNNs) have shown remarkable success in learning from graph-structured data. However, their application to directed graphs (digraphs) presents unique challenges, primarily due to the inherent asymmetry in node relationships. Traditional GNNs are adept at capturing unidirectional relations but fall short in encoding the mutual path dependencies between nodes, such as asymmetrical shortest paths typically found in digraphs. Recognizing this gap, we introduce Commute Graph Neural Networks (CGNN), an approach that seamlessly integrates node-wise commute time into the message passing scheme. The cornerstone of CGNN is an efficient method for computing commute time using a newly formulated digraph Laplacian. Commute time is then integrated into the neighborhood aggregation process, with neighbor contributions weighted according to their respective commute time to the central node in each layer. It enables CGNN to directly capture the mutual, asymmetric relationships in digraphs. Extensive experiments confirm the superior performance of CGNN. Source code of CGNN is anonymously available here." }, "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", "Message Passing", "Commute Time", "Node Classification" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/6bc6f7cbcd52cd57b6b144506252118318511eea.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": "Commute 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": "Please refer to the weaknesses above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Writing quality is good. The paper is well-structured, and clearly written.\n2. SOTA performance. The paper demonstrates the state-of-the-art performance on mainstream detection and segmentation datasets, such as COCO and ADE20K, which is impressive.\n3. Versatility of the method. The paper shows the simplicity of CR2PQ, which can be easily integrated into a variety of popular representation learning frameworks, such as mask-based learning, contrastive learning, and distillation methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the Continuous Relative Rotary Positional Query to enhance dense visual contrastive learning by improving pixel/patch correspondence across different views. It addresses limitations in existing self-contrasting methods by transforming discrete positional embeddings into continuous representations. The proposed CR2PQ enables more effective patch-level representation learning, achieving state-of-the-art results and faster convergence in detection and segmentation tasks on the COCO dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Reliance on random cropping. Although random cropping can increase the variability of the input, its results may still be limited by the randomness of the cropping. In extreme cases, it may result in almost no overlap between the generated views, affecting the learning effect of the model.\n2. Computational complexity. Complex matrix operations are required when calculating relative position embedding and rotating embedding, which increases the burden in scenarios with limited computing power.\n\nP.S. There is an error in Figure 1. [CLS] should be global information, while patch is local information." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "-Why use a pretraining network for the teacher? You are comparing with other baselines which some of which learn everything from scratch. This seems like a logical thing to try, have you tried that?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "-The empirical results are good and outperform previous SOTA.\n\n-I think this paper can be worthwhile to accept, I'm willing to improve my score based on the author's reply." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a distillation technique where a student is densely trained to match teacher features. The novelty comes from using 2D RoPE in the network as well as a cross-attention module with relative positional information. They show good empirical results on detection and segmentation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-L142: Relative positional encoding = RoPE?\n\n-L161: W_{pos} v.s. P_{pos} ?\n\n-The notation in equation 1 is confusing. It is as if the patches don’t interact with each other. I would use a new variable to define a patch representation. Also if f_\\theta denotes the ViT, why does it take z as input, which already contains the linear layer on the left side of the equation but not on the right side. I think the notation should be made more precise.\n\n-Equation 2 has some n and m mixed.\n\n-L219: “we set each patch size of the view A as 1”, but in L227 p_A (the patch size) is defined?\n\n-L228: There is a sentence “Since we set each grid size of the anchor view as 1.” What is that supposed to mean?\n\n-L297: If I’m not mistaken, the definition of q doesn’t make sense.\n\n-The first stated contribution is using 2D RoPE for SSL based methods. Then, in L358, shoud state “We also evaluate the detection and segmentation without pretraining i.e. directly using 2D RoPE”. First, that entry is only in Table 1 and not Table 2. Second, I think you should also independently show empirical evidence of your 2 first contributions (2D RoPE and cross-attention module) and report results for that.\n\n-In general, I think the paper could be more explicitaly precise with how sizes/positions are encoded e.g. is it relative to the original image input grid or relative to the crop?\n\n\n\nMinor:\n\n-L082: “as the downstream task only input”\n\n-If I’m not mistaken, there is a problem with sentence at L203 starting with i.e." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. What is the performance of the CR2PQ backbone performance on some strong detectors, such as DINO or Co-DETR?\n\n2. CR2PQ requires the teacher model to provide contrastive pairs, however, the performance does not improve as the model becomes larger (ViT-L vs ResNet50). The reviewer wonders about the performance of a larger model for the student. Does this approach work for a larger backbone as a student, such as ViT-L/ViT-G? The authors are suggested to validate the scalability of the method.\n\n3. Some small mistakes\n\n- The font of the paper is different from other papers. Should it be correct? \n\n- line 274, there is an overlap between the table and the caption." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. CR2PQ introduces a pioneering method for dense visual representation learning by utilizing continuous relative rotary positional embeddings, which is a significant departure from traditional discrete embeddings.\n\n2. The method achieves state-of-the-art results across various benchmarks, including object detection and segmentation tasks on COCO and semantic segmentation on ADE20K, outperforming previous leading methods by a considerable margin.\n\n3. The introduction of a positional-aware cross attention module enhances the learning of semantic information without incurring significant additional computational costs. CR2PQ's use of rotary positional embeddings makes it robust to various view augmentations, including random cropping, which is a common challenge in contrastive learning methods.\n\n4. The paper supports the method's strengths through extensive experiments and ablation studies, providing a thorough analysis of CR2PQ's performance under different conditions and configurations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. The paper introduces Continuous Relative Rotary Positional Query (CR2PQ), a novel method for dense visual representation learning.\nCR2PQ addresses the challenge of establishing pixel/patch correspondence across different views in dense contrastive learning (DRL) by transforming discrete positional embeddings to continuous representations.\n\n2. It utilizes a rotary positional embedding to represent the relative positions between two views and reconstructs the latent representations of one view from another through a rotary positional query.\n\n3. The method simplifies the dense contrastive learning paradigm by making it correspondence-free and integrates easily into various representation learning frameworks.\n\n4. Extensive experiments on standard datasets demonstrate state-of-the-art (SOTA) results, outperforming the previous SOTA method (PQCL) significantly in detection and segmentation tasks on COCO with improved mAP scores." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Experiments. The author should provide more scales of backbone to validate the scalability of the method. Most experiments are conducted on ViT-S. The reviewer understands the efficiency of the experiments, however, there should be some experiments on larger backbones." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- In Table 1, what does the row \"RoPE\" exactly correspond to? A ViT-S/16 equipped with rotary positional embedding, randomly initialized and finetuned on the downstream task?\n\n- In Table 4, what does the row \"EMA update (Contrastive)\" exactly correspond to? Is the teacher randomly initialized?\n\n- At line 219. it is mentioned that the patch size of view A is set to 1, but then it is set to $p_{A}$. Can you clarify this?\n\n- At line 227: I suggest using another notation for $p_{A}$ as the patch size, as it is confusing." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The proposed self-supervised framework for dense visual representation learning is novel.\n- The method elegantly eliminates the need to establish explicit correspondence between local features across views by leveraging relative positional cues.\n- The performance on dense downstream tasks is thoroughly evaluated, showing faster convergence and achieving state-of-the-art results on standard benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel self-supervised framework for dense visual representation learning, which avoids the need for explicit dense correspondences between local features across views. Instead, the framework reframes the task as predicting local representations from one view to another, guided by relative positional cues. It integrates rotary positional embeddings within the student model and distills knowledge from a pre-trained, frozen teacher model. This approach yields faster convergence and improved performance on standard benchmark evaluations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The method differs from existing baselines in three key ways: (1) the use of rotary positional embeddings, (2) the use of a pre-trained, frozen teacher model, and (3) the proposed pretext task. This makes it challenging to assess the contribution of each component to the overall performance. Specifically, the fairness of the experimental setup is questionable, as other methods are trained from scratch while CR2PQ benefits from a pre-trained teacher. More ablation studies are needed to separate the impact of each element.\n\n- Overall, the writing is difficult to follow, with multiple notation inconsistencies, typos, and signs of negative vertical spacing used to fit within the page limit.\n\n- Equation 1 is misleading/incorrect as it suggests that the representation of a single patch is independent of its context.\n- Equation 2: The angle of the key seems incorrect.\n- Line 210: The image dimensions are inconsistent with line 157.\n- Line 214: Inconsistent use of $\\mathbf{p}{a}$ and $\\mathbf{p}{A}$.\n- Line 234: The notation is inconsistent with the left side of Equation 3.\n- Table 1: Framwork $\\rightarrow$ framework.\n- Figure 1: There seem to be inconsistencies in the notations used within the figure and also with respect to the method section.\n- \"pertaining\" $\\rightarrow$ \"pretraining\"/\"pre-training\" (11 occurrences).\n- Line 86: exhausted $\\rightarrow$ exhaustive.\n- Line 161: $\\mathbf{W}{pos}$ $\\rightarrow$ $\\mathbf{P}^{i}{pos}$." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024crpq,\ntitle={{CR}2{PQ}: Continuous Relative Rotary Positional Query for Dense Visual Representation Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3l6PwssLNY},\nnote={under review}\n}" }, "abstract": { "value": "Dense visual contrastive learning (DRL) shows promise for learning localized information in dense prediction tasks, but struggles with establishing pixel/patch correspondence across different views (cross-contrasting). Existing methods primarily rely on self-contrasting the same view with variations, limiting input variance and hindering downstream performance. This paper delves into the mechanisms of self-contrasting and cross-contrasting, identifying the crux of the issue: transforming discrete positional embeddings to continuous representations. To address the correspondence problem, we propose a Continuous Relative Rotary Positional Query ({\\mname}), enabling patch-level representation learning. Our extensive experiments on standard datasets demonstrate state-of-the-art (SOTA) results. Compared to the previous SOTA method (PQCL), our approach achieves significant improvements on COCO: with 300 epochs of pretraining, {\\mname} obtains \\textbf{3.4\\%} mAP$^{bb}$ and \\textbf{2.1\\%} mAP$^{mk}$ improvements for detection and segmentation tasks, respectively. Furthermore, {\\mname} exhibits faster convergence, achieving \\textbf{10.4\\%} mAP$^{bb}$ and \\textbf{7.9\\%} mAP$^{mk}$ improvements over SOTA with just 40 epochs of pretraining." }, "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": [ "Self-supervised learning", "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/27b268506d9ce023797408062676ad8dc9be0dd5.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": "CR2PQ: Continuous Relative Rotary Positional Query for Dense Visual Representation Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. I believe that the \"cef\" measure in Table 1 doesn't provide a fair comparison, as there is no direct relation between communication cost and accuracy. \n2. It would be nice to see more experimental support, including diverse datasets and non-IID scenarios with different data heterogeneity levels (α = 0.05, 0.5, 0.1).\n3. Also the authors should consider to include one or two standard FL baseline like SCAFFOLD, FedProto, FedTGP, to better demonstrate method's superiority." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper proposes an innovative approach for federated learning, which dynamically initializes effective parameters for new clients and utilizes Learngene concept to reduce communication overhead and strengthen privacy.\n2. The results show that the performance of the proposed method is comparable with the baselines.\n3. The paper is well-structured." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript proposes a framework, called DFL2G, to address two main challenges in federated learning: (1) initialization of the client model parameters for new \"agnostic\" clients and (2) to reduce communication overhead between clients and server during training process. The framework consists of three modules: Learngene Smooth Learning, Learngene Dynamic Aggregation, and Learngene Initial Agnostic Model, to effectively address these challenges. Experimental results demonstrate that the approach effectively reduces communication cost while maintaining comparative classification accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of convergence proof and theoretical support.\n2. The experimental results are limited. Further the authors have not considered different heterogeneous settings in their experiments.\n3. There is no comparison with the baselines having similar objectives (e.g., FedProto, FedTGP)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "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": "1. **Seemingly Effective Reduction of Communication Costs:**\n \n D2FL seemingly lowers communication overhead in federated learning where instead of transmitting full model updates, local updates are compressed into lightweight \"learngenes,\" which are then shared with the server. For a fixed communication budget, the tradeoff is improved. This is shown in experimental work\n \n2. **Efficient Initialization of Agnostic Client Models:**\n \n The framework leverages accumulated knowledge from participating clients to generate and store learngenes in a central pool. When new or agnostic clients join the network, they can initialize their models by inheriting these learngenes, facilitating rapid and effective model initialization. \n \n3. **Improved Privacy Preservation:**\n \nBy avoiding the direct sharing of global models and instead using condensed learngenes, D2FL offers improved safety against standard gradient attacks unlike FedAvg. The authors also highlight that the \"privacy\" means defense against gradient based attacks only." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce **D2FL**, a novel method designed to address the challenge of initializing local models for agnostic clients in federated learning without necessitating the sharing of a global model. Leveraging the **Learngene paradigm**, D2FL focuses on the rapid initialization of agnostic models through the use of \"learngenes.\" These learngenes encapsulate essential model knowledge, allowing new or agnostic clients to initialize their local models efficiently by inheriting this distilled information. The primary claims of D2FL include reduced communication overhead and enhanced privacy compared to the standard Federated Averaging (FedAvg) approach. By minimizing the need to transmit large model updates and avoiding the distribution of a global model, D2FL aims to achieve more scalable and privacy-preserving federated learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Ambiguous Notation for Agnostic Clients:**\n \n The notation used to represent agnostic clients, particularly in lines 128-129, is unclear. \n \n2. **Scalability Concerns Due to Server-Side Storage Overhead:**\n \n The server maintains K cluster models, which introduces significant storage overhead. As the number of clusters increases, the storage requirements may become prohibitive, raising concerns about the scalability of D2FL in large-scale federated learning environments. This limitation is not adequately addressed or acknowledged in the paper. This is especially relevant when comparing with other baselines\n \n \n3. **Insufficient Explanation of the Likelihood Function for FIM Computation:**\n \n The **Fisher Information Matrix (FIM)** is utilized within the framework, but the paper does not explicitly explain the likelihood function used to compute it 202-203. \n \n4. **Complexity of the Learngene Concept:**\n \n As there are multiple procedures happening in the paper, the introduction and explanation of the Learngene concept are convoluted, making the paper difficult to follow. It required multiple reading to understand some concepts. The authors should simplify the presentation of this concept, possibly by providing more intuitive explanations or systematically develop concepts to improve comprehension.\n \n5. **Unclear Combined Loss Function:**\n \n In line 230, the paper presents a combined loss function where the same weight parameter λ controls multiple aspects of the loss. The interaction and impact of λ on different loss components are not clearly delineated. Also the ablation studies do not incorporate the impact of the hyper parameter adjustment of these seperate learngene and elastic gene loss functins\n \n\n\n 6. **Ambiguities in Experimental Figures and Tables:**\n \n **Figure 4:** The dataset and model used in this figure are not clearly specified. Additionally, the performance of D2FL in low epoch regions (e.g., epochs less than 10) is smaller than some baselines other methods that perform better under these conditions. This needs to be acknowledged.\n \n **Table 4:** The table does not include standard deviations. Furthermore, it fails to separately evaluate the impact of elasticity and the Learngene component, despite elasticity being a core component of the paper. Same hyper parameter controls both the loss function so it is difficult to establish the impact of these seperate loss functions. This omission makes it challenging to determine the individual contributions of each component to the overall performance.\n \n **Table 5:** Similar to Table 4, Table 5 lacks descriptive information about the datasets used and the statistical measures reported.\n\n7. **Absence of Theoretical Convergence Guarantees:**\n \n The paper does not provide any theoretical analysis or proofs to support the convergence of the Learngene-based initialization method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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": 1 }, "strengths": { "value": "- The paper proposes \"collaborating, condensing, initializing\" steps analogous to the Learngene paradigm. \n- The topic of dynamic agnostic federated learning is important. \n- The provided empirical results cover various settings and baseline methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies dynamic agnostic federated learning, specifically on initializing the client models (by using the learngene paradigm) and achieving better communication overhead while protecting the privacy of the models. They propose DFL$^2$G, which consists of smooth updating, dynamic aggregation, and initial agnostic model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Readability**: \n - There are many mistakes, both in the text and notations, creating obstacles for the reader. \n - $\\mathcal{X}_{k,i}$: why do you need $k$ here? The local datasets $\\mathcal{X}_i$ are not being clustered. \n - Eq.8: why multiplier? \n - [Line 240]: $\\sum_{l=1}^L \\xi_{k,i}^{(l)} = 1$. How does this sum up to 1? It does not seem to be valid. \n - [Line 229]: Overall, you have the following objective function: \n\\begin{equation}\n\\mathcal{L}\\_{all} = \\lambda \\mathcal{L}\\_{gen} + \\lambda \\mathcal{L}\\_{elg},\n\\end{equation}\nwhich gives \n\\begin{equation}\n= \\lambda \\mathcal{L}\\_{cls} (\\mathcal{X}\\_{k,i}) + \\lambda^2 \\|\\| \\theta\\_{k,i} - \\Theta\\_{k} \\|\\|_2 + \\lambda^2 \\|\\| \\theta\\_{k,i}^{'} - \\Theta_k ||_2, \n\\end{equation}\nand it has issues in the formulation. \n - Typos in lines: 198, 199, 201, 226 (what is the second loss function?), 243 (different subscripts), 272, 283 (why j? you can stick to k.), 313, etc. \n\n- Section 2.4. Problems in the SVD decomposition and formulation. How can you set the data dimension $d$ to 5? $d$ can not equal some other value than its original value. \n\n- Privacy analysis. For a fair comparison with other baseline methods, you need to leverage all available information to reconstruct the samples $\\mathcal{X}_i$. Since clients are sharing $V_i$'s with the server, which can aid your reconstruction objective you have (Eq. 12), using the iDLG objective solely is not fair; therefore, it raises a question regarding the results in the paper (Figure 5). \n\n- The number of local epochs is huge (line 335, local epochs = 10), which should not be the case in heterogeneous FL since it makes the clients overfit to their local data. \n\n- The proposal of a new metric. Why propose a metric if you use it only in one table (Table 1)? Also, it is better to see the Acc. measures in Table 1. \n\n- Performance curve comparison (Figure 4). The figure doesn't correspond to what is reported in the table, which questions the study's validity. Also, the proposed method has a high variance (deviation) compared to other methods, which doesn't necessarily mean the method outperforms others. The baseline methods do not improve, having a straight-line performance (FedLP, Flearngene). \n\n- Table captions should be on top. \n- Consider citing other works using \\citep{}." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "How you update the cluster was not specific in algorithm 1. As a new agnostic client join the network, it is added to the nearest cluster as stated in line 18 of Algorithm 1. However, as new clients involve the cluster should be updated. Or is it the cluster only built at the beginning?" }, "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 an innovative solution through the introduction of the Learngene framework. By integrating Learngene into the Dynamic Agnostic Federated Learning paradigm, the authors enable efficient model initialization and communication, particularly for agnostic clients that join the system dynamically.\n\nThe experimental results are compelling, demonstrating a significant reduction in communication costs while maintaining or even enhancing model accuracy. This highlights the framework's ability to improve both scalability and performance in federated learning environments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is aiming at addressing two key challenges in Federated Learning (FL): \n1) privacy leakage during client-server communication, and \n2) communication overhead in transmitting model updates. \nTo tackle these issues, the authors propose the Learngene framework for Dynamic Agnostic Federated Learning (DAFL). The Learngene framework introduces a mechanism for compressing model updates into learngenes, which capture the most important information while reducing data transmission and mitigating the risk of privacy leakage. Additionally, the framework supports dynamic client participation, allowing clients to join and leave the system flexibly without compromising performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Assume a one-shot dataset in the client. This assumption allows for efficient clustering and model initialization but may limit the framework’s flexibility in handling the common dataset with more samples. \n2) Lack of Dynamic Cluster Management: The paper does not address how to manage clusters when they become too large or too small. In cases of high data heterogeneity, more clusters are required to accurately represent the diversity among clients. However, the framework does not discuss mechanisms to dynamically adjust the number of clusters based on client performance, data distribution, or scalability concerns. \n3) Insufficient Privacy Guarantees: The paper does not provide strong privacy guarantees. The only implication we have based on your illustration is that \"iDLG cannot recover the feature $X \\in R^d $ given learngene\". \nMoreover, the privacy protection is questionable when considering the specifics of the Singular Value Decomposition used in the framework. your $X_i \\in R^{1\\times d}$, $X_i = U_i \\Sigma_i V_i^T$. $U \\in R^{1\\times1}$, $\\Sigma \\in R^{1\\times d}$ diagonal matrix. Therefore, there are only 2 unknown numbers to recover $X_i$. if ignoring the scale ($U \\in R^{1\\times1}$), there are only one number left to recover your $X_i$, which would be easy. \nBesides, The dimensions are not clearly explained for SVD here. Your $X_i$ should be a matrix $X_i \\in R^{1\\times d}$ \n\n\nPresentation: \n1) Your citation format is incorrect for the entire paper. In latex, most of your citations should be \\citep{}. and will be rendered \"FL (McMahan et al. 2017)\". \n2) Since you still have space, I suggest that your algorithm should be placed in the main body of the paper. Because it provides a more general view of how you integrate Learngene smooth learning, learngene dynamic aggregation, and learngene initial agnostic model into one framework.\n3) your algorithm line 4. The tilde of $\\theta$ is in the wrong place. \n4) #276 your mentioned $d=5$. Does this mean that your private data $X \\in R^d = R^5 $, If so, is this a typo here?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose DFL$^2$G, a \"Collaborating & Condensing & Initializing\" dynamic federated learning framework inspired by Learngene, aiming to achieve low-cost communication, robust privacy protection and effective initialization of agnostic models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024dflg,\ntitle={{DFL}\\${\\textasciicircum}2\\$G: Dynamic Agnostic Federated Learning with Learngene},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3l9NRfezlo},\nnote={under review}\n}" }, "abstract": { "value": "Dynamic agnostic federated learning is a promising research field where agnostic clients can join the federated system at any time to collaboratively construct machine learning models. The critical challenge is to securely and effectively initializing the models for these agnostic clients, as well as the communication overhead with the server when participating in the training process. Recent research usually utilizes optimized global model for initialization, which can lead to privacy leakage of the training data.\nTo overcome these challenges, inspired by the recently proposed Learngene paradigm, which involves compressing a large-scale ancestral model into meta-information pieces that can initialize various descendant task models, we propose a \\textbf{D}ynamic agnostic \\textbf{F}ederated \\textbf{L}earning with \\textbf{L}earn\\textbf{G}ene framework. The local model achieves smooth updates based on the Fisher information matrix and accumulates general inheritable knowledge through collaborative training. We employ sensitivity analysis of task model gradients to locate meta-information (referred to as \\textit{learngene}) within the model, ensuring robustness across various tasks. Subsequently, these well-trained \\textit{learngenes} are inherited by various agnostic clients for model initialization and interaction with the server. Comprehensive experiments demonstrate the effectiveness of the proposed approach in achieving low-cost communication, robust privacy protection, and effective initialization of models for agnostic clients." }, "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", "Low-cost Communication", "Learngene" ] }, "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/5c3f3b5292f82e612c8385a0db63b455a726e2ea.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": "DFL$^2$G: Dynamic Agnostic Federated Learning with Learngene" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How does ProDMP generate smooth trajectories based on the predefined conditions of the initial state? Please give detailed justification and explanation.\n\n2. Could the author provide a detailed explanation of how a meta-learning problem can contribute to simulating new scenarios?" }, "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 shows a clear motivation for initial state uncertainty and data limitation, which are all critical problems in related research fields.\n\n2. Consider the \"node-level latent features,\" which is, to the best of my knowledge, a novel method for solving such a problem.\n\n3. The results of the new simulation task in the paper are convincing for the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a graph network simulator for mesh-based simulation on material study. The framework is constructed on a meta-learning problem and applies conditional Neural Processes to address data limitations. This paper shows both qualitative and quantitative experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some methodology details are unclear, especially in the \"Probabilistic Dynamic Movement Primitives\" section and \"Meta-Learning and Graph Network Simulators.\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 model's architecture is not clearly explained, and it is unclear why certain modules are necessary. For example, from the results, it seems that MGN, even without history information, can surpass M3GN in performance. This raises questions about the value of incorporating historical information in M3GN. Moreover, the experimental results do not clearly demonstrate the necessity or advantages of using a meta-learning scheme. A thorough analysis on how meta-learning benefits model performance would be valuable, including ablation studies comparing model performance with and without meta-learning.\n2. The authors claim that the baseline MGN does not incorporate historical information, which appears inaccurate. In certain datasets, MGN does include history. For a fair comparison, the MGN baseline should also be evaluated with historical data to assess its impact on performance.\n3. The results section only reports the average MSE across all time steps. It would be helpful to provide a comparison of MSE over the number of prediction steps, as this would give insight into the model's performance stability over time as claimed in the paper.\n4. Based on Figure 3, the proposed M3GN method does not appear to use ground truth collider information. If this is the case, does the collider state being predicted by the mode? How accurate is the collider state prediction, especially when history steps are limited? Additionally, including collider ground truth (as in MGN) is actually intuitive and makes sense, as the primary goal of developing a simulation model is to understand how a solid deforms under varying contact forces and obstacle displacements. Predicting these external forces may not be necessary for achieving this objective.\n5. It would be informative to visualize the node-level latent task descriptions learned by the model. Such visualizations could help in understanding how task-specific information is represented.\n6. The datasets used in this paper have relatively small node counts compared to those in previous MGN studies or those used in other related papers. When the number of nodes increases significantly, it is concerned that M3GN may struggle due to the large number of historical steps required. Comparing M3GN’s memory usage with MGN’s would provide a more comprehensive evaluation. \n7. The authors consider each trajectory as a separate task with varying context sizes. However, this approach may not align with the broader goals of meta-learning, as tasks are typically defined by consistent properties such as the same material setting. Currently, the meta-learning setup seems more focused on adapting to different context sizes rather than generalizing across diverse tasks.\n8. As the input context size changes, will the number of predicted steps vary as well? If so, the model’s ability to generalize to different context sizes is unclear, and it may not be as flexible as MGN in this respect. Any experiments or evaluation on this aspect? Additionally, splitting single data points into multiple input-output sets seem to increase the effective amount of training data for M3GN, potentially creating an unfair comparison with MGN which use less training data.\n9. The authors do not specify how material properties are incorporated. Also, it is unclear whether the test data involve material properties that are in-distribution or out-of-distribution relative to the training data. Providing this information is crucial for evaluating the model's generalization capabilities.\n10. The authors mention that material node features are not added to M3GN. Given that these features enhance MGN's performance, it would be useful to understand the rationale for this exclusion and perform related ablation study.\n11. Although the authors mention other methods in related work besides MGN, these methods are not included in the baselines. Some of these methods have better accuracy and efficiency. Including these additional baselines would provide a clearer view of M3GN’s comparative performance.\n12. Will the data used in this study be publicly available? Making the dataset accessible would facilitate further research and replication studies." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper takes a novel approach to enhancing rollout stability by predicting entire future mesh states, and it incorporates a meta-learning scheme to improve adaptability within the simulation framework." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Movement-primitive Meta-MeshGraphNet (M3GN), a model for simulating object deformations in data-limited scenarios. M3GN combines meta-learning and movement primitives to improve the adaptability and accuracy of Graph Network Simulators (GNSs) by framing mesh-based simulation as a meta-learning task." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the approach appears novel, the rationale behind certain modules in the model is unclear, and the results do not provide sufficient evidence to justify their inclusion. Also, the paper is not clearly written and sometimes hard to follow. The detailed comments and suggestions are listed 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. What is the timestep for simulation?\n\n2. A figure illustraing all the relation and symbols of input, output can be added. Fig.3 Right is not information for undertanding the task setting." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This work aims to address two important problems in learning-based simulation:\n\n1. It treats the simulation as a trajectory-level meta-learing problem and use trajectory history as the context to predict future trajectories.\n\n2. It mitigates the problem of error accumulation by using ProDMP to directly predict the full simulation trajectories.\n\nThe paper is well structured and written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a graph network simulator that combines movement primitives and trajectory-level meta-learning. The network uses the simulation history as the context information to predict the deformation for objects with unknown properties. They also use probabilistic dynamic movement primitives to represent the future trajecteries and directly predicts the full simulation trajectories instead of iteratively predicting the next-step. Experiments show that it outperforms STOA in different simulation tasks. Abalation studies validate the effectivenss of the design choice." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some descriptions are unclear and some important details are missing.\n(1) in line 242, \"graph edges between the deformable object and collider are added based on physical proximity to model interactions\n between objects.\" what is the physical proximity exactly? Since the deformation mesh node position for the end timestep is unknown, I suppose we cannot use that to compute the distance. Whether this edge creation is done only for known timesteps or if it's updated during prediction?\n\n(2) in line 231, why is the term c_1y_1(t) + c_2y_2(t) only depending on the inital conditions? What is the representation of the pre-computed basis fuction \\phi?\n\n2. More detailed description of the training/val/test split should be added. Specify how trajectories are divided between training, validation, and test sets. What are different between training and test? Clarify if test trajectories involve different objects, material properties, or initial conditions than training trajectories. In the limitation part, it is claimed ''We currently consider each trajectory as a task, and require initial states of this trajectory as a context set during inference.\"\n\n3. Since the method needs a trajectory with simulated states as context, the author better include a runtime comparison between your method (including context computation) and traditional simulators for predicting the same number of future timesteps and discuss the trade-offs between computation time and accuracy compared to traditional simulators." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Strength:\n1. Adopting meta-learning to deal with dynamic prediction tasks is novel, especially the concept of regarding each tajectory as a new task is interesting. \n\n2. The authors consider past information and the eventual state of the collider as the condition to predict the subsequent movement trajectory, which make the network infer the future from the past rather than remember the dynamic behaviour of a certain material. In addition, predicting the whole rest path by a single forward pass could significantly improve the efficiency, compared with previous Graph-based single timestep prediction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper first propose a meta-learning framework to efficiently learn generalizable mesh-based dynamic prediction tasks. Different from previous graph neural simulators which predict the state updates in a step-by-step manner, the proposed M3GN targets to predict the whole trajectories by a conditional neural process to effectively diminish the error accumulation issue." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weakness:\n1. This paper is highly related to the Graph-based Neural Simulators. However, in the related work section, the latest advancements in this field are not included, and most of the work discussed is from 2023 or earlier. This could make the paper appear somewhat outdated. I believe this section could benefit from a more comprehensive overview of the field, especially more works from 2024. Below are two of the latest advancements about Graph Network Simulators that I recommend the authors to discuss them in Section 2.1 ,or better, use them as baselines for comparison. However, given the tight rebuttal timeline, it is also tolerated that concurrent works were not included for comparison.\n\n (1) \"DEL: Discrete Element Learner for Learning 3D Particle Dynamics with Neural Rendering\" 2024 .. This work integrate traditional Newton mechanics into the graph network design to benefit from mechanics priors for longer term prediction.\n\n (2) \"Equivariant graph neural operator for modeling 3d dynamics\" 2024 .. This paper deal with dynamic prediction tasks as trajectory-level rather than next-step level by operator learning, which is somewhat relavent with this reviewing work. Also, it handle the equivariant issues. \n\n2. For Equation 3, does it use past trajectory collider states when encoding z because I saw that you seem to only use the latest state, or does it rely solely on the historical information of the deformed object? I believe it would be more reasonable to use all the historical information of the collider here as well, since the deformation of the mesh is passive. \n\n3. If this method is trained on an elastic dataset, can it generalize directly to elastoplastic materials? I believe it would be worthwhile to discuss the generalization across different materials in the experiments, rather than limiting it to variations in mechanical parameters within the same material. \n\n4. Line 276 mentions that the context information z is concatenated with the node features. Is the same z concatenated to each node?\n\n5. Finally, the neural network predicts a set of weights, and the shape of the weight matrix is 𝑇, 𝐷,3. Which basis functions are these weights applied to in order to obtain the predicted trajectory? Are they precomputed from the historical trajectory? If yes, how?\n\nIn appendix A.2 \"Initially, we integrate a relative goal position as part of the node weights w\" What's the exact mean of the relative goal position? \n\nI will raise the score if most of concerns are well addressed by the authors." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a latent task-specific Graph Network Simulator, which improves over existing learned simulators by framing mesh-based simulation as a meta-learning problem." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024latent,\ntitle={Latent Task-Specific Graph Network Simulators},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3lDxKQepvn},\nnote={under review}\n}" }, "abstract": { "value": "Simulating object deformations is a critical challenge in many scientific domains, with applications ranging from robotics to materials science. \nLearned Graph Network Simulators (GNSs) are an efficient alternative to traditional mesh-based physics simulators. Their speed and inherent differentiability make them particularly well-suited for inverse design problems such as process optimization.\nHowever, these applications typically offer limited available data, making GNSs difficult to use in real-world scenarios. We frame mesh-based simulation as a meta-learning problem and apply conditional Neural Processes to adapt to new simulation scenarios with little data. In addition, we address the problem of error accumulation common in previous step-based methods by combining this approach with movement primitives, allowing efficient predictions of full trajectories. We validate the effectiveness of our approach, called Movement-primitive Meta-MeshGraphNet (M3GN), through a variety of experiments, outperforming state-of-the-art step-based baseline GNSs and step-based meta-learning 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": [ "Graph Network Simulators", "Graph Neural Networks", "Meta-Learning", "Neural Processes", "Deformable Object Simulation", "MeshGraphNets" ] }, "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/9e4c58013535c2e948cb54b42f34a4300aa2307e.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/8f6093e48c557d36cb1bee23fb67c7c20a50bee8.zip" }, "title": { "value": "Latent Task-Specific Graph Network Simulators" }, "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. The result of pairwise ConMix shown in Table3 on CIFAR-10 is better than the result of ConMix with M=500 presented in Figure 2.Is this reasonable? In my understanding, the former is equivalent to ConMix with a larger M on CIFAR-10.\n2. Have you conducted additional experiments on balanced models of other methods as ConMix-B to support the opinion about robustness?\nQuestion3. Are there other clustering methods being studied besides k-means?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThe author has conducted comprehensive experiments, compared with multiple clustering algorithms, and prove the effectiveness of ConMix under long-tailed distribution.\n2.\tReasonable theoretical analysis is given to verify that ConMix can implicitly achieve the loss-balance.\n3.\tContributions of different elements in ConMix are studied through extensive experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new method called ConMix for dealing with the long-tailed problem of deep clustering. A major challenge in long-tailed deep clustering is how to deal with class imbalance in a dataset without label information. ConMix solves this problem through an innovative approach to mixed representations in contrastive learning to enhance deep clustering performance in the case of long-tailed distributions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The representation synthesis part is supposed to be represented more intuitively, which may be a little confusing at first reading." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Conduct additional experiments on large and complex datasets like ImageNet to validate the effectiveness and generalization capability of ConMix.\n\nEnhance the discussion on method interpretability by providing more empirical analysis regarding its impacts.\n\nProvide detailed descriptions of the experimental setup and hyperparameter selections to improve transparency and reproducibility of the research." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The introduction of ConMix as a contrastive mixup method specifically designed for long-tailed deep clustering is a notable contribution to the field. The approach is innovative, extending mixup techniques into the realm of unsupervised learning.\n\nThe authors provide a theoretical foundation for their method, demonstrating how it can implicitly balance losses across head and tail classes. This theoretical insight is valuable and adds depth to the paper.\n\nThe evaluations on various benchmark datasets and the assertion of outperforming existing methods lend credibility to the proposed approach. The performance metrics presented seem robust." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel method, ConMix, aimed at addressing the challenges of long-tailed distributions in deep clustering. The authors argue that existing deep clustering approaches typically assume balanced class distributions, which is not the case in many real-world datasets. ConMix leverages a contrastive mixup strategy to enhance representation learning, theoretically proving its effectiveness in rebalancing class losses without the need for label information. The method is evaluated on benchmark datasets, demonstrating superior performance over existing state-of-the-art approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Diversity of Datasets: The experiments are limited to a few benchmark datasets, lacking validation of the method’s effectiveness on more complex and diverse datasets. It is recommended to conduct experiments on larger image classification datasets such as ImageNet to thoroughly evaluate the model’s generalization ability and practicality.\n\nInterpretability of the Method: Although theoretical proofs are provided, the interpretability of how ConMix specifically affects the model learning process remains insufficient. Consider adding comparative experiments to illustrate the specific impacts of ConMix under varying conditions (e.g., different long-tail ratios) to enhance the depth of the paper.\n\nDetails of Experimental Setup: The experimental section lacks detailed descriptions of hyperparameter choices and training specifics, which could affect the reproducibility of results. It is suggested to include these details in the methodology section to assist other researchers in understanding and replicating the experiments." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Could the authors elaborate on the statement in Line 220: “…. Equivalent to implicitly sampling different weights from the beta distribution”. Do the authors refer to the mixing coefficient in the original mixup formulation? While the reviewer understands that the cardinality of the set U_m follows a beta distribution, the final mixup representation will be the mean representations of these samples, which is different from the mixing coefficients. \n\nFurther, could the authors comment on the performance of the proposed approach in the balanced setting and on the results in Table 3 with regards to the pairwise mixup and the variability of the results?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Limited work has been done on considering imbalance in deep clustering, with most approaches adopting a balanced assumption, and approaches addressing this shortcomings are thus of significance to the community.\n\nWhile mixup on a representations has previously been integrated into contrastive learning (also mixing multiple samples), there is a certain novelty of leveraging this in the clustering setting to address class imbalance, which is further supported by the theoretical analysis.\n\nThe proposed approach is simple and appears to be effective in the settings considered in this work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes to leverage mixup to improve deep clustering approaches for imbalanced datasets. In particular, a multi-sample mixup is incorporated into the SimCLR loss. Instead of just contrasting two augmentations of the same sample, a random subset of samples are selected and the mean representations of their two augmentations are contrasted. A theoretical analysis is performed that shows, under a certain set of simplifications, that this procedure increases the loss of the underrepresented classes. Further, empirical evaluation demonstrate that the scheme can outperform alternative approaches in the imbalanced setting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the author’s main focus is on the imbalanced setting, it would be beneficial to also include comparisons in a balanced setting to be able to judge the overall ability of the method.\n\nThe overall clarity of Section 3.3. can be improved. How are the “stochastically assigned tags” selected? Does each sample have a certain probability of being included (independent of each other)? If that is the case, what is the probability set to? Also, in line 215, the notation of the cardinality of the set is not aligned with Eq. 3.\n\nThe comparison to pairwise ConMix (standard manifold mixup) in Table 3 is not clear. It appears that the pairwise mixup obtains equivalent results to ConMix w/o SDCLR warmup and it is unclear if pairwise ConMix leverages SDCLR warmup here. Also, is this pairwise ConMix directly leveraging the mean representation of the pair or do the authors create a convex combination with weights sampled from a beta distribution? \n\nAs deep clustering methods tend to be a bit less stable than supervised models, some measures of variability/statistical significance would be beneficial in Table 3." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We innovatively propose a ConMix method that can effectively address long-tailed deep clustering." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024conmix,\ntitle={ConMix: Contrastive Mixup at Representation Level for Long-tailed Deep Clustering},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3lH8WT0fhu},\nnote={under review}\n}" }, "abstract": { "value": "Deep clustering has made remarkable progress in recent years. However, most existing deep clustering methods assume that distributions of different clusters are balanced or roughly balanced, which are not consistent with the common long-tailed distributions in reality. In nature, the datasets often follow long-tailed distributions, leading to biased models being trained with significant performance drop. Despite the widespread proposal of many long-tailed learning approaches with supervision information, research on long-tailed deep clustering remains almost uncharted. Unaware of the data distribution and sample labels, long-tailed deep clustering is highly challenging. To tackle this problem, we propose a novel contrastive mixup method for long-tailed deep clustering, named ConMix. The proposed method makes innovations to mixup representations in contrastive learning to enhance deep clustering in long-tailed scenarios. Neural networks trained with ConMix can learn more discriminative representations, thus achieve better long-tailed deep clustering performance. We theoretically prove that ConMix works through re-balancing loss for classes with different long-tailed degree. We evaluate our method on widely used benckmark datasets with different imbalance ratios, suggesting it outperforms many state-of-the-art deep clustering approaches. The code has been submitted to the supplementary file." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "deep clustering", "long-tailed deep clustering", "unsupervised 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/469504b6d02f88e296ac2ca87feb399da7390ddf.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/bcc5091ad7aa2f2f6737e4206dba246c391f7bc2.zip" }, "title": { "value": "ConMix: Contrastive Mixup at Representation Level for Long-tailed Deep Clustering" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "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": "1. How does the algorithm's performance vary with different types of divergence functions beyond total variation, such as Kullback-Leibler divergence?\n\n2. Would the RTZ-VI-LCB framework be adaptable to handle more complex multi-agent settings with more than two players?\n\n3. How sensitive is the model’s performance to variations in the clipping parameter $C_r^*$, and what guidelines can be provided for choosing this parameter effectively?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper provides a rigorous theoretical framework, including upper and lower sample complexity bounds, which supports the robustness and efficiency claims of the RTZ-VI-LCB algorithm.\n\n2. The design of RTZ-VI-LCB is explained in a step-by-step manner, making it easy to follow the rationale behind each component, such as the use of lower confidence bounds and two-player-wise rectangularity.\n\n3. The paper adapts the robust Bellman equations specifically for two-player games, enhancing the clarity and relevance of the methodology in the context of MARL." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses robust multi-agent reinforcement learning (MARL) in two-player zero-sum Markov games (TZMGs) by introducing the RTZ-VI-LCB algorithm, a sample-efficient approach to handle offline settings with environmental uncertainties. The algorithm improves robustness by applying value iteration with data-driven penalties and establishing sample complexity bounds without requiring full state-action space coverage." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper assumes that both players in the two-player zero-sum game have identical uncertainty sets (same divergence function for both players). This simplifies the model but may limit its applicability to real-world scenarios where players could have different levels of uncertainty.\n\n2. The penalty term introduced in the RTZ-VI-LCB algorithm is crucial for the robust value estimation, but the paper does not clearly explain how the penalty is calibrated or how different choices of penalty function influence the algorithm’s performance.\n\n3. The paper assumes that historical datasets can be treated as independent samples after applying subsampling techniques, but it does not fully address the potential temporal dependencies within offline data." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why target two-player zero-sum games? Is there any special structure that helps the results, which hinders the authors from considering more general general-sum multi-agent games?\n\nOther minors:\n1) For presentation, as actually the max-player and min-player enjoys very similar formulation, algorithm update rules, and others, the presentation is a little bit redundant. It will be better to only write one time of them, such as equations 8(a), 8(b) can be represented as one if we let the min-player's everything be its negative version. The same for equation 9, 10, 18, the two terms in 22, and etc." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This is the first work that targets offline settings for robust MARL problems, which is an interesting topic.\n2. It provides both upper and lower bounds for understanding this problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work focuses on developing provable algorithm for distributionally robust multi-agent reinforcement learning in the face of environmental shift, in offline setting using only a history dataset. Considering two-player zero-sum games, it proposes RTZ-VI-LCB with an upper bound and a lower bound for this problem." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing and presentation need to be revised a lot. A lot of parts of the paper are similar to prior art. For instance, the two-fold sampling method in Algorithm 1 is almost the same as Algorithm 3 in [1]. Although cited the prior works, the algorithm needs to be rewritten entirely.\n2. The contributions are a little bit overclaimed from the reviewer's viewpoint. In line 104, this work claims that \"To the best of our knowledge, this is the first time optimal dependency on actions {A, B} has been achieved\". While the concentrability coefficient also involves potential terms of A and B. So it is better to also say this is only for offline settings.\n3. Some writing issues such as in line 107. The \"transition kernel\" does not need to be solved, it seems to need to be revised to \"RTZMG\". In line 113, \"across a range of uncertainty levels\", it seems there is something missing in this half sentence.\n4. In the discussion part after showing the theorems, the reviewer highly suggests that the author check the claims again. For instance, in line 511-512, it seems the upper bound and lower bound do not match in $H$ even if $\\min\\\\{\\sigma^+, \\sigma^- \\\\} \\geq \\frac{1}{H}$. The upper bound has $O(H^5)$, while the lower bound has $O(H^4)$? So it is not optimal yet, which is also claimed in the second paragraph of the discussion.\n\n[1] Li, Gen, et al. \"Settling the sample complexity of model-based offline reinforcement learning.\" The Annals of Statistics 52.1 (2024): 233-260." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. In remark 1, the paper mentions that the coefficient $C_r^\\star$ could be $\\frac{AB}{A + B}$. Given that the sample complexity result is $\\tilde{O}\\left(\\frac{C_r^*(A + B)}{\\epsilon^2}\\right)$, does this imply that the complexity is reduced to $\\tilde{O}\\left(\\frac{A B}{\\epsilon^2}\\right)$ in terms of $A$ and $B$, which is the same as the result in DR-NVI? \n2. How should we compare the term $\\min\\left(f(\\sigma^+,\\sigma^-),H\\right)$ in the upper bound and the term $\\min\\left(1/\\min(\\sigma^+,\\sigma^-),H\\right)$ in the lower bound? Additional discussion on this comparison would clarify the practical implications of the upper bound's tightness relative to the lower bound.\n3. From the similarity between the lower bounds of RTZMGs and RZMGs (Shi et al., 2024b), I assume that RTZMGs are not significantly easier than RZMGs. Given this, is it feasible to extend the RTZ-VI-LCB algorithm to more than two players?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-written and easy to follow. The problem of efficiently finding the robust NE in RTZMGs is significant for the field. The theoretical results are strong, as the sample complexity of the RTZ-VI-LCB algorithm nearly matches the lower bound for this problem class. Additionally, the lower bound analysis indicates that RTZMGs are not notably easier than traditional TZMGs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed an algorithm RTZ-VI-LCB, designed to efficiently find the robust Nash Equilibrium(NE) in Robust Two-player Zero-sum Markov Games (RTZMGs). The authors employ confidence bounds innovatively in the algorithm, enabling it to achieve a sample complexity close to the lower bound except for the order of the horizon." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are still some technical problems to justify, which will be discussed 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": 2 }, "primary_area": null, "questions": { "value": "Can you please clarify the new technical ideas in the proof as compared to the work of Blanchet et al and She et al?\n\nCan you please clarify the relationship of the algorithmic ideas to prior work, highlighting which components are natural extensions and which represent new algorithmic ideas?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The question of robust learning in strategic multi-agent settings has received considerable attention in recent years. This paper builds on recent work, providing a clear contribution by combining technical and algorithmic ideas from recent work to provide tighter and more general results than the state-of-the-art. \n\nAdditionally, the work provides interesting lower bounds, whose proofs provide important new insight for the area." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies robust two-player zero-sum Markov games. Recent papers have provided near optimal sample complexity bounds in this setting under partial and limited coverage of historical data individually, but cannot handle both settings simultaneously. This work provides an algorithm that can achieve near-optimal sample complexity under partial and limited coverage simultaneously while also providing information theoretic lower bounds." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The algorithmic novelty in the work is not clear. A core component of the algorithm is a natural extension of work by Li et al the setting in this paper. From my read, the key algorithmic novelty is primarily in the penalty term. \n\nThe technical novelty of the paper is not clearly presented. It seems that key components of the proof follow recent work (e.g. much of the work in step 1 of the proof follows closely the approach of Shi et al.). That said, there are certainly new ideas in the proof, it is just that the paper does not do a good job of highlighting the new techniques in the analysis. \n\nNumerical comparisons to the work of Blanchet et al and Shi et al are not included. Such comparisons would increase the potential impact of the paper and highlight the extent to which the improvement in theoretical bounds represent empirical improvements." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Based on the discussion of the paper's strengths and weaknesses, I have the following questions for the authors:\n\n1. The authors focus on the finite-horizon setting. Can the methodology presented in the paper be extended to analyze the infinite-horizon setting, as in [1]? Additionally, why did the authors choose to focus on the finite-horizon case rather than the infinite-horizon scenario?\n\n2. The algorithmic framework follows from [1]. What are the specific technical challenges in extending the techniques of [1] from standard zero-sum Markov games to robust zero-sum Markov games?\n\n\n [1] Yan Y, Li G, Chen Y, et al. Model-based reinforcement learning is minimax-optimal for offline zero-sum markov games[J]. arXiv preprint arXiv:2206.04044, 2022.\n\n [2]Shi, Laixi, and Yuejie Chi. \"Distributionally robust model-based offline reinforcement learning with near-optimal sample complexity.\" Journal of Machine Learning Research 25.200 (2024): 1-91.\n\n [3]Li G, Shi L, Chen Y, et al. Settling the sample complexity of model-based offline reinforcement learning[J]. The Annals of Statistics, 2024, 52(1): 233-260." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The primary strengths of this paper can be summarized in the following two aspects:\n\n1. This paper introduces the first algorithm that achieves optimal sample complexity with respect to the dependence on action spaces. \n2. The paper offers a comprehensive analysis of robust tabular zero-sum Markov games, presenting both upper and lower bounds on the sample complexity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a robust model-based algorithm for offline two-player zero-sum Markov games (RTZMGs), effectively addressing the challenges of learning under partial coverage and environmental uncertainty. The key contributions of the paper are as follows:\n\n- The authors introduce the robust tabular zero-sum Markov game framework by extending the standard tabular zero-sum Markov game to a robust setting. Under this framework, they propose a new algorithm, RTZ-VI-LCB, which integrates robust value iteration with a data-informed penalty term to estimate robust Nash equilibria.\n- The authors provide a finite-sample complexity analysis for RTZ-VI-LCB, demonstrating its optimal dependency on the number of actions. This represents the first set of optimal sample complexity bounds for RTZMGs.\n- The authors establish a lower bound on the sample complexity for learning RTZMGs, confirming the tightness of their upper bound and demonstrating the near-optimality of RTZ-VI-LCB across varying levels of uncertainty." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are two major weaknesses from my perspective:\n\n1. The authors do not discuss whether a Nash equilibrium exists under their definition of the robust zero-sum Markov game. It is well known that in robust Markov games, the existence of a Nash equilibrium can be affected by the choice of uncertainty sets and specific problem settings. Therefore, I believe it is essential to provide a discussion on the existence of Nash equilibrium within their framework.\n2. Another weakness is the limited technical novelty of the work. The presentation in Sections 3.1 and 3.2 closely resembles that of [1], and the overall methodology appears to be a direct combination of [1] and [2]. The primary contribution seems to be the incorporation of the two-fold subsampling trick from [3] to sharpen the sample complexity bounds." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We design an algorithm that first achieves the optimal upper bound under partial coverage and environment uncertainty in robust two-player zero-sum Markov games." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024sample,\ntitle={Sample Efficient Robust Offline Self-Play for Model-based Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3lXZjsir0e},\nnote={under review}\n}" }, "abstract": { "value": "Multi-agent reinforcement learning (MARL), as a thriving field, explores how multiple agents independently make decisions in a shared dynamic environment. Due to environmental uncertainties and fluctuations, policies in MARL must remain robust to tackle the sim-to-real gap. Although robust RL has been extensively explored in single-agent settings, it has seldom received attention in self-play, where strategic interactions heighten uncertainties. We focus on robust two-player zero-sum Markov games (TZMGs) in offline RL, specifically on tabular robust TZMGs (RTZMGs) with a given uncertainty set. To address sample scarcity, we introduce a model-based algorithm (*RTZ-VI-LCB*) for RTZMGs, which integrates robust value iteration considering uncertainty level, applying a data-driven penalty term to the robust value estimates. We establish the finite-sample complexity of RTZ-VI-LCB by accounting for distribution shifts in the historical dataset, without requiring for full state-action space coverage. To the best of our knowledge, we provide the upper bound in RTZMGs, which first achieves optimal sample complexity on the dependency of action spaces. Our algorithm is capable of learning under partial coverage and environmental uncertainty. An information-theoretic lower bound is developed to show that learning RTZMGs is at least as difficult as standard TZMGs when the uncertainty level is sufficiently small. This result confirms the tightness of our upper bound, which is near-optimal for the big uncertainty level, except for the horizon length." }, "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": [ "robust Markov games", "self-play", "distribution shift", "model uncertainty", "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/8c126e65d2bc6caf6732b97185caf0b5b8f18dbf.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": "Sample Efficient Robust Offline Self-Play for Model-based Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. What are the differences between square adversarial patches and physically realizable attacks? \n2. Why is it necessary to design defense algorithms specifically for these attacks, and what are the limitations of existing defense methods ?\n3. What is the purpose of designing a binary mask? Could you please explain?\n4. The location of the mask is randomly selected, and then gradient information is used to determine the final patch. What is the difference between this approach and selecting the mask first followed by a random selection of the patch? Is there any advantage to this method ?\n5. Why is the adversarial training method presented in this paper inferior to L_\\infty-bounded adversarial training when applied to clean data?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1.\tThe topic studied in the paper is practical.\n2.\tThe proposed method demonstrates a degree of generalization, as it does not rely on specific attack algorithms.\n3.\tThe proposed method is effective against common adversarial attack algorithms.\n4.\tThe experiments conducted are relatively comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Early efforts have primarily focused on defending against adversarial patches, leaving adversarial training (AT) against a broader range of physically realizable attacks underexplored. In this work, the authors address this gap by proposing a unified AT method to defend against various physically realizable attacks. They introduce PBCAT, a Patch-Based Composite Adversarial Training strategy, which optimizes the model by combining small-area gradient-guided adversarial patches with imperceptible global adversarial perturbations that cover the entire image. This design enables PBCAT to defend not only against adversarial patches but also against unseen physically realizable attacks, such as adversarial textures. Extensive experiments across multiple settings demonstrate that PBCAT significantly enhances robustness against various physically realizable attacks compared to state-of-the-art defense methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper lacks novelty.\n2. The authors should emphasize why standard adversarial training cannot effectively address physically realizable attacks and highlight the advantages of the proposed method presented in this paper. \n3. In lines 251-253, the authors' findings seem meaningless, as unlimited adversarial noise will inevitably lead to a decline in training performance.\n4. Although the training cost of PBCAT is comparable to that of standard training, it still demands additional computational resources due to the gradient post-processing steps (partial partitioning and 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": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the Weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The method is simple and effective.\n- The experimental results and ablation studies are convincing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a adversarial training method to defend against physically realizable attacks. Specifically, they propose a new adversarial patch attack and use them to train the model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It is curious that the proposed methods work for naturalistic patch attacks. Experiments on defending naturalistic patch attack will strengthen the paper.\n- No black-box experiments are conducted. For example, FastRCNN trained with the proposed method against different datasets and attacks using other surrogate models such as Yolo.\n- Hyper-parameter tuning and training time is a concern" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- The authors mention physically realizable attacks that extend beyond adversarial patches. Why should these represent distinct attacks if they are computed to fool the same model? Adversarial patches could potentially encompass also features of an adversarial t-shirt, as they are capable of generalizing and representing any potential adversarial texture. For instance, at the end of Section 2.3, the authors suggest that real-world adversarial patches may not generalize well to other types of physical attacks, why?\n\n- The adversarial training is applied only for inf-norm bounded attacks. It would be interesting to explore SOTA patch and texture attacks bounded on different norms. What about the robustness against L-2 Attacks? How much it is the model cpaable of etending the robustness against L-2 Attacks?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- As remarked by different experiments, the proposed method increases the robusteness over different attacks. \n\n- Overall I think that the results are quite intersting, it provides a quite large gap above other strategies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel adversarial training method designed to defend against various physically realizable attacks on object detection tasks. The perturbation for generating adversarial examples during training includes a global perturbation, constrained by an \nℓ-inf norm with a small budget applied across the entire image, and a local patch, randomly positioned within the bounding box. This local patch is composed of sub-patches, with only some selected to inject a larger budget constraint." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The approach may impact accuracy sometime, especially when dealing with large datasets like COCO, as shown in Table 5. However, the effectiveness in terms of improved robustness is noteworthy.\n\n- The authors could have added metrics on training costs in the table to better clarify possible efficiency with respect to other training strategies" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 unclear whether the patch selection calculation (i.e., the $l_2$ norm calculations) is performed on the clean image or the adversarial image (the one containing the adversarial patch).\n* Could you please clarify this?\n* Additionally, what is the rationale behind choosing a square-shaped mask?\n* Have you considered experimenting with different norms beyond the $l_2$ norm?\n\n2. In the model training, are the weights initialized to random values or pre-trained weights? If random initialization is used, the object detector may risk overfitting on the Inria dataset, which contains only a few hundred images. This could explain the inconsistencies observed between the results on MS-COCO and Inria.\n\n3. In line 345, the total number of sub-patches is set to $n^2=64$ , and in lines 238-239, you mention that the top half are selected, indicating that 32 patches are chosen. However, in the ablation study regarding the number of sub-patches used during the selection process (Table 3), only a single value (16) is presented as a portion of the sub-patches, since using 64 means utilizing the entire set. This leads me to infer that 16 is deemed the optimal value. Does using 32 sub-patches result in better performance? It would be beneficial to explore additional values in this experiment.\n\n4. Could you provide some insights into the results presented in Table 2, particularly concerning the \"Global\" component? I find it challenging to understand why the \"Global\" component enhances robustness against AdvTexture and AdvCat attacks, given the significant differences in perturbation styles between them. Additionally, why does robustness decrease against AdvTexture when the Patch and Partition components are added (Lines 3 and 4)?\n\n5. Following the above question, in line 465 it is stated that “Partition” denotes the patch partition strategy. What is the strategy other than “Gradient”? what does Line 4 in Table 2 mean?\n\n5. While I acknowledge that the paper focuses on patches attached to objects, it would be beneficial to evaluate the proposed approach against attacks that place patches in different locations (e.g., DPatch [5]) and to study the effect of the \"Global\" component on these attacks. Demonstrating the ability to mitigate the impact of such patches could significantly enhance the paper's contributions." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper addresses a practical yet underexplored topic: adversarial training for defending object detection models against realizable attacks.\n2. The evaluation setup is well-detailed, and the provided code ensures easy reproducibility.\n3. The proposed method achieves excellent performance in terms of adversarial robustness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce a patch-based adversarial training technique designed to improve the robustness of object detection models against both patch-based and more recent texture-based attacks. The method involves two types of perturbations: local perturbations applied to the attacked object and a global perturbation affecting the entire image. The global perturbation is aimed at enhancing the robustness against texture-based attacks. In their evaluation, the authors compare their technique to one adversarial training (AT) approach and several non-AT methods across three patch-based attacks. They also present ablation studies to assess the impact of various hyperparameters. Finally, the evaluation is extended to other object detection models to demonstrate the method's broader applicability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Incomplete literature review – while the authors state that there are no previous works that specifically propose patch-based AT for object detection, a more in-depth review of the literature would have revealed that techniques such as Ad-YOLO [1] and PatchZero [2] already exist (and should be compared to). Additionally, including comparisons to more recent non-AT methods (e.g., PatchBreaker [3], NAPGuard [4]) would strengthen the paper's overall contribution.‏\n\n2. Lack of novelty – The proposed method appears relatively simple, primarily combining existing techniques adapted for object detection without introducing substantial new contributions, aside from the patch partitioning and selection strategy.\n\n3. Experiments - While the authors conduct a relatively comprehensive evaluation, several aspects are lacking:\n\n* Models: Since the focus is on person detection, which typically involves real-time scenarios, the evaluation should prioritize low-latency models (e.g., one-stage detectors) rather than slower ones like Faster R-CNN. Including YOLO models, particularly the most recent versions, would have been more relevant, as they are widely used in real-time object detection.\n* \"Clean\" results: While the authors acknowledge the performance drop on clean images as a limitation, the degradation in accuracy is significant, especially when compared to (Li et al. 2023) in Tables A1, 5, and 6. This raises concerns about whether the improved robustness stems from a robustness-accuracy trade-off. A more fair comparison would require matching the AP on clean images across methods before assessing robustness. \n* Results discussion: The results are presented with limited interpretation. The discussion would benefit from addressing edge cases and explaining unintuitive findings (as highlighted in question 4 below).\n\n4. Presentation - the submission is held back by the writing quality, particularly in the method section, mainly focused around the partially existing formulaic descriptions. For instance, the number of selected sub-patches should be parametrized (with an accompanying equation or algorithm) to better align with the presentation of the ablation study in Section 4.3.2.\n\nMinor comments:\n- Algorithm 1 – the use of $m$ and $m_p$ is confusing.\n- The placement of the tables on Page 9 makes them hard to read.\n- Best “Clean” performance should also be marked with bold.\n\n[1] Ji, N., Feng, Y., Xie, H., Xiang, X., & Liu, N. (2021). Adversarial yolo: Defense human detection patch attacks via detecting adversarial patches. arXiv preprint arXiv:2103.08860.\n\n[2] Xu, K., Xiao, Y., Zheng, Z., Cai, K., & Nevatia, R. (2023). Patchzero: Defending against adversarial patch attacks by detecting and zeroing the patch. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 4632-4641).\n\n[3] Huang, S., Ye, F., Huang, Z., Li, W., Huang, T., & Huang, L. (2024). PatchBreaker: defending against adversarial attacks by cutting-inpainting patches and joint adversarial training. Applied Intelligence, 54(21), 10819-10832.\n\n[4] Wu, S., Wang, J., Zhao, J., Wang, Y., & Liu, X. (2024). NAPGuard: Towards Detecting Naturalistic Adversarial Patches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 24367-24376).\n\n[5] Liu, X., Yang, H., Liu, Z., Song, L., Li, H., & Chen, Y. (2018). Dpatch: An adversarial patch attack on object detectors. arXiv preprint arXiv:1806.02299.\n\n‏" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "The first work showing strong robustness against various physically realizable attacks" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024pbcat,\ntitle={{PBCAT}: Patch-Based Composite Adversarial Training against Physically Realizable Attacks on Object Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3lZd6eoPJz},\nnote={under review}\n}" }, "abstract": { "value": "Object detection plays a crucial role in many security-sensitive applications, such as autonomous driving and video surveillance. However, several recent studies have shown that object detectors can be easily fooled by physically realizable attacks, \\eg, adversarial patches and recent adversarial textures, which pose realistic and urgent threats. Adversarial Training (AT) has been recognized as the most effective defense against adversarial attacks. \nWhile AT has been extensively studied in the $l_\\infty$-bounded attack settings on classification models, \nAT against physically realizable attacks on object detectors has received limited exploration. \nEarly attempts are only performed to defend against adversarial patches, leaving AT against a wider range of physically realizable attacks under-explored.\nIn this work, we consider defending against various physically realizable attacks with a unified AT method. \nWe propose PBCAT, a novel Patch-Based Composite Adversarial Training strategy. PBCAT optimizes the model by incorporating the combination of small-area gradient-guided adversarial patches and imperceptible global adversarial perturbations covering the entire image. With these designs, PBCAT has the potential to defend against not only adversarial patches but also unseen physically realizable attacks such as adversarial textures.\nExtensive experiments in multiple settings demonstrated that PBCAT significantly improved robustness against various physically realizable attacks over state-of-the-art defense methods. Notably, it improved the detection accuracy by 29.7\\% over previous defense methods under one recent adversarial texture attack." }, "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": [ "adversarial robustness", "object detection" ] }, "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/092a436a6d1ec5d3fed0de6615f072af2eb4aa31.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": "PBCAT: Patch-Based Composite Adversarial Training against Physically Realizable Attacks on Object 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": 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. Interpretability of Scoring: Could the authors clarify how the gap between teacher and student outputs specifically reflects drug-likeness, possibly by linking it to characteristics like toxicity markers or functional groups?\n2. Hyperparameter Sensitivity: How sensitive is the model to masking ratios in atom/bond modeling tasks?" }, "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 offers a scalable approach for drug-likeness screening, with practical applications in drug discovery and unsupervised molecular learning.\n2. By using 2D molecular graphs instead of SMILES, the approach effectively reduces biases commonly associated with SMILES-based drug-likeness scoring.\n3. The method consistently demonstrates superior performance compared to baseline models, highlighting its robustness and effectiveness.\n4. The knowledge distillation approach proposed in the paper might be an effective way to address challenges with unbalanced datasets in drug discovery, where true positives are often limited." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents an unsupervised method for predicting drug-likeness in molecules that exploits 2D features of molecules. It uses a knowledge distillation approach with two models: a \"teacher\" model trained on a large dataset of molecules, which learns molecular topology through tasks like masked atom and bond prediction, and a \"student\" model trained only on real drugs. The student model mimics the teacher’s output on drug-like molecules but diverges on non-drug molecules, allowing for a drug-likeness score based on the difference between the models’ outputs. Experimental results show that this method outperforms existing models and is less affected by biases, offering a potentially more accurate way to determine drug likeliness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The scoring method relies solely on the difference between teacher and student models. Including additional criteria, such as molecule toxicity features, could improve robustness.\n2. While the model leverages 2D molecular graphs, drug effectiveness often depends on 3D molecular interactions with proteins, which this paper does not address as a limitation.\n3. To assess the model's true potential in drug discovery, testing on novel, unseen datasets and conducting out-of-distribution benchmarks would be valuable.\n4. In practical applications like drug discovery, an interpretability analysis would be beneficial to understand the model’s 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": 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 authors should better clarify what the novelty of this work is, also accounting for the comments above.\n- The authors should introduce more baselines, in particular focusing on:\n - State-of-the-art molecular generative methods (e.g., based on graph-based representations) used to estimate likelihoods, instead of SMILES-based.\n - Other self-supervised methods used to learn general chemical representations, and to define the chemical space.\n - Outlier detection methods used to define novelty. \n\nIn this context, the authors should better clarify the original contributions proposed by 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 topic of the paper is relevant, as an improved quantification of drug-likeness can accelerate the drug discovery process and enable other approaches.\n- The method is clearly explained.\n- Analysis and ablation studies help develop an understanding of the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on drug-likeness prediction based on chemical structures. Instead of framing the problem as a supervised task or likelihood-based estimation, this work proposes an approach based on self-supervised learning followed by knowledge distillation. The proposed method ios compared against several baselines, including supervised classification, likelihood-based (RNN), and QED. The approach is tested on multiple datasets. Multiple analysis and ablation studies are conducted, providing further insights into the results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main limitations of this work are related to its novelty, lack of baselines, and limited clarity on its overall positioning. \n- First of all, in the introduction and motivation, the paper distinguishes itself from other unsupervised-based approaches based on the fact that previous work leverages SMILES representations, while this work leverages 2D graphs. However, it is actually possible (and typically done) to compute likelihoods based on 2D graph representations. Indeed, this is typically one of the main ways graph (and molecule) generative methods are evaluated (see, e.g., Diamant et al., 2023 ICML). It is in general well known that for molecular tasks, graph-based representations outperform SMILES-based representations, both for supervised and generative tasks (see, for example, leaderboard https://ogb.stanford.edu/docs/leader_graphprop/). Therefore, using graph-based representations (which have been state-of-the-art for years) instead of SMILES-based representations does not seem to be particularly novel.\n- This paper introduced a self-supervised framework that appears to be very similar to previous work (see \"Evaluating Self-Supervised Learning for Molecular Graph Embeddings\", NeurIPS 2023 for some examples). In this context, the choice of the self-supervised model introduced in this paper appears not novel and arbitrary. \n- This method is framed as novel compared to existing methods based on outlier-based estimation. However, the proposed approach is actually an outlier estimation technique, given that the drug-likeness score is obtained as difference between a model trained on the whole chemical space, and a model trained only on drug-like (i.e., \"known\") molecules. Therefore, more advanced outlier estimation methods should be evaluated.\n\nOverall, it is not clear what the contribution and novelty of this paper is. Additionally, several critical baselines 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": 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 weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is well written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors proposed a 2D-based unsupervised drug-likeness prediciont method. They performed knowledge distribution by pretraining a teacher model on both positive and negative molecules and futher trained a student model on positive drug-like molecules only, and further minimized the embedding between teacher model and student model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are some major concerns about this paper:\n1. The performance of the RNN paper looks great enough according to Table 1, even in the BondError dataset proposed by the authors, the RNN performance is pretty great. I see the main disadvantage of the RNN method is about its bias on the SMILES length. Thus, the authors should proposed some new datasets which contain molecules with different scales of SMILES lengths. Even though the authors showed the comparsion between RNN and their method on different scales of SMILES lengths in Figure 5, which partially address this concern, it's still not that complete. And the authors didn't display the number of molecules for different lengths.\n2. The baseline methods are too weak. The RNN method was way too old. Even the SMILES-BERT was trained 5 years ago. I wonder if the authors would use any transformer for comparison.\n\nThere are some other minor concerns about this paper:\n1. The score is based on atom embedding level, why not consider bond embedding as well?\n2. Be more careful about the potential data leakage problem, even though it might be hard to avoid when there is pretraining stage. Consider scaffold split.\n3. There are two $L_{mam}$ in Formula (1)\n4. Only positive examples are used in the training of the student model, what if introduce some negative examples and perform constrastive 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": 1 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could you please consider a comparison with the baselines Molformer, Graphformer and ImageMol when they are finetuned on the druglikeliness prediction tasks?\n\nCould you please compare to the GraphMVP methods?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper was well written, the related works and the motivation behind their methods is well explained. \nThe idea of using the difference between the teacher models and student models likeliness prediction is interesting.\nInteresting analysis on the bias of RNN toward short sequences." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses drug-likeness prediction challenges and introduces a novel knowledge distillation approach. In this method, a teacher model is pretrained using 2D molecular graphs with atom/bond masking predictive modeling, trained on a large dataset comprising both drugs and non-drugs. The student model, by contrast, is trained solely on drugs, separate from the teacher's dataset. The final drug-likeness prediction is based on the difference in likelihood predictions between the teacher and student models. \n\nThe authors evaluate their method using standard benchmarks, comparing it to five baselines. The baselines include two classes: supervised approaches (QED, a graph neural network (GCN), and a recurrent neural network (RNN)) and unsupervised methods (GlocalKD and HimNet). In four subsets of the FDA-approved drugs dataset, the proposed approach significantly outperforms these baselines. An ablation study is also conducted to examine the contributions of pretraining and distillation, alongside analyses highlighting the RNN’s bias toward shorter drug molecules. While the source code is not yet available, the authors have committed to open-sourcing it in the future." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although this work targets a specific molecular property prediction task, it does not thoroughly discuss or compare against a substantial body of research in molecular representation learning. For instance, methods based on molecular fingerprints and GNNs, such as *ADMET Property Prediction through Combinations of Molecular Fingerprints* ([arXiv:2310.00174](https://arxiv.org/abs/2310.00174)), have shown strong results in ADMET prediction and could be readily adapted to tasks like drug-likeness prediction. Additionally, recent advancements in pretrained models—such as *Molformer* ([Nature](https://www.nature.com/articles/s42256-022-00580-7)), *Graphormer* ([GitHub](https://github.com/microsoft/Graphormer)), and *ImageMol* ([GitHub](https://github.com/HongxinXiang/ImageMol))—would be valuable baseline comparisons for the present study.\n\nThe novelty of the proposed pretraining tasks also appears limited, as atom and bond masking in graph pretraining has become a widely adopted approach. For example, *GraphMVP* ([OpenReview](https://openreview.net/pdf?id=xQUe1pOKPam)) employs similar masking strategies to pretrain GNNs, covering both 2D and 3D graphs, with masking applied to parts of 2D graphs as well.\n\nFurthermore, the source code has not been made publicly available, hindering reproducibility of the experimental results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024unsupervised,\ntitle={Unsupervised 2D Molecule Drug-likeness Prediction based on Knowledge Distillation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3lfSk8NWWp},\nnote={under review}\n}" }, "abstract": { "value": "With the research significance and application value, drug-likeness prediction aims to accurately screen high-quality drug candidates, and has attracted increasing attention recently. In this regard, dominant studies can be roughly classified into two categories: (1) Supervised drug-likeness prediction based on binary classifiers. To train classifiers, the common practice is to treat real drugs as positive examples and other molecules as negative ones. However, the manual selection of negative samples introduces classification bias into these classifiers. (2) Unsupervised drug-likeness prediction based on SMILES representations, such as an RNN-based language model trained on real drugs. Nevertheless, using SMILES to represent molecules is suboptimal for drug-likeness prediction, which is more relevant to the topological structures of molecules. Besides, the RNN model tends to assign short-SMILES molecules with high scores, \nregardless of their structures. In this paper, we propose a novel knowledge distillation based unsupervised method, which exploits 2D features of molecules for drug-likeness prediction. The teacher model learns the topology of molecules via two pre-training tasks on a large-scale dataset, and the student model mimic the teacher model on real drugs. In this way, the outputs of these two models will be similar on the drug-like molecules while significantly different on the non-drug-like molecules. To demonstrate the effectiveness of our method, we conduct several groups of experiments on various datasets. Experimental results and in-depth analysis show that our method significantly surpasses all baselines, achieving state-of-the-art performance. Particularly, the prediction bias of SIMILES length is reduced in our method. We will release our code upon the acceptance of our paper." }, "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": [ "Drug-likeness Prediction", "Molecule Representation", "Molecular Property Prediction" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f551d5df651e2426c956085f20f8b60d5e3f29a1.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": "Unsupervised 2D Molecule Drug-likeness Prediction based on Knowledge Distillation" }, "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": 1 }, "primary_area": null, "questions": { "value": "NA" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Using untrained message passing layers for GNN can be a computationally efficient approach, which is an important topic in the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work explores the use of untrained message passing layers in GNN for link prediction tasks. The authors showed that, experimentally, untrained message passing layers provides competitive performance when compared against fully trained layers for link prediction. The authors also provided a simple theoretical analysis to justify their claims." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weakness:\n\nOverall, the presented work gives the impression of an early draft, and I find it challenging to fully assess its contributions in its current form. Below are some clear issues:\n\n1. The presented “theoretical results” are poorly organized, and it is very challenging to judge its correctness given there is no clear distinction between the authors’ contributions and existing results. To be honest, I am not very sure what is the theoretical contributions provided by the authors. The results rely on oversimplified assumptions (e.g., orthonormality line 304); and the authors were linking random things together (e.g., PageRank line 348). It is very challenging for me to decipher what the authors what to convey here.\n\n2. It does not seem like the experimental results support the authors’ claim. In many cases, the untrained variant of the network performs very poorly, especially in Hits@100 dataset. If the authors want to claim the simplified network performs very well, the paper should be written as such.\n\n3. Lack of ablation experiments, intuitive results on synthetic datasets, or example results/visualizations on representative datasets.\n\n4. The presented figure 1 is of low quality. Maybe at least show the standard deviation across different runs?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Can you comment on why orthogonality would be expected in homophilic networks?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is written well and the presentation is good.\n\nThe analysis of the inter layer values to path-based measures is nice, though not unexpected.\n\nExperimental results have been carried out on the usual link prediction benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes untrained and linear message passing layers for graph neural networks for the task of link prediction. Theoretical analysis relates the values computed at the intermediate layers to path-based and random walk based connectivity measures. An assumption is made regarding the orthogonality of the initial node features. Experimentally, the method is shown to be comparable to trained and non-linear layers in a GNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The idea of untrained and linear layers (as acknowledged by the authors) has previously appeared for node classification in Wu et al, ICML 2019. So, the idea has limited novelty.\n\nThe assumption of orthogonality may not always hold especially under conditions of homophily where neighboring nodes have similar features.\n\nLink prediction has been studied in many works and the impact of another paper is 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": 3 }, "primary_area": null, "questions": { "value": "* While running the code I noticed that the inference for the untrained layers uses considerably more memory than even training the full GNN. So much so that I ran out of memory on my laptop. I’m wondering, could your code actually be used for much larger graphs?\n* The meaning of this sentence was a bit unclear to me: “ Since the simplified architectures consist of UTMP layers followed by a trainable linear layer, the consideration of UT models which do not include the linear layer also covers all possible ablation studies” Could you clarify?\n* Some other suggestions for improvements were already described together with the weaknesses. \n\nIn light of the weaknesses I mentioned, I tend towards a reject. Especially the empirical analysis has to be improved considerably." }, "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, this is the first application of untrained message-passing layers to link prediction.\n* The empirical results show that untrained layers perform reasonably well.\n* Some theoretical observations are included to complement the empirical analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates the use of untrained message-passing layers for link prediction tasks. Both a completely untrained model and a model with a trainable layer after the message passing layers are compared to standard trained message-passing layers. The authors provide theoretical observations to support their empirical analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Lack of novelty: Most of the work builds directly on top of Wu et al. and mirrors many parts of it. The architecture and setup are almost exactly the same, and even some claims, like the benefit of efficiency and interpretability, are taken straight from there. This is not to say that they are not true, but for example in the case of interpretability, there is not much evidence provided beyond the fact that the architecture is simpler. \n\n* The theoretical contribution in the paper feels somewhat limited, with unclear takeaways. Although the authors aim to support their work with theory, the analysis falls short of substantiating the core claims. For instance, the authors state that “Our theoretical analysis further provides insights into the effectiveness of widely used node initialization schemes such as one-hot encodings and random features.” However, as this theoretical analysis is restricted to feature vectors that are pairwise orthonormal - often true for one-hot encodings and random features - it doesn’t convincingly explain why these should be more effective than other potential initializations that lack this precondition. Additionally, this analysis seems somewhat peripheral to the main focus of the paper: comparing untrained and trained MPNNs. To strengthen this aspect, I suggest clearly outlining the theoretical contributions by structuring them into theorems with proofs and more directly linking them to the paper’s central claim.\n\n* The main point of the paper is the fact that untrained message-passing layers perform very well in comparison to their trained counterparts. To evaluate this properly, a good benchmarking framework is necessary that guarantees that the difference in predictive performance is coming from what the authors claim it’s coming from, and especially that the comparison is fair. This is where I have my main problem with this paper. No established benchmarking framework is used, and almost no attention is paid to the fact that link prediction tasks are notoriously hard to evaluate. I would recommend the authors to consult recent works like [1] and [2], which go into more detail on various problems, but let me name some important ones here: Looking at the provided code, it looks like negative sampling of edges is done randomly, which is likely to cause bad performance on these tasks. This is problematic because it’s not clear if untrained layers really perform that well in comparison or if the training procedure was just not good. From Table 2, I can tell that in several cases, the final trained linear layer (which was also trained with random negative sampling) performs worse than the untrained one. This is really surprising to me and could be due to the negative samples that were not chosen well. While looking at the code, I also noticed that the test dataset for the untrained variants is not the same as for the trained ones because the random link split in the dataset transform is initialized with `num_val=0.00, num_test=0.1` in contrast to `num_val=0.05, num_test=0.1` for the trained counterpart. While I don’t expect this one to make a huge difference, it just goes to show that the benchmarking is not done thoroughly enough to warrant the claims made in the paper. Getting link prediction right is actually quite hard, considerably more so than for node and graph-level prediction tasks, and a paper that builds on top of these results that much should put more scrutiny into it. My proposal is this: Use an existing benchmarking framework. This also makes it possible to compare the results to other papers and to run with more recent datasets from ogb, which are completely missing from this analysis. On a side note, I think that these would be quite important as they are bigger and could demonstrate the claimed scalability advantage of untrained layers.\n\n[1] Li, Juanhui, et al. \"Evaluating graph neural networks for link prediction: Current pitfalls and new benchmarking.\" Advances in Neural Information Processing Systems 36 (2024). \n\n[2] Zhu, Jing, et al. \"Pitfalls in link prediction with graph neural networks: Understanding the impact of target-link inclusion & better practices.\" Proceedings of the 17th ACM International Conference on Web Search and Data Mining. 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": "See weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper introduces untrained message-passing architectures, extending existing research from node classification to link prediction. This perspective is relatively novel and addresses the computational limitations of GNNs. The untrained models, by eliminating learnable parameters, are shown to be faster and more resource-efficient, making them suitable for large-scale applications.\n\n\n2. Theoretical results provide a deeper understanding of how UTMP layers approximate traditional path-based link prediction metrics (e.g., random walks and common neighbors), making the models highly interpretable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores the use of untrained message-passing layers (UTMP) for link prediction tasks in graph neural networks (GNNs). The authors propose simplifying GNN architectures by removing trainable parameters and nonlinear components, resulting in interpretable and computationally efficient models. Their research finds that these simplified architectures can often outperform or match the performance of fully trained GNNs in link prediction tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited baselines are compared. Path-based methods and edge-wise methods should be compared. This doesn't mean the authors should change them into untrained models and do comparison. The author discuss the theoretical relationship with path-based methods, so the emperical comparison is needed to validate the theory.\n\n2. Limited datasets are included. Large datasets like OGB datasets are not included so the application is limited. \n\n3. For non-attributed graphs the results of original GNNs are much better than on the attributed graphs, compared to S-models and UT-models. Can the authors do some ablations to discuss this observation? Maybe it's because of the one-hot encoding? Can the authors show the results of one-hot encoding in attributed graphs?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Untrained message passing layers in graph neural networks outperform trained counterparts for link prediction, offering efficiency and interpretability, especially with high-dimensional features." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024link,\ntitle={Link Prediction with Untrained Message Passing Layers},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3llRc6oXEW},\nnote={under review}\n}" }, "abstract": { "value": "In this work, we explore the use of untrained message passing layers in graph neural networks for link prediction. The untrained message passing layers we consider are derived from widely used graph neural network architectures by removing trainable parameters and nonlinearities in their respective message passing layers. Experimentally we find that untrained message passing layers can lead to competitive and even superior link prediction performance compared to fully trained message passing layers while being more efficient and naturally interpretable, especially in the presence of high-dimensional features. We also provide a theoretical analysis of untrained message passing layers in the context of link prediction and show that the inner product of features produced by untrained message passing layers relate to common neighbour and path-based topological measures which are widely used for link prediction. As such, untrained message passing layers offer a more efficient and interpretable alternative to trained message passing layers in link prediction 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": [ "graph neural networks", "untrained message passing layers", "link prediction", "path-based similarity measures" ] }, "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/7270511aba665d3ddc04383bc3048715e869dc40.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": "Link Prediction with Untrained Message Passing Layers" }, "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. I wonder whether the proposed method is effective for other types of scenes. Exploring and presenting results on different scene categories would be beneficial.\n2. I suggest to incorporate user studies to enhance the robustness of the experimental validation.\n3. The appendix seems to be directly attached after the references. It might be more appropriate to format the appendix according to the conference guidelines, ensuring it is properly integrated or submitted as a supplementary document as required." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed Prompt Balance and Characteristics Prominence strategies contribute to enhancing the quality of generated scenes, particularly in handling complex scenes with multiple instances.\n2. The method is relatively simple and does not require additional training, making it efficient and practical." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a Training-free Triplet Tuning (T³-S2S) method for sketch-to-scene generation, aiming to enhance the quality of generated scenes from sketches without additional training. The authors identify challenges in existing diffusion models related to imbalanced prompt energy and value homogeneity in the cross-attention mechanism, which lead to missing or coupled instances in complex scenes. To address these issues, they introduce two strategies: Prompt Balance and Characteristics Prominence. Additionally, they incorporate Dense Tuning from Dense Diffusion to refine attention maps. The proposed method is evaluated qualitatively on game scene generation tasks, demonstrating improvements in generating detailed, multi-instance scenes that align with input sketches and prompts." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed Prompt Balance and Characteristics Prominence strategies appear to be incremental improvements on existing techniques such as TopK methods. Dense Tuning is adapted from Dense Diffusion. Therefore, the paper may lean more towards engineering optimization rather than presenting significant novel techniques.\n2. The experimental comparisons use models based on different versions of Stable Diffusion, with the proposed method and T2I Adapter using SDXL (a more advanced model), while Dense Diffusion is based on SD v1.5. Since SDXL offers higher generation quality than SD1.5, this results in an unfair comparison. Considering that this work draws inspiration from Dense Diffusion, it would be more appropriate to either adapt Dense Diffusion to SDXL or apply the proposed method to SD1.5 to ensure a fair evaluation.\n3. The experiments focus primarily on game scenes, with limited variety in the types of scenes presented. Some examples are repeated in the paper, and the lack of diverse examples may not fully demonstrate the method's generalizability to other contexts. \n4. The paper focuses heavily on qualitative analysis and lacks quantitative experiments. \n5. Line 360 and 361: citations of SDXL and ControlNet are incorrect." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Since this method can enhance the performance of existing controllable generation models without the need for training, is it possible to include experimental analyses based on more diverse existing models, such as the T2I-Adapter (https://arxiv.org/abs/2302.08453) that the article compares?\nIf additional experiments could be added to demonstrate this, it would better showcase the universal applicability of the method proposed in the article, significantly enhancing its contribution.\n\n2. Can the method proposed in this article be applied to a wider range of scenarios? From the experimental results presented in the article, it seems that simply altering the prompt should be sufficient to produce a more diverse array of scene images. I hope to see more experimental outcomes. Is it possible to incorporate a wider variety of input types (for example, black and white hand-drawn sketches)? Can the output be expanded to include a richer variety of scene images (for example, real-world images), not just limited to the isometric view of game scenes?\nParticularly, the experimental results in Figure 6 of the article suggest that the method's control over the style of generated images should stem solely from the specific input prompt \"Isometric view of game scene.\" Therefore, if the article could showcase more use cases across a variety of application scenarios, it would greatly increase the value of the method proposed in the article." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This article addresses a very meaningful area, utilizing the approach of scene generation through sketches.\n\n2. The method requires no training and can be seamlessly integrated into existing controllable generation models, thereby enhancing the models' generative performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This article primarily focuses on the task of sketch-to-scene generation. Specifically, for complex scenes with multiple detailed objects, previous methods sometimes miss small or uncommon instances. Addressing this issue, the article proposes a training-free triplet adjustment method for sketch-to-scene generation. This method can be directly applied to existing SDXL and ControlNet models, enabling them to effectively tackle the multi-instance generation problem, including prompt balancing, feature highlighting, and dense adjustment. The proposal is made after reviewing the entire cross-attention mechanism. This solution revitalizes the existing ControlNet model, allowing it to effectively handle multi-instance generation, involving prompt balancing, feature highlighting, and dense adjustment. Experiments demonstrate that this triplet adjustment method enhances the performance of existing sketch-to-image models, enabling the generation of detailed, multi-instance 2D images that closely follow input prompts and enhance visual quality in complex multi-instance scenes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experimental content of this article is somewhat lacking in depth. From the presented experimental results, it only includes the generation of isometric view of game scene using colored sketch lines as control inputs. This raises the question of whether the method proposed in this article has universal applicability to scene sketch generation tasks.\n\n2. According to the description in this article, this training-free method at the feature level should be universally applicable to various controllable generation models. The method proposed in this article is based on ControlNet, but the article does not sufficiently discuss whether this approach would still be effective when integrated into other controllable generation 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": 2 }, "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": "1. The paper provides a thorough analysis of text embeddings and the value matrices in cross-attention, revealing the issues of prompt energy imbalance and value matrix homogeneity.\n\n2. The proposed method seemly addresses the challenges associated with sketch-to-scene generation involving multiple subjects." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper identifies a phenomenon in sketch-to-scene generation where certain instance are not effectively represented in complex, multi-instance scenes. The authors attribute this issue to the imbalance of prompt energy and the homogeneity of value matrices. To address these challenges, the paper proposes a method that incorporates prompt balance, characteristic prominence, and dense tuning for sketch-to-image generation in complex scenes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My primary concern lies with the clarity of the writing and the evaluation methodology:\n\n1. The writing lacks clarity and the notation system is somewhat confusing.\n2. Although the method does not require training, the optimization during inference may increase time demands, yet the paper does not report specific inference times.\n3. The results are solely qualitative, lacking quantitative comparisons. The effectiveness of the proposed modules is not validated with measurable metrics, raising concerns about potential cherry-picking in qualitative assessments. While it may be challenging to establish reasonable quantitative benchmarks for this task, there even is no mention of conducting user studies to support the findings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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) The prompt balance mechanism could also be applied to text-to-image models. I recommend testing the effectiveness of the prompt balance mechanism on some multi-instance text-to-image benchmarks. Specifically, you could compare it with training-free optimization algorithms designed for multi-instance T2I scenarios, such as AAE [1] and StructureDiff [2].\n\n(2) When performing the characteristics prominence mechanism, the paper states that the optimal setting for K is 2. When the number of instances exceeds K, how does your characteristics prominence mechanism ensure that each instance is effectively enhanced? More specifically, how does it handle multiple instances that belong to the same category but have different fine-grained attributes? For example, when dealing with five cats—one red, one blue, one green, one black, and one white.\n\n(3) I have some questions regarding specific details. When performing characteristics prominence, should the operation be applied on cross_attention(I, T) or on I + cross_attention(I, T), where I is the image feature and T is the text embedding?\n\n(4) I believe further experiments on widely used multi-instance generation benchmarks are necessary to better demonstrate the effectiveness of your method.\n\n[1] Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models.\n\n[2] TRAINING-FREE STRUCTURED DIFFUSION GUIDANCE FOR COMPOSITIONAL TEXT-TO-IMAGE SYNTHESIS." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Methodologically, T3-S2S demonstrates certain innovations compared to previous approaches, and the results presented by the authors indicate that this method indeed yields notable improvements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an approach, termed Triplet Tuning for Sketch-to-Scene (T3-S2S), a training-free method designed to enhance T2I generation capabilities in multi-instance scenarios, effectively mitigating challenges such as instance omission. The T3-S2S method introduces a prompt balance mechanism to automatically balance the energy of each instance within text embeddings and incorporates a characteristics prominence mechanism that strengthens cross-attention by highlighting Top-K indices within each channel, ensuring that essential features are more robustly represented based on token sketches. Methodologically, T3-S2S demonstrates certain innovations compared to previous approaches, and the results presented by the authors indicate that this method indeed yields notable improvements. Therefore, at this stage, I would rate this paper as [weak accept]. However, I believe the current experimental results are still insufficient; please see the question section for details. I look forward to the authors' responses." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I believe the current experimental results are still insufficient. First, the paper does not include comparisons with state-of-the-art (SOTA) methods on popular benchmarks. Additionally, it lacks quantitative comparison results, which are essential for a comprehensive evaluation of the method’s effectiveness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024tss,\ntitle={T\\${\\textasciicircum}3\\$-S2S: Training-free Triplet Tuning for Sketch to Scene Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3m6VqesEMw},\nnote={under review}\n}" }, "abstract": { "value": "Scene generation is crucial to many computer graphics applications. Recent advances in generative AI have streamlined sketch-to-image workflows, easing the workload for artists and designers in creating scene concept art. However, these methods often struggle with complex scenes with multiple detailed objects, sometimes missing small or uncommon instances.\nIn this paper, we propose a Training-free Triplet Tuning for Sketch-to-Scene (T$^3$-S2S) generation after reviewing the entire cross-attention mechanism. This scheme revitalizes the existing ControlNet model, enabling effective handling of multi-instance generations, involving prompt balance, characteristics prominence, and dense tuning. \nSpecifically, this approach enhances keyword representation via the prompt balance module, reducing the risk of missing critical instances. It also includes a characteristics prominence module that highlights TopK indices in each channel, ensuring essential features are better represented based on token sketches. Additionally, it employs dense tuning to refine contour details in the attention map, compensating for instance-related regions.\nExperiments validate that our triplet tuning approach substantially improves the performance of existing sketch-to-image models. It consistently generates detailed, multi-instance 2D images, closely adhering to the input prompts and enhancing visual quality in complex multi-instance scenes." }, "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": [ "Sketch-to-scene generation", "training-free diffusion model", "cross-attention mechnism" ] }, "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/07c94f04ac11a467d698d51b9d76862811d9d612.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": "T$^3$-S2S: Training-free Triplet Tuning for Sketch to Scene 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": 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": "Listed in the weaknesses section itself.\n\n- Make section 3 more clear by adding more details.\n- Give better reasoning for using the 3 dimensions of naturalness, coherence and plausibility.\n- measure the reliability of GPT-4o evaluations." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Originality: The paper presents a novel approach to simulating human-like daily activities using a desire-driven framework inspired by Maslow’s hierarchy of needs. Unlike traditional AI agents that rely on specific instructions or task-based rewards, the Desire-driven Autonomous Agent (D2A) framework introduces intrinsic motivation as the driving factor. This approach is unique in that it models a human-like motivational system, enabling the agent to select actions autonomously based on intrinsic desires rather than predefined goals.\n\nQuality: The paper is good quality with descriptive figures and clear results.\n\nClarity: The paper is well structured with a clear distinction provided between their proposed method and past methods and how theirs performs better.\n\nSignificance: This work holds significance for fields focused on human-like AI, agent-based simulations, and real-world applications requiring adaptive behaviour. By adopting an intrinsic motivation model, this framework could be used for applications such as social robotics, interactive gaming, and assistive technology, where human-like adaptability and engagement are essential." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel framework, the Desire-driven Autonomous Agent (D2A), designed to simulate human-like behaviour in daily activities. Traditional task-oriented AI agents primarily operate based on specific instructions or external rewards, but this approach often limits their ability to display intrinsic motivations similar to humans. This paper proposes an alternative: a motivation system inspired by Maslow’s hierarchy of needs, enabling the agent to autonomously generate and select activities based on desires like social interaction, self-care, and personal fulfilment. The D2A framework uses a system of 11 dimensions representing different human desires. The agent evaluates its current state and intrinsic motivations to decide on activities that align with its desires, generating more coherent and varied behaviours compared to existing models. The study uses Concordia, a text-based simulator, where a Game Master provides the environmental context for D2A to interact in a simulated household environment. The results suggest that D2A successfully simulates human-like activities by aligning with intrinsic motivations, which opens new avenues for developing desire-driven agents in various applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are a few weaknesses.\nIn Section 3, Problem Formulation, the math is not very clear, and it is also not explained in more detail how the activity distribution could be generated.\n\nIn Section 6.3.1, Naturalness, Coherence and Plausibility are used to evaluate the activity sequences, but these three dimensions seem to have been picked arbitrarily and I am not sure if they are enough to rigorously test the outputs.\n\nEvaluation is done using GPT-4o but how are we to ensure that these evaluations can be taken at face value. I think the paper would do well to do a human verification of these evaluations and how reliable they are." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "* My foremost question would be about the practical use of generating human daily activity data. I do not know this subject, and it might be the focus of an entire subcommunity which I am not familiar with. It would be important for the authors to elaborate on this point, and I am ready to reconsider my score if the demonstration is convincing;\n* What do the statistics of the generated activities look like?\n* Is there a list of predetermined activities one may do? (predefined action space?)\n\n## Notes\n\n* I think that the naturalness of activities might also come from the fact that desires are satisfied for quite some time giving the agent the opportunity to concentrate on more diverse activities than in the baselines.\n\n## Suggestions\n\n* I believe this paper should cite Park et. al 2023 (https://arxiv.org/abs/2304.03442), a seminal work in human behavior simulation, and Colas et. al. 2023, which presents an intrinsically-motivated agent operating on other principles than Maslow's hierarchy (https://arxiv.org/abs/2305.12487)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "## Originality\n\nThe paper is quite original, as I have not seen Maslow's hierarchy of needs used in the context of a text agent.\n\n## Quality\n\nThe paper is well-written, experiments are quite well-designed, several seeds are provided to account for variability. The figures are nice, and the main one does a good job of summarizing how the agent works. The results of the paper support the claims made in the introduction and the abstract.\n\n## Clarity\n\nThe paper was easy to follow and the points are clear and easy to grasp.\n\n## Significance\n\nThe paper demonstrates a recipe for creating more realistic human daily trajectory activities. I can see the type of agent developed here being useful for other applications, such as creating LLM-based non-playable characters in video games." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is concerned with simulating realistic trajectories of human activities in household environments. The authors introduce a text-based home environment as well as an LLM-based agent whose outputs aim to be activity traces as diverse as possible.\n\nInspired by Maslow's theory of needs, their agent incorporates 11 desire scalars (with each a target value, variable across agents). These desires are split according to the levels of Maslow's hierarchy. The desired levels, current levels, previous activities, and other environment information is provided to the LLM-agent in a tree-of-thought framework to generate the next activity.\n\nThe authors find the generated trajectories are deemed more likely by judge LLMs, and decrease dissatisfaction compared to other LLM-agent baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* I am not completely convinced of the end-goal of this paper, specifically, building sequences of human activities. I see the authors justifying this goal in the potential for generating data for psychological, economic or sociological academic study. However, the validity of the generated behavior with respect to at least one downstream application is not investigated in the paper. How to make sure the data generated is useful in these contexts?\n* The introduction also briefly argues that building agents with behaviors aligning with human ones will guarantee their intelligence (the Turing test argument). But unfortunately this does not seem to be the case; I cannot see how giving agents simulated desires will make them score higher on GSM8K for instance.\n* A human-judge evaluation of the validity of the AI judge would be nice (although I am still pretty convinced of the comparison results)\n* I think there is a methodological flaw in the design of the human reference agent: the human is (as far as I understood) given a list of the desires, and the criteria is whether the agents are able to minimize dissatisfaction for those same desires. A better reference would be a dataset of real human activities on a stay-at-home day;" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "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": "- The proposed framework allows agents to operate based on intrinsic motivations, which is a significant departure from existing task-oriented AI agents that rely on explicit instructions or external rewards. \n- The paper is well written and has nice figures.\n- The authors conducted a comprehensive comparative analysis with three baseline approaches (ReAct, BabyAGI, and LLMob) to evaluate the effectiveness of their framework. The development of a flexible text-based activity simulator using Concordia components is another strength of the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a Desire-driven Autonomy framework for LLM-based agents to simulate human-like daily activities. The framework is inspired by Maslow's theory of needs and includes 11 dimensions of human-like desires. The Desire-driven Autonomous Agent (D2A) operates based on intrinsic motivations, autonomously proposing and selecting tasks that fulfill its motivational framework. The authors developed a flexible text-based activity simulator using Concordia components, supporting various agent types and textual environments for reliable interaction and evaluation. They conducted simulations in a detailed textual home environment with a Game Master providing relevant observations. The experiments demonstrated that D2A generates appropriate activities effectively and efficiently, achieving described desires. A comparative analysis with three baseline approaches (ReAct, BabyAGI, and LLMob) showed that D2A generates more natural, coherent, and plausible daily activities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I think there are some weaknesses in this paper:\n- Limited Technological Innovation. The paper primarily focuses on the conceptual framework and theoretical underpinnings of the desire-driven autonomy approach. While the idea of using intrinsic motivations inspired by Maslow's theory of needs is innovative, the technological implementation details might not be as groundbreaking or novel within the recent advancements in the field of AI and LLMs. The work seems in the flow of LLM agents, while I think it is more of a LLM project rather than a technologically-solid paper.\n- Although the paper demonstrates the effectiveness of the D2A agent in a specific textual home environment, there might be questions about its generalization and scalability to other environments or domains. \n- The authors have established a set of concepts, such as human needs, desires, characteristics, and values, to guide the model's behavior. Although I understand the authors' intent to direct the generation of human-like daily activities, I am uncertain about the definition and composition of these intermediate variables. They are determined by the authors without sufficient psychological support or experimental validation to substantiate the overall design. How is the overall pipeline designed? I.e., for the five-level Maslow model, why you further define 11 desire dimensions?\n- The experiment is limited. The testing environment is confined to a single room containing a kitchen, living area, bedroom, and bathroom. It is unclear whether there is any randomness or variation between each epoch's setup, such as the rooms in the house or the items within the room. The experiments should be conducted in a wider variety of settings to minimize the impact of environmental bias and to get general ideas. How many different scenes or settings are included in the experiment? Also, for LLMs, have the authors studied how the prompt design will influence the overall 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": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please answer the issues mentioned in the weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**1. Proactive Action Based on Intrinsic Motivation**: The proposed D2A framework demonstrates an ability to proactively initiate actions driven by intrinsic motivations. Through the integration of a value system and a desire-driven planner, the framework establishes a dynamic interaction between desires and actions, wherein lowered desire values trigger corresponding activities to restore balance. This mechanism, though relatively simple, allows the agent to engage autonomously in daily activities in a manner that mirrors proactive human behavior, setting it apart from purely reactive or task-driven models.\n\n**2. Human-Inspired Intrinsic Motivations Across Life Dimensions**: Unlike recent approaches that focus on intrinsic motivations for exploration or collaboration, the D2A framework offers a multi-dimensional model inspired by human needs. By integrating eleven desire dimensions (e.g., physiological, social, and self-fulfillment needs), D2A provides a broader, more human-like motivational structure. This approach goes beyond typical reward-driven or exploratory motivations by simulating daily life activities that dynamically balance internal desires, reflecting human motivation patterns more authentically. This novelty enhances the agent's potential for replicating realistic, human-inspired behaviors within single-agent environments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the Desire-driven Autonomous Agent (D2A), a framework for simulating human-like daily activities based on intrinsic motivations rather than explicit tasks or external rewards. Inspired by Maslow's Theory of Needs, D2A prioritizes actions that fulfill a hierarchy of desires (e.g., physiological, social, self-actualization), allowing it to autonomously select actions that align with its motivational framework. This desire-driven approach contrasts with traditional task-oriented agents by focusing on fulfilling internal motivations, which provides the agent with the capacity for more adaptable and human-like behavior." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**1. Inconsistent Application of Maslow’s Theory**: While D2A is presented as being inspired by Maslow’s Theory of Needs, it does not implement the theory’s hierarchical structure. According to Maslow, higher-level desires are pursued only once lower-level needs are met, but D2A treats each desire independently, allowing the agent to pursue higher-level needs without satisfying foundational ones. This weakens the theoretical foundation and could make the agent's behavior feel less authentically human-like.\n\n**2. Overly Simplistic Desire-Action Dynamics**: Despite its innovative multi-dimensional motivational structure, the D2A framework uses a straightforward linear deduction of desire values to simulate the fluctuation of needs. This approach falls short of capturing the organic variation in human desires, which often intensify or wane in response to context, time, or recent actions. By exploring more complex dynamics—such as non-linear decay, situational adjustments, or time-of-day cycles—the framework could better reflect realistic interactions between desires and actions. The current simplicity reduces the authenticity of the agent's behavior, limiting the depth of its desire-driven model.\n\n**3. Narrow Focus on Physiological Desires in Single-Agent Setting**: The experimental results in the paper primarily emphasize physiological desires, with minimal exploration of higher-level motivations such as social connectivity or self-actualization. Additionally, the study tests only one agent, limiting insights into potential interactions or complex social behaviors that might arise in multi-agent settings. These restrictions make the results narrowly focused and reduce the paper's ability to demonstrate the full range of behaviors that D2A could potentially simulate, ultimately constraining the framework’s demonstrated impact.\n\n**4. Opaque GPT-4o Evaluation Methodology**: The evaluation of human-likeness using GPT-4o lacks transparency. Key details—such as the prompts used, scoring consistency, and validation of the assessment criteria—are not provided, making it difficult to gauge the robustness of the results. This lack of methodological clarity limits confidence in whether the evaluation effectively captures nuanced human-like behavior or merely reflects surface-level patterns.\n\n**5. Limited Qualitative Evaluation of Generated Action Sequences**: The paper lacks an in-depth qualitative analysis of the action sequences generated by D2A, making it difficult to assess the framework’s true effectiveness. The brief, simplistic sequence shown in Appendix O does not illustrate the model’s potential for creating complex, dynamic routines, leaving the impact of D2A’s design unclear. More detailed, varied sample sequences, along with thoughtful discussion, would provide stronger evidence of D2A's capabilities in simulating realistic human-like behaviors." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A desire-driven autonomy framework to guide an agent to simulate human-like daily activities" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024simulating,\ntitle={Simulating Human-like Daily Activities with Desire-driven Autonomy},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3ms8EQY7f8},\nnote={under review}\n}" }, "abstract": { "value": "Existing task-oriented AI agents often depend on explicit instructions or external rewards, limiting their ability to be driven by intrinsic motivations like humans. In this paper, we present a desire-driven autonomy framework to guide a Large Language Model based (LLM-based) agent to simulate human-like daily activities. In contrast to previous agents, our Desire-driven Autonomous Agent (D2A) operates on the principle of intrinsic desire, allowing it to propose and select tasks that fulfill its motivational framework autonomously. Inspired by the Theory of Needs from Maslow. A.H., the motivational framework incorporates an understanding of human-like desires, such as the need for social interaction, personal fulfillment, and self-care. Utilizing a desire-driven task generation mechanism, the agent evaluates its current state and takes a sequence of activities aligned with its intrinsic motivations. Through simulations, we demonstrate that our Desire-driven Autonomous Agent (D2A) generates coherent, contextually relevant daily activities while exhibiting variability and adaptability similar to human behavior. A comparative analysis with other LLM-based frameworks demonstrates that our approach significantly enhances the rationality of the simulated activities." }, "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": [ "desire;autonomy;daily activities;" ] }, "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/a0643fadb5187142e3a64870cef981e0b8314a56.pdf" }, "presentation": null, "primary_area": { "value": "applications to robotics, autonomy, planning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Simulating Human-like Daily Activities with Desire-driven Autonomy" }, "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 check 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": "1. The paper presents a rigorous, information-theoretic framework for disentangling data for different domains and modalities under conditions where MNI is unattainable.\n\n2. Based on the theoretical analysis, the authors designed a two-step training algorithm, and provided an optimality guarantee for this proposed method theoretically." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a self-supervised learning approach to disentangle shared and modality-specific information in multimodal data. The authors also explain that the optimal case in prior work is not doable based on Minimum Necessary Information (MNI), as the comprehensive analysis of the optimality. The experiments on vision-language and molecule-phenotype retrieval tasks demonstrate its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weakness:\nFor the experiments on synthetic data, they only contain the data for unattainable MNI. The synthetic experiments for attainable MNI are also essential here to demonstrate the reliability of the theoretical bound. My suggestion is to add Gaussian noise produced by different variances as the degrees of MNI attainable or unattainable.\n\n\n1. In Figure 4 (b), the superiority of DISENTANGLEDSSL is not significant compared to other baselines, such as JointOPT. The hyperparameter front of JointOPT is closer to the front of DISENTANGLEDSSL.\n\n2. Table 1 shows the three versions of DISENTANGLEDSSL are unstable. For example, DISENTANGLEDSSL (shared) can achieve promising performance on MOSI while failing on MUSTARD. It would be better to explain why such a phenomenon exists, probably focusing on the main difference between these three versions and properties of multi-modal data. \n\n3. The performance gain shown in Table 2 is quite limited. Could the author explain if small improvements might be important in this domain? In addition, it would be better if the authors could provide the reason for this tiny improvement. \n\n4. The paper does not satisfy the page limitation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Although in most cases DisentangledSSL-both achieves highest performance, there are also several tasks where either DisentangledSSL-shared or DisentangledSSL-specific alone achieves top performance (and sometimes by quite a large margin over DisentangledSSL-both, such as Mustard). Do you observe any trend / suggest any guidelines for when it is best to use each configuration?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The paper introduces a new perspective in how to perform effective multimodal representation learning. Specifically, the paper demonstrated that disentanglement of modality-specific and shared information into 2 separate representations can be effective in improving downstream task performance, which seems to be original and novel.\n\n2. The authors provided extensive theoretical justifications and proofs for their proposed approach.\n\n3. The experiments are quite comprehensive. They include both synthetic and real-world datasets, and the proposed method is evaluated against several multimodal SSL baselines on different tasks from drastically different domains. The proposed method achieved top performance in almost all tasks.\n\n4. All experiment details are presented in the Appendix (high reproducibility)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors proposed a novel self-supervised representation learning method for multimodal representation learning. The proposed method disentangles the representation of multimodal information into shared information and modality-specific information, and uses a separate neural network to learn each representation. The authors supported their proposed method through extensive theoretical analysis and proofs. The proposed method was evaluated on both synthetic datasets and real-world multimodal datasets, where the proposed method achieved top performance over baselines in most tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the extensive use of information-theory notations allowed rigorous proof of the proposed approach, it isn't really easy for readers who isn't interested in the proofs and just want to learn about the concrete algorithm/methodology. Perhaps the authors should consider including a Pseudocode/Alg block either in the main text or in Appendix to clearly demonstrate the training process." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Please respond to the weaknesses listed above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "It makes sense to learn disentangled shared and modality-specific representations of multimodal data. \n\nThe theoretical analysis to evaluate the quality of disentanglement, which can be applied even in cases where Minimum Necessary Information (MNI) is unattainable, is provided.\n\nThe efficacy of DISENTANGLEDSSL is demonstrated across a range of synthetic and real-world multimodal datasets and tasks, including prediction tasks for vision-language data and molecule-phenotype retrieval tasks for biological data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors believe that \"shared information between modalities is precisely what is relevant for downstream tasks\" and \"the modality gap ... results in misalignment between modalities, restricting the application of these methods in numerous real-world multimodal scenarios.\" Hence they motivate \"the need for a disentangled representation space that captures both shared and modality-specific information\". To address this need, they propose DISENTANGLEDSSL, an information-theoretic framework designed to learn disentangled shared and modality-specific representations of multimodal data. The authors have conducted a theoretical analysis to evaluate the quality of disentanglement, which can be applied even in cases where Minimum Necessary Information (MNI) is unattainable. The efficacy of DISENTANGLEDSSL is demonstrated across a range of synthetic and real-world multimodal datasets and tasks, including prediction tasks for vision-language data and molecule-phenotype retrieval tasks for biological data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Some of the key assertions motivating this work are inaccurate, and the proposed graphical model is flawed, which together raise questions about the validity of the resulting framework.\n\nFor instance, the statement \"shared information between modalities is precisely what is relevant for downstream tasks\" is not universally accurate. What is relevant depends heavily on the specific downstream tasks. Furthermore, \"shared information\" is a vague term, especially when applied across modalities, making it problematic in this context.\n\nThe assertion that \"the modality gap ... results in misalignment between modalities, restricting the application of these methods in numerous real-world multimodal scenarios\" is also oversimplified. The modality gap is not the sole or primary cause of misalignment; other factors such as alignment methods, distributional shifts, domain shifts, sampling biases, and more can contribute significantly.\n\nI suggest the authors to refine their motivation to more accurately reflect the complexities of multimodal relationships, while still maintaining the core idea of disentangling shared and modality-specific information.\n\nThe graphical model in Figure 2, intended to represent the generative process (as claimed in line 128), has two major issues. First, the authors have conflated the generative and inference processes. Only the pathway from Z to X represents the generative model, while the path from X to \\hat{Z} corresponds to inference, which should not be included in the data-generating graphical model. This issue is straightforward to resolve by removing the inference process from the model, which would not affect the algorithm or results, or to create separate diagrams, one for the generative process and the other for inference to clarify.\n\nThe second issue is more severe: the assumption of a shared latent variable Z_c existing between two modalities may not hold. This assumption lacks foundation, as it is more likely that the relationship takes the form of two variables Z_c^1 and Z_c^2, with potential connections between them: (1) Z_c^1 -> Z_c^2, (2) Z_c^1 <- Z_c^2, or (3) or Z_c^1 <-C-> Z_c^2, depending on the modalities involved. This discrepancy significantly undermines the justification for the proposed algorithm. I encourage the authors to discuss the implications of using separate but potentially connected latent variables (Z_c^1 and Z_c^2) instead of a single shared Z_c, and how might this change affect the proposed algorithm and results. This could lead to a more nuanced and realistic model of multimodal relationships." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Are there any alternative approaches for implementing the $I(Z^1;\\hat{Z}_{c}^{1*})$ in $L_s^1$? If so, how does their performance compare to that of the orthogonal loss implementation?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1.The idea of applying conditional entropy bottleneck (CEB) in multi-modal self-supervised learning is novel.\n\n2.Comprehensive theoretical analysis is conducted to prove the optimality of both modality-specific and modality-shared information.\n\n3.DisentangledSSL demonstrates superior performance across tasks, including vision-language prediction and molecule-phenotype retrieval." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the disentanglement of modality-shared and modality-specific information in multi-modal self-supervised learning and proposes DisentangledSSL. From an information-theoretic perspective, the modality-shared information is optimized using a conditional entropy bottleneck (CEB). Correspondingly, the authors formulate an optimization problem to isolate modality-specific information. Theoretical analysis of the optimality of each disentangled representation, particularly when Minimum Necessary Information is unattainable, along with experimental results, demonstrates the superiority of DisentangledSSL." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.How does the optimization objective for the modality-shared representation change from Eq. (1) to Eq. (3)? Why is $I(X^1;X^2)$ omitted?\n\n2.Related studies, such as CoCoNet [1] and SimMMDG [2], which also address the separation of modality-shared and modality-specific information, are recommended for inclusion and comparison. Note that SimMMDG [2] can be easily adapted to a self-supervised setting by substituting supervised contrastive learning with the original self-supervised counterpart.\n\n3.Ablation studies on the impact of $\\beta$ and $\\lambda$ in the MultiBench datasets should be provided to demonstrate the importance of separating modality-specific and modality-shared information in real-world applications. \n\n4.Some typos.\n1)$Z^2$ should be $X^2$ in Eq.(6) and in the equation below Eq.(6).\n2)$I(X^1;Z^1)$ should be $I(X^1;X^2)$ in Figure 3.\n \n[1] Li J, Qiang W, Zheng C, et al. Modeling multiple views via implicitly preserving global consistency and local complementarity. TKDE, 2022.\n[2] Dong H, Nejjar I, Sun H, et al. SimMMDG: A simple and effective framework for multi-modal domain generalization. NIPS, 2023." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose DisentangledSSL to separate modality-specific information from shared information in multimodal data, providing theoretical guarantees and strong empirical performance, especially when Minimum Necessary Information is unattainable." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024an,\ntitle={An Information Criterion for Controlled Disentanglement of Multimodal Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3n4RY25UWP},\nnote={under review}\n}" }, "abstract": { "value": "Multimodal representation learning seeks to relate and decompose information inherent in multiple modalities. By disentangling modality-specific information from information that is shared across modalities, we can improve interpretability and robustness and enable downstream tasks such as the generation of counterfactual outcomes. Separating the two types of information is challenging since they are often deeply entangled in many real-world applications. We propose $\\textbf{Disentangled}$ $\\textbf{S}$elf-$\\textbf{S}$upervised $\\textbf{L}$earning (DisentangledSSL), a novel self-supervised approach for learning disentangled representations. We present a comprehensive analysis of the optimality of each disentangled representation, particularly focusing on the scenario not covered in prior work where the so-called $\\textit{Minimum Necessary Information}$ (MNI) point is not attainable. We demonstrate that \\algo successfully learns shared and modality-specific features on multiple synthetic and real-world datasets and consistently outperforms baselines on various downstream tasks, including prediction tasks for vision-language data, as well as molecule-phenotype retrieval tasks for biological data." }, "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 Representation Learning", "Disentanglement", "Self-Supervised Learning", "Information Theory" ] }, "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/1cbbdc92eae7216190180703b0608ff22b851ac2.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": "An Information Criterion for Controlled Disentanglement of Multimodal Data" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How the score-based and constraint-based methods are used in the extendable PC algorithm?\n- Are the more existing work that incrementally learn the structure? What are their 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": "+ structure learning is a challenging problem due to exponentially large search space, and the paper makes progress in speeding up the search without sacrificing accuracy.\n+ the proof is solid and the presentation is clear.\n+ strong experimental results in accuracy and running time." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an iterative/incremental algorithm to learn the structure of a Bayeisan network from observed data.\nThe algorithm added one variable at a time, and modify the previously learn structure accordingly.\nThe novelty is to prove that adding one variable only leads to deletion of previous edges, and therefore trim the search space of possible networks.\n\nIn particular, Lemma 1 shows that if the new variable Y satisfies some properties, a certain kind of edges in the old graph can be safely deleted. The paper introduced constrained-based and score-based approaches to utilize this lemma. The results is a reduction of number of CI tests.\n\nExperiments on 13 datasets shows that the running time is significantly reduced without compromising accuracy of the learned structure." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The scope of the paper (symbolic Baysian network) may not fit ICLR conference (representation learning).\n- Insufficient discussion of related work. There shall be a large number of related work on incremental structure learning, while the submission only cite two most relevant ones ((Kocacoban & Cussens, 2019) and (Alcobe, 2005)). This makes the contribution of the submission less clear." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Is the Extendable P-map learner in Algorithm 3 implemented using Algorithm 1?\n2. What is Sepset(X,Y)?\n3. An intuitive explanation for why the iterative learning algorithm outperforms PC algorithm while learning the complete graph is recommended.\n4. How is this work positioned relative to the prior works that study the information-theoretic limits of Bayesian structure learning?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The problem studied in the paper is well-motivated and will be of interest to ML community. Recycling the existing information when new variables are added or revealed provides a novel perspective to the structure learning problem. The iterative structure learning algorithm is a valuable contribution. Experiments convincingly demonstrate the computational efficiency of the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides a novel perspective to the Bayesian structure learning problem with an efficient mechanism to add new variables to an underlying graphical structure. Specifically, the learning strategy hinges on efficiently incorporating a new variable $Y$ into an existing Bayesian network ${\\cal G}$ over the set of variables ${\\cal X}$, which results in an. updated Bayesian network $\\bar {\\cal G}$ on the augmented set of variables ${\\cal X} \\cup Y$. This learning strategy is further extended to provide a novel learning paradigm for structure learning for Bayesian networks. Experiments demonstrate significant computational efficiency over existing state-of-the-art algorithms." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I found the discussions and organization of Section 3 to be more convoluted than necessary. In particular, it would be helpful to have the relevance of Algorithms 2-4 explictly elucidated." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 have the following comments about the paper:\n\n1. Performance guarantees such as error analysis, or on how to check whether the algorithm has converged to the correct graph structure are not discussed. \nA discussion on how to verify the correctness of the P-map finder algorithm used should be added. If the iterative approach is used, how does the error accumulate?\nAre the conditions in Lemma 1 and 2 sufficient for the P-map finder algorithm output the true structure?\n\n\n2. The key assumption that when a new variable $Y$ is added to the existing set $X$, that no new edges get assigned between the elements of $X$, this assumption needs further explanation. This seems to be a necessary condition of the algorithm to have low computational cost. In many situations, the introduction of a new variable might introduce new dependencies between existing nodes, e.g., in root-cause analysis, causal learning, molecular prediction, and others. Also, such situations could occur in time evolving DAGs. Further discussion on these will illustrate the applicability of the proposed method to different problems. \n\n\n3. Minor Comment:\ni. In abstract, algorithm named PC might not be known to general readers. Similarly, FCI. Some of these acronyms are not defined." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Strengths:\n1. Extendable structure learning for Bayesian networks is studied and two new approaches are proposed. \n2. The new approaches achieve much lower runtime than relearning the graphs without prior structure. \n3. Numerical results are presented on many graph datasets showing significant speedup." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the structure learning problem in Bayesian networks, i.e.. learning a directed acyclic graph (DAG) which defines the conditional probability distribution over the given variables. Particularly, an extendable structure learning strategy is proposed to update an existing Bayesian network graph efficiently, when a new variable is introduced. Two approaches (constraint-based and score-based) are discussed for extendable structure learning, which leverage the previous graph structure, and have much lower computational cost compared to relearning the graph. It is then shown that these procedures can be used to design an iterative algorithm for structure learning. Numerical results on many graph datasets illustrate the performance of the proposed method, and shows significant speedup over the approach where the graph is learned from scratch." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weakness:\n1. Performance guarantees for the proposed method are not presented.\n2. Details about a key assumption can be further 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": 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": "- Is there any bound on the total number of CI tests required for running Algorithm 3?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper is well organized. The notations and definitions are mostly self contained.\n- The authors provide experiments on multiple datasets, and show that their approach save significant runtime because of the fewer number of CI tests required in the proposed iterative method. \n- The proposed method seems straightforward and easy to implement in practice." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an extendable stracture learning method for Bayesian networks, which updates an existing network by adding new variables. The authors also propose a iterative approach for structured learning by starting with a small set and adding the remaining variables to the P-map graph. The authors run the extendable PC algorithm on multiple datasets, and show that their approach requires fewer number of CI tests compared to the original approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am not an expert on this line of literature. My main concern is that the paper seems very heuristic with no formal guarantees: \n- While the experimental results look good, I wonder if the proposed method of extendable PC has any consistency, faithfulness, or optimality guarantee. \n- The results in Table 2 - 3 suggest that extendable PC always has a better runtime with fewer number of CI tests compared to PC. Can that be proved?\n- The result in Table 5 shows that the proposed iterative PC does not always require fewer CI tests compared to PC. Under what statistical or topological conditions will that happen? In my opinion this is be the bigger risk, because without any theoretic characterization, this shows the possibility that the proposed approach does not generalize." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) A simple iterative type algorithm for learning bayesian networks, which is computationally efficient. \n2) Insights on unidentifiability and its relationship to faithfulness of the graph. \n3) Backed up by theoretical claims" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an efficient method for updating Bayesian network structures as new variables are introduced, eliminating the need for retraining from scratch. The approach reduces computational costs by up to 1300x without sacrificing accuracy. The authors also propose an iterative strategy that builds the network iteratively, thereby offering runtime benefits comparable to common algorithms like PC while maintaining accuracy. This scalable approach is well-suited for real-time and high-dimensional applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "n/a" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024extendable,\ntitle={Extendable and Iterative Structure Learning for Bayesian Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3n6DYH3cIP},\nnote={under review}\n}" }, "abstract": { "value": "Learning the structure of Bayesian networks is a fundamental yet computationally intensive task, especially as the number of variables grows. Traditional algorithms require retraining from scratch when new variables are introduced, making them impractical for dynamic or large-scale applications. In this paper, we propose an extendable structure learning strategy that efficiently incorporates a new variable $Y$ into an existing Bayesian network graph $\\mathcal{G}$ over variables $\\mathcal{X}$, resulting in an updated P-map graph $\\bar{\\mathcal{G}}$ on $\\bar{\\mathcal{X}} = \\mathcal{X} \\cup \\{Y\\}$. By leveraging the information encoded in $\\mathcal{G}$, our method significantly reduces computational overhead compared to learning $\\bar{\\mathcal{G}}$ from scratch. Empirical evaluations demonstrate runtime reductions of up to 1300x without compromising accuracy. Building on this approach, we introduce a novel iterative paradigm for structure learning over $\\mathcal{X}$. Starting with a small subset $\\mathcal{U} \\subset \\mathcal{X}$, we iteratively add the remaining variables using our extendable algorithms to construct a P-map graph over the full set. This method offers runtime advantages comparable to common algorithms like PC while maintaining similar accuracy. Our contributions provide a scalable solution for Bayesian network structure learning, enabling efficient model updates in real-time and high-dimensional settings." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "structure learning", "Bayesian networks", "iterative" ] }, "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/0524315b7e6ebf64389c11e1270a208f362fc639.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": "Extendable and Iterative Structure Learning for Bayesian Networks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "There is a natural hierarchy among some of the classes, such as Building, Terminal (subtype of Building), Tower (subtype of Building). Did you consider defining an explicit class taxonomy rather than just a flat list of classes? A taxonomy would enable natural expansion to additional classes and potentially resolve ambiguities such as incorrectly declaring a false alarm when a Terminal is labeled as a Building.\n\nDuring data collection, why were the videos so short, 15 sec? Since the cameras are fixed, it seems straightforward to collect long videos e.g. hours in order to capture a diversity of long-range airport activities.\n\nHow many airports are in the dataset? How many unique camera views? Many important details are missing.\n\nWhat proportion of the dataset was used for fine-tuning vs. testing? Were less-frequent classes handled differently than more common ones?\n\nThe IOU results on Person are the lowest of any any category, despite the maturity of person detectors. Why? Is it the small pixel size of people in this dataset? What would happen if the images were up-sampled 2X or 4X?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Demonstrating the low accuracies of popular segmentation techniques on a real-world problem is a significant contribution. The challenges of this dataset are also found in many surveillance datasets, hence addressing them through research on this dataset could have impact in many surveillance domains.\n\nThe distribution of annotations and pixels by class is useful to see, Fig. 5. The dataset seems to follow a long-tail distribution which is very common in real-world settings like this, whereas more contrived datasets often have a relatively uniform distribution.\n\nThe annotation compactness metric is well formulated, based on a standard spatial moment calculation. It would be interesting to see a comparison to other surveillance segmentation datasets.\n\nThe evaluation is quite thorough, testing a variety of current approaches for both image and video segmentation. The low accuracies of all of the algorithms indicate the difficulty of the problem and the utility of the dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new dataset consisting of 250 short videos of outdoors airport scenes, with segmentation annotations of 18 categories on all frames. A variety of recent image and video segmentation methods are fine-tuned and tested on the dataset, yielding relatively low accuracies on most categories. The dataset is characterized by various analytics and a few metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There is essentially no stated motivation for semantic segmentation at airports. The one mention of this, in the first paragraph is very vague. It would be more convincing to enumerate a set of use cases motivating the need for accurate segmentation rather than bounding boxes.\n\nThe Intro should clarify that only outdoor scenes are included in the dataset, since there are numerous video surveillance datasets that include indoor scenes are airports and other large facilities.\n\nDownselecting from 5000 collected videos to 250 included videos is a huge reduction. Was this primarily motivated by the cost of creating ground-truth segmentations? The methods used for data selection are not described.\n\nThe data annotation section describes the process for annotating one image, but does not mention how video is annotated. In video from a fixed camera, a single annotation of a fixed object e.g. a Building should be transformable to subsequent video frames without editing. Was this method used? Even for movers, the annotation on the previous frame can be copied to the current frame and adjusted, greatly reducing effort and inter-frame annotation variability.\n\nCreating segmentation annotations manually is expensive on images, even more so on video as performed here. The dataset is much smaller than VSPW in terms of number of videos and classes, partly because of the narrower problem domain.\nImage coherence, Eq. 1, does not seem to be an advantage nor disadvantage. With a moving camera, image coherence would be very low, for example. A large number of movers would yield low coherence. Similarly, label coherence is a function of camera and object movement, not just label spatial consistency across frames. The purpose of these measures, and the comparison to other datasets, is very unclear and not well motivated. I’d suggest removing this section, or, more significantly, reformulating these metrics to measure how consistent an object label remains across video frames.\n\nThe topic of this paper is rather specific and seems more appropriate for a computer vision venue such as WACV, or AVSS (the conference with the same acronym as the 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": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to the weakness part." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. It provides insights into potential issues in airport semantic segmentation, such as extreme multi-scale, intra-class diversity, and inter-class similarity.\n2. The proposed benchmark has been evaluated against various state-of-the-art (SOTA) models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel benchmark for Airport video semantic segmentation and introduces a semantic segmentation algorithm based on a 3D airplane model. The authors identify key challenges in airport scenes, including extreme multi-scale variation, intra-class diversity, and inter-class similarity. Addressing these issues, they propose a benchmark that is evaluated from various perspectives. First, they conduct a statistical analysis by measuring class distribution, inter-frame coherence, and compactness. By applying the proposed method to various models, the authors demonstrate the increased difficulty of their benchmark compared to existing public datasets. To further assess generalizability, they train models on the proposed dataset and compare performance with the VSPW dataset. To this end, the authors present a 3D airplane model-based algorithm tailored specifically for airport segmentation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It would be helpful to specify the diversity principles in “Data collection” in detail.\n\n2. A reference to AnyLabeling in the “Data annotation” should be included.\n\n3. The analysis of whether the proposed benchmark can cover intra-class diversity, a key challenge in airport semantic segmentation, is insufficient. It would be beneficial to examine various aspects, such as color distribution and feature distribution within the same category, to provide a more comprehensive analysis.\n\n4. There is a need to analyze whether the proposed benchmark addresses inter-class similarity.\n\n5. I have doubts about the actual relationship between compactness and segmentation difficulty. For example, in Table 2, while there is a large gap in results between \"Runway\" and \"Liaison Road,\" the difference in compactness is not significant. Additionally, while there is a small gap in results between \"Runway\" and \"Person,\" the difference in compactness is large.\n\n6. For the generalizability analysis experiment, comparing similar categories from another dataset (such as Cityscape) would be more helpful.\n\n7. The author suggests that a model performing well on AVSS is likely to achieve favorable segmentation results on other datasets (page 8, lines 430-431). However, it would be useful to have an experiment to verify this claim. For example, by comparing the performance differences between the top 3 models with the highest performance and the bottom 3 models with the lowest performance on AVSS, it could provide valuable insights into whether models that perform well on AVSS have a similar tendency on other datasets.\n\n8. There are no qualitative results for the proposed 3D airplane-based algorithm for airport semantic segmentation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "No." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "In conclusion, the main contributions are as follows:\n1. this paper establish the novel AVSS, providing a benchmark for airport semantic segmentation.\n2. this paper evaluate the generalizability of AVSS and 18 SOTA segmentation algorithms on AVSS.\n3. this paper propose principles for designing airport video semantic segmentation models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, it introduces the first large-scale Airport Video Semantic Segmentation dataset (AVSS) for airport surveillance. AVSS comprises 18 common semantic categories at airports, and 250 videos, totaling over 140,000 frames with accurate manual annotations. AVSS covers a wide range of challenges for airport video surveillance, such as extreme multi-scale, intra-class diversity, inter-class\nsimilarity, etc. The authors analyze statistical information and evaluate 18 state-of-theart (SOTA) semantic segmentation algorithms on AVSS. The significant performance degradation indicates that current models are far from practical application. Furthermore, this paper discuss how to develop video semantic segmentation algorithms for airport surveillance and the generalizability of AVSS to other tasks\nand datasets. AVSS serves as a research resource for airport semantic segmentation and a robustness evaluation tool for segmentation algorithms in practical applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weakness can be as follows:\n\n1. In table 1, the paper shows the coherence is higher than other datasets. But for the cohenrence metric equaiton (1) (2), it's not clear how to compute the corresponding pixels between difference frames; in addition, higher coherence means the data variety is low for a video squezze, as the AVSS datasets only has 250 videos, does it mean it has only 250 scenes with just changing people, plane etc but the background is fixed. From this view, the variety of the AVSS dataset can be low.\n2. For the 4.3 GENERALIZABILITY, the first sentence is \"test the model trained on AVSS on VSPW\", is it a typo? As the table 4 shows \"The classes IoU of SOTA models trained on AVSS, evaluated on AVSS and VSPW.\" For this experiment, it is necessary to conduct the experiment comparing the model trained on VSPW and test on AVSS and VSPW." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Refer to weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ The dataset is based on novel airport scenes, providing a new perspective to evaluate VSS models and contributing a valuable resource for future researchers.\n+ The dataset is manually labeled, ensuring high-quality annotations.\n+ The experiments reveal the limitations of existing models on this dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new dataset for airport video semantic segmentation (AVSS) with manually labeled masks. It then benchmarks current VSS models on this dataset, highlighting a significant drop in performance when applied to AVSS and providing future insights for designing suitable VSS models for this domain." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The scale of AVSS (only 250 videos) is relatively small compared to current VSS datasets.\n- The related work section should include more recent advancements in VSS models and benchmarking should add some recent works, such as:\n - Mask propagation for efficient video semantic segmentation\n - Pay attention to target: Relation-aware temporal consistency for domain adaptive video semantic segmentation\n - Temporal-aware Hierarchical Mask Classification for Video Semantic Segmentation" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024avss,\ntitle={{AVSS}: a new benchmark for airport video semantic segmentation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3nkIRKh3Sk},\nnote={under review}\n}" }, "abstract": { "value": "Airport video semantic segmentation is fundamental to airport surveillance applications, yet there currently lacks a specialized benchmark and algorithms for this task. In this paper, we introduce the first large-scale Airport Video Semantic Segmentation dataset (AVSS) for airport surveillance. AVSS comprises 18 common semantic categories at airports, and 250 videos, totaling over 140,000 frames with accurate manual annotations. AVSS covers a wide range of challenges for airport video surveillance, such as extreme multi-scale, intra-class diversity, inter-class similarity, etc. We analyze statistical information and evaluate 17 state-of-the-art (SOTA) semantic segmentation algorithms on AVSS. The significant performance degradation indicates that current models are far from practical application. Furthermore, we discuss how to develop video semantic segmentation algorithms for airport surveillance and the generalizability of AVSS to other tasks and datasets. AVSS serves as a research resource for airport semantic segmentation and a robustness evaluation tool for segmentation algorithms in practical applications. AVSS is available at www.agvs-caac.com/avss/avss.html." }, "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": [ "Airport Ground", "Semantic Segmentation", "Video Surveillance" ] }, "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/668d8d696b379d3ab42065f7f7800acb33d0b13c.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": "AVSS: a new benchmark for airport video semantic segmentation" }, "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": "- missing ']' in Fig. 1 dihedral term.\n- If the paper aims to emphasize the Path Weisfeiler-Lehman (PWL) capacity, it should evaluate the capacities of classical models, such as DimeNet, GemNet, and MACE. For reference, see the Geometric Weisfeiler-Lehman (WL) paper." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "It proposes path complex-based message passing (PCMP) with some detailed graph theories, and achieves good results on some molecular tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Path Complex-based Message Passing (PCMP) and achieves promising results on molecular property prediction benchmarks. However, there are some major weakness in this paper including **lack of literature review, novelty and evaluation on current benchmarks**, and need to be further revised." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Incorporate geometric features like bond, angles, dihedrals and improper angles in the modeling is not novel, as seen in Fig.2 of [1] and Fig.2 of [2]. There are also many works already involved these many-body systems.\n- Although the paper introduces the concept of a “path complex,” the actual features used, which are bond distances and angles (Table 5), are standard. Similar methods already exist, such as hierarchical message passing for updating geometric embeddings in [2]. Also, models like NequIP, Allegro, and MACE employ **many-body expansions** in message passing, leveraging **tensor products** to incorporate higher-order geometric tensors (a path fusing process [3]), which are beyond the basic features mentioned in this paper. PCMP maybe a subset of the tensor products. A more comprehensive literature review is needed for the authors [4, 5].\n- The primary weakness is that the authors claim that they are inspired by MD force fields, but **no results on any standard MD benchmark, such as MD17, rMD17, MD22 are provided**. Not to mention the conduction of MD simulations further driven by this MLFF. Since the paper focuses on molecular 3D structures, **it’s a necessity to proof invariance or equivariance**, which are missing here. In contrast, this paper provides a list of the graph path theories, appears more relevant to topological graphs and is insufficient to address geometric graphs. Furthermore, despite claiming that the method “enables systematic exploration of connectivity and interaction for analyzing complex systems and networks,” **there are no experiments on these tasks supporting this claim**.\n- The GEM paper is relatively old and actually we do not need to generate 3D structures using RDKit from smiles for the 2D molecule datasets. Beside the datasets mentioned in the above question, **numerous 3D molecular datasets for DFT-level property prediction, such as QM9 (with 12 targets), OC20, OE62, and PCQM4Mv2, are available**. I strongly recommend evaluating PCMP on these benchmarks for a more comprehensive assessment.\n\n[1] Wang, Yusong, et al. \"Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing.\" Nature Communications 15.1 (2024): 313.\n\n[2] Pei, Hongbin, et al. \"Hago-net: Hierarchical geometric massage passing for molecular representation learning.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 13. 2024.\n\n[3] https://docs.e3nn.org/en/latest/api/o3/o3_tp.html\n\n[4] Zhang, Xuan, et al. \"Artificial intelligence for science in quantum, atomistic, and continuum systems.\" arXiv preprint arXiv:2307.08423 (2023). **Section 5.2**\n\n[5] Han, Jiaqi, et al. \"A Survey of Geometric Graph Neural Networks: Data Structures, Models and Applications.\" arXiv preprint arXiv:2403.00485 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Overall, I believe the authors present a good tool to predict molecular properties. However, the following questions should be addressed to clarify key aspects and improve rigor:\n\n1. How do you compare the accuracy of your model with the more recent models, M3GNet, MACE, or EquiformerV2?\n\n - In addition, you only compared the accuracy, how about the speed compared with other models?\n\n2. Is it possible to expand this framework to periodic systems, i.e. inorganic materials?\n\n3. **Interpretability of Path Features**: Could the authors explore or comment on how path features across different orders contribute to the final prediction? Are there plans to visualize or interpret specific paths in relation to molecular properties?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Originality: The PCMP model presents a unique innovation in molecular property prediction by applying path complexes, which go beyond conventional graph-based representations. The incorporation of multi-order path complexes allows for capturing high-order interactions like bond angles and dihedral angles. This approach offers a fresh perspective on molecular graph representation, making PCMP stand out from other GNN-based models that primarily rely on node and edge interactions.\n* Quality: The paper provides a thorough experimental validation, comparing PCMP with a diverse set of baseline models, including both pretrained and non-pretrained GNNs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel approach called Path Complex Message Passing (PCMP) for molecular property prediction using geometric deep learning. Unlike traditional graph neural networks (GNNs) that operate on molecular graphs, PCMP employs path complexes that capture multi-body interactions in molecules through paths of various orders. The model’s hierarchical message-passing mechanism updates high-order paths first, followed by lower-order paths, facilitating effective feature communication between these paths. Extensive experiments on benchmark datasets demonstrate that PCMP achieves state-of-the-art results in molecular property prediction, showcasing its potential to model complex molecular interactions comprehensively." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Clarity: I suggest the authors to hide some of the technical details in the Appendix.\n* Experimental Limitations: Although the experiments are extensive, the paper could benefit from a more diverse set of benchmarks. The current datasets focus primarily on small to medium-sized molecules, while macromolecular structures, such as proteins, are absent from the evaluation. In addition, the test of efficiency, i.e. speed of training or inferences, 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": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Can the authors provide empirical validation on MM force field datasets to substantiate their claim of mimicking MM force fields?\n2. Are there specific path orders or features that consistently contribute more to accurate predictions? Could this be visualized or quantified?\n3. Can the model be used to capture the improper angle? Is it possible to implement this? \n4. For Figure 4, does the path complex graph alleviate the over-squashing? If so, can the author provide an empirical study on this, for example, compare the effective resistance.\n5. So the current experiments contain only validation from small molecules dataset, what about for large molecules? How does the model scale to larger molecules?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. **Innovative Modeling of Molecular Interactions**: PCMP introduces path complexes, offering an approach to capture both local and global molecular features that go beyond traditional graph representations.\n2. **Methodological Rigor**: The hierarchical message-passing mechanism is well-detailed, showing how path complexes of different orders contribute to the molecular representation.\n3. **Thorough Ablation Studies**: The authors provide in-depth ablation studies that highlight the importance of various path orders and message-passing mechanisms, strengthening the evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents **Path Complex Message Passing (PCMP)**, which introduces path complexes to model intricate chemical and non-chemical interactions for molecular property prediction. The model demonstrates promising results on molecular benchmark datasets and aims to provide a more detailed molecular representation by incorporating high-order interactions in a path complex framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Unsubstantiated Claim on Force Field Mimicking**: A major claim of the paper is that PCMP mimics the molecular mechanics (MM) force field, yet no benchmarks or empirical results using MM force field datasets (e.g., MD17, MD22) are provided to substantiate this claim. Without benchmarking against MM force fields, the claim appears unsupported, and this oversight detracts from the paper's validity in this area.\n \n2. **Computational Complexity**: The inclusion of path complexes, especially higher-order ones, is likely computationally demanding. However, the authors do not provide insights into potential trade-offs, such as runtime or scalability on larger datasets.\n\n3. **Lack of Equivariance in Model Design**: Given the model’s target application in molecular property prediction, its architecture does not incorporate rotational or translational equivariance, which would enhance its ability to handle spatial molecular data more robustly. Adding equivariant layers could make the model better suited to capturing geometry-sensitive properties.\n\n4. **Interpretability**: The model’s complex hierarchical structure might hinder interpretability, as it’s not clear which paths contribute most significantly to predictions or whether high-order interactions have consistent relevance across datasets. A comprehensive interpretability study is recommended." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1) What is the time and memory complexity of constructing and operating on path complexes compared to traditional graph-based methods? How does the method scale with molecule size?\n\n2) Related to above. Could you provide empirical justification for using 3-path complexes as the maximum order? What happens with higher orders?\n\n3) Could you provide a brief description of each dataset's characteristics in the main paper?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper displays technical rigor in developing its mathematical foundations, establishing formal proofs for path complex properties and their relationship to molecular structures. The proposed path complex representation is well motivated from a chemistry perspective, making explicit connections to molecular force fields and showing how different path orders correspond to physical properties (bond lengths, angles, and dihedral angles). The architecture is novel. While limited, the experimental results do show consistent improvements across multiple benchmark datasets, and the ablation studies help in distilling the contribution of different components of the model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Path Complex Message Passing (PCMP), a method for molecular property prediction that represents molecules using path complexes of different orders to capture various aspects of molecular structure (bond lengths, angles, and dihedral angles). The work is heavily theoretical. The method is tested on five subsets of MoleculeNet, showing improvements over baseline models. Most of the papers content is theoretical development and proofs, with a relatively brief experimental section." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper suffers from several weaknesses. In the Introduction, the authors refer to quite 'old' papers. The experimental validation is notably thin compared to the extensive theoretical development, taking up only about 2 pages of the 10-page paper. The empirical improvements, while consistent, are relatively modest and don't seem to justify the substantial complexity introduced by the method. There's inadequate discussion of computational overhead and scalability. The path complex representation likely introduces significant computational costs, but this isn't analyzed. Even though datasets are described in the appendix, authors haven't included any brief description of the dataset, or the task at hand: for example, for QM9, one cannot know against which property the authors are regressing against from the main paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024path,\ntitle={Path Complex Message Passing for Molecular Property Prediction},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3nwlXtQESj},\nnote={under review}\n}" }, "abstract": { "value": "Geometric deep learning (GDL) has demonstrated enormous power in molecular data analysis. However, GDL faces challenges in achieving high efficiency and expressivity in molecular representations when high-order terms of the atomic force fields are not sufficiently learned. In this work, we introduce message passing on path complexes, called the Path Complex Message Passing, for molecular prediction. Path complexes represent the geometry of paths and can model the chemical and non-chemical interactions of atoms in a molecule across various dimensions. Our model defines messages on path complexes and employs neural message passing to learn simplex features, enabling feature communication within and between different dimensions. Since messages on high-order and low-order path complexes reflect different aspects of molecular energy, they are updated sequentially according to their order. The higher the order of the path complex, the richer the information it contains, and the higher its priority during inference. It can thus characterize various types of molecular interactions specified in molecular dynamics (MD) force fields. Our model has been extensively validated on benchmark datasets and achieves state-of-the-art results.\nThe code is available at \\url{https://anonymous.4open.science/r/Path-Complex-Neural-Network-32D6}" }, "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 Property Prediction; Path Complex; Geometric Deep Learning; High-order Interaction; Low-order Interaction" ] }, "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/ce1d2e8f99cb7682646d8a3c5aa2bc1499c52534.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": "Path Complex Message Passing for Molecular Property Prediction" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is clearly motivated, the problem of hallucinations in LLM reasoning has not been explored as much.\n- The paper presents the problem, the methods and the baselines clearly.\n- The results on the three datasets, MATH, blocksworld and boardgameQA are comprehensive and provide sufficient evidence of the utility of the method.\n- The authors also present sufficient ablations of their method, with varying versions of the curriculum used." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the problem of hallucinations in reasoning for large language models. Specifically, in methods that use expert iteration for improving this reasoning in LLMs. The authors propose adding a Refusal option to the EI pipeline and rewarding the refusal when the problem has a certain level of difficulty (length of reasoning is used as a proxy for difficulty). Based on the reward that a response gets, the data for the next iteration of EI is selected accordingly. To balance refusal and improvements in reasoning, the authors propose using a curriculum to balance the two objectives. Based on measures of accuracy and refusal rate, the authors show the effectiveness of their method compared to baselines and ablations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Missing citations for central EI works: STaR (https://arxiv.org/abs/2203.14465), Training Chain-of-Thought via Latent-Variable Inference (https://arxiv.org/abs/2403.04642)\n- What other measures of difficulty are possible (very recent paper that the authors can choose to ignore since it came out after the paper deadline: https://arxiv.org/abs/2410.04707)? Can a linear probe decode difficulty? A discussion is needed.\n- Similarly, are there other mechanisms of knowing when to say I dont know possible? Is a linear probe enough for refusal? Like in Language models mostly know what they know (https://arxiv.org/abs/2207.05221).\n- Where could the current framework of length based difficulty fail? When sampling multiple times, what is the variation in the number of steps needed to reach the answer for a problem?\n- Are the number of iterations for Auto-CEI and EI etc matched? More generally, how does the amount of training data / compute required vary across the method and baselines? I think some of the implementation details could be moved to main in the revision. This would help readability.\n- How could this method be extended beyond expert iteration, to more online methods of learning like PPO or REINFORCE?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Why the process of update R is called curriculum learning? IMO, I think it is more like a search process and curriculum learning is about first learning elementary stuffs and then complex stuffs." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Using expert iteration is a good idea.\n2. The authors consider the trade off between avoiding hallucinations and avoiding laziness, which is important.\n3. This paper aims to avoiding reasoning hallucinations and teaching the model to say IDK to reasoning tasks which beyond its ability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an automatic curriculum expert iteration (AUTO-CEI) to enhance LLM reasoning and align responses to the model's capabilities. Enable the model to answer within its limits and decline when tasks exceed them.\nExpert iteration can explore the reasoning trajectories near the LLM policy, guiding incorrect paths back on track to reduce compounding errors and improve robustness.\nAUTO-CEI uses the length of reasoning steps to measure difficulty and designs an automatic curriculum for expert iteration that rewards correct reasoning.\nAUTO-CEI can automatically estimates the boundary of LLMs' reasoning capacity to achieve a reasonable alignment to maximize capacity and control behaviors." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although considering the trade off between helpfulness and laziness, the paper control this trade off by hyper-parameters, instead of proposing some principles or methods to chose the optimal hyper-parameters.\n2. The evaluation metrics in the paper are incomplete; for example, IDK only measures the proportion of times the model outputs \"IDK\" without assessing the accuracy of those IDK responses.\n3. The experiments in the paper may involve unfair comparisons, as AUTO-CEI conducts multiple searches for EI, resulting in significantly more training steps and data compared to other methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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 MATH dataset has manually labeled question difficulty; it would be helpful to include a figure illustrating the correlation between generated text length and question difficulty as a validation of this signal.\n\n2. Is the method effective across datasets with different distributions?\n\n3. There are many established metrics for measuring uncertainty with different confidence levels, such as AP used in R-tuning. The current paper only reports results for the binary case, i.e., answering or refusing to answer. I am curious about the performance of this method at different confidence levels." }, "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 presents a reasonable motivation that previous approaches to addressing hallucination problems often suffer from overcorrection, leading to overly conservative responses from LLMs on many questions.\n\n2. The paper introduces an innovative reward mechanism that dynamically balances the choice between responding or refusing to answer based on question difficulty. Subsequent experiments demonstrate that this method effectively handles the trade-off between answering and refusing. Figure 3 also illustrates that as the generated text length increases, more questions are declined.\n\n3. The writing is clear, and the figures are well-designed, enhancing the overall readability of the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper enhances LLM calibration ability in reasoning tasks by mitigating hallucinations and reducing lazy behaviors. The method uses a curriculum of expert iteration, balancing answering the question and saying 'I don't know' by adapting the reward system according to the difficulty of reasoning tasks. It shows great balancing in assertiveness and conservativeness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper’s reward uses the length of the generated text to measure the level of the exploration. This limits the method’s generalizability and scalability, as it’s challenging to find a universal text processing metric for different types of text (e.g., code, tables). Furthermore, reward parameters such as c1 and c2​ also vary across datasets, which makes me concerned about the method’s effectiveness in situations where training and testing distributions are not similar. \n\n2. The experimental results show that even the best method in this paper achieves overall accuracy significantly lower than Vanilla EI. While I understand that incorporating refusal may lead to a drop in accuracy, achieving performance closer to Vanilla EI would be more compelling, as the current method shows that adding a refusal option will significantly reduce the accuracy." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 weaknesses.\n\nAdditionaly,\n\n(1) Have you tested the correlation between the number of reasoning steps and difficulty, as this is a key assumption of AUTO-CEI? If not, I suggest conducting a test experiment on the MATH dataset, considering it has manually labeled difficulty tags.\n\ntypos:\nline 286 N -> K" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* This paper is well-written and presents clear ideas.\n\n* The idea of aligning responses to the model's capabilities–assertively answering within its limits and declining when tasks exceed them is novel.\n\n* The motivation behind AUTO-CEI is straightforward and strong." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Automatic Curriculum Expert Iteration (AUTOCEI) to enhance LLM reasoning and align responses to the model's capabilities–assertively answering within its limits and declining when tasks exceed them, so as to mitigate hallucination and laziness in reasoning tasks. Through experiments on BoardgameQA, MATH and Blocksworld with Llama-3.1-8B-instruct, The authors demonstrate the effectiveness of AUTO-CEI, achieving superior alignment by effectively balancing assertiveness and conservativeness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* AUTO-CEI introduces additianal training overheads. Considering that the process of AUTO-CEI includes EXPERT ITERATION (each iteration needs large amount of resampling), the additional training overhead can not be ignored. I suggest the authors align the sampling cost between different baselines.\n\n* Limited validation across models. The effectiveness of AUTO-CEI is validated only on Llama-3.1-8B-instruct. Further exploration is needed to assess the generalizability to other models.\n\n* The performance is significantly affected by the hyperparameter λ. Table 2 shows that the performance of AUTO-CEI fluctuates greatly under different λ, which can lead to a considerable amount of additional cost during actual usage." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Auto-CEI pushes and estimates the limits of LLM reasoning capacities and aligns LLM's assertive and conservative behaviours according to these limits for more reliable reasoning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024automatic,\ntitle={Automatic Curriculum Expert Iteration for Reliable {LLM} Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3ogIALgghF},\nnote={under review}\n}" }, "abstract": { "value": "Hallucinations (i.e., generating plausible but inaccurate content) and laziness (i.e. excessive refusals or defaulting to \"I don't know\") persist as major challenges in LLM reasoning. Current efforts to reduce hallucinations primarily focus on factual errors in knowledge-grounded tasks, often neglecting hallucinations related to faulty reasoning. Meanwhile, some approaches render LLMs overly conservative, limiting their problem-solving capabilities. To mitigate hallucination and laziness in reasoning tasks, we propose Automatic Curriculum Expert Iteration (Auto-CEI) to enhance LLM reasoning and align responses to the model’s capabilities--assertively answering within its limits and declining when tasks exceed them. In our method, Expert Iteration explores the reasoning trajectories near the LLM policy, guiding incorrect paths back on track to reduce compounding errors and improve robustness; it also promotes appropriate \"I don't know\" responses after sufficient reasoning attempts. The curriculum automatically adjusts rewards, incentivizing extended reasoning before acknowledging incapability, thereby pushing the limits of LLM reasoning and aligning its behaviour with these limits. We compare Auto-CEI with various SOTA baselines across logical reasoning, mathematics, and planning tasks, where Auto-CEI achieves superior alignment by effectively balancing assertiveness and conservativeness." }, "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", "Reasoning", "Hallucinations", "Laziness", "Alignment" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f1170ffa46567404483005fcf7740f374314589b.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": "Automatic Curriculum Expert Iteration for Reliable LLM Reasoning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. in table2, Why are there no variance experiments to illustrate the stability of various metrics?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The method introduces a novel approach to targeted unlearning by updating a single targeted layer using a one-time gradient computation, which is distinct from more common methods that require iterative model updates across multiple layers.\n\n2. The paper presents two new metrics, layer importance and gradient alignment, to determine the optimal layer and gradient direction for unlearning, enhancing the targeted precision of the process.\n\n3. The experiment was sufficient for me." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an innovative approach to the issue of machine unlearning, which involves removing the influence of specific data subsets from trained machine learning models without retraining from scratch." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Table 2: Performance overview of different unlearning methods on UnlearnCanvas. in this table, My intuition is that there is a lack of variance experiments, that is, running multiple rounds to see the best and worst performance of the algorithm." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Based on the weaknesses part, here are some corresponding suggestions:\n\n1. **Incorporate a Comprehensive Related Work Section**\n \n If it is available, add a dedicated Related Work section that reviews pertinent literature on machine unlearning and saliency-based methods. \n\n2. **Enhance Clarity in the Single Layer Update Methodology**\n \n The methodology for selecting and updating the single targeted layer is not clearly explained, potentially causing confusion among readers. Please follow the weakness part to provide more clear explanations.\n \n3. **Strengthen and Expand the Experimental Evaluation**\n \n Based on the weakness part, could you provide more numerical results on VLM task, and do more experiments under previous evaluation metrcis on image classfication task.\n\n4. **Improve Formatting and Structural Consistency**\n \n The paper's formatting, such as line spacing between titles and sections, lacks consistency, which can detract from readability and professionalism." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Innovative Saliency-Based Approach to Machine Unlearning**\n \n The paper introduces a novel saliency-based method specifically designed to address the machine unlearning problem. The authors present SLUG technique, which efficiently removes targeted information by updating only a single designated layer of the model through a one-time gradient computation. This method offers a streamlined solution compared to traditional unlearning techniques that often require extensive model modifications and incur high computational costs.\n\n2. **Comprehensive Validation Across Diverse Downstream Tasks**\n \n The effectiveness of the proposed SLUG method is thoroughly validated across three distinct downstream tasks, demonstrating its versatility and robustness: CLIP-Based Image Classification, Stable Diffusion-Based Image Generation and Vision-Language Models (VLM)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel saliency-based method for the machine unlearning task. The proposed approach, named Single Layer Unlearning Gradient (SLUG), effectively removes targeted information by updating only a single specific layer of the model through a one-time gradient computation. Compared to traditional unlearning techniques, SLUG significantly reduces computational costs while ensuring minimal impact on the model's performance for unrelated content.\n\nThe authors evaluate SLUG using metrics such as low computational cost, effective unlearning, and utility retention. They demonstrate the method's efficacy across three downstream tasks: CLIP Zero-Shot Classification, Generative Models on UnlearnCanvas benchmark and Vision-Language Models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Lack of Related Work Discussion**\n \n The paper does not include a comprehensive review of related work. This omission makes it difficult to contextualize the proposed method within the existing body of research and to understand how it compares to or improves upon previous approaches in machine unlearning and saliency-based methods.\n\n2. **Insufficient Clarity in Single Layer Update Methodology**\n \n The description of the **Single Layer Unlearning Gradient (SLUG)** method lacks clarity, particularly in the selection and updating of the single targeted layer. This can lead to confusion among readers regarding the following aspects:\n \n - **Balancing Equations (7) and (8)**: The paper does not adequately explain how these equations balance the unlearning process. Additional textual explanations are needed to clarify the interplay between these equations and their role in achieving effective unlearning.\n \n - **Computation of Single Gradient Direction**: The rationale behind choosing the gradient direction based on the initial parameters is not sufficiently elaborated. More detailed explanations are necessary to justify this choice and its impact on the unlearning process.\n \n - **Consistency in Parameter Updates**: Although the authors emphasize updating parameters in a single layer, this point is not clearly reiterated in Section 3.2. Ensuring consistent emphasis throughout the methodology section would enhance understanding.\n\n3. **Limited and Inadequate Experimental Evaluation**\n \n The experimental results presented in the paper are not particularly compelling, and the evaluation metrics used are insufficiently comprehensive. Specific issues include:\n \n - **Unlearning for CLIP (Section 4.2)**:\n - **Optimal Results Visualization**: The results for different learning rates are not clearly highlighted. Using color-coding to indicate the best-performing results would improve readability and interpretation.\n - **Evaluation Metrics Consistency**: The paper does not maintain consistency with established definitions for classification unlearning tasks, such as those outlined in \"Model Sparsity Can Simplify Machine Unlearning.\" Aligning the evaluation metrics with these definitions would strengthen the validity of the results.\n \n - **Unlearning for Stable Diffusion (Table 2)**:\n - **Limited Performance Advantages**: Beyond demonstrating efficiency, the method does not show significant advantages in other performance metrics. This limitation raises questions about the overall effectiveness of SLUG in this context.\n \n - **Application to Vision-Language Models (VLMs)**:\n - **Lack of Reported Data**: Although the paper highlights the application of the unlearning method to VLMs and mentions corresponding evaluation metrics, it fails to report the actual data results. This absence undermines the persuasiveness of the claims regarding the method's effectiveness in VLMs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 check the questions in the weaknesses above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-organized and easy to follow. This approach achieves effective unlearning with just a single gradient update on one layer, demonstrating remarkable efficiency, particularly in the context of large models.\n\n2. In the proposed approach, the author employs the diagonal of the Fisher information matrix to approximate layer importance, thereby enhancing interpretability.\n\n3. The author conducted extensive experiments on large-scale multimodal models including CLIP, Stable Diffusion, and VLMs, demonstrating the wide applicability of the proposed approach and empirically demonstrating its advantages in balancing efficiency and model utility. And the author provided complete code that supports reproducibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author proposes a method that only requires one-time gradient calculation to update a single layer of the model for achieving unlearning. By approximating the importance of the measurement layer using the diagonal of the Fisher information matrix and balancing gradient alignment, the author selects a single target layer and finally updates its parameters in a single step to achieve the desired outcome. The effectiveness of this approach is validated through extensive experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed scheme only updates the most important layer to achieve excellent forgetting effects. Although the experimental results can provide an empirical guarantee for forgetting, intuitively there must be residual information in the remaining layers. From the experimental results, the difference in importance between layers is not large. Hence, it feels more reasonable to update as many layers as possible while maintaining model performance. It is better to add more discussions.\n2. The design of the approach requires access to all forgotten and retained data. However, the targeted domain involves relatively large datasets, requiring substantial storage space. If complete access to the data is not feasible, could this negatively impact the effectiveness of the scheme?\n3. The paper's description of layer selection is not clear enough, and I did not correspond the graph well with the Pareto optimal set. I cannot clearly understand how the author balances importance of layers and grade alignment.\n4. From the experimental results of unlearning for stable diffusion, it can be seen that unlearning leads to a slight decrease in the quality of image generation. \n5. The experiment of unlearning for VLM lacks quantitative analysis and only shows examples. Adding quantitative analysis will provide clearer evidence for the method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "## Questions\n\n1. **Retain Set Curation:** Could the authors provide a detailed explanation of how the retain set is curated? Clarifying this process is essential for reproducibility and assessing the method's robustness.\n2. **Iterative Update Performance:** It is recommended to report the performance of the iterative update version of SLUG. If performance metrics decline, this could highlight underlying foundational issues that need to be addressed." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "## Strengths\n\n1. **Balanced Unlearning and Performance:** The proposed method effectively balances the unlearning process with the model's general performance, addressing a crucial trade-off in model management.\n2. **Computational Efficiency of SLUG:** SLUG requires gradient computation only once, offering two significant advantages:\n - **Faster Computation:** Reduces overall computation time.\n - **Prevention of Over-Unlearning:** Minimizes the risk of excessively removing learned information.\n3. **Generalization Across Models:** SLUG demonstrates effectiveness not only on stable diffusion models but also yields promising results on Vision-Language Models (VLMs), showcasing its potential for broader applicability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a method called Single Layer Unlearning Gradient (SLUG), aimed at addressing the challenges of unauthorized generation of privacy-related and copyright-infringing contents. SLUG is designed for unlearning of targeted information from trained machine learning models, requiring only a single gradient computation (and then reuse it) and updating only one layer of the model. This approach minimizes computational costs and maintains the model’s overall utility, particularly for unrelated tasks.\nThe method has been tested with popular models like CLIP and Stable Diffusion, demonstrating superior efficiency and effectiveness compared to existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Weaknesses\n\n1. **Dependence on Retain Set:** SLUG relies on a retain set to preserve general performance. The methodology for curating this set is critical, yet the paper lacks sufficient discussion or guidelines to ensure reproducibility.\n2. **Incomplete Computational Time Analysis:** While Table 1 presents a computation time comparison, the analysis based on $O(N_f + N_r)$ overlooks key factors:\n - **Iterative Parameter Updates:** SLUG requires iterative updates of model parameters as described in Equation 9.\n - **Layer Importance and Gradient Alignment:** The time associated with determining layer importance and performing gradient alignment is not accounted for, potentially underestimating the actual computational cost.\n3. **Insufficient Evaluation on VLMs:** The claims regarding SLUG's performance on VLMs are not fully substantiated. More comprehensive experiments are necessary to convincingly demonstrate its superiority in this context." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose an efficient unlearning method for targeted information removal from multi-modal foundation models using single layer update with one-time gradient calculation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024targeted,\ntitle={Targeted Unlearning via Single Layer Unlearning Gradient},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3p4raemLAH},\nnote={under review}\n}" }, "abstract": { "value": "The unauthorized generation of privacy-related and copyright-infringing content using generative-AI is becoming a significant concern for society, raising ethical, legal, and privacy issues that demand urgent attention. Recently, machine unlearning techniques have arisen that attempt to eliminate the influence of sensitive content used during model training, but they often require extensive updates in the model, reduce the utility of the models for unrelated content, and/or incur substantial computational costs. In this work, we propose a novel and efficient method called Single Layer Unlearning Gradient (SLUG), that can unlearn targeted information by updating a single targeted layer of a model using a one-time gradient computation. We introduce two metrics: layer importance and gradient alignment, to identify the appropriate layers for unlearning targeted information. Our method is highly modular and enables selective removal of multiple concepts from the generated outputs of widely used foundation models (e.g., CLIP), generative models (e.g., Stable Diffusion) and Vision-Language models. Our method shows effectiveness on a broad spectrum of concepts ranging from concrete (e.g., celebrity name, intellectual property figure, and object) to abstract (e.g., novel concept and artistic style). Our method also exhibits state-of-the-art efficiency with effective unlearning and retention on the comprehensive benchmark UnlearnCanvas. Our code is available at https://anonymous.4open.science/r/SLUG-6CDF" }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Machine unlearning", "multi-modality", "CLIP", "vision-language model (VLM)", "stable diffusion", "privacy protection", "copyright protection", "trustworthy and safe machine 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/a36a86b5db4a3df168395e176bcecc3c7f1035c3.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": "Targeted Unlearning via Single Layer Unlearning Gradient" }, "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 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": "The BRAL-T framework is quite effective considering good performance on some standard datasets. Overall, the paper is easy to follow and well explained." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a data selection method named as TrustSet which takes into account uncertainty, diversity and class distribution. TrustSet in combination with reinforcement learning based sampling policy introduced BRAL-T framework which extends the benefits of TrustSet to select meaningful data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It’s not clear why partcularly GradNd score is used for TrustSet extraction in case of large datasets and complex models. GranND scores appear to be highly sensitive towards small changes in model parameters and may also add high computational overhead.\n\nPerhaps large datasets such as ImageNet could be used to test the performance and time complexity of the proposed method in an image classification setup.\n\nIt’d be interesting to understand the motivation to select negative Wasserstein distance as reward function since it’s considerably computationally expensive and might turn out to be challenging in high-dimensional spaces. It’d also be useful to compare the proposed framework to a few recent baselines as well." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* I find some of the experimental settings to be confusing. As stated in Appendix B, the active learning experiments involved querying nearly the entire unlabeled datasets—for instance, 5,000 out of 5,436 for BreakHis, 5,000 out of 5,132 for PneumoniaMNIST, and 4,000 out of 4,695 for Waterbird. This large query budget effectively undermines the purpose of active learning. I understand these settings follow [1], but I would appreciate it if the authors could explain their rationale for adopting such an approach.\n\n* In my humble opinion, Figure 1 seems to be a generic framework of active learning with RL, and does not provide much insight for the proposed BRAL-T method. I would suggest the author add more information to Figure 1 or merge it with Figure 2.\n\n* I would kindly suggest the authors double-check their citations. For example, CoreSet is referred to [2] instead of the original paper, and both BADGE and KMeans refer to the same paper [3].\n\n\n[2] Zhan, Xueying, et al. \"A comparative survey of deep active learning.\" arXiv preprint arXiv:2203.13450 (2022).\n[3] Ash, Jordan T., et al. \"Deep batch active learning by diverse, uncertain gradient lower bounds.\" arXiv preprint arXiv:1906.03671 (2019)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The paper introduces a novel TrustSet approach designed to balance class distribution and mitigate the long-tail problem present in CoreSet. In addition, incorporating reinforcement learning to extend the properties of TrustSet, based on labeled data, to unlabeled data is a promising direction in active learning.\n\n* The authors perform comprehensive experiments across eight active learning benchmarks and various long-tailed active learning / fine-tuning tasks. The authors also perform rigorous baseline comparisons and ablation studies, demonstrating that each component of the framework contributes meaningfully to performance improvements. \n\n* The presentation is overall well-organized, with detailed figures and algorithms that illustrate each component of the proposed methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Batch Reinforcement Active Learning with TrustSet (BRAL-T) framework, which combines batch active learning (BAL) and reinforcement learning (RL). This method introduces TrustSet, which selects a balanced subset of labeled data that improves model performance, especially for data with long-tail distribution. To adapt TrustSet for unlabeled data, BRAL-T uses a Reinforcement Learning (RL) sampling policy trained on the labeled set to select unlabeled samples approximating the qualities of TrustSet data. The authors validate the performance of BRAL-T across several image classification and fine-tuning benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* To my understanding, TrustSet is a pruning/selecting strategy for **labeled data**, and the benefits highlighted in the abstract \"selects the most informative data from the labeled dataset, ensuring a balanced class distribution to mitigate the long-tail problem\" and \"optimizes the model’s performance by pruning redundant data and using label information to refine the selection process\" (L018-L022) both apply to the **labeled data**. However, the authors do not provide theoretical proof or experimental results to substantiate these claims regarding the **labeled data**. \n I acknowledge the authors provide some discussion about BRAL-DiffSet in Section 5.4. However, in my personal opinion, this ablation study validates the effectiveness of EL2N score on active learning, rather than the effectiveness of TrustSet on the selection of labeled data. Given that TrustSet is a key contribution of this work, I believe it is crucial to validate its effectiveness through empirical evidence or theoretical analysis. \n\n* Setting the number of candidate actions $A_c$ to a fixed number might be an issue, particularly for imbalance/long-tail datasets. In particular, if the clustering divided the unlabeled $U$ into $C$ as expected with each cluster primarily containing samples from the same class, the resulting $c$ clusters would be highly imbalanced. In such case, setting $A_c$ to a fixed number will make the sub-clusters, i.e., the candidate actions, $U_c^a$ to be imbalanced as well across $c$ clusters, which seems to contradict the objective of achieving balanced data selection. Can the authors elaborate on why choosing a fixed value for $A_c$?\n\n* It seems that the experiments do not contain any RL-based active learning baselines. Could the authors elaborate on why TAILOR [1] is not included, as it also aims to find class-balanced examples in active learning by incorporating RL?\n\n* Some presentation issues. For example, most citations in Section 2 and Section 5 are in double brackets; The legends in Figure 4 are barely readable; The captions are above tables in the main manuscript but are below tables in the Appendix; the baseline methods WAAL, LossPrediction, and SIMILAR are not discussed in Related Works." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": 1 }, "primary_area": null, "questions": { "value": "Please see “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. Defining state space with $L$ and $U$ is novel.\n2. Extensive experiments on long-tail datasets are interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes the Batch Reinforcement Active Learning with TrustSet (BRAL-T) ensuring a class-balanced sampling for long-tail problem. By introducing RL in batch active learning scenario, TrustSet selects high-quality samples from the unlabeled data. Extensive experiments demonstrate the effectiveness of the proposed BRAL-T." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Understanding and reading research papers is quite challenging. The notation should be well-defined in Section 3.\n2. TrustSet naively incorporates the concepts of class-balanced and curriculum learning into the existing methods such as GradNd and Super Loss.\n3. There are cases where a subset $S$ is subsampled from $L$. I’m curious about the intuition behind this approach and why the entire set $L$ isn’t used." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "No" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. It is a good trial to adopt the concept of reinforcement learning to help address batch active learning.\n2. This paper is structured, and well-written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed an RL-based method to address a classical setting, i.e., active learning. Specifically, this paper designed the reward function, state space, and action space to help the sample query stage. The experimental results were conducted among 5 benchmarks, validating its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Insufficient Experimental Support on the Design of Curriculum Learning Mechanism When learning the TrustSet, this paper adopts some concepts from curriculum learning, i.e., super loss, to help address the uncertainty of the selected samples, which stands strongly on an argument \"...Data with high GradNd scores tend to be difficult and uncertain samples...\", and this paper uses the results in Figure 3 to claim the rightness of this conjecture. However, compare to the version in CVPR’24, the authors delete the pilot experimental parts, (Figure 3) in original version, which further weakens the claim since no pilot experiment is built. Most importantly, we are not clear whether this issue only occurred for the GradNd-based methods since only this one is shown (What if other methods, like uncertainty score, BatchBald, GrandMatching, Submodular,..., have no such issue?), which I believe needs more comparing methods to help support such motivation to use the SuperLoss.\n\n2. Unclear performance bound about the TrustSet. The aim of this paper is to construct and optimize the selected samples, treating them as the best suited during each query. This also suggests that using all samples could be the best. Therefore, there should be a theoretical performance bound about the proposed RL method, which aims to approximate the optimal \"TrustSet\". Besides, does TrustSet have to be built by GraNd? Why not other methods? This paper needs to make a clear illustration for that.\nThe whole design of RL framework is built for the cluster-wise groups instead of each sample, which is obviously different from those traditional methods focusing on the sample score. It also means that the proposed RL method is only applicable to the group-level selection, which, intuitively, may not be effective when the number of samples in the unlablled pool is small (since each sample counts).\n\n3. I do not see clear novelty and value of such RL-based design. As this paper claims in Related Work about \"Active Learning with RL\", the deficiency of some methods using accuracy metrics are \" ...the relationship between the target model’s predictions and the training set is complex, making it difficult to train the RL policy...\" However, the reward function in the proposed framework is not based on an intuitive evaluation criteria but a TrustSet, which is not as measurable as the former. So how such a design could be more \"simple\" than those criteria-based methods to train the RL policy? Besides, the most important of RL is the design of the reward function, but I cannot see the clear value and novelty of designing such this reward function.\n\n4. Based on the results in Figure 4, this method works not as promising as expected. Besides, where is the results of Tiny-ImageNet with progressive data volume?not 2%-3%\n\n5. The time complexity analysis of this RL-based method lacks comparison to other methods, and the experimental results about that shall be presented (such as training time, FLOPS...)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024labeled,\ntitle={Labeled TrustSet Guided: Combining Batch Active Learning with Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3qDB9j6p3S},\nnote={under review}\n}" }, "abstract": { "value": "Batch active learning (BAL) is a crucial technique for reducing labeling costs and improving data efficiency in training large-scale deep learning models. Traditional BAL methods often rely on metrics like Mahalanobis Distance to balance uncertainty and diversity when selecting data for annotation. However, these methods predominantly focus on the distribution of unlabeled data and fail to leverage feedback from labeled data or the model’s performance. To address these limitations, we introduce TrustSet, a novel approach that selects the most informative data from the labeled dataset, ensuring a balanced class distribution to mitigate the long-tail problem. Unlike CoreSet, which focuses on maintaining the overall data distribution, TrustSet optimizes the model’s performance by pruning redundant data and using label information to refine the selection process. To extend the benefits of TrustSet to the unlabeled pool, we propose a reinforcement learning (RL)-based sampling policy that approximates the selection of high-quality TrustSet candidates from the unlabeled data. Combining TrustSet and RL, we introduce the **B**atch **R**einforcement **A**ctive **L**earning with **T**rustSet (**BRAL-T**) framework. BRAL-T achieves state-of-the-art results across 10 image classification benchmarks and 2 active fine-tuning tasks, demonstrating its effectiveness and efficiency in various 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": [ "Active learning", "Reinforcement Learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/085f25b2456a6d17e9cb4e08dd04d2f395ef88b5.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": "Labeled TrustSet Guided: Combining Batch Active Learning with Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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 about KANs makes them a compelling choice for this application in particular: encoding numeric tabular data to pass into a downstream transformer? E.g., what theoretical properties are especially relevant for this application? Why not use them in other parts of the model?\n\n- Did you evaluate other numeric encoders within the same framework as your model, such as linear, MLP, and/or piecewise linear encoders?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- Exploring the potential uses of KANs in tabular modelling is a relevant current topic.\n- The model shows some promise in outperforming other NN models.\n- Code is provided to reproduce some of the results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a discriminative model for tabular data that is distinct from previous models in that it uses a KAN for embedding numeric features rather than linear or MLP layers. Its performance is compared to a selection of GBDT and NN tabular models. In the experiment settings evaluated, TabKANet consistently outperforms the other NN models but is mostly outperformed by the GBDT models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "On the whole, I think paper has a poor contribution due to issues with the evaluations and a lack of depth in justifying the proposed modelling choices. The overall novelty of the method is limited, being a relatively simple combination of existing modelling components (KANs, batch normalization, and transformers), and I don't think the rest of the paper has the depth or soundness to make up for that.\n\n- No hyperparameter tuning is done for baseline models. This is especially problematic for GBDT models, which require hyperparameter tuning for a practical comparison. This is out of line with prior research (e.g., see Gorishniy et al. 2021) and severely limits the utility of the comparison with GBDTs.\n\n- The NN models being compared to are not state-of-the-art, so the comparison to them does not indicate much about the paper's contribution. I would have at least liked to have seen TabR and TabPFN included, and other recent models that are widely used as baselines such as FT-Transformer and MLP-PLR would have been welcome.\n\n- The discussion of why KANs should be used in this area is lacking in depth. The paper doesn't provide theoretical or empirical insights into how they would be useful for tabular data in particular. Instead, there are just vague references to their flexibility. Ablations are also not provided to compare the contribution of the KAN part alone versus other encoders.\n\n- Using batch normalization instead of layer normalization is a fairly trivial tuning choice, and the discussion on page 5 is unclear and does not provide significant technical insights to justify treating it as an important decision.\n\n- Parts of the paper contain misleading claims:\n\t- The introduction indicates that existing attempts to use transformers in tabular modelling are using them to encode categorical variables, when in fact there's a much wider variety of transformer models for tabular data (some of which are given in the Related Work). In general, the proposed model is not contextualized with respect to the entire range of existing tabular transformer models.\n\t- The introduction also claims that the proposed model achieved \"identical performance\" to GBDTs on almost all datasets - the performance was not identical, and was lower on average in most cases (evaluation issues notwithstanding).\n\t- \"Current scientific research has not yet proposed a simple, stable, and universal numerical embedding module\" - this is a very strong claim that is not justified. MLPs, linear layers, and piecewise linear encodings arguably satisfy these criteria just as well as the proposed solution." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* How were the datasets for the evaluation selected?\n* How where hyper-parameters tuned for all models?\n* How well does the model perform when removing the KAN layer?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper discusses encoding of continuous values for tabular classification. This is a hot topic, and the combination with the recently proposed KAN architecture is timely.\nThe model is evaluated on a wide variety of tasks and the datasets are discussed in detail." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a model for tabular data based on the Transformer architecture, similar to TabTransformer, but including a KAN layer for encoding continuous features. The model is evaluated on binary and multi-class classification tasks as well as regression and is found to outperform deep learning baselines and in some cases GBRT models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "# Main concern\nGiven that the novelty of the method is relatively small, the empirical evaluation of the method is critically important. However I have several concerns regarding this:\na) It's unclear how the datasets for evaluation were selected. There is many established benchmark suites, such as the AutoML benchmark, TabZilla, OpenML-CC18 and the Grinsztajn collection. Not adhering to a standard benchmark suite allows for cherry-picking of dataset.\n\nb) Since the emphasis of the paper is on the encoding of continuous features, I think FTTransformer, and Gorishniy et al: On Embeddings for Numerical Features in Tabular Deep Learning are critical baselines to compare against.\n\nc) Some important ablations are missing; in particular, what happens if the KAN layer gets replaced by simple input scaling? This seems to be different form TabTransformer, which skips the transformer entirely for continuous data. Also, Table8 shows a big improvement between the LN and BN versions. An obvious comparison here would be to TabTransformer with BN.\n \n\nd) The paper doesn't describe how hyper-parameters for the methods were tuned.\n\ne) Some deep baselines are missing. While it's not reasonable to compare against all publications, I would suggest comparing against TabR and TabPFN (even when subsampled to a maximum of 3000 samples, TabPFN performance is still strong, see McElfresh.\n\n\nI think the clarity of the paper could also be improved, and the novelty of the KAN is overstated, in particular wrt existing neural networks and \"On Embeddings for Numerical Features\".\n\n# Other Concerns\n## Overselling KAN\nThe paper states in several places that KANs are more powerful than MLPs; however, that is not strictly the case. A KAN can easily be represented with an MLP by constraining the MLP structure and potentially changing the activation function. For example a KAN with piecewise linear splines with r pieces is equivalent to an MLP with ReLU activations where each node is replaced by a small neural network with r nodes. This view calls into question claims like line 63 \"This feature offers neural networks\n more flexible performance compared to Multilayer Perceptron\" and Line 115 \"rigidity in MLP\".\nAlso, neural networks with spline activation functions have long been studied, and it's unclear what (if any) novelty can be attributed to KANs.\n\n## Minor suggestions\nLine 028 \"ordered different features\" is hard to read and not very clear.\n\nLine 028: add citations for most commonly used and oldest business data format. I am not convinced by these claims. A lot of data is actually in spreadsheets and relational databases, neither of which are tabular data in the sense of the standard ML datasets. NoSQL data is also extremely common.\n\nLine 036: Citing Hollman for the prevalence of GRBT is strange, since TabPFN is a deep model that clearly outperforms GRBT.\n\nLine 043: None of the three arguments for neural networks seems sound, and I am a firm believer in using neural networks on tabular data. 1) is vague, 2) it is unclear what is meant by scalability and how it relates to multimodality 3) unsupervised schemes exist for tree-based models, though they are not as common.\n\nLine 070: It's unclear what is meant by \"business structure framework\"\n\nLine 215 1): It's unclear wrt to what baselines you are discussing improvements. Both LN and BN are common in transformer models. Is this wrt KAN or wrt TabTransformer?\n\nLine 218 3): This is just concatenating features, right?\n\nFigure 2: The meaning of d is unclear, it seems to be embedding size. However, it's unclear why the output would be reshaped to (m+n) * d? Is this just to input into the MLP? Also, the figure shows multiple rows in input and output, but the model operates on one row at a time, right?\n\nLine 223: It's unclear what's meant by \"subconscious best solution\".\n\nLine 377: \"predict auc scores\" I think you mean predict the target and compute AUC scores?\n\n## Typos\n\nLine 027 \"The tabular\" -> \"a tabular\"\n\nLine 030 \"medicineI\" ->\"medicine\"\n\nLine 053 \"they have used Transformers\" -> Transformers have been used.\n\nFigure 1: \"Nurmerical\" -> \"Numerical\"\n\nLine 212: Numercal -> \"Numerical\"\n\nAcknowledgements contain author instructions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- General question. Why are numerical and categorical data processed separately? Would it make sense to explore a combined representation that integrates both types of features?\n\n- Line 221: The statement *\"Firstly, normalization for numerical items is crucial, which is essential to avoid gradient explosion, especially with real-world data.\"* raises a question: Why not simply normalize the numerical data before feeding it to the neural network? Using batch normalization introduces additional learnable parameters and can significantly affect training, as the normalization depends on batch size. Although this question is somewhat addressed later (Line 238), it introduces another point of confusion: *\"Repeatedly pairing numerical normalization results and category features will bring additional training data.\"* This may be more accurately described as data augmentation rather than additional data, which could be explored through an ablation study.\n\n- Line 223: Could you clarify the phrase *\"This is a subconscious best solution\"*? The entire sentence is somewhat confusing and may need rephrasing for clarity.\n\n- Line 255: The sentence about data splitting is unclear. Did you apply cross-validation, or was the data split into three groups: training, testing, and validation?\n\n- Tables 2 and 3: What is the motivation for separating the neural network and machine learning baselines? A unified comparison would improve clarity.\n\nMinor Comments\n\n- In the motivation, you mention that self-supervised learning is a major strength of deep learning approaches. How could TabKANet be adapted for use in self-supervised settings?\n\n- Section 3: The statement \"As mentioned in Sec. 2.1, GBDTs outperform NNs in table modeling tasks because of the skewed or heavy-tailed features in table information\" reads as a hypothesis rather than a definitive fact. Could this be clarified?\n\n- Line 267: *\"MLP is a traditional deep learning model consisting of multiple layers of neurons.\"* How are you defining a \"traditional\" deep learning model here? A more specific description might be helpful." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper addressing an important issue of the structured tabular learning \n- Authors utilize various tabular datasets from different domains" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces TabKANet, a model that integrates Kolmogorov-Arnold Network (KAN) and Transformer architectures to improve the handling of structured tabular data. The authors demonstrate that TabKANet outperforms selected deep learning baselines in various tasks. However, its performance gains over decision tree-based ensemble methods, such as Gradient Boosted Decision Trees, are minimal." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The explanation of KAN and the model’s overall structure is brief and lacks clarity. The paper’s organization could be improved to aid readers in understanding the background and methodology.\n\n- The proposed method, TabKANet, offers only marginal improvements over existing models on key metrics for several datasets. For instance, on 2 out of 6 datasets, the AUC improvement compared to CatBoost is very small, with a difference of only **0.003**. This casts doubt on the claim from the abstract that *\"Its performance is comparable to or surpasses that of Gradient Boosted Decision Tree models (GBDTs).\"* Overall, such minor performance gains may not justify the added complexity of the approach.\n\n- The paper omits several recently proposed methods for deep tabular learning, such as TabPFN, GANDALF, DeepTLF, and more in [1]. Including these baselines would provide a more comprehensive comparison and better contextualize the performance of TabKANet. \n\n- The evaluation of TabKANet could be strengthened by using widely recognized benchmarks, such as Tabzilla, which is well-accepted by the community and would provide a more robust assessment of the model's effectiveness.\n\n\n[1] Borisov, Vadim, Tobias Leemann, Kathrin Seßler, Johannes Haug, Martin Pawelczyk, and Gjergji Kasneci. \"Deep neural networks and tabular data: A survey.\" IEEE transactions on neural networks and learning systems (2022)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce TabKANet, a novel model that leverages a KAN-based Numerical embedding module and Transformer to overcome neural networks' limitations in tabular data. It achieves performance comparable to or exceeding GBDT models on various datasets." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tabkanet,\ntitle={Tab{KAN}et: Tabular Data Modeling with Kolmogorov-Arnold Network and Transformer},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3qDhqj6qfu},\nnote={under review}\n}" }, "abstract": { "value": "Tabular data is the most common type of data in real-life scenarios. In this study, we propose the TabKANet model for tabular data modeling, which targets the bottlenecks in learning from numerical content. We constructed a Kolmogorov-Arnold Network (KAN) based Numerical Embedding Module and unified numerical and categorical features encoding within a Transformer architecture. TabKANet has demonstrated stable and significantly superior performance compared to Neural Networks (NNs) across multiple public datasets in binary classification, multi-class classification, and regression tasks. Its performance is comparable to or surpasses that of Gradient Boosted Decision Tree models (GBDTs). Our code is publicly available on GitHub: https://github.com/AI-thpremed/TabKANet." }, "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": [ "Tabular Data Modeling; Kolmogorov-Arnold Network; Numerical Feature Embedding" ] }, "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/5f3beb5443afd972cd5e1ed1d6a376afad0fc0d8.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": "TabKANet: Tabular Data Modeling with Kolmogorov-Arnold Network and Transformer" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Can we repeat the produce of finding A->B'-C for each segment recursively to obtain A->E'->D'->B'->F'->G'-C (maybe a longer path), such that there is no essential barrier existing?\n2. How can we use the input mode connectivity to give a picture of the decision boundary of DNNs?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This topic is interesting. Investigating the model connectivity in the input space could help us to shape the decision boundary of DNNs.\n2. The insight of mode connectivity is indeed an intrinsic property of high-dimensional geometry is important, as it might be able to explain various phenomena lied in the field of model connectivity, such as wide neural networks are easier to satisfy mode connectivity after accounting for permutation invariance.\n3. The potential application towards adversarial examples is insightful, which might explain the existence of adversarial examples." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors identify an interesting phenomenon, namely input mode connectivity, where samples with similar predictions could be approximately linearly interpolated such that the interpolated sample remains a low loss." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The major issue of this work is that the investigation is not in-depth enough. For example, in Fig. 3, the path A->B'->C and A->B->C look similar but they differ significantly in terms of mode connectivity. \n - How should we quantify such differences? or why the small difference B'-B (as shown in right bottom of Fig. 1) is significant in terms of model connectivity? \n - Here is another example, in the adversarial example part, why real-adversarial pair shows a large barrier than real-real pair? An intuitive explanation is at least expected.\n\nThese are all important questions and represent the motivation why we are interested in investigating the mode connectivity in input space.\n\n2. Their theory cannot explain the phenomenon they discovered. Their conjecture is only able to explain the mode connectivity for untrained NNs with infinity large input dimension. However, in their experiments, two real images with similar predictions are usually not connected unless another intermediate point is found, say B'. Clearly, their theory cannot explain the realistic scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Do you have any intuition whether results would fundamentally change if other architectures were considered? E.g. [1] find that mode connectivity in the parameter-sense is influenced by the choice of architecture, i.e. there is different behaviour for vision transformers or multi-layer perceptrons.\n2. I’m a bit confused regarding the adversarial example detector; Are you comparing loss barriers after a single iteration of your algorithm in both cases, or after two iterations in case of the adversarial setting? I thought that in both cases the barriers became very small? \n\n[1] Disentangling Linear Mode-Connectivity, Altintas et al., 2023" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The role of the input when it comes to loss behaviour is somewhat understudied and the authors develop new ideas in this direction while keeping things very analagous to the results observed for parameter loss landscapes.\n2. The authors give further credibility to their results by mathematically proving them in an idealized setting assuming independence. While this is not realistic, I do find the argument of the authors convincing that correlations in this case will most likely help connectivity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies input space mode connectivity, where instead of studying the loss surface as a function of the parameters, the authors consider varying the inputs (for a fixed set of parameters). More specifically, the authors investigate paths connecting two sets of inputs and the resulting behaviour of the loss, in complete duality to the well-known mode connectivity in parameter space. Several choices for “modes” in this context are explored: (1) Validation data points that achieve very low loss, (2) a validation data point and an adversarial example optimized towards the same class but based on a datapoint of different class and (3) synthetic “optimal\" points optimized to maximise a given logit. For all these cases, the authors show how simple piecewise linear paths suffice to connect such points, while limiting the barrier to very small values. Adversarial examples exhibit larger barriers in a significant manner, allowing a detector to leverage this difference to classify whether an image is adversarial or not." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The biggest weakness is the lack of motivation presented in this paper for input space connectivity. Why is this an interesting quantity to study? In case of the parameter loss landscape where this notion originated, the motivation was from an optimisation point of view; is SGD attracted to a convex region of the parameter space? Does it find isolated minima or are there entire regions of low loss? There might be good motivations for input space connectivity as well (I’m not an expert in this area) but the paper does not do a good job at presenting it in its current form. In general I also have no intuition whether it is surprising that real-real images can be connected with two segments or not, etc. \n\n2. I like the idea of using the difference in barrier between real-real and real-adversarial inputs, but as usual in adversarial robustness, I think that new threat models need to be investigated when taking this idea into account. I.e. can one now develop adversarial examples that are designed to mimick the barrier of real examples, thus fooling the new classifier? I don’t expect the authors to necessarily develop such an algorithm but this possibility should at least be discussed in the paper.\n\n3. I also have a hard time interpreting the adversarial example results. It is not that surprising that it requires more segments to connect things properly compared to the real-real scenario (the image is still very different after all). How does this compare to simply two images coming from different classes and their barriers in between? \n\n4. The writing of the paper is not very satisfying. The notation is rather sloppy (e.g. \"0.1*MSE for image deviation and 1e-7*high-frequency penalty\"), optimization details are listed without defining them (e.g. high-frequency penalty). The actual algorithm to obtain a piece-wise linear curve is never properly defined." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- I don't understand the algorithm presented in Section 4.2.1. In my understanding, the claim is that if one of the endpoints is from an adversarial attack, then the method of optimizing $B$ to find a connected (piecewise linear) path will not work. However, the algorithm of detecting adversarial attack is to use the interpolation loss curve and logits as the input and do a classification. I don't see how this proposed algorithm is related to the finding.\n- The wording of Section 4.3 is too confusing that I totally can not understand. What does \"natural datasets for untrained models\" mean? What does \"starting from Gaussian noise\" mean (start from it and do what?) What does \"high-frequency penalization\" mean?\n- In Conjecture 5.1, why does the condition has \"for any probability $0 < p < 1$\", but $p$ is never used in the statement? What does \"almost always connected\" mean?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This phenomenon is interesting and potentially valuable for understanding the behaviour of deep neural networks and the geometry of high-dimensional spaces.\n- This finding has a practical useness that it can be used to detect adversarial attacks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an interesting phenomenon, that given two input points $A,C$ which are classified as the same class by the model (i.e. have similar output), there exists a point $B'$, such that the linear interpolation between $A,B'$ and $B',C$ are all classifed as the same class. The authors refer to this phenomenon as \"input space mode connectivity\", despite they are not actually linear connected. The authors also presented a analysis under a very strong and unrealistic condition, conjecturing that this phenomenon is intrinsic to high-dimensional spaces." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The presentation is kind of confusing. See Questions for concrete issues.\n- In my understanding, the finding in this paper is not really linear connectivity as the path found by the proposed method is actually a piecewise lienear path with 2 pieces. This makes the title and introduction kind of misleading. \n- The experiments are only on a few image classification tasks and models. It is not clear if this phenomenon is general enough. \n- It seems that the thoretical explanation presented in Section 5.2 (even if we omit the too strong assumption of randomized labeling of each grid) only explains why there can exist a connected path, but does not guarantee that the path is linear, which is inconsistent with Conjecture 5.1.\n- In Conjecture 5.1, why do you need to assume \"a subset of input space $X' \\subseteq X$\"? Does the conjecture hold even if $X'$ itself is unconnected? Moreover, there is a statement \"two random inputs $x_0, x_1 \\in X'$\". What is the distribution of the randomness? If it is uniform, then there must be extra constraints on $X'$ (such as compactness), since not every subset of a Euclidean space has a uniform distribution. (For example, you can not draw two points uniformly from a 2d plane)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We extend the concept of loss landscape mode connectivity to the input space of deep neural networks." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024input,\ntitle={Input Space Mode Connectivity in Deep Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3qeOy7HwUT},\nnote={under review}\n}" }, "abstract": { "value": "We extend the concept of loss landscape mode connectivity to the input space of deep neural networks. Mode connectivity was originally studied within parameter space, where it describes the existence of low-loss paths between different solutions (loss minimizers) obtained through gradient descent. We present theoretical and empirical evidence of its presence in the input space of deep networks, thereby highlighting the broader nature of the phenomenon. We observe that different input images with similar predictions are generally connected, and for trained models, the path tends to be simple, with only a small deviation from being a linear path. Our methodology utilizes real, interpolated, and synthetic inputs created using the input optimization technique for feature visualization. We conjecture that input space mode connectivity in high-dimensional spaces is a geometric effect that takes place even in untrained models and can be explained through percolation theory. We exploit mode connectivity to obtain new insights about adversarial examples and demonstrate its potential for adversarial detection. Additionally, we discuss applications for the interpretability of deep networks." }, "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": [ "mode connectivity", "input space", "deep learning", "adversarial detection", "interpretability", "percolation theory" ] }, "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/b43c68644eb8bc147d6ba96272238c7c7a370d0d.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": "Input Space Mode Connectivity in Deep 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": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I am not familar with this field at all and I cannot evaluate the novelty of this work (especially over its predecessor / related works). It is an interesting read and I didn't identify significant issues. The paper is well motivated with solid analysis and experiments. My major concern is that this paper only evaluates on toyish environments and may not generalize well to real-world scenarios." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well-motivated and novel. It successfully extends the Dreamer framework to handle asymmetric information in RL, presenting a novel ACPOMDP framework. It shows the effectiveness of using privileged Information.\n- The results from Safety-Gymnasium benchmarks show that AsymDreamer outperforms baseline models in both task performance and safety metrics, especially in complex scenarios like 3D navigation. The thorough ablation studies are also conducted.\n- The paper includes rigorous theoretical analysis and compares ACPOMDP to standard CPOMDP." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents AsymDreamer, an approach based on the Dreamer framework that specializes in exploiting low-dimensional privileged information to build world models, thereby enhancing the prediction capability of critics. AsymDreamer employs the Lagrangian method to incorporate safety constraints. This paper formulates the proposed approach as an Asymmetric CPOMDPs (ACPOMDPs) framework. Experiments on the Safety-Gymnasium benchmark demonstrate that AsymDreamer outperforms existing approaches in both performance and safety." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am not in this field so I am unable to evaluate the signficance of the proposed method and ACPOMDP problem/analysis.\n\nSome of my concerns include:\n- Over-reliance on privileged information: AsymDreamer’s performance relies on privileged information during training, which may not be available or hard to simualte in real-world environments. It potentially leads to bad performance or compromised safety in environments with limited or unavailable privileged information.\n\n- Extension to real-world environments: The experiments on Safety-Gymnasium benchmark are a bit toyish (even the most challenging 3D navigation one, although it may also be the cases for the related work). Additional testing in real-world environments would be beneficial to demonstrate the 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": 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. Please clarify the similarities and differences between the proposed method and SafeDreamer[1].\n2. The reviewer is confused about what the global state in your privileged world model input is and where the privileged information (the low-dimensional vector from Appendix E) is used.\n\n[1] Huang, Weidong, et al. \"Safe dreamerv3: Safe reinforcement learning with world models.\" arXiv preprint arXiv:2307.07176 (2023)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The theoretical analysis of ACPOMDPs in this paper is comprehensive. The authors propose the ACPOMDPs framework and provide a detailed analysis showing that asymmetric inputs reduce the number of critic updates and lead to a more optimal policy compared to the standard CPOMDPs framework.\n2. The authors introduce privileged information into the RSSM world model, enhancing the model's imaging capabilities and consequently improving the performance of the safe policy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose AsymDreamer, a Dreamer-based safe reinforcement learning framework that utilizes low-dimensional privileged information to construct world models. The world model in AsymDreamer features two branches: the Privileged World Model, which takes a handcrafted low-dimensional vector as input, and the Observation World Model, which uses partially observed images (64x64 RGB). Additionally, the authors formulate their approach within the framework of Asymmetric CPOMDPs (ACPOMDPs) and integrate AsymDreamer with the Lagrangian method. Empirical results show that AsymDreamer outperforms existing safe reinforcement learning methods on the Safety-Gymnasium benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The methodology in this paper is too similar to prior work[1], and lacks sufficient novelty. In my view, the only difference is the inclusion of privileged information in the modeling of the RSSM world model.\n2. The authors are encouraged to include an ablation study to analyze the impact of adding privileged information to the observation input on the baseline algorithm.\n3. There are some typos in the paper that need to be corrected in the next version (e.g., line 400 and figure 2).\n\n[1] Huang, Weidong, et al. \"Safe dreamerv3: Safe reinforcement learning with world models.\" arXiv preprint arXiv:2307.07176 (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": "1.\tI think in Figure 1, the encoder, decoder, and hidden state are different for the two world models. So, it would be clearer to use different colors of notations for them to avoid confusion.\n2.\tWhy not use SafeDreamer as a baseline?\n3.\tIn Figure 4, why do all baselines have constant cost? Don’t they vary across training steps? What is the target cost limit?\n4.\tWhat does the red solid line mean in Figure 4 and Figure 5?\n5.\tThere is no explanation of the baselines. What does OSRP mean? It would be better to include simple descriptions of each baselines in the appendix.\n6.\tThe ablation study results in Figure 5 raise a lot of questions. I think the authors also feel surprised that the privileged world model fails to train a viable cost predictor when taking privileged information as input. In addition, DreamerV3 is much better than AsymDreamer(S) in terms of the reward. This is also hard to interpret. I think the authors should do more experiments to explain these results to provide insights into the model. Otherwise, there seems no clear message that can be summarized in this paper." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tThe idea of using privileged information in the Dreamer structure is interesting. Separating observation modeling and task-centric prediction modeling avoids the potential trade-off between these two tasks. This also allows the observation world model to capture more detailed observation information, thus enabling the actor model to achieve better performance with richer input features.\n2.\tThe paper is generally well-written and well-organized. Figure 1 clearly shows the training pipeline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the safe RL problem that struggles with performance degradation and often fails to satisfy safety constraints. They attribute this problem to the lack of necessary information in partial observations and inadequate sample efficiency. Specifically, they exploit low-dimensional privileged information to build world models, thereby enhancing the prediction capability of critics. The authors propose Asymmetric Constrained Partially Observable Markov Decision Processes, a relaxed variant of CPOMDPs. The key distinction is that ACPOMDPs assume the availability of the underlying states when computing the long-term expected values. To ensure safety, they employ the Lagrangian method to incorporate safety constraints. The experiments conducted on the SafetyGymnasium benchmark demonstrate that the proposed approach outperforms existing approaches dramatically in terms of performance and safety." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\t**Justification of using privileged information**. It seems that the motivation for using privileged information is this sentence “Since training is often conducted in simulators, there is potential to leverage privileged information during training to reduce uncertainty from partial observations”. Does the proposed method have potential applications beyond simulations, such as in real-world scenarios? In addition, I am not sure if it is fair to compare with other methods that do not use privileged information.\n2.\t**Evaluation results**. The experimental evaluation is only conducted on 4 tasks, including one self-made task. The authors may need to include all the remaining tasks in the Safety Gymnasium. \n3.\t**Missing baseline on the same benchmark** Although this paper proposes a model-based method, I think it is still meaningful to compare with some well-known methods on the same benchmark. This website contains some results that can be used as reference: https://fsrl.readthedocs.io/en/latest/tutorials/benchmark.html. \n4.\tThere is no clear evidence to show the benefit of using privileged information. The results in Figure 5 need more investigation and explanation. Otherwise, it is hard to summarize the main conclusion 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": 3 }, "primary_area": null, "questions": { "value": "It would be great if the authors could address the weaknesses I outlined above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The proposed method of combining DreamerV3 with two world models and Lagrangian methods, which is novel and well-motivated.\n* The paper presents empirical results that compare AsymDreamer with many relevant model-based and model-free baselines in several tasks, achieving competitive performance on the Safety-Gymnasium benchmark.\n* The paper is mostly well-written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this work, the authors address the challenge of exploiting asymmetric inputs under the CPOMDPs framework. They propose AsymDreamer, a new algorithm that uses privileged information to build a world model for the critic. They also introduce the ACPOMDPs framework, an extension of CPOMDPs allowing asymmetric inputs for the actor and critic. Theoretically, asymmetric inputs reduce critic updates and lead to a better policy. AsymDreamer constructs two world models, one for the actor based on historical information and another for the critic with privileged information. It is integrated with the Lagrangian method and shows competitive performance on the Safety-Gymnasium benchmark and strong adaptability to complex scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The work is built on top of SafeDreamer, incrementally adding another world model for critic, and differences in terms of the loss optimized are not sufficiently highlighted.\n* The DreamerV3 is basically combinations of existing methods like CPOMDPs, Augmented Lagrangian, and RSSM. While this is not a problem in itself as they are common in model-based reinforcement learning, there is very little discussion on why these particular methods were chosen and what possible alternatives from the literature exist and whether they might yield better results.\n* Lack of introduction to baseline algorithms. And in the PointGoal2 scenario, the performance results seem to be different from SafeDreamer reported in the paper.\n* Stability verification needs to be considered. How the hyperparameters in equation (8), (9) are selected, and certain ablation experiments need to be conducted.\n* Whether modeling two world models will lead to a significant increase in learning time and the effectiveness of model learning also need to be considered.\n* Privileged information and partial observations should be considered in more environments. A single scenario in QuadrotorGoal1 cannot be trusted to determine whether privileged information enhances performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024asymdreamer,\ntitle={AsymDreamer: Safe Reinforcement Learning From Pixels with Privileged World Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3rSeDrPj4B},\nnote={under review}\n}" }, "abstract": { "value": "Safe Reinforcement Learning from partial observations frequently struggles with rapid performance degradation and often fails to satisfy safety constraints. Upon deeper analysis, we attribute this problem to the lack of necessary information in partial observations and inadequate sample efficiency. World Models can help mitigate this issue, as they offer high sample efficiency and the capacity to memorize historical information. In this work, we introduce AsymDreamer, an approach based on the Dreamer framework that specializes in exploiting low-dimensional privileged information to build world models, thereby enhancing the prediction capability of critics. To ensure safety, we employ the Lagrangian method to incorporate safety constraints. Additionally, we formulate our approach as an Asymmetric CPOMDPs (ACPOMDPs) framework and analyze its superiority compared to the standard CPOMDP framework. Various experiments conducted on the Safety-Gymnasium benchmark demonstrate that our approach outperforms existing approaches dramatically in terms of performance and safety." }, "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": [ "Safe Reinforcement Learing; World Model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/996199ad2305ae0b92f476f04af6cb25003b37f0.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": "AsymDreamer: Safe Reinforcement Learning From Pixels with Privileged World 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": { "value": "the first two step, i.e. the mining and verifying, can be directly applied to mine and search for text embedding that generates malicious contents." }, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weakness part." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. This paper studies a critical area. The capability of these text2image models enables unauthorized usage of these models to generate harmful or disturbing content.\n2. Experiments exhibit good overall performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Dark Miner for mitigating unsafe content generation in text-to-image diffusion models. Dark Miner involves a recurring three-stage process of mining, verifying, and circumventing. It's designed to iteratively identify and suppress unsafe content generation. Comprehensive experiments demonstrate the superior performance of Dark Miner." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The performance of Dark Miner is largely limited by the image pool related to c. That is to say, a lot of text that leads to concept c, which is not linked to the images in the pool, will not identified by the mining step.\n2. Questionable performance by CLIP in Section 3.2.2. Can you discuss how well CLIP performs this task, as this is critical to your methods, I believe this is an important part of the ablation study. \n3. Also, CLIP is used in both the model to be erased (i.e. SD) and Section 3.2.2 to identify the concept. However, texts that circumvent the defense will lead to images that cannot be identified by the CLIP.\n\n4. Methods are only evaluated on two SD models. Will this method generalize well beyond the SD families?\n5. Lack of ablation on the three steps of Dark Miner. For instance,\n * how well does the method perform with and without the verifying step?\n * Ablation over the size of the image pool\n * Ablation over different parameters and configurations of optimizing for the embeddings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 classes are used in NudeNet classifier and threshold?\nSLD and ESD original paper test their FID score on COCO-30k which is from COCO 2024 validation dataset, why authors didn’t use the same one?\nAccording to another relevant paper in the task of nudity elimination as following: Li, Xinfeng, et al. \"SafeGen: Mitigating Unsafe Content Generation in Text-to-Image Models.\" arXiv preprint arXiv:2404.06666 (2024). In the Table 3, they also give the FID score on COCO 2017 validation dataset, achieving 20.36 and 20.31 on ESD and their proposed method SafeGen, which is better than the author’s implementation. We suggest author should revisit their FID calculation on ESD on COCO 2017." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Authors point out the limitation of existing method and provide theoretical analysis.\nAuthors propose an iterative scheme to mine the text c that with maximum likelihood to generate unsafe content while previous method usually predefined such text c. \nAuthors propose a method to avoid overly concept removal by verifying the embedding before circumventing. Authors apply CLIP model to extract delta features from reference image and generated image, then a new metric based on cosine similarity of delta features.\nAuthors make effect to test proposed method against 4 different attacks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed Dark Miner to eliminate unsafe content generation of T2I models. It searches and verifies the embeddings which contain the unsafe concept and reduces unsafe generation by adjusting LoRA adapter of T2I models.\nThe paper mainly made following contributions:\n1.\tPoint out the previous methods fail to avoid unsafe generation on out-of-distribution prompts and easily being tricked by attacking methods.\n2.\tPropose Dark Miner which includes three stage to reduce the optimal embeddings related to unsafe concept and maintain generation ability on benign prompts.\n3.\tEvaluate efficient of proposed compared with 6 SOTA methods and conduct 4 SOTA attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In section 3.2.2, authors use prompt “a photo” to generate reference image and compare the distance with target image generated by mined embedding. However, due to the randomness of reference image, the difference between reference image and target image is always high. In an extreme case, when there are benign images or non-explicit sexual related images in the image pool, the verifying and circumventing will be affected.\n \nAnd the prompt of target image has no relation with prompt of reference image. Another potential drawback is that, all the concepts from “unsafe” prompt are shifted away to random direction without guidance. Therefore, the image generated by “unsafe” prompt will not maintain any semantic information from its prompt even there are safe content expressed by “unsafe” prompt. The generation will be random which might degrade the utility of model.\n\nIn evaluation metrics, author use the mean classification score to evaluate inappropriateness by NudeNet, however, author didn’t mention what classes and threshold used in their implementation. And it is better if authors could also show the number of images being classified as inappropriate image instead of classification score.\n\nFor the CLIP score, the proposed method has relative low performance compared with other SOTA 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": 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. The paper claims that existing methods fail to guarantee safe generation for unseen texts in the training phase. Does the proposed Dark Miner can provide such guarantee? If so please provide more details or discussion regarding this.\n\n2. In the mining stage of DM2, can DM2 obtain the novel harmful prompts? E.g. this work shows that text-to-image models also suffer from Multimodal Pragmatic Jailbreak prompts [1]. Can the Jailbreak prompts also be learned by the proposed DM2? Some discussion about the limitation of the mining stage will be helpful to the readers.\n\n[1] Liu, Tong, et al. \"Multimodal Pragmatic Jailbreak on Text-to-image Models.\" arXiv preprint arXiv:2409.19149 (2024).\n\n3. How the proposed approach performs if the system is black-box? Is it still feasible?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The research topic in this paper is relevant to the community.\n2. Organization of the paper is relatively clear, even not perfect.\n3. Experimental details are clearly stated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes, Dark Miner, an approach designed to address unsafe content generation in text-to-image diffusion models. Unlike existing methods that mainly adjust generation probabilities for known unsafe textual inputs, Dark Miner emphasizes minimizing unsafe generation probabilities for unseen or adversarial prompts. This is achieved through a recurring three-stage process: mining embeddings with high probabilities for unsafe content, verifying them, and circumventing unsafe generation pathways. Some experimental results demonstrate that Dark Miner outperforms six state-of-the-art methods in erasing and defending against unsafe content." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The author claims that the following as one of the contributions:\n“Based on the total probabilities, we analyze the reason why existing methods cannot completely erase concepts for text-to-image diffusion models and are vulnerable to attacks.”\nI did not find a (sub)section for this part, although some discussion can be found in some paragraphs.\nThe reason cannot be found in conclusion section either.\n\n2. Missing related work: The proposed Dark Miner aims to emphasize unsafe generation for unseen or adversarial prompts. One of the previous work [1] also handle unseen or adversarial harmful prompts via building a robust detection space, which is missing in the related work.\n\n[1] Liu, Runtao, et al. \"Latent guard: a safety framework for text-to-image generation.\" ECCV, 2024.\n\n3. Experiments: Previous work show their effectiveness on many harmful concepts. This work only conducts experiments on two. The generalisation of the proposed approach remain to be further verified. The previous approach also shows very different formance on different harmful concepts.\n\n4. The proposed approach cannot handle prompts to generate biased images, e.g. gender bias? Bias is also a harmful concept in responsible text-to-image generation [2,3].\n\n[2] Li, Hang, et al. \"Self-discovering interpretable diffusion latent directions for responsible text-to-image generation.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\n[3] Friedrich, Felix, et al. \"Fair diffusion: Instructing text-to-image generation models on fairness.\" arXiv preprint arXiv:2302.10893 (2023).\n\nMinor issues:\n- The citation format is incorrect across the paper.\n- Confusing annotation, e.g. 0c in Equation 8" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 help to address weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper presents a clear logical flow and is well-written.\n- Experiments demonstrate the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the model editing task of the text-to-image diffusion model. To address the challenges in erasing all unsafe prompts, this paper proposes a method called Dark MINER. This method consists of three steps: 1) mining the potential embeddings related to the unsafe images, 2) assessing whether the potential embeddings effectively induce the model to generate unsafe images, 3) if the mined embedding is effective, this paper conducts the erasing process. In step 3, to protect the generation of safe concepts, this paper also incorporates the regularization in three kinds of concepts: a predefined anchor concept, a null concept, and a concept '-c' that is defined as 'unsafe embedding * -1'." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Please clarify the difference from below existed studies [1,2].\n \n2. In Table.1, the attack success rate of UnlearnDiff in the Violence concept is still 79%. Please provide an explanation for this. Similar phenomenon appear in P4D with the Violence concept (ASR is 46%) and the Church concept (ASR is 49).\n \n3. In Eq.8, this paper proposes three regularization terms to protect the generation of safe concepts, but lacks of related ablation experiments to assess the effect of these three terms.\n\n[1] RACE: Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model, ECCV 2024\n\n[2] Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models, ECCV2024" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A new method to erase unsafe concepts and defend against attacks for text-to-image diffusion models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024dark,\ntitle={Dark Miner: Defend against unsafe generation for text-to-image diffusion models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3rUAS7HCKE},\nnote={under review}\n}" }, "abstract": { "value": "Text-to-image diffusion models have been demonstrated with unsafe generation due to unfiltered large-scale training data, such as violent, sexual, and shocking images, necessitating the erasure of unsafe concepts. Most existing methods focus on modifying the generation probabilities conditioned on the texts containing unsafe descriptions. However, they fail to guarantee safe generation for unseen texts in the training phase, especially for the prompts from adversarial attacks. In this paper, we re-analyze the erasure task and point out that existing methods cannot guarantee the minimization of the total probabilities of unsafe generation. To tackle this problem, we propose Dark Miner. It entails a recurring three-stage process that comprises mining, verifying, and circumventing. It greedily mines embeddings with maximum generation probabilities of unsafe concepts and reduces unsafe generation more effectively. In the experiments, we evaluate its performance on two inappropriate concepts, two objects, and two styles. Compared with 6 previous state-of-the-art methods, our method achieves better erasure and defense results in most cases, especially under 4 state-of-the-art attacks, while preserving the model's native generation capability. Our code can be found in Supplementary Material and will be available 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": [ "Text-to-Image Diffusion Models", "Unsafe Generation", "Concept Erasure" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/3d1af118dcf2b14ce995c3402e7362f459be3da6.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/f58504f2c8c28270f0ea6f93e59c988381fb1d6a.zip" }, "title": { "value": "Dark Miner: Defend against unsafe generation for text-to-image 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": 3 }, "primary_area": null, "questions": { "value": "1. Related to the weakness (2), would swapping I_src and I_tar, i.e. using the deformed image and sketch as input with the original image as ground truth, give better results? \n2. How are the hyperparameters 2.5, 0.25, and 273 in Equation 10 decided? How do the choices on these values affect the results?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The method is based on diffusion models, which are shown to produce higher-quality and more diverse images than GANs. Therefore, it is unsurprising that the proposed method outperforms the previous approaches based on GANs. \n2. The method can edit an image based on partial input sketches without masks. It is more user-friendly than ControlNet and inpainting. \n3. The model is trained using hand-drawn sketches, reducing the gap between training and test-time conditions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript presents a method, called DiffStroke, for image editing based on partial sketch inputs using diffusion models. The method consists of two components, the trainable plug-and-play Image-Stroke Fusion (ISF) module and a mask estimator. The ISF modules fuse the sketch input encodings with the source image features. The mask estimator estimates a mask based on the input sketch to prevent alternation in irrelevant areas. Experimental results demonstrate that the proposed method outperforms previous methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main weakness is the limited novelty. There are multiple existing mask-free sketch-based image editing methods. The main difference between the proposed method and the existing methods is that the proposed method uses the diffusion model while the existing methods are based on GANs. This contribution is not significant enough for a top conference like ICLR. \n2. The proposed method uses free-form deformation to generate training data, limiting the editing capability to simple deformations and resulting in visually unpleasant images. Many result images shown in the paper are not aesthetically pleasing, for instance, the face in Figure 6 and the mug in Figure 2 do not look natural. \n3. The proposed method is not useful in practice. The technique employs a heavy text-based diffusion model while only being able to produce simple deformations. Users can get a similar effect with PhotoShop in a few simple 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": 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": "**Experimental results**\n\n1.\tIn Fig. 4 and Fig. 5, the predicted masks are suggested to be given.\n2.\tI wonder how the prompts affect the output? Given the same image and sketch, using different prompts will generate the same image or diverse image. Will wrong or imprecise prompts fail the method?\n3.\tHow to calculate the mask without ISF in the ablation study of Fig. 6?\n4.\tAlthough the authors discuss the reason why SketchEdit shows poor results on human faces, it is still valuable to conduct comparisons with SketchEdit on the official images provided by SketchEdit. \n5.\tThe proposed method is claimed to be `plug-and-play`. However, there is no validation about this. For example, what is the performance when applied to controlnet?\n\nSome unclear details or minor issues:\n\n1.\tLine 258, why use $S_{tar}-S_{src}$? This will lead to some -255 values in the images? Should it be $max(0, S_{tar}-S_{src})$?\n2.\tLine 272, $h_i^{src}$ and $h_i^s$ should be $\\mathbf{h_i}^{src}$ and $\\mathbf{h_i}^{s}$\n3.\tLine 286 and Line 295, MLP is misleading, since the implementation of $f_{MLP}(.)$ is a one-layer CNN. It should be $f_{CNN}(.)$\n4.\tLine 318 and Line 373, `Eq. 3` should be `Eq. (3)`\n5.\tLine 327, it is unclear why t=273 is used? In what part of the algorithm it is used? Only for extracting the features from $z_{src}$? Why not using the same noise level as the main network for feature fusion?\n6.\tLine 335, `mask mask` should be `mask`\n7.\tLine 466, `the metrics of the metrics`" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**Originality** This methods propose a ISF block, to fuse the features of original images and stoke images to predict the editing mask and guide the diffusion-based conditional image generation process. With this design, the users are free from drawing masks, which make this method user-friendly. I think this task is interesting and valuable.\n\n**Performance** Generally, I found this method perform good. Even without masks, this method outperforms many other mask-based methods, and show more flexibilities through more precise mask prediction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a DiffStroke framework for mask-free image editing with sketches. It designs an image-stroke fusion module to fuse the features of original images and stoke images to predict the editing mask and guide the diffusion-based conditional image generation process. This method is plug-and-play and shows good performance over other mask-based or mask-free image editing methods. The main contribution is that the proposed method is make-free, which can save many users’ efforts during editing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Experimental results** Although most of the results look reasonable, some cases have limitations. For example, in the Fig. 4, in the last case (church), the unedited part in the left of the sketches is incorrectly edited into a tree. It seems that the predicted mask is not precise enough and would harm the content of the original image in the unedited region. \n\nThen, another limitation of this paper, as also discussed in the paper, is that this paper is designed to only edit the shape of the objects under a small distortion. To me, the proposed methods handle some image morphing or image warping operations, lacking the generative abilities of the original diffusion models. It cannot handle large region changes such as replacing some objects or creating some new objects. I believe this greatly limits the application scenarios of the method.\n\nIn addition, the proposed method is claimed to be `plug-and-play` (Line 120). However, I didn’t find validations about this." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "This technology can be used for creating deepfakes." }, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please find the questions in the Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The motivation is clear, the experiments are thorough, and the presentation is clear and easy to understand.\n- Both qualitative and quantitative comparisons with state-of-the-art methods demonstrate the effectiveness of the proposed methods.\n- There are ablation studies that confirm the efficacy of the proposed module." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents DiffStroke, a mask-free method for high-quality image editing using only partial sketches. The proposed approach includes a trainable plug-and-play image-stroke fusion (ISF) module and a mask estimation module aimed at addressing the limitations of previous techniques. Experimental results demonstrate that DiffStroke outperforms existing methods in both simple and complex stroke-based editing tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Experiments from the ablation studies show that providing masks from estimation is important. Creating a mask is not as difficult as drawing sketches; for example, it can be easily obtained from users by scribbling. Masks can also be derived from sketches by identifying the outline boundaries. What would be the differences between these approaches? It would be beneficial to compare the proposed method by modifying it to accept both masks and sketches as inputs, using the outline boundaries directly obtained from the given sketch as the mask.\n\n- The method is built upon a sketch-based T2I adapter with modifications to the feature infusion. Compared to the T2I adapter, the form of information fusion does not seem fundamentally different; it simply passes the features through transformer layers to fuse them and uses an additional MLP for mask estimation. Could you elaborate on the differences and possibly conduct ablation studies on the structure of the ISF blocks? It would be helpful to see a comparison with minimal changes made by adapting the idea based on the current T2I adapter structure.\n\n- The paper did not provide diverse generation results or failure cases. It would be helpful to show diverse results with the same sketch and prompt, the same sketch with different prompts, or the same prompt with different sketches. Additionally, what would happen without prompts? In the provided results in Figure 6, the prompt appears quite lengthy, making the editing process less user-friendly.\n\n- Some edited results appear less natural than those from the T2I adapter or ControlNet. It would be beneficial to conduct user studies comparing the results from different methods in terms of consistency and naturalness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 weaknesses 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 and easy to follow.\n\n(+) The results are visually pleasing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel plug-and-play module, namely the Image-Stroke Fusion (ISF) module, for mask-free image editing using strokes. The ISF module contains a simple concatenation of the source image features extracted by the T2I-adapter and the sketch information, and its shallowest layer is used for mask prediction. Experimental results are visually pleasing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(-) My major concern is that the technical contributions are thin and the methodology is a bit too straightforward. The proposed method relies heavily on T2I-adapter and the proposed Image-Stroke Fusion (ISF) module does not demonstrate many architectual contributions as it contains a simple concatenation of the source image features extracted by the T2I-adapter and the sketch information, and the mask prediction branch is also straightforward. There are not many novel insights here as well (Why the ISF can improve quality? What is the merit of the proposed mask estimation strategy compared to existing ones?). The editing through masking is also a common technique in diffusion-based models.\n\n(-) The data preparation part follows a similar strategy in previous works and also limits the scope of sketch editing to shape changes.\n\n(-) The paper claims that the module is plug-and-play but there are only experiments with T2I-adapter, experiments on several other models are required to justify this claim." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A conditional control diffusion model for high-quaility, mask-free image manipulation with partial sketches." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024diffstroke,\ntitle={DiffStroke: High-Quality Mask-free Image Manipulation with Partial Sketches},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3rnraGvyNr},\nnote={under review}\n}" }, "abstract": { "value": "Sketches offer a simple yet powerful way to represent object configurations, making them ideal for local image structure manipulation. Traditional methods often treat sketch-based editing as an image inpainting task, requiring both user-provided strokes and masks, which hinders the user experience. Although recent mask-free stroke-based editing methods are more convenient, they often produce significant artifacts or unintentionally modify irrelevant regions. To overcome these challenges, we propose DiffStroke, a mask-free method for high-quality image editing using only partial sketches. Trainable plug-and-play Image-Stroke Fusion (ISF) modules and an effective mask estimator are developed to address the limitations of previous conditional control diffusion models in preserving style consistency and protecting irrelevant areas. The ISF modules fuse stroke encodings with source image features as input conditions, enabling DiffStroke to control local shapes while preserving overall style consistency. The mask estimator automatically predicts masks to preserve irrelevant regions without the need for manual input. Specifically, DiffStroke blends the estimated clean latent image with the encoded source image using the predicted mask, with the mask estimator trained to minimize the error between the blended result and the latent target image. Experimental results on natural and facial images demonstrate that DiffStroke outperforms previous methods in both simple and complex stroke-based image editing tasks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Image manipulation", "sketch-based image editing", "mask-free", "diffusion model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/61cb065b01665d54724bf27af028b535f91732d8.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/e6512213a70e67b61888227d52734bb7f8f4cbad.zip" }, "title": { "value": "DiffStroke: High-Quality Mask-free Image Manipulation with Partial Sketches" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "How does the computational complexity scale with audio length and batch size compared to baseline methods?\nHow robust is the method to different audio qualities or noise levels? Was this tested?\nWhat is the impact of different positional embedding choices on the final performance? While rotary embeddings performed best, is there a theoretical justification for this?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper introduces an unbiased sliced Wasserstein RBF (USW-RBF) kernel that effectively handles temporal information across modalities while avoiding dimensionality curse issues that affect traditional Wasserstein distances.\n\nStrong Theoretical Foundation: Provides formal proofs for the kernel's properties (positive definiteness, unbiasedness).\nDemonstrates convergence rate for Monte Carlo estimation.\n\nComprehensive Evaluation: Tests on multiple datasets (AudioCaps and Clotho). Uses both automatic metrics and human evaluation\nIncludes detailed ablation studies for various components.\n\nAchieves state-of-the-art performance on multiple metrics" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces ACUS (Audio Captioning with Unbiased Sliced Wasserstein kernel), a novel framework that addresses exposure bias and temporal misalignment issues in audio captioning systems. The key technical contribution is the development of an unbiased sliced Wasserstein RBF (USW-RBF) kernel equipped with rotary positional embeddings, which effectively measures similarity between acoustic and linguistic features while preserving temporal information. Experimental results on AudioCaps and Clotho datasets demonstrate significant improvements over state-of-the-art methods across multiple metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "No analysis of computational overhead from the USW-RBF kernel\n\nUnclear how the method performs on longer audio sequences\n\nWhile ablation studies are included, there's limited discussion of how sensitive the method is to various hyperparameters\nCould benefit from more guidance on hyperparameter selection for new datasets.\n\nLacks detailed analysis of failure cases" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. There are some minor errors, like in line 187, it should be \\( \\nu = \\frac{1}{N} \\sum_{j=0}^N \\delta_{z_y^j} \\), not \\( \\nu = \\frac{1}{M} \\sum_{j=0}^N \\delta_{z_y^j} \\), to make two empirical distributions have the same number of supports. I did not thoroughly inspect every math part in this paper, but I think authors could check the whole paper again thoroughly. Also, in conclusion, it should be \"unbiased kernel\", not \"unbias kernel\". (No worries, they are just minor error, but it is better to correct for clarity.)\n\n2. As I mentioned in weakness part, the reported improvements over baselines appear modest. Could you provide more analysis on how the proposed method performs in more challenging scenarios (e.g., multi-speaker or noisy environments) to better highlight its strengths? If not, do you believe that temporal information is important in audio captioning task? \n\n3. The application area of this work seems limited. Do you have any plans to extend this work for multilingual audio captioning or automatic speech recognition? If so, how might the kernel method adapt to language diversity in audio processing? how you modify the text embedding under multilingual scenario? \n\n4. Since you use stochastic decoding strategies in inference stage, which may lead to high computational costs, and the reported score in your results is not very decent, we might not need such a high-cost method to get a minor improvement. Thus, could you provide more details on the differences in diversity and quality of captions generated by your approach?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The introduction of the unbiased sliced Wasserstein RBF (USW-RBF) kernel, which captures temporal information across audio and text modalities, is an advancement. By accounting for temporal alignment, it addresses limitations in prior contrastive methods that often ignore the temporal structure of audio data.\n\t\n2. ACUS effectively addresses exposure bias—a common issue in captioning tasks—by combining the USW-RBF kernel with stochastic decoding methods. This approach ensures that generated captions maintain diversity and relevance across varying contexts.\n\t\n3. The ACUS framework enhances not only the length and diversity of captions but also their semantic alignment with audio events. By capturing temporal details, it generates more descriptive and meaningful captions, which are validated by both quantitative metrics and qualitative assessments.\n\n4. The paper thoroughly derives and proves the properties of the USW-RBF kernel, reinforcing its validity for multimodal tasks. Additionally, by introducing a practical approach to reducing exposure bias, it offers a methodological contribution that may extend beyond audio captioning to other sequence generation tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel approach for audio captioning to address issues related to exposure bias and temporal misalignment. Here’s a summary of the key contributions:\n1. Unbiased Sliced Wasserstein RBF Kernel (USW-RBF): The authors propose a novel kernel method to accurately measure cross-modal similarity between audio and textual data. This kernel, equipped with rotary positional embedding, captures temporal information more effectively than traditional methods, addressing limitations like dimensionality and temporal distortion.\n2. Mitigating Exposure Bias: ACUS employs stochastic decoding techniques, such as nucleus and top-k sampling, to reduce exposure bias during the inference stage, enhancing caption diversity and quality. This is achieved by leveraging the USW-RBF kernel to improve alignment between generated captions and audio inputs.\n3. Extensive Evaluation and Compatibility: The framework’s efficacy is validated through experiments on two datasets, AudioCaps and Clotho. Results demonstrate improved caption length, lexical diversity, and self-retrieval accuracy, and compatibility with diverse encoder-decoder architectures.\n\nIn essence, ACUS represents an advancement in audio captioning by integrating the unbiased USW-RBF kernel with stochastic decoding, leading to more descriptive and temporally coherent audio captions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the paper introduces a promising method, the improvements observed in the current experiments are relatively modest. To more rigorously validate the effectiveness of your approach, I recommend evaluating your model on additional, more challenging benchmarks, which may make this work more convincing if your method reach higher score in these benchmarks, such as:\n\na. SUPERB: This benchmark includes a broad range of speech processing tasks, covering content, speaker, semantics, and paralinguistics. It would provide a comprehensive baseline, helping clarify how well your model generalizes across core speech tasks.\n\nb. Dynamic-SUPERB: This benchmark extends SUPERB with instruction-tuning and zero-shot tasks, pushing models to handle more complex and varied speech processing scenarios. Testing on Dynamic-SUPERB could demonstrate your method’s robustness and adaptability in handling multi-task and instruction-following requirements, offering deeper insights into its generalization capabilities.\n\nc. SpeechCaps: Given the emphasis in your work on speaker-specific and temporal information, SpeechCaps offers a relevant test for multi-talker and speaking style captioning. Its focus on speaker and prosodic information could highlight the strengths of your model in more intricate, real-world audio scenarios, such as multi-speaker dialogues and expressive speech.\n\n2. Authors provide a detailed explanation of the USW-RBF kernel. However, it lacks sufficient details on how this kernel is integrated within the overall model architecture. You can try these to make it better, such as:\n\na. Integration Details: Please provide a clearer, step-by-step description of how the USW-RBF kernel is incorporated into the model pipeline. \n\nb. Diagram or Flowchart: Consider adding a diagram or flowchart that visualizes the integration process, illustrating where and how the USW-RBF kernel interacts with audio and textual embeddings within the architecture." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- How does the performance of USW-RBF compare with other non-Wasserstein-based kernels for audio captioning tasks?\n- What are the potential trade-offs between the accuracy improvements and the computational costs introduced by Monte Carlo sampling and stochastic gradient optimization in ACUS?\n- Why was the rotary positional embedding favored over other encoding techniques, and could alternative embeddings further enhance the results?\n- Why audio captioning? Can the method be useful to other tasks? Audio understanding? Speech Recognition?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper is decently written and easy to follow for an expert. However, I would like to mention it might be difficult for a traditional audio captioning community to read. A bit more background on the 1, the unbiased sliced Wasserstein RBF kernel would have been appreciated. The equations could have been improved by defining the notations better. Fro examples, if I want to read Eqtn 9., I need to find what the notations mean somewhere else in the paper.\n- The problem handled in the paper is new. The approach is also novel. ACUS combines the USW-RBF kernel with stochastic decoding methods like nucleus and top-k sampling to alleviate exposure bias during inference. A lot of work in audio captioning ideally propose new architectures. This paper brings a fresh perspective in the problem space.\n- The evaluation is sound. The 2 usual benchmark datasets are used and it is also combined with human evaluation. The metrics of descriptiveness, correctness, and fluency are good metrics for comparison as ideal benchmark metrics seem to have saturated and require style memorization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new framework for audio captioning designed to mitigate exposure bias and improve temporal cross-modal similarity measurement. The authors claim that traditional audio captioning models trained via maximum likelihood estimation face exposure bias, leading to \"degeneration\" in generated captions. This paper introduces the unbiased sliced Wasserstein (USW-RBF) kernel equipped with rotary positional embedding to capture temporal information across acoustic and linguistic modalities, thereby reducing exposure bias. The authors show improvements on benchmark datasets (AudioCaps and Clotho)" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The abstract says\" \"Prior works propose the contrastive method to deal with caption degeneration. However, the contrastive method ignores the temporal information when measuring similarity across acoustic and linguistic modalities, leading to inferior performance.\" -- which contrastive method and how does it ignore the \"the temporal information when measuring similarity across acoustic and linguistic modalities\"? This first line of the paper is very difficult to understand.\n- The Monte Carlo sampling and stochastic gradient optimization may increase computational costs, potentially impacting efficiency in real-world large-scale applications.\n- While I understand that the authors focus on Enc-Dec framework, a good number of baselines were missed for evaluation. ACUS can act as complimentary to most other methods proposed in literature as all methods require an audio encoder and a language decoder (including prefix based architectures). Thus, some baselines were missed. See [1,2] as examples and papers compared in [1,2].\n- The analysis section is just ablations. A deeper analysis section (see questions below) would have strengthened the paper.\n\n\n### Citations\n\n[1] https://ieeexplore.ieee.org/abstract/document/10448030. \n[2] https://ieeexplore.ieee.org/abstract/document/10096877." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. Could the authors provide a more detailed explanation of the differences between the two types of AAC architectures? Additionally, could the proposed method be adapted for application within prefix-tuning structures?\n2. What is the increase in computational cost introduced by the framework? For example, how much additional inference time is required when using stochastic decoding to generate $\\mathcal{B}$ candidate captions?\n3. Considering the advanced reasoning and generative capabilities of large language models, frequently used in AAC tasks, could the proposed approach be adapted to work alongside LLMs to achieve higher-quality captions?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper introduces a novel temporal-similarity score, utilizing the unbiased sliced Wasserstein RBF (USW-RBF) kernel with rotary positional embeddings, to mitigate exposure bias in audio captioning models. Unlike prior research, which has not employed the USW-RBF kernel for cross-modal similarity calculation, this study leverages it to capture temporal dynamics more effectively.\n2. The proposed framework is adaptable to a wide range of existing AAC models, with experimental results underscoring its effectiveness in improving model performance.\n3. Comprehensive qualitative and quantitative experiments support the method's efficacy. Ablation studies comparing various similarity metrics further highlight the advantages of the USW-RBF kernel over alternative approaches." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel framework tackling the training-inference mismatch in automated audio captioning (AAC) by introducing a temporal-similarity score based on the unbiased sliced Wasserstein RBF (USW-RBF) kernel with rotary positional embeddings. By integrating this score with a stochastic decoding strategy, the approach effectively addresses caption degeneration issues encountered during inference. Experimental results on established AAC benchmark datasets demonstrate notable improvements in model performance, validated through both quantitative and qualitative metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tCombining the USW-RBF kernel with stochastic decoding strategies may lead to high computational costs. For example, at inference time, generating $\\mathcal{B}$ candidate captions through stochastic decoding results in increased computational time. While the paper demonstrates effectiveness, it lacks a detailed discussion of computational overhead in training and inference.\n2.\tThe performance increase with the proposed approach is relatively minor. According to the original paper, EnCLAP-large achieved SPIDEr scores of 49.5 and 27.8 on AudioCaps and Clotho, while the proposed method reached only 50.0 and 27.5, making the claimed improvement less convincing. According to this comparison, the exposure bias is not that important in AAC tasks.\n3.\tThe application scope of this work appears limited, focusing primarily on AAC tasks. The authors have not explored the framework’s performance on other audio-text multimodal tasks, such as audio-text retrieval, and automatic speech recognition. For instance, can the proposed temporal-similarity score enhance the temporal reasoning capability of the CLAP model?\n4.\tThe study does not deeply explore the sensitivity of the framework to key hyper-parameters, such as the coefficient $\\alpha$ in the objective function or the number of Monte Carlo samples $L$ used for the USW-RBF kernel.\n5.\tThe paper dedicates considerable space to explaining the USW-RBF kernel but provides a limited description of how it integrates with the model itself. For example, it’s unclear whether the text embedding is derived from the penultimate layer of the text decoder or from another layer.\n6.\tAlthough the paper separates AAC models into encoder-decoder and prefix-tuning architectures, with experiments performed only on the encoder-decoder type. The difference between these two types of architecture is not substantial. Both approaches essentially share the same structure of an audio encoder and a text decoder.\n(Minor problems - Line 44: the former architecture → the latter architecture)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 proposing the unbiased kernel is the core technical contribution in the paper, how much gain is there from a biased kernel to a unbiased kernel ?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper propose the unbiased sliced Wasserstein kernel framework to improve the audio captioning performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is written to solve the exposure bias problem in audio captioning.\nThey propose the unbiased sliced Wasserstein RBF kernel, which is a better cross-modality similarity measure.\nTogether with the contrastive learning method, gains are observed in the audio captioning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The motivation of the paper does not sound convincing.\nExposure bias is a general problem to auto-regressive networks. \nIt is not a critical problem for audio captioning.\nIn general, exposure bias can be mitigated by a better training of model. \nI believe using a larger decoder will ease the problem a lot.\nSpecifically, I do not see in the paper how much the exposure bias problem is harming the audio captioning performance.\nEven if we regard this as a serious problem, reinforcement learning (RL) should be a popular way to solve it as RL trains the model according to its inference output. The paper doesn't discuss about RL and address the audio captioning problem in a very narrow perspective." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We develop the audio captioning with an unbiased sliced Wasserstien kernel to alleviate caption degeneration" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024acus,\ntitle={{ACUS}: Audio Captioning with Unbiased Sliced Wasserstein Kernel},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3sf7SpOYIe},\nnote={under review}\n}" }, "abstract": { "value": "Teacher-forcing training for audio captioning usually leads to exposure bias due to training and inference mismatch. Prior works propose the contrastive method to deal with caption degeneration. However, the contrastive method ignores the temporal information when measuring similarity across acoustic and linguistic modalities, leading to inferior performance. In this work, we develop the temporal-similarity score by introducing the unbiased sliced Wasserstein RBF (USW-RBF) kernel equipped with rotary positional embedding to account for temporal information across modalities. In contrast to the conventional sliced Wasserstein RBF kernel, we can form an unbiased estimation of USW-RBF kernel via Monte Carlo estimation. Therefore, it is well-suited to stochastic gradient optimization algorithms, and its approximation error decreases at a parametric rate of $\\mathcal{O}(L^{-1/2})$ with $L$ Monte Carlo samples. Additionally, we introduce an audio captioning framework based on the unbiased sliced Wasserstein kernel, incorporating stochastic decoding methods to mitigate caption degeneration during the generation process. We conduct extensive quantitative and qualitative experiments on two datasets, AudioCaps and Clotho, to illustrate the capability of generating high-quality audio captions. Experimental results show that our framework is able to increase caption length, lexical diversity, and text-to-audio self-retrieval accuracy. We also carry out an experiment on two popular encoder-decoder audio captioning backbones to illustrate that our framework can be compatible with a diversity of encoder-decoder architectures." }, "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 captioning", "exposure bias", "multimodal 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/a546d3352b9ff0753d387a59d665375778fd2114.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": "ACUS: Audio Captioning with Unbiased Sliced Wasserstein Kernel" }, "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": [ "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) When participants read words aloud, the movement of their vocal tract can influence the EEG recordings. Could the authors address this by using visual cues and having participants read the cues silently without the movement of the vocal tract?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1) This paper tackles a highly under-explored area, largely limited by the scarcity of curated datasets, by introducing a publicly available bilingual brain-to-speech dataset that holds significant potential for advancing research in this field.\n\n2) The authors propose the MoBSE framework for brain-to-speech synthesis, which achieves improved performance over the FastSpeech 2 baseline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents CerebroVoice, a bilingual brain-to-speech synthesis dataset featuring stereotactic EEG recordings of Chinese and English words and digits. The dataset is benchmarked for two key tasks: speech synthesis and voice activity detection. Additionally, the authors introduce a novel framework, Mixture of Bilingual Synergy Experts (MoBSE), which employs low-rank expert weights tailored for language-specific decoding tasks. The proposed MoBSE framework demonstrates superior performance compared to the baseline FastSpeech 2 model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The authors explain the advantages of stereotactic EEG over ECoG; however, these invasive methods have limited practicality due to the complexity of data collection. It would be beneficial if the authors addressed why surface EEG, a non-invasive alternative, was not used instead in their study.\n\n2) In Subject 1, electrodes were implanted in the right hemisphere, while in Subject 2, they were implanted in the left. However, both hemispheres could contribute to speech production, suggesting that electrodes should ideally be placed in both hemispheres for each participant. Additionally, data collection was limited to only two participants, which restricts the generalizability of the models built with this dataset.\n\n3) The paper uses only one baseline, based on the FastSpeech 2 architecture, which is primarily designed for text-to-speech tasks. However, there are existing models in the literature for synthesizing speech from invasive and non-invasive multi-channel EEG signals, such as [1], [2], and [3], etc. These models could have been used as baselines for more comprehensive benchmarking of the dataset and comparison with the proposed MoBSE framework.\n\n4) Although the paper focuses on speech synthesis and reports using a Hifi-GAN vocoder for generating speech, it does not present any results for the synthesized audio output. To fully assess the quality of the reconstructed speech, it is essential to include both subjective evaluations (such as mean opinion score) and objective metrics (like mel cepstral distortion and root mean squared error).\n\n5) The model architecture presented in Figure 3 is unclear. FastSpeech 2 typically processes text inputs, yet the authors are instead feeding multi-channel EEG signals to the model. The method for obtaining sEEG embeddings from these multi-channel EEG signals is not explained. Additionally, Figure 3 (c) lacks details regarding the structure of the Universal Expert module.\n\n\n\nReferences: \n\n[1] Metzger, Sean L., Kaylo T. Littlejohn, Alexander B. Silva, David A. Moses, Margaret P. Seaton, Ran Wang, Maximilian E. Dougherty et al. \"A high-performance neuroprosthesis for speech decoding and avatar control.\" Nature 620, no. 7976 (2023): 1037-1046.\n\n[2] Kim, Miseul, Zhenyu Piao, Jihyun Lee, and Hong-Goo Kang. \"BrainTalker: Low-Resource Brain-to-Speech Synthesis with Transfer Learning using Wav2Vec 2.0.\" In 2023 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI), pp. 1-5. IEEE, 2023.\n\n[3] Lee, Young-Eun, Seo-Hyun Lee, Sang-Ho Kim, and Seong-Whan Lee. \"Towards voice reconstruction from EEG during imagined speech.\" In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 5, pp. 6030-6038. 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": [ "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": "- Why do the authors argue that other datasets can not be used for VAD if the labels for that task are obtained automatically?\n- The authors assert that the audio quality was assessed and the recordings edited accordingly during the data curation process. Was this task performed subjectively? Who was in charge of this task? \n- Specifications of audio recording equipment were not included, which is relevant to analysis results and prevent biases in case future data fusion tests can be performed.\n- The authors presented independent results per subject. Were these results obtained using a single model trained with data from the two subjects, or were also two models trained (one per patient)?\n- Results regarding LFS, HGA, and BBS signals are confusing. There is no apparent coherence regarding frequency bands or between subjects' behavior. Why do the authors consider that these experiments provide a benchmark in this field, considering the scarcity of subjects, which limits the power of any analysis? \n- The organization of the paper could be improved. The meaning of LFS, HGA, and BBS features and the relevance of their evaluation should be presented in section 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 relevant topic and introduces a new open dataset that can help advance a field far from being consolidated and where the data is highly costly and complex to acquire. It also provides relevant measures that can help objectively evaluate the improvement of further approaches in this field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a data set consisting of pairs of stereotactic EEG and speech signals recorded simultaneously and a set of experiments in the context of brain-to-speech synthesis aiming to provide a benchmark for further research in this area. The dataset comprises sEEG and speech signals from two participants, and the protocol included the repetition of auditory stimuli in two languages. The paper also analyses the voice activity detection problem from the sEEG signals. The paper uses a similar architecture to that of the FastSpeech2 TTS model but substitutes phoneme embeddings with a sEEG embedding layer and proposes an alternative way to codify the language information into the network through a MLP layer that weights the feature representation of the network depending on a one-hot-encoding vector that indicates one of two possible languages in the dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The general novelty of the work is limited. The introduced dataset is valuable and constitutes a significant contribution to the academic community because of its complexity, but with such a limited number of participants in the study, it is hard to consider this work a valid benchmark for the task. Moreover, the proposed mixture of bilingual synergy experts component is not presented clearly, and the whole pipeline is not well 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": { "value": "While the authors mention addressing ethical concerns related to patient privacy, I still believe it should be reviewed by the ethics committee, as it is very sensitive to make invasive human neural data publicly available. There is insufficient discussion on data storage security measures, access controls, and compliance with data protection regulations such as GDPR." }, "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": 4 }, "primary_area": null, "questions": { "value": "Did you perform any statistical significance testing to confirm that the improvements of MoBSE over FastSpeech2 are meaningful?\n\nIs there a reason why raw sEEG data is not provided alongside the processed data, allowing researchers to perform custom preprocessing and explore different frequency bands?\n\nHow and why is positional encoding used in the MoBSE framework? Can you provide more insight into its implementation?\n\nAre there any samples of the reconstructed speech available for qualitative assessment?\n\nHow is VAD accuracy measured exactly? I'm trying to figure out if you chose a window of silence vs speech? how long was the window?\n\nHave you considered combining electrode data from both subjects to create a \"super subject\" to enhance coverage?\n\nThanks," }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "-The authors introduce CerebroVoice, a publicly accessible sEEG dataset tailored for neural to speech synthesis and Voice Activity Detection (VAD). This is particularly significant given the scarcity of publicly available sEEG datasets and benchmarks, providing a valuable resource for researchers to compare and validate their methods, fostering progress in brain-computer interface applications.\n\n-By incorporating bilingual data, specifically focusing on a tonal language, Chinese Mandarin, the dataset opens new avenues for research, addressing the complexities associated with tonal languages in brain-to-speech synthesis.\n\n-The methodology for data acquisition is thoroughly and clearly explained, ensuring transparency.\n\n-The authors introduce a Mixture of Experts (MoE)-based framework for neural-to-speech synthesis, which improves bilingual decoding by dynamically organizing language-specific experts. This novel approach outperforms the FastSpeech2 baseline, demonstrating its effectiveness.\n\n-The authors address important ethical concerns related to patient privacy and the sensitive nature of invasive neural recordings, demonstrating a strong commitment to ethical research practices." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces CerebroVoice, a new dataset for bilingual brain-to-speech synthesis and Voice Activity Detection (VAD) using stereotactic EEG (sEEG). It includes recordings from two bilingual participants who read Chinese Mandarin words, English words, and Chinese Mandarin digits. The authors developed a novel method called Mixture of Bilingual Synergy Experts (MoBSE) that uses a language-aware dynamic organization of low-rank expert weights and tested it against the FastSpeech2 baseline, setting a new benchmark for their dataset. They found that MoBSE performs better than FastSpeech2 in producing speech from neural recordings. Additionally, they reproduced three existing VAD methods and established benchmarks for VAD using CerebroVoice. The dataset is publicly available on Zenodo, and the preprocessing code can be found on GitHub." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-The CerebroVoice dataset is limited by its small size, featuring only two participants and a repetitive, narrow vocabulary, which restricts its generalizability and raises concerns about potential overfitting. Its focus on simple speech synthesis tasks diminishes its flexibility for broader neuroscience research areas such as brain decoding and semantic reconstruction. Additionally, the task design lacks originality, as many similar speech synthesis/reconstruction objectives have been addressed in previous studies [1, 2, 3], Most existing invasive datasets can be requested from the authors while non-invasive ones are generally publicly available, reducing the novelty of CerebroVoice’s contribution to the field. To enhance its impact, the authors could consider expanding the dataset with more participants and a more diverse vocabulary and/or task in future work.\n\n-The GitHub repository lacks implementations of the proposed models, hindering reproducibility and preventing other researchers from building upon the work. It would be beneficial for the authors to include model implementations, training scripts, and detailed documentation in their GitHub repository. \n\n -It is unclear how FastSpeech2 was adapted to produce audio from sEEG signals. The paper does not provide a detailed explanation of the training procedures, architectural changes, or loss functions used in adapting this text-to-speech model for brain-to-speech synthesis. Providing specific details about these adaptations would make the methodology more understandable and reproducible.\n\n-The architecture of the experts within the MoBSE framework is not clearly explained, leaving gaps in understanding how the model functions. It does not specify how many experts were used in the MoBSE framework and lacks ablation studies to justify this choice, hindering the evaluation of the model's components.\n\n-The evaluation primarily uses Pearson Correlation Coefficient (PCC). Including additional metrics like ESTOI (Extended Short-Time Objective Intelligibility) would provide a more comprehensive assessment of speech synthesis quality. This is a very common metric in speech synthesis/reconstruction tasks.\n\n[1] M. Angrick, M. Ottenhoff, S. Goulis, A. J. Colon, L. Wagner, D. J. Krusienski, P. L. Kubben,\nT. Schultz, and C. Herff, “Speech synthesis from stereotactic EEG using an electrode shaft dependent multi-input convolutional neural network approach,” in 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)\n\n[2] Verwoert, Maxime, et al. \"Dataset of speech production in intracranial electroencephalography.\" Scientific data 9.1 (2022): 434.\n\n[3] Akbari, Hassan, et al. \"Towards reconstructing intelligible speech from the human auditory cortex.\" Scientific reports 9.1 (2019): 874." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "Yes, 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. Line 162: What does “a Python-scripted audio playback and sEEG-marking mechanism” mean? At the onset of audio stimuli (not the participant’s audio), the system sends a marker to ths sEEG recordings to identify the onset of audio stimuli." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "**Significance**: Open-source sEEG speech datasets are rare. Their publishing of the dataset (Line 035) is good news for the community as it will lower the entry threshold for future research. Additionally, they demonstrate how different sEEG features (e.g., LFS, HGA, BBS) affect the performance of brain-to-speech synthesis and voice activity detection. These results may help future works on speech decoding.\n\n**Clarity**: The text has a good structure and is well-written. The figures also help in understanding the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce a novel dataset, CerebroVoice (publicly available), for bilingual brain-to-speech synthesis and a neural architecture MoBSE, which utilizes a language-aware prior dynamic organization for efficient handling of language-specific decoding tasks.\n\n**Dataset**: The audio stimulus set contains `50` different stimuli, including 30 Chinese Mandarin words, 10 Chinese Mandarin digits, and 10 English words. For each trial, one randomly selected audio stimulus is played; then, the patient is asked to repeat that word (or digit). The dataset includes `1600` trials (i.e., 29 trials per Chinese Mandarin word, 48 trials per Chinese Mandarin digit, and 24 trials per English word). In each trial, two kinds of brain responses are recorded, including listening and reading. Each trial lasts either `4` or `5` seconds and is paired with the corresponding audio recording.\n\n**Model**: The authors propose MoBSE, which is similar to `model ensemble`. MoBSE uses an additional gating module to support the dynamical fusion of the outputs from different experts.\n\n**Experiment**: Previous methods (e.g., FastSpeech2, EEGNet, STANet, EEGChannelNet) are compared. Besides, the authors conducted different ablation studies regarding sEEG settings (sEEG feature, subject, word categories, etc.).\n\n**In summary, it seems like a dataset paper.**" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Major**\n1. Why is common average referencing, instead of laplacian reference, used in BrainBERT[1] (for listening decoding) or bipolar reference used in Du-IN[2] (for speech decoding)? Could you provide brain-to-speech synthesis results based on either laplacian reference or bipolar reference? Although previous studies[3] on speech synthesis use common average referencing + HGA, the speech synthesis task has a trivial solution (the mel-spectrum distribution of human speech is easy to regress). Maybe I’m wrong, but with these additional results, we can gain a deeper understanding of the dataset. Could the authors include the results of brain-to-speech synthesis (i.e., Table 1) baesd on the preprocessed data after either laplacian reference or bipolar reference?\n\n2. How about the results of word classification? CerebroVoice dataset includes at least `24` trials per words, it should be able to evaluate 30-way classification task (i.e., 30 Chinese Mandarin words). Could the authors include results on word-classification tasks (e.g., 30-way on Chinese words, 10-way on Chinese digits, 10-way on English words)?\n\n**Minor**\n1. Line 90: Additional publications the authors should be aware:\n - In Du-IN (https://arxiv.org/abs/2405.11459), their preprocessed dataset is open available.\n\nCould the authors summarize these works in Table 1?\n\n2. Line 99: Additional publications the authors should be aware:\n - In Feng et al. (https://www.biorxiv.org/content/10.1101/2023.11.05.562313v3), they also explore speech decoding based on tonal language (i.e., Chinese Mandarin).\n\nCould the authors summarize these works in the Related Works?\n\n**Reference**\n\n[1] Wang C, Subramaniam V, Yaari A U, et al. BrainBERT: Self-supervised representation learning for intracranial recordings[J]. arXiv preprint arXiv:2302.14367, 2023.\n\n[2] Zheng H, Wang H T, Jiang W B, et al. Du-IN: Discrete units-guided mask modeling for decoding speech from Intracranial Neural signals[J]. arXiv preprint arXiv:2405.11459, 2024.\n\n[3] Chen J, Chen X, Wang R, et al. Subject-Agnostic Transformer-Based Neural Speech Decoding from Surface and Depth Electrode Signals[J]. bioRxiv, 2024." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present CerebroVoice, the first public sEEG dataset for bilingual brain-to-speech synthesis and voice activity detection. Our MoBSE model shows significant performance improvements. We providing insights for brain-computer interfaces" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024cerebrovoice,\ntitle={CerebroVoice: A Stereotactic {EEG} Dataset and Benchmark for Bilingual Brain-to-Speech Synthesis and Activity Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3sfOGsBh85},\nnote={under review}\n}" }, "abstract": { "value": "Brain signal to speech synthesis offers a new way of speech communication, enabling innovative services and applications. With high temporal and spatial resolution, invasive brain sensing such as stereotactic electroencephalography (sEEG) becomes one of the promising solutions to decode complex brain dynamics. However, such data are hard to come by. In this paper, we introduce a bilingual brain-to-speech synthesis (CerebroVoice) dataset: the first publicly accessible sEEG recordings curated for bilingual brain-to-speech synthesis. Specifically, the CerebroVoice dataset comprises sEEG signals recorded while the speakers are reading Chinese Mandarin words, English words, and Chinese Mandarin digits. \nWe establish benchmarks for two tasks on the CerebroVoice dataset: speech synthesis and voice activity detection (VAD). For the speech synthesis task, the objective is to reconstruct the speech uttered by the participants based on their sEEG recordings. We adopt FastSpeech2 as the baseline model and propose a novel framework, Mixture of Bilingual Synergy Experts (MoBSE), which uses a language-aware dynamic organization of low-rank expert weights to enhance the efficiency of language-specific decoding tasks. The proposed MoBSE framework achieves significant performance improvements over FastSpeech2 across all subjects, producing more natural and intelligible reconstructed speech. \nThe VAD task aims to determine whether the speaker is actively speaking. In this benchmark, we adopt three established architectures and provide comprehensive evaluation metrics to assess their performance. Our findings indicate that low-frequency signals consistently outperform high-gamma activity across all metrics, suggesting that low-frequency filtering is more effective for VAD tasks. This finding provides valuable insights for advancing brain-computer interfaces in clinical applications. \nThe CerebroVoice dataset and benchmarks are publicly available on Zenodo and GitHub for research purposes." }, "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-to-speech Synthesis", "Voice Activity Detection", "Stereotactic Electroencephalograph", "Bilingual and Tonal Speech", "Brain Computer Interface" ] }, "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/d9e45815b32e825d17ebd876b5955b2eee0bfcfc.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/8543ad10a7556610526d084d5c7fcb17fc562103.pdf" }, "title": { "value": "CerebroVoice: A Stereotactic EEG Dataset and Benchmark for Bilingual Brain-to-Speech Synthesis and Activity 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": 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": "See weaknesses.\nHow are the speech and text tokens interleaved to form training samples? What are the details of this data creation process?\nHow does the model benefit from interleaved speech and text modalities? \nHow do you deal with the different sampling rates and information granularities between speech and text tokens during the process?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The ASR-based speech tokenizer achieves semantic information preservation and decent speech audio reproduction at the same time.\n2. The low-bitrate speech tokenizer and the text-to-token model effectively use the existing large amounts of text data to synthesize large amounts of speech tokens, which saves resources to collect large amounts of speech audio data and improves the language model's speech performance after pretraining." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a speech-text pretraining process for scaling speech-language model training without acquiring large amounts of speech audio data. The process mainly includes an ASR-based low-bitrate speech tokenizer and a text-to-speech-token model to produce large quantities of speech tokens for speech-text pertaining. The pre-trained model is fine-tuned on spoken dialog datasets and shows competitive performance compared to existing SOTA models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weaknesses are mainly in terms of paper writing and presentation. \n1. The paper mentions \"we are first to use supervised semantic tokens for SpeechLMs\". However, one of the baselines, Mini-Omini also uses a whisper-based speech tokenizer. \n2. The details on how the speech and text modalities are interleaved are missing. \n3. As an important part of the process, the details of the text-to-token model are missing—for example, model architectures, training schemes, etc.\n4. The large amounts of speech tokens generated by the text-to-token model are still from existing datasets and speech-synthesized audio from text. How is this process different from generating speech tokens from synthesized speech audio using large amounts of text? For example, llama-omni also uses cosy-voice to synthesize speech audio to augment training data. What's the innovation here between text-to-speech-to-token and text-to-token?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "While the whisper ASR model has achieved excellent performance on a range of tasks, it does have its limitations, especially with regards to unseen or low-resource languages. That is not an issue for this paper which seems to focus on English (although there was quite a bit of Chinese data used as well). Have the authors given any thought as to how to extend this work to cover more languages?\n\nAre there any plans to open source the tokenizer, text-to-token model, or the speech LM itself?\n\nAlso, it would be nice if the authors could describe the amount of computation required to pretrain the speech LM." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper is a nice contribution to the very hot topic of speech LMs. By developing an effective speech tokenizer and text-to-tokenizer model the authors are able to create a very large speech language model that produces impressive results on a wide range of tasks. The authors perform extensive experiments and ablation studies on the speech tokenizer, speech generator (decoder), and the speech LM. The model is able to achieve strong performance on both spoken language modeling and spoken question answering tasks. Finally, when fine-tuned on dialogue data, the model does well on a spoken chat-bot task." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is about scaling up data to train large speech language models. The authors present a method for tokenizing speech using the Whisper encoder and demonstrate their tokenizer retains semantic information as well is fine-grained information for good quality speech generation. They also describe a method for training a text-to-token model. With these, they are able to tap into large resources of text data to generate synthetic training data, which they interleave with other conventional text and speech/text sources to pre-train a speech LM. By fine-tuning the LM on a dialogue corpus they demonstrate a speech chat-like capability. Extensive experimentation is performed, and the speech pretrain method does quite well on a range of tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although this is not necessarily a weakness, this paper seems very strong on the engineering side and a little weaker on the novelty side of things. The recipe the authors put forward consists of three separate steps 1) tokenizer, 2) text-to-token model 3) pretrain speech LM. While the authors build a strong tokenizer based on the Whisper model, the approach is not especially novel as it is built on top of a strong speech recognition model. Likewise the use of a TTS corpus to learn a text-to-token model is a nice approach, but has been done before to learn similar kinds of models (e.g., Hsu et al., Text-Free Image-to-Speech Synthesis Using Learned Segmental Units, 2020). Finally, the interleaving of different kinds of text and speech data to pretrain an LLM with an additional token vocabulary is not especially novel. However, while these points are arguably true, I find it impressive that the authors have put all the pieces together to create a very strong speech LM." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please consider clarifying the following questions in the final revision, (apologies if I missed some details that are in the paper):\n\nSection 2.1 and table 1:\n- I’m assuming that you measured WER in content/semantic retention - is that right?\n- How did you measure the ground truth (Whisper?)? \n- The title says ‘Speech Tokenizer Results” but the Quality also measures your speech decoder.\n- Is the speech decoder single/multi-speaker? Please add information regarding speaker identity preservation. \n- What was the training data of the speech decoder? (single or multi speaker?)\n- Missing citations (e.g. Expresso). \n- LS stands for Librispeech? this isn’t stated.\n\nTable 2:\n- How do you measure WER? (those tokens into the bottleneck layer of the quantized Whisper? Decode the audio using the speech decoder and apply the regular Whisper?)\n\nSection 2.2:\n- Regarding \\eta, please clarify if that is the ratio of text that is replaced or the ratio that audio will take out of the final sequence. I assume it is the first based on section A.1 but better for it to be clear in the main text too.\n\nSection 2.3.2\nIn the main text, please add that you used GPT-4 to filter examples, shorten the responses, and avoid outputting text that cannot be read aloud (help the reader understand it from the main text). Also consider adding optimization details on the finetuning step (lr, batch size, etc')\n\n\nSection 3.1: \n- Please help the reader and add in the main paper that the GPT-4 content quality is on a scale of 1-10.\n- The response of the SpeechLM was converted to text before it moved to GPT-4, right? How was the conversion performed? (Quantized whisper? your SpeechLM? Decoded and then used regular whisper?)\n- Two comments regarding the ASR-WER metric in Table 4. First, it is more of a content quality metric rather than a speech quality metric. Secondly, as there are many ways to answer a question correctly, I suggest moving to ASR-ROUGE or ASR-BLUE instead.\n\nTable 3: \n- what’s the difference between \\emptyset and “-“?\n\n\nComments:\n\n- Fig1a is hard to understand at first glance - specifically, that the yellow tokens are replaced with speech tokens. Adding a color legend (Yellow: SpeechTokens, Cyan: TextTokens) would make it easier to understand.\n\n- The second row in Fig 2a needs to be clarified. “Text—Audio token—>Text-to-token LM” can be perhaps: “Text — (TexttoTokenLM)—> Audio Tokens”.
\n- Semantic tokens determine the pacing of speech, which is a part of the speaker’s prosody. Your synthetic audio tokens are not conditioned on a speaker, so you are likely to get the semantic tokens of an average speaker. It is fine overall.\n\n- Regarding audio tokenization - consider reducing the dimension (e.g. from 1024 to 8/16) before quantization to prevent codebook collapse. 
\n- Moreover, I suggest applying text-tokenization algorithms (BPE?) on the speech units, to produce variable-length representations with a more balanced distribution, and further shorten the audio sequence.\n\nTypos:\n- Line 514: AudioLLM->AudioLM\n- Line 516: Moshi Citet->citep" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Supervised speech tokenizers are a great way to distill the content from audio. Audio is high-dimensional, and using text and a low bitrate bottleneck to focus on content is a good idea, suitable for SpeechLMs. \n- Training a “TTS” model to generate synthetic audio tokens is interesting - as it doesn’t require generating the final audio (high-bitrate, compute-intensive, issues with OOD synthetic data). 
Instead, they generate latent audio tokens that focus on content. \n- the interleaving (replacing spans of text with its synthetically produced speech tokens) is interesting as it forces the model to learn alignment between text and audio tokens. It was also shown to be effective in practice.\n- The ability to perform text-guided response (which is a kind of chain-of-thought) is interesting.\n- The ablation study was done well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the important problem of closing the gap between textLMs and SpeechLMs, to enable a spoken conversation with an AI model.\nThe paper leverages the recently proposed *supervised* semantic tokens (introducing a discrete bottleneck in ASR models) which show better alignment with text. \nMoreover, they train a text to audio tokens to enable the generation of synthetic audio tokens based on high-quality texts. \nThey suggest randomly replacing textual tokens with the corresponding synthetic speech tokens (resulting in an interleaved sequence), which helps to align the text and audio tokens. \nThey train large SpeechLMs on diverse text/audio inputs (audio only, text only, interleaved, [text,audio] and [audio,text]), and show convincing results on SLM, SQA. \nThey perform supervised finetuning on a proprietary (?) spoken dialogue dataset, and evaluate their model as a spoken chatbot using GPT-4 as a judge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Several methodological evaluation details are missing (what was measured and how was it computed), mostly in Section 2.1 and Table 1 (See questions). 
Whenever you report some metric with an intuitive non-exact name (e.g., Content Preservation - LS), you should explain somewhere it more precisely (e.g., Content Preservation: We run our quantized whisper on the LS (LibriSpeech) dataset to generate text and report the WER to the GT transcript)
I understand that there’s a space limitation, but this is important. \nI've listed some specific details I found missing in the Questions section. I suggest adding a short sub-paragraph that describes the evaluation methodology (defines all datasets+metrics being used) or adding those details into the main text within the relevant sections. If space is an issue, you can add those into an appendix section.\n\n- Currently there's no sample page (unless I missed something). Consider creating a sample page with samples on speech continuation (audio prefix, audio GT continuation, and the model's audio continuation). Also consider adding examples of spoken question answering (audio question, audio GT answer, the model's prediction). Examples from the spoken chatbot evaluation would also be great. Moreover, you can visualize how the interleaved samples sound like (paragraph with audio tokens that were decoded in it)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1 Why is the method pre-trained on Chinese data?\n2. Why is GPT-4 used for scoring?\n3. Why is whisper used as the encoder? did it perform better than other encoders?\n4. It is stated that the model can do streaming, was this evaluated?\n5. Why are the ablation also done on a 1.5B model?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. A novel approach for expanding the training corpora of SpeechLMs.\n2. The method is easy to implement, and thus can be expanded to other methods. \n3. State of the art results on sTopic-StoryCloze, sStoryCloze, Web Questions, Llama Questions and TriviaQA." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for scaling SpeechLMs using a new pre-training scheme called \"Synthetic Interleaved Data\". In this scheme, a text to token lM is first trained on supervised data, and then used to expand text-based datasets by predicting the speech tokens directly from the text data. For pre-training, only spans of text are converted to speech tokenization and text and speech are interleaved with one another. After this pre-training, the model is trained in the usual SpeechLM format. The strength of this approach is the ability to generate a large-scale dataset from text-corpora. Furthermore, this method shows strong results on speech-understanding and generation datasets.\n\nOverall, the method presented in this paper and novel, and has an interesting contribution. On the other hand, the writing is unclear and the evaluations are somewhat lacking. I thus recommend to borderline reject this paper." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing is unclear for most parts of the paper. While Synthetic Interleaved Data is the main contribution of the paper, it is not clear what this means from the abstracts and introduction. Furthermore, the main explanation of this is just a small part of the paper. I would suggest reducing the length and condensing the section regarding speech tokenization (as this is a well established concept) and increasing the amount of detail in the section regarding Synthetic Interleaved Data. I would also suggest adding a better summerization of this concept in the introduction.\n2. While there is a good ablation analysis, there aren't any explanations for why the architectural / training parameters where chosen as they where. I suggest to add a dedicated subsection or add this into the methodology section where the parameters are introducted.\n3. The training datasets are lacking in clarity:\n- For table 1, is it unclear on which dataset it was evaluated and why MOSNet scores are so low. \n- It is unclear what Supervised speech-text data are used to train the model.\n- It is unclear what datasets areused to train the text to speech tokens model. \n- It is unclear what datasets are used to fine-tune the tokenization encoder and decoder.\nThese should be added in the section specifying the training pipeline or in a dedicated table / figure.\n4. The experimental results are lacking in clarity:\n- the origin of the baseline numbers in all tables is lacking, are these from other papers or from independently evaluating? I would suggest adding these directly to the table or in the caption.\n- In table 3, speechGPT and Spectron are speech to speech methods, while the results are stated in speech to text.\n- in Table 1 MOSNet was used while in table 4 UTMOS is used. This reason for this should be explained in the paper or have uniformity between them. \n5. The paper is lacking some evaluations:\n- Human evaluations of speech quality, such as MOS or MUSHRA evaluations where humans will rate the speech quality of the proposed method compared to the baseline.\n- Evaluation on other tasks, such as speech continuation, reconstruction and TTS for the full method. Speech continuation and Reconstruction results at least should be added, while TTS might be left for future work." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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. I'm curious if there is any additional filtering process for the text input into the text-to-tokens model, as there are many texts that cannot be synthesized, such as code or mathematical formulas.\n\n2. Why is it required for the encoder and decoder to be causal when training the speech tokenizer? During the inference stage, speech are segmented into 2s-chunks for inference, which does not require a streaming speech tokenizer." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This work demonstrates the need to use interleaved speech-text data for cross-modal pre-training, paving the way for more effective speech-text pretraining. Previous pretraining methods have typically relied on paired ASR or TTS data, which is limited in scale; or independently utilized unsupervised speech and text data, which cannot model the dependency between the two modalities. I believe that this work makes a great contribution to the filed of speech-text pretraining.\n\n2. Although the use of interleaved data has been proven effective in the field of image-text pretraining [1], it has not been explored in the field of speech-text pretraining. In contrast to the vision field, web data can naturally form interleaved image-text data, but it is difficult to collect real data with interleaved speech and text. This work proposes a novel method to synthesize pseudo-interleaved data using a text-to-tokens model, and through thorough experimentation, demonstrates the effectiveness of synthesized data and observes that scaling synthesized data continues to provide benefits.\n\n3. This work is very solid and well-motivated. The paper is well-structured, with a clear presentation of the methodology, experiments, and results. This work also reports state-of-the-art performance in speech language modeling and spoken question answering.\n\n[1] Chameleon team. Chameleon: Mixed-Modal Early-Fusion Foundation Models. arxiv: 2405.09818." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work present a novel approach for scaling speech-text pre-training by leveraging large-scale synthetic interleaved data derived from existing high-quality text corpora. This work utilized existing text-to-speech (TTS) datasets to train a text-to-token language model, which is used to synthesize 600B tokens of speech-text interleaved data. Experiments have demonstrated the effectiveness of incorporating interleaved speech-text data, which can effectively align speech and text. Furthermore, this work constructs a speech instruction dataset, SpeechDialog-90K, to fine-tune models into a chatbot model, which can directly generate speech responses without intermediate text response and significantly improve the previous SOTA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is a lack of performance evaluation for the text-to-tokens model. For example, after converting a piece of text into tokens and then decoding it into speech using a vocoder, what is the ASR-WER of the resulting speech? This result is necessary to demonstrate the semantic representation capability of the tokens generated by the text-to-tokens model.\n\n2. Based on my experience, tokens generated by text-to-models lacks diversity, and the speech instruction dataset SpeechDialog-90K in the SFT stage is also synthesized by TTS, so I am concerned about whether the model can understand real speech input. I checked the evaluation datasets in this work, all of which were synthesized through TTS, lacking evaluation on real speech input (such as AIRBench [2]).\n\n3. The quality of the output speech is not satisfactory, as evidenced by the poor ASR-WER in Table 4. In comparison to llama-omini, which was only trained on 100 hours of speech data, this model was trained on a much larger scale of 700k hours of speech data. The author needs to provide a reasonable explanation for why the ASR-WER is so poor.\n\n[2] Yang, Qian, et al. AIR-Bench: Benchmarking Large Audio-Language Models via Generative Comprehension. arxiv 2402.07729." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024scaling,\ntitle={Scaling Speech-Text Pre-training with Synthetic Interleaved Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3tukjsVyrE},\nnote={under review}\n}" }, "abstract": { "value": "Speech language models (SpeechLMs) accept speech input and produce speech output, allowing for more natural human-computer interaction compared to text-based large language models (LLMs).\nTraditional approaches for developing SpeechLMs are constrained by the limited availability of unsupervised speech data and parallel speech-text data, which are significantly less abundant compared to text pre-training data, thereby limiting the scalability of SpeechLMs as LLMs.\nWe present a novel approach for scaling speech-text pre-training by leveraging large-scale synthetic interleaved data derived from existing high-quality text corpora.\nOur method employs a supervised speech tokenizer derived from an automatic speech recognition (ASR) model (e.g. Whisper) by incorporating a vector-quantized bottleneck into the encoder. In this process, we create tokenizers with various sampling rates ranging from 50Hz to as low as 6.25Hz. This supervised training approach results in discrete speech tokens with strong semantic preservation even at lower sampling rates, while still maintaining speech reconstruction quality.\nBy synthesizing speech-text data from existing text pre-train corpora with a text-to-token language model and scaling our pre-training to 1 trillion tokens, we achieve state-of-the-art performance in both speech language modeling and spoken question answering, improving performance on spoken questions tasks from the previous SOTA of 13\\% (Moshi) to 31\\%.\nWe further demonstrate that by fine-tuning the pre-trained model with speech dialogue data, we can develop an end-to-end spoken chatbot achieves competitive performance comparable to existing baselines in both conversational abilities and speech quality, even operating exclusively in the speech domain." }, "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; speech language model; spoken chatbots" ] }, "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/a570512a55f674448429815bfac08a506151c1f7.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": "Scaling Speech-Text Pre-training with Synthetic Interleaved 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": 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": "None noted / 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 method dissociates neural-only, vs. behavior-only, vs. shared neural-behavioral spaces.\n- BRAID has the ability to take in measured inputs. However, this is not providing any conceptual advance from existing methods, since they are simply fed into the RNNs with an additional input transformation in the form of K." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "BRAID aims to distinguish the effect of measured inputs towards neural dynamics, while dissociating shared neural-behavioral dynamics, neural-only dynamics, and behavior-only dynamics. It does so by training a series of recurrent neural networks: (1) Stage 1 trains a shared recurrent neural network that outputs both neural activity and behavior for (a) 1-step ahead and (b) multi-step ahead prediction, (2) Stage 2 does similarly but for neural dynamics only, and (3) Stage 3 does so for behavioral dynamics only. The authors show better correlation coefficients with the neural and behavioral activity in validation datasets as compared to some baseline comparisons." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- BRAID is very similar to DPAD (Sani et al., 2024), but with the addition of measured inputs, behavior-only dynamics, and an additional network that predicts m-steps ahead. None of these additions are conceptual advances, and are extremely straightforward additions to an existing method.\n- The paper claims that \"BRAID disentangles [input] dynamics from intrinsic dynamics\", however the contribution of the inputs is not analyzed further at all - can one effectively disentangle input dynamics from intrinsic dynamics via this approach?\n- The modeling strategy is multi-stage and quite involved, with multiple RNNs being trained without fully going into the utility of each one. The 'RNN_{fw}' models seem to be forecasting, but why is it necessary to have a separate RNN for forecasting when the 'RNN_1' model is a dynamical model that should be in theory capable of predicting m-steps forward in time?\n- The R^2 should be reported throughout the paper instead of Pearson's; the R^2 is more standard in this field, and takes into account the predictability using the mean value of the signal.\n- There is no attempt at interpretability of the underlying dynamics and the contribution from different sources as identified by this method.\n- While the authors show that BRAID performs with higher behavior reconstructions than TNDM as shown in the Appendix, this is very much to be expected since TNDM does not optimize separately for behavior reconstruction, as BRAID does. Similarly, DPAD does not either (and does not take in inputs). However, these are very simple to add to both of these methods, and thus do not provide fair comparisons in their existing form." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* Could the authors attempt to make the writing clearer with regards to explaining the model and the various training stages? Perhaps a summary as a table might help.\n* Is it possible to show results when ablating for the second and third (post-hoc) stages of training? Apologies if I've missed this. Also, could the authors comment on what happens when just training the model end-to-end?\n* Could the authors consider experiments that show the scaling performance of the model vs number of neurons, and attempt to compare with LFADS (and CEBRA) if possible?\n* Could the authors comment on multi-session models, generalisation to unseen sessions?\n* Apart from neural predictivity and behaviour decoding, could the authors perhaps use a similarity metric to explicitly compare the learned and true dynamical systems (e.g. using [Dynamical Similarity Analysis](https://openreview.net/forum?id=7blSUMwe7R))? This should be fairly easy for the synthetic tasks, and should be possible in case of the neural data as well.\n* It would be interesting to see the model's performance on a dataset involving multiple brain regions, but I understand that this would take more time to run and depend on the availability of a dataset.\n* I spotted a potential typo on Line 073 (\"prepossessing\" -> \"preprocessing\"?), and another minor issue on Line 111 (\"intrinsic representation of dynamic.\" -> \"dynamics\"?)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The method is well-motivated given prior work, and the causal formulation makes it amenable to real-time inference, which is a strong asset.\n* The figures are neat and the experiments are comprehensive. The writing is mostly clear, but I have some comments on explaining the method in a slightly clearer manner (see Weaknesses).\n* The method performs well not only on synthetic tasks designed to show its efficacy over several baselines, but also on decoding real neural data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new deep learning method to jointly model neural and behavioural data by explicitly modelling task inputs, in order to better model and disentangle input effects and intrinsic neural dynamics that are predictive of behaviour. The learning of the intrinsic dynamics is enabled by the use of a forecasting objective, i.e., $m$-step-ahead neural and behavioural prediction. The method involves 2 (or 3) RNN models, and optimisation is done in multiple stages: a pre-processing stage to filter out behaviours that are not relevant to recorded neural dynamics (used instead of actual behaviour for training), a stage to learn the neural dynamics predictive of behaviour, and a stage to learn any residual neural dynamics. The proposed method outperforms several baselines at predicting neural activity and decoding behaviour, and is also amenable to real-time predictions due to its causal formulation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The model architecture and the learning stages are quite complicated, and the writing makes this hard to understand in some places. If I've understood this correctly, there are mainly two RNNs: one to predict the next timestep neural activity given current activity and observed inputs, and another to predict activity $m$ timesteps into the future using just observed inputs. Each of these RNNs' parameters are split into up to 3 subsets, which are optimised sequentially: one to learn behaviourally relevant neural dynamics, one to learn any residual dynamics, and another to learn behaviour that is not encoded in the neural dynamics. This is in addition to an RNN that preprocesses inputs.\n\n While this preprocessing RNN has been ablated for, can the authors comment on how one identifies in practice whether or not to use the second and third stages of training (RNN2 and RNN3), which were mentioned to be optional? In general, some additional clarity in the writing here would be appreciated.\n\n* It would be important to see how the method scales with the number of neurons – based on the details in the appendix, it seems like the maximum dimensionality explored here (in the neural data case) is around 45. Perhaps the authors could run an experiment comparing decoding performance and neural predicitivity for differing numbers of neurons from the same dataset to show this (and also comment on the time taken to train BRAID).\n\n* From the experiments it seems that BRAID is mainly a single-session model. It is well-known that there can be a lot of variability in neural activity across sessions as animals learn to perform the task better or due to some representational drift. This does not seem to be addressed in the paper as far as I can see, but could the authors comment on how BRAID generalises to unseen sessions, and also specifically for the later parts of a session when training on the initial parts (one of the 5 folds)?\n\n* While the experiments are comprehensive and baselines have been compared against, I think comparisons with LFADS (and if possible, CEBRA) could be useful here – both these approaches seem slightly less involved in terms of training complexity, but would represent baselines when ablating for explicit input modelling (LFADS) or explicit dynamics modelling (CEBRA). The idea here is also that LFADS is designed to infer inputs to the dynamical system, so it might perform better than the extended TNDM baseline included currently." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Some of the following questions relate directly to the weaknesses raised above. \n- How does your predictor relate to sequential posterior in variational inference?\n- Monkey experiment: What do x^1 and x^2 look like in this task? Is there any non-encoded activity x^3?\n- Could you provide more details on the \"automatic\" selection (L313)? Is it simply picking the best performing?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The work tackles a very relevant question in the field of statistical neuroscience of disentangling the relationships between behavior, neural activity and inputs, and shows originality in doing so with careful modeling. \n- The modeling is well-formulated in the main text, and made clear for the reader in the appendix and through provided code.\n- The decomposition into multiple stages and sub-components to encourage learning behavior-relevant activity is interesting and novel (to my knowledge).\n- The authors showcase their model on synthetic and real experimental data, showing strong performance in both. \n- The authors help support the significance of their results by (1) comparing them against many baselines and extensive ablations of their model, and (2) performing many runs, providing error bars for all numerical results. \n- The metrics are meaningful and bypass common non-identifiability problems (such as considering the eigenvalues of A)\n- The figures are clean and easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors tackle the problem of disentangling intrinsic from input-driven neural dynamics underlying behavioral measurements. They do so through a novel nonlinear model called BRAID. Their approach is to conceptualize the intrinsic as the generative (i.e. forward) dynamics, and the input-driven as the (posterior-) predictor dynamics. The authors model each of these components with nonlinear DNNs for flexibility, allowing each to be learned. Importantly, they devise a multi-stage training procedure that prioritizes learning behavior-relevant dynamics, with neural reconstruction placed second. They show how this approach can help disentangle the neural dynamics directly relevant to behavior, in both synthetic experiments and monkey motor reaching neural activity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am putting a score of 6 (marginal accept) but I would be willing to increase it if the authors can help address my weaknesses/questions, and in particular the \"major\" ones below. \n\nMajor:\n- I am not convinced by the claims of real-time inference in the discussion, and the lack of surrounding literature on the predictor components of the model. Variational inference approaches for sequential models similarly model the (sequential, i.e. filtering) posterior and optimize it jointly with the generative model in the ELBO. I believe the similarities could be actually quite exact, in which case a fairer depiction of the relationship with VI is necessary. On that note, Table 1 refers to your method as pure LL optimization, but perhaps it could be more easily interpreted through an ELBO objective.\n- Most of the results are numerical, and the paper lacks a bit in alternate results such as posterior trajectories or visual reconstruction. \n- In line with the previous comment, the monkey experiment results are numerical and would benefit from some post-training analysis if the goal is to show the model as a modeling and analysis tool. \n\nMedium:\n- Appendix should be for additional but not necessary details to understand the paper. One example of this is the simulations in section 4.1. The notation from L312 follows the appendix where you introduce $f_{C_z}$ and $\\nu, \\bar\\nu$ -- I would make it consistent with eq. (1) or add eq (A.11) to the main text.\n- Identifiability: the authors discuss behavioral- vs neural-relevant activity, but having nonlinear A, Cs are another type of inter-dependence, which could be discussed further.\n- Monkey experiment: numerical performance results are good but sometimes lack transparency in their presentation. For instance, U-BRAID does do better on neural, but that is by construction and does not put BRAID in any lesser light. The corresponding paragraph (3 of section 4.2) however does not acknowledge this better performance. Similarly, the authors refer to mmPLRNN as having an \"unfair advantage\" (L451). I would remove this.\n- The relationship with probabilistic formulations is skimmed but still alluded to with the ELBO/LL in Table 1. I would expand further on these ideas. \n\nMinor:\n- Using $\\cdot$ (\\cdot) instead of $.$ (dot) is more standard for place-holder variables in functions (L191, L312, L1015)\n- Dependency on x^1 on Stage 2 could be made more explicit in the text\n- Table 2: bold entries per metric would be more representative" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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.\tIn Figure 1, should there be a connection between stage 2 and behavior Z_k if you are learning C_z(2) in Stage 2 2b? If so, what is the difference between X_k(1) and X_k(2) since both are connected to neural activity via [C_y(1), C_y(2)] and to behavior via [C_z(1), C_z(2)]?\n2.\tIn equation 1, the model can learn behavior dynamics that are predictable from the input but are not encoded in the recorded neural activity, but how do you make sure that the behavior prediction is not dominated by the input? In Figure 3a, there is a high correlation between your input target position and your behavior cursor position and velocity. If so, can the model learn any information encoded in neural activity?\n3.\tThe model also has an input term in modeling neural activity y, what does this input mean? Because you have an input in latent space that encodes intrinsic contribution, then you also have the same input to encode input-driven contribution. \n4.\tHow do you define your initial states x0?\n5.\tI am curious, how do you ensure that the model learned all encoded information and the residual is the non-encoded information since the model is nonlinear with high flexibility? \n6.\tFollowing Q5, you mentioned you trained the model until convergence, could you show your learning curve against epochs with zooming in the last epochs if necessary?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "BRAID considers input-driven neural and behavioral dynamics and dissociates them. The authors performed ablation studies to shown which parts of BRAID contributed to the neural and behavior forecasting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduced a novel framework, BRAID, modeling nonlinear neural and behavioral dynamics with external input. This model dissociates behaviorally relevant neural dynamics, neural specific and behavioral specific dynamics, and outperforms baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the model has been shown to be highly efficient in modeling neural spiking data, it has not been tested on other modalities, such as widefield calcium imaging.\n2. Although the authors mentioned that their model is not to infer unmeasured input, I still think this may be a weakness of this model, because (for example) the neural and behavioral dynamics can be encoded by unmeasured input." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024braid,\ntitle={{BRAID}: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3usdM1AuI3},\nnote={under review}\n}" }, "abstract": { "value": "Neural populations exhibit complex recurrent structures that drive behavior, while continuously receiving and integrating external inputs from sensory stimuli, upstream regions, and neurostimulation. However, neural populations are often modeled as autonomous dynamical systems, with little consideration given to the influence of external inputs that shape the population activity and behavioral outcomes. Here, we introduce BRAID, a deep learning framework that models nonlinear neural dynamics underlying behavior while explicitly incorporating any measured external inputs. Our method disentangles intrinsic recurrent neural population dynamics from the effects of inputs by including a forecasting objective within input-driven recurrent neural networks. BRAID further prioritizes the learning of intrinsic dynamics that are related to a behavior of interest by using a multi-stage optimization scheme. We validate BRAID with nonlinear simulations, showing that it can accurately learn the intrinsic dynamics shared between neural and behavioral modalities. We then apply BRAID to motor cortical activity recorded during a motor task and demonstrate that our method more accurately fits the neural-behavioral data by incorporating measured sensory stimuli into the model and improves the forecasting of neural-behavioral data compared with various baseline methods, whether input-driven or not." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Deep learning", "Dynamic modeling", "Sensory stimuli", "RNN", "Intrinsic", "Behavior" ] }, "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/fe1071b2c453fa5e7cd35c024ad10a45fc21f1e5.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": "BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral 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": 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": "1. The denoising mechanism requires the addition of a tanh function. Could this cause performance degradation? Is it always applicable to general ML tasks?\n2. There are no empirical results for cut layers below L-2. Exploring these results could help elucidate the proposed methods' limitations.\n3. Could the authors provide more insight into why scaling does not mitigate FSHA effectively?\n4. Could the authors discuss the limited performance improvement of the denoising mechanism on Laplacian noise?" }, "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 provides a theoretical proof explaining how the proposed denoising methods reduce MSE under certain conditions, effectively denoising while preserving privacy guarantees. The paper covers both a simple identity layer and a more complex linear layer with non-linear activation, with implications for both classification and broader applications.\n2. It discusses how the two denoising methods approximate DP-SGD and contribute to privacy.\n3. Simulations are conducted for both linear and non-linear cases. Results indicate that the denoising methods exhibit different characteristics at varying noise scales, validating the theoretical claims and offering guidance on the application of each method.\n4. The authors conduct practical experiments across five datasets and various ML tasks, including image classification, recommendation, and language modeling. The results show significant improvements, with accuracy close to the clean model performance.\n5. The paper examines how hyperparameter settings (e.g., learning rate, weight decay, and optimizer choice) impact denoising performance in practical CIFAR-10/100 image classification tasks.\n6. Empirical studies indicate that the proposed method, combined with noise injection, effectively mitigates FSHA.\n7. Code is available for review and extensive discussion & extra experiments are available in appendix" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces two denoising techniques—scaling and random masking—for Differential Privacy (DP) within the Split Learning framework. These methods aim to preserve security guarantees while maintaining model accuracy. The authors focus on theoretical contributions supported by extensive simulation and empirical studies, demonstrating that the proposed techniques enhance the accuracy of split neural network classification under DP. Additionally, the paper shows that the resulting deep neural networks are resilient to state-of-the-art hijacking attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The theorem and its proof are limited to L-1 layers, which restricts the scope.\n2. While simulation largely supports the theoretical findings, discrepancies are observed between simulation and practical results. For example, scaling outperforms masking in simulation, though this does not hold consistently in practice.\n3. Performance degradation occurs when the cut layer is set below L-1.\n4. The paper does not explore hybrid applications of the two denoising methods.\n5. The empirical study tests only a limited set of noise-injection parameters (e.g., noise scale = 0.7). Exploring denoising limitations would be informative.\n6. The empirical study also uses a limited range of cut-layer settings.\n7. The paper evaluates only a limited range of attack types.\n8. The techniques are demonstrated only in a two-party split learning setup, despite their potential applicability to Split Federated Learning, which has broader real-world use." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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) The literature review could be expanded to include more recent studies, providing a broader context for the proposed techniques.\n2) Additional details on the experimental setup would enhance reproducibility and transparency, allowing other researchers to validate the findings.\n3) The paper could benefit from more comprehensive discussions on the limitations of the proposed methods and potential future work.\n4) A more detailed analysis of the impact of varying noise levels on training accuracy could provide deeper insights into the practical applications of the proposed techniques." }, "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 addresses a critical issue in Split Learning regarding data privacy and accuracy, which is highly relevant in today’s data-driven environment.\n2) It introduces two novel denoising techniques—scaling and random masking—that show significant promise in improving training performance under noisy conditions.\n3) The theoretical analysis is well-supported by experimental results, enhancing the validity of the proposed methods.\nThe clarity of the writing and organization of the content facilitate understanding, making complex ideas accessible to a broader audience." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents innovative denoising techniques for Split Learning to enhance training accuracy while preserving data privacy against reconstruction attacks. The authors propose scaling and random masking methods, demonstrating their efficacy through theoretical and experimental results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The experimental validation may lack diversity in the datasets used, potentially limiting generalizability.\n2) Comparisons with existing methods could be more robust to highlight the advantages of the proposed techniques.\n3) The paper could benefit from clearer explanations of the denoising algorithms for readers unfamiliar with the underlying concepts." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. Is there any conclusion on the choice of parameters (i.e. \\lambda, p). Authors mostly use \\lambda=0.1 or 0.2, p = 0.1 or 0.2. How did author choose on this setting? Does the theoretical analysis provide guidance on how we should choose \\lambda and p?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors aim to demonstrate that the proposed denoising methods (scaling and random masking) improve test accuracy through both theoretical analysis and empirical evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper discusses the use of scaling and random masking as a denoising method for noise-injected networks in the context of split learning. The authors present both theoretical and empirical evidence showing that applying these denoising techniques during the training phase improves testing accuracy. Additionally, the study demonstrates that the proposed denoising methods enhance the network’s resilience against feature space hijacking attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I believe the focus of the paper is misaligned with the target. First, I think the motivation was made clear: to protect models against reconstruction attacks, noise is often added to intermediate representations (IRs), which leads to a drop in accuracy. This creates a privacy-accuracy or robustness-accuracy trade-off. The primary goal of the paper should be to demonstrate that the proposed methods (scaling and random masking) improve this trade-off, but unfortunately discussions about this trade-off is completely missing.\n\nThe authors dedicate substantial space to showing that the denoising methods preserve accuracy when noise is injected. However, this seems unnecessary. It is intuitive that denoising in noise-injected networks would help maintain accuracy—a rather trivial observation. Variance scales as the square of the variable scaling, so scaling down naturally reduces variance. Similarly, adding random masking during training makes the network more robust to noise. This is intuitive and easy to understand. \n\nWhat the authors really want to show is that the proposed methods also helps in improve network privacy (or at least no degrade on it) so that a better trade-off become possible. This is critical but unfortunately, the authors spend so less effort on it. I expected to see quantitative measurements against multiple types of attacks, but only the feature space hijacking attack is covered, while other attacks mentioned in the introduction are ignored. Additionally, quantitative results are lacking. \n\nAs such, I feel authors fail to demonstrate a better accuracy-privacy trade-off, leading to their contribution unjustified.\n\nAdditionally, they are also minor issues in both theoretical and empirical study. In my understanding, the denoising methods are applied during the training, so with and without this it will lead to different weights. I don't think it is reflected in theoretical analysis. In the empirical analysis part, I don't think it is the best choice to show graphs of accuracy by training epoch as it takes a lot of space. Instead, I would expect to see much richer results from using different parameter settings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024accurate,\ntitle={Accurate Split Learning on Noisy Signals},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3vE4B61VSw},\nnote={under review}\n}" }, "abstract": { "value": "Noise injection is applied in Split Learning to address privacy concerns about data leakage. Previous works protect Split Learning by adding noise to the intermediate results during the forward pass. Unfortunately, noisy signals significantly degrade the accuracy of Split Learning training. This paper focuses on improving the training accuracy of Split Learning over noisy signals while protecting training data from reconstruction attacks. We propose two denoising techniques, namely scaling and random masking. Our theoretical results show that both of our denoising techniques accurately estimate the intermediate variables during the forward pass of Split Learning. Moreover, our experiments with deep neural networks demonstrate that the proposed denoising approaches allow Split Learning to tolerate high noise levels while achieving almost the same accuracy as the noise-free baseline. Interestingly, we show that after applying our denoising techniques, the resultant network is more resilient against a state-of-the-art attack compared to the simple noise injection approach." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Split Learning", "Denoising techniques" ] }, "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/7ea82d9fc9b7d035076a18775693a1507cb599d4.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/58baa0cb9796821cb4c90445d312589b76397971.zip" }, "title": { "value": "Accurate Split Learning on Noisy Signals" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Please look at the weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The work provides a list of theoretical results. I can only check some of them and believe they are correct. \nIt proposes a more realistic setting for the NTK." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work focus on two problems: under what conditions the training error bound and the generalization error bound are obtained and go to zero in the NTK setting. The paper disproves the work of Arora and then provides proofs for better results, particularly for the case the initial values of training weights do not decrease with the sample size." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The crucial condition for Theorem 2.3 hold is $\\kappa = O(n^{\\alpha})$ with $\\alpha <0$, that only appears its proof in the Appendix. Based on the condition, the considered setting is very specific, that is the initial weights depend on $n$. Let $n$ go to infinity, the $\\|\\mathbf{y}\\|^2$ grows as $n$, the distance from initial weights to its desired destination increases as $n$ increases. So intuitively what is surprise from the result: the training weight cannot converge to the weight in which model has zero loss?\n\nIt is often that the paper states some theoretical results depending on some conditions, then those conditions happen under assuming another set of conditions? Could the author state everything together so that readers can check them all? For example, Proposition 3.1 states condition 1 and 2 hold, when requiring other conditions. In Proposition 3.1, $\\epsilon$ appears, does $\\epsilon$ affect the conditions on $m, n$ and other parameters in the proofs of other theoretical results? \n\n\nThe paper is very long, full of technical, not well-organized. There is no explanation for intuition behind the proofs and the proof's structure. For example, equation (24) appears at the beginning of the Appendix's proof refers to equation (138), which is at 20 pages later. It is not easy for reader to read 50 pages of proof with that presentation. It is unfair to reject the paper just due to poor presentation, but it makes reader hard to verify technical results to be certain that they are all correct. \n\nIn short, could the authors make a table comparing this work and the work of Arora in term of conditions and results and then highlight the technical advancement of this work, also give a discussion about all related parameters in one. Since for this kind of theoretical work, the most difficult task is to be certain that all conditions do not conflict each other and cover a wide range of cases. \n\nMinors\n1. Unconventional notation $\\{i\\}$ for the set $\\{1,2,\\ldots, i\\}$\n2. Math notations are not consistent, not defined, $\\Omega(1), O(n), \\Omega(n), \\Theta(n)$, $o(1)$ with respect to $n$.\n3. I could not find the proof of Proposition 3.1???\n4. Line 375, how large is $\\alpha$ acceptable, since $m$ depends on $n$ in other results?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I would like the authors to response to the weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The orginaztion of the paper is clear and easy to read.\n- Revising the impact of initialization to the generalization error bound is an interesting problem and extends the existing theory of neural tangent kernel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper revises the Neural Tangent Kernel (NTK) in analyzing the generalization performance of neural networks at the infinite width limit, focusing particularly on the impact of the scale $\\kappa$ of initialization over the bounds for optimization and generalization. Showing that the error bounds in the previous works Arora et al. (2019a;b) actually does not hold in the case $\\kappa = o(1)$ which is necessary for the bound to be non-vanishing. Then, this paper revises the previous result, establishing optimization and generalization bounds independent of $\\kappa$. Basing on the new error bounds, the paper discusses the generalization error bounds with arbitrary initialization. Finally, numerical simulations are provided to verify the theory." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "### Correctness\n\nOne major problem is the correctness of the results in this paper. \nIn the proof of Lemma 4, the paper applies the Markov’s inequality for $\\mathbf{a}$ with fixed $\\check{W} \\in \\Gamma$ to prove Eq. (41), and then use (41) for $W(k)$ subsequently, which is shown to lie in $\\Gamma$. However, since $W(k)$ is random and dependent on $\\mathbf{a}$, there is no guarantee that Eq. (41) holds for $\\check{W}$ replaced with $W(k)$. To make this approach work, a uniform version of Eq. (41) over $\\Gamma$ is needed. Given that Lemma 4 is not well justified, the main results in this paper are not well-supported.\n\n### Novelty\n\nThe results regarding $\\kappa = \\Theta(1)$ are not novel and are slight modifications of the existing results. While the proof seems to be long, they are mostly minor revisions of the existing proofs.\nAlso, this paper considers the simplest setting of a two-layer ReLU network with only the first layer trainable. However, as existing NTK theory (for example, [1]) can deal with more general settings such as multi-layer networks, extensions to these settings should also be considered.\n\nMoreover, regarding the generalization error bounds, this paper misses some related literature studying the NTK regime in terms of kernel regression [2,3], where sharper bounds are established and also do not depend on the scale of initialization (as the NTK in this setting does not depend on the scale).\nI think a more detailed comparison with the existing works should be provided.\n\n\n### Minor\n\nSome notations are used without definition and are not consistent. For example, $H^*$ is used in Section B.2 but defined in Section C, which seems to refer to $H^\\infty$ in the main text.\n\nWhen reviewing this paper, I also find a very recent paper concerning the impact of initialization [4]. A comparison with this paralleling work would be benefitial to the readers.\n\n### References\n\n[1] Zeyuan Allen-Zhu, Yuanzhi Li, and Zhao Song. A convergence theory for deep learning via over-parameterization. In International Conference on Machine Learning, pages 242–252. PMLR, 2019.\n\n[2] Namjoon Suh, Hyunouk Ko, and Xiaoming Huo. A non-parametric regression viewpoint: Generalization of overparametrized deep relu network under noisy observations. In International Conference on Learning Representations, 2021.\n\n[3] Tianyang Hu, Wenjia Wang, Cong Lin, and Guang Cheng. Regularization matters: A non-parametric perspective on overparametrized neural network. In International Conference on Artificial Intelligence and Statistics, pages 829–837. PMLR, 2021.\n\n[4] On the Impacts of the Random Initialization in the Neural Tangent Kernel Theory, Guhan Chen, Yicheng Li, Qian Lin. https://arxiv.org/abs/2410.05626" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Could the authors provide a complete proof of Theorem 3.1?\n\nCould the authors clarify the formula numbering and proof structure in Appendix B in a future version?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The revised NTK framework proposed in this paper solves the dependency problem between the NTK initialization scale and the sample size. In addition, the numerical experiments in this paper support improving the proposed regularization method in improving generalization performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the optimization and generalization problems of over-parameterized neural networks, proposes a revised NTK framework that eliminates the dependency between initialization scale and sample size, and verifies its experimental results on benchmark datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The writing of the paper is not clear, especially in Appendix B, where the proof structure and formula numbering are very confusing and difficult for readers to understand. In addition, the lack of proof of Theorem 3.1 in the appendix makes it impossible to confirm the correctness of the theoretical results of the paper. Overall, the paper is not yet in a submittable state. I suggest authors complete any unfinished parts and thoroughly reorganize the paper to improve its clarity and readability before submitting it to a future conference." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 note that the generalization upper bound in Thm 2.5 is \n\n$$O\\left( \\sqrt{ \\frac{2(y - u(0))^T (H^\\infty)^{-1} (y-u(0))}{n}} + \\sqrt{\\frac{\\log {\\frac{n}{\\lambda_0 \\delta}}}{n}}\\right),$$\n\nwhich is about $O(1)$ since $y - u(0)$ is a $n$-length vector. However, during these years, I have seen lots of recent work which provide tighter upper bounds for generalization error of networks based on NTK (e.g., $O(n^{-\\frac{d+1}{2d+1}})$,[Suh et al., 2021], [Hu et al., 2021],), which shows that maybe $O(1)$ is not tight enough. At the very least, an upper bound on the generalization error that does not decrease with increasing sample size n is hard to consider as tight. I think maybe it is a better choice to compare the result with theirs.\n\nIn Table 1, I see that this paper compared \"Original CMD\" with \"Revised CMD\", which is generally the same (e.g., 0.5998 with 0.5997). Table 1 also compares the generalization upper bound, which is common due to the removal of the initialization scale factor. So I am slightly confused by the meaning of Table 1. In Figure 2, I think the experiment should represent more details to make Figure 2(b) reliable (e.g., the stopping time of training). \n\nAdditionally, I don't understand the y-axis of Figure 2(c). I thought CMD is time-invariant and should not change with training, so I am confused about what the y-axis of Figure 2(c) represents. I would appreciate it if you could correct my misunderstanding." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper extends the work of Arora et al. 2019, removing the dependency of initialization on sample size. The authors provide a practical approach that aligns with real-world initialization schemes. This contribution enhances the applicability of NTK theory in real-world scenarios, and can be extended to future research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper mainly builds on the results in Arora et al. 2019, focusing on the optimization and generalization of over-parameterized neural networks. The paper addresses a limitation in NTK-based analysis in Arora et al. 2019, which requires the scaling of initial parameters to decrease with respect to the sample size, a condition that contradicts practical initialization schemes. To resolve this issue, the authors try to removes the dependency of initialization on sample size, and extend to applying the method in real practice." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The generalization upper bound in Theorem 2.5, approximated as $O(1)$, may not be sufficiently tight compared to more recent works that provide tighter bounds, such as Suh et al. (2021) and Hu et al. (2021), which suggest a bound of $O(n^{-\\frac{d+1}{2d+1}})$. A bound that does not decrease with sample size $n$ raises concerns about its effectiveness. At the same time, the characterization of generalization capability in this paper is precisely based on this upper bound, which makes me feel that it is not solid enough.​\n\n\n## References\n\n[Suh et al., 2021] Suh, N., Ko, H., and Huo, X. (2021). \"A non-parametric regression viewpoint: Generalization of overparametrized deep ReLU network under noisy observations.\" In _International Conference on Learning Representations_.\n\n[Hu et al., 2021] Hu, T., Wang, W., Lin, C., and Cheng, G. (2021). \"Regularization matters: A nonparametric perspective on overparametrized neural networks.\" In _International Conference on Artificial Intelligence and Statistics_, pp. 829–837. PMLR." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024revised,\ntitle={Revised {NTK} Analysis of Optimization and Generalization with Its Extensions to Arbitrary Initialization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3vSN5Oumob},\nnote={under review}\n}" }, "abstract": { "value": "Recent theoretical works based on the neural tangent kernel (NTK) have shed light on the optimization and generalization of over-parameterized neural networks, and partially bridge the gap between their practical success and classical learning theory. However, the existing NTK-based analysis has a limitation that the scaling of the initial parameter should decrease with respect to the sample size which is contradictory to the practical initialization scheme. To address this issue, in this paper, we present the revised NTK analysis of optimization and generalization of overparametrized neural networks, which successfully remove the dependency on the sample size of the initialization. Based on our revised analysis, we further extend our theory that allow for arbitrary initialization, not limited to Gaussian initialization. Under our initialization-independent analysis, we propose NTK-based regularizer that can improve the model generalization, thereby illustrating the potential to bridge the theory and practice while also supporting our theory. Our numerical simulations demonstrate that the revised theory indeed can achieve the significantly lower generalization error bound compared to existing error bound. Also importantly, the proposed regularizer also corroborate our theory on the arbitrary initialization with fine-tuning scenario, which takes the first step for NTK theory to be promisingly applied to real-world applications." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "neural tangent kernel", "optimization", "generalization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/e47cb047a5cbf65d99915fa2ee4231803e758a51.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/c2ba95d4b5fbe5f12c3c458f009aaefae4d86c89.zip" }, "title": { "value": "Revised NTK Analysis of Optimization and Generalization with Its Extensions to Arbitrary Initialization" }, "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": 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": "# High-level questions\n1. Definition 3. Why is KLoM called KL-divergence of Margins? I couldn't find the reasoning.\n2. Line 321 and Alg A.1. Assuming that dataset $S$ has distinct datapoints, isn't $S_{finetune}$ the same as $S$, implying that Alg A.1 is simply matching the model's outputs to the oracle? Then I don't understand the novelty of Oracle Matching.\n3. Line 364. Why replace the first term with $\\beta$? Linear $\\beta$ is exactly the definition of the estimator $\\hat{f}$, which is only an approximation to $f_x$. It'd make sense to explicitly state that DM-DIRECT approximately simulates the oracle outputs. You could then argue that even this approximate simulator works well empirically.\n\n# Low-level questions\n1. Line 173. Is this estimator/datamodel only for a single $x$? Does this mean that distinct inputs might require distinct datamodels?\n2. Equation 2. What is the approximation in? Is it in some measure of distributions?\n3. Figures. Why are there multiple points on the plot for each legend entry?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is written extremely well, and has a natural flow to it. When reading I thought of many questions and added annotations, only to find them answered in the next paragraph or section. E.g. the authors introduce Oracle Matching with the strong assumption of oracle access to data attributor $f^{oracle}$ (Section 4.2), and immediately discuss how to simulate such an oracle without such an access (Section 4.3). In Section 5 as well, the paper attends to a natural question readers could have: is oracle matching useful when the problem is easy to solve with gradient descent. This is good writing, and I appreciate the authors' efforts in putting themselves in the readers' shoes.\n\nIn sum, the Oracle Matching and Simulation methods are intuitive yet thoroughly tested on image classification tasks. These methods are original as well, and it is surprising to me that linear datamodels work so well empirically. I'd like to see this paper accepted, and the research directions it inspires." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the problem of _machine unlearning_ (removing the effect of a few training data points \"forget set\" on the model's outputs) with by reducing it to the problem of _data attribution_ (predicting the effect the training set on the model's outputs). With this, the paper proposes a meta-algorithm Datamodel Matching (DMM) that gets predictions from data attribution on all-but-forget set and finetunes the model to match the predictions. A new unlearning metric KL Divergence on Metrics (KLoM) is also introduced. Finally, the paper presents experiments on unlearning in image classification tasks, showing that DMM is better at unlearning and faster than naive-retraining on the all-but-forget-set." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is unclear that linear datamodels extend to other kinds of tasks, e.g. language modeling or regression problems. I believe this to be a major weakness of the paper. While linear datamodels lead to simple algorithms in this paper, the previous work [1] does not have a good argument for why linear datamodels work [1; Section 7.2]---in fact Figure 6 of [1] display imperfect matching using linear datamodels. It'd be useful to mention this limitation in this manuscript as well, and discuss the limitation's impact to machine learning.\n\n# Suggestions:\n1. Line 156. It'd be useful to the reader to add a citation on differential privacy, e.g. one of the standard works like [2].\n2. Line 176. $\\hat{f}$ should have output range in $\\mathbb{R}^k$ since the range of $f_x$ is in $\\mathbb{R}^k$. \n3. Line 182. \"show\" -> \"empirically show\".\n4. Definition 3. Write safe, $S_F$, and input $x$ explicitly in KLoM, otherwise KLoM$(\\mathcal{U})$ looks like KLoM of the unlearning function across _all_ safe functions and inputs. I'm curious why the authors wrote KLoM$(\\mathcal{U})$.\n5. Add a Limitations section.\n\n[1] Ilyas, A., Park, S. M., Engstrom, L., Leclerc, G., & Madry, A. (2022). Datamodels: Predicting predictions from training data. arXiv preprint arXiv:2202.00622.\n[2] Dwork, C., & Roth, A. (2014). The algorithmic foundations of differential privacy. Foundations and Trends® in Theoretical Computer Science, 9(3–4), 211-407." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. This question is regarding the GD baseline. Given a model $\\theta_{full}=A(S)$ trained on a full dataset S (including the retain set $S_R$), GD minimizes the loss on the retain set using gradient descent, starting from $\\theta_{full}$. Since $S_R$ was already in the dataset S that the model was trained on, this essentially involves further training on $S_R$, which could lead to overfitting without effectively forgetting $S_F$. This may explain why, in Figure 3, GD performs similarly to 'Do Nothing' and, on average, even worse. My interpretation is that further training on $S_R$ may not significantly alter a model that was sufficiently trained on S if the loss is strongly convex and could lead to overfitting (and hence degrading performance on the validation) in more complex landscapes. Could the authors clarify if there’s any specific benefit of GD for unlearning that I might be overlooking?\n\n2. The authors mention the possibility of having duplicates across the forget and retain sets on page 6 when discussing the drawbacks of Gradient Ascent. Given that $S_F$ and $S_R$ are sets and $S_R$ is defined as $S \\backslash S_F$, I don't see how this duplication is possible. Could the authors clarify this?\n\n3. Could the authors provide more interpretation of the results in Figure 2? How do you interpret the changes and fluctuations in the red and gray lines?" }, "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. It provides a solid introduction to Unlearning, offering a detailed overview of related work, recent advancements, and existing challenges. The motivation is well-described, and the flow effectively positions this paper within the broader field, helping readers understand its scope and contributions better. The use of data attribution to approximate the Oracle model is particularly compelling." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new Unlearning algorithm called Datamodel Matching (DMM) to unlearn a forget set by fine-tuning a model that has been pre-trained on a larger dataset including the forget and retain set. They use predictive data attribution to approximate the oracle model---the one that is retrained from scratch on the retain set and hence has not seen the forget set at all. Predictive data attribution learns datamodels for each input x to simulate how a model trained on the retain set would behave on x. With this approximation, DMM then applies Oracle Matching to align the model's output distribution with that of the oracle. They introduce KLoM for measuring the unlearning quality and show empirically that their algorithm outperforms previous gradient-based algorithms, achieving a lower KLoM and quickly approaching oracle-level accuracy with only a fraction of the retraining time." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experimental results, while supportive of the algorithm’s effectiveness, are somewhat limited in scope. The analysis focuses primarily on CIFAR-10 and an ImageNet subset (Living-17), with figures 1 and 3 illustrating outcomes only on these datasets. Expanding the experiments to include additional datasets would enhance the generalizability of the findings. On the Retain set, Oracle Matching gets close to the Oracle in CIFAR-10 but this is not the case with Living-17 which means that while the algorithm can reduce KLoM on the forget set, it does not work as well on the retain set. More discussion on this observation can improve the understanding of the algorithm’s limitations, and I'm interested to know how this observation extends to other tasks.\n\n2. The paper could benefit from clearer explanations and interpretations of the figures and baselines. The discussion around interpreting each figure is limited. Additionally, some of the baseline methods are not well-defined; for instance, SCRUB is introduced without a description, and GD (Gradient Descent) is not explained well until page 9 (also see my question on GD in the Questions).\n\n3. Minor typo: The RHS of equation (2) should be dependent on x.\n\n4. Calculating one $\\beta$ vector for each x in the dataset appears computationally intensive. Although the authors mention that this is a one-time process that amortizes over unlearning requests, further discussion on its computational cost relative to the unlearning phase would be helpful." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "line 246 - specify $\\text{safe}(S_F)$ - how do you know this safe(S_F)? How did you calculate this? \n\npage 6 - line 317 - this description is clear but what exactly distinguishes your approach from distillation. The comparison between your method and a distillation unlearning approach such as [Efficient Two-stage Model Retraining for Machine Unlearning](https://openaccess.thecvf.com/content/CVPR2022W/HCIS/papers/Kim_Efficient_Two-Stage_Model_Retraining_for_Machine_Unlearning_CVPRW_2022_paper.pdf)\n\nFigure 3 - For this evaluation, did you retrain the model and then computed the unlearned model using oracle matching?\n\nline 365: in this formulation, to create the datamodels for remaining data, you used the datamodel of whole dataset minus the datamodel of forget set. if my understanding is correct, doesn't it add the additional computation because you need to estimate two datamodels first. \nalso can you prove this formulation result in an exact or good estimation of datamodel for remaining data? \n\nline 374: the true oracle output - I am just worried about the closeness of the proxy to actual retain datamodel to the.\n\nline 404 - , the datamodel generalizes well to new forget sets in practice. very interesting how do you demonstrate this ?\n\nORACLE MATCHING FOR LINEAR MODELS: my understanding is that the quality of unlearning for the oracle matching heavily is influenced by the unlearning quality of approximated oracle model . it would be interesting to investigate the influence of that model." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The proposed approach is both innovative and highly impactful, offering substantial reductions in unlearning time. The authors demonstrate a profound understanding of unlearning methods based on gradient descent and fine-tuning. They delve into the challenges of these methods, particularly examining their impact on the model's performance after unlearning. Although gradient-based approaches are among the most effective unlearning methods, they often degrade the predictive performance of the model considerably—a drawback that the authors have thoroughly investigated and discussed.\n\nThe concept of Oracle Matching, combined with the use of DataModels, greatly reduces time complexity. This approach shows considerable promise for improving the efficiency of unlearning processes without compromising model performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the concept of Oracle Matching to significantly reduce the time complexity of unlearning processes, achieving this in a fraction of the time required for traditional retraining or fine-tuning while maintaining model performance as close as possible to that of full retraining. The authors utilize the concept of \"DataModels\" to efficiently approximate a proxy for the oracle, leveraging this proxy within the Oracle Matching framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Grammatical errors:\n- Abstract - line 2\n- The abstract and introduction was written in rush. please \n\n\nEarly introduction of specific notions:\n- such as $S$, $S_R$ in the beginning of the paper without providing the backgrounds and clearly stating them, confuses the reader. (Intro - second paragraph)\n- line 74 - trained model $\\theta$\n\nAmbiguity:\n- line 49 - Simple models // what is considered to be a simple model?\n- line 59 - variety of empirical evaluations and benchmarks // what are these benchmarks? either needs to mention at least one of evaluation criteria, or rephrase the sentence.\n\nIncorrect Statement:\n- line 63 - , fine-tuning-based methods typically employ.... - incorrect, the simple fine tuning only focus on the remaining datapoints and fine tune the model on $S_R$. If there is a paper that conducts the fine tuning in the way you mentioned in this line, you need to point it out, but otherwise, check this paper [\"Model Sparsity Can Simplify Machine Unlearning\"](https://openreview.net/pdf?id=0jZH883i34)\n\nIncorrect notation:\n- line 152 - the notation for approximate unlearning is the same as exact unlearning.\n\nline 83 - Empirically, we find that .... - I was expecting to see a comparison between your method and unlearning distillation approach, but you didn't made that comparison." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "principled and practical algorithm for unlearning training data by leveraging recent advances in predictive data attribution" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024machine,\ntitle={Machine Unlearning via Simulated Oracle Matching},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3vXpZpOn29},\nnote={under review}\n}" }, "abstract": { "value": "Machine unlearning---efficiently removing the effect of a small \"forget set\" of training data on a pre-trained machine learning model---has recently attracted significant research interest. Despite this interest, however, recent work shows that existing machine unlearning techniques do not hold up to thorough evaluation in non-convex settings. In this work, we introduce a new machine unlearning technique that exhibits strong empirical performance even in such challenging settings. Our starting point is the perspective that the goal of unlearning is to produce a model whose outputs are *statistically indistinguishable* from those of a model re-trained on all but the forget set. This perspective naturally suggests a reduction from the unlearning problem to that of *data attribution, where the goal is to predict the effect of changing the training set on a model's outputs. Thus motivated, we propose the following meta-algorithm, which we call Datamodel Matching (DMM): given a trained model, we (a) use data attribution to *predict* the output of the model if it were re-trained on all but the forget set points; then (b) *fine-tune* the pre-trained model to match these predicted outputs. In a simple convex setting, we show how this approach provably outperforms a variety of iterative unlearning algorithms. Empirically, we use a combination of existing evaluations and a new metric based on the KL-divergence to show that even in non-convex settings, DMM achieves strong unlearning performance relative to existing algorithms. An added benefit of DMM is that it is a meta-algorithm, in the sense that future advances in data attribution translate directly into better unlearning algorithms, pointing to a clear direction for future progress in unlearning." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "machine unlearning", "data attribution", "training data attribution", "privacy" ] }, "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/344c7c5cbf130d5c09e10cafe4fd79478fd29985.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": "Machine Unlearning via Simulated Oracle Matching" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "1. What does it mean by justified approach through the lens of probability? (ln.76). Please be more concrete.\n2. Please fix formatting errors in the related work section, the dataset section, section 5.1, etc. \n3. \"Given new inputs, ... refining out candidate sets and improve LLM's accuracy for subsequent generation.\" (ln. 178-179). Just to clarify, is this referring new inputs from D_small or D_large? Does that mean after each sample, you need to know the correctness of the produced output?\n4. line 209 ChatGPT typo. \n5. line 231 incomplete sentence. \n6. What is P? Are you using P as a random variable, a model or a notation for probability? What do you mean \"Assume P is the LLM\" (ln. 232)?\n7. Is X a single feature or a feature class? or a variable? It is sometimes referred to as a features and other times as features (Section 4.1), but in many equations X is referred to as a variable (e.g., X=x_1 in Eq. 5).\n8. Section 4.3: What is K, what is a category? When is the concept of category introduced? What is the difference between a category and a label? Is Y a category or a label?\n9. Undefined symbols: What is $K$, $c$, $\\vec{Y}_G$, $q$, $\\vec{y}$?\n10. Is $Y_i'$ a vector or a value? It is referred to as a vector in lines 305-306 but a value 1 in line 313. Additionally, could you clarify what does it mean by \"If $Y_i'=1$, the corresponding potential candidate set is the first row of $\\vec{Y}_{\\text{Updated}}$\"? Does the value of $Y_i'$ correspond to \"first row\" of the matrix? Why?\n11. \"i.i.d. sample from the large size dataset\" (ln.321) --> do you mean uniformly sample?\n12. Table 1. Bolding is mentioned in the caption but no values are bolded in the table.\n13. \"The problem with this method is its dependence on multiple sources of paths; even slight changes in one source’s prediction can drastically impact the final prediction.\" (ln. 374-375). Why is this the case? Self-consistency takes the mode of multiple reasoning paths. Furthermore, the \"using a single query minimises the uncertainty\" (ln. 371-372) is not well-justified.\n14. All of the figures and Table 1 are not referred to in the text of the paper." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The main idea of the paper, which states that memorizing false positive samples would help prediction performance, is interesting and can be a good direction for improving test-time techniques to improve model performance. The datasets and baselines examined in the experiments section is fairly comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The main motivation behind this paper is memorizing past mistakes would make LLMs not make the same mistakes again. The authors propose Memorizable Prompting (MP), which allows LLMs to understand response dependence patterns and store them in a memory bank to prevent repeating false positive predictions. The memory bank is constructed using a small labeled set of samples." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The justification for the Memorization masking assumption is confusing (line 225-234). Please lay out the specific constraints assumed made by this claim, and deduce the final assumption. If I understand correctly, this assumption is claiming that assuming if the model has learned the characteristics of a category well enough, then it should reliably predict the same Y' distribution for each Y regardless of the input feature X. The current justification is not clearly or rigorously written and need further editing.\n\nSection 4 is hard to follow and not rigorously written. X and Y are sometimes referred to as variables and some other times as values/examples. For example, equation (4) does not \"simplify\" to equation (5) when the given assumption (ln. 258) is Y=1 for a given X=x1; this is only correct if the given assumption is X=x1 and Y=1. \n\nSection 4.3. It is not clear how M is used in the transformation from $\\vec{Y}$_Query to $\\vec{Y}$_True. Additionally, \"Our goal is to design a prompting scheme to enable LLMs to generate the correct annotation for each q from the corresponding $\\vec{y}$\" (line 303-304). What exactly is the prompting scheme? It is not discussed or included in the appendix.\n\nOverall, there are fundamental flaws in the probabilistic framework and problem formulation. The core mathematical error lies in the proposed marginalization over Y: $P(Y'|X,\\vec{Y}) = \\sum_Y P(Y'|Y,X,\\vec{Y})P(Y|X,\\vec{Y})$. This equation is fundamentally incorrect as Y represents a ground truth label, not a random variable. The main objective to \"obtain\" $P(Y|X,\\vec{Y})$ through $P(Y|Y',X,\\vec{Y})$ shows a fundamental misunderstanding of the causal relationship in supervised learning. THe prediction Y' should not influence the probability of the true label Y. The learning objective should instead be formulated as maximizing the probability of correct classification, $\\arg\\max_G P(Y=y^*|X,Y)$. These issues reflect a fundamental misunderstanding of probabilistic modeling, making the proposed method mathematically unsound. While the research direction may be promising, the current formulation requires substantial revision to establish a valid theoretical foundation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Refer to Weaknesses" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The motivation of this paper is explicitly to make the LLMs remember the dependencies of responses and prevent making false positive answers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The entire paper is more of a technical report in that it improves upon the prompt strategy, which aims to make the LLMs remember the dependencies of responses and prevent making false positive answers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The entire paper is more of a technical report in that it improves upon the prompt strategy; however, it is difficult to read the entire paper. Even after reading the whole paper, I don't even know what kind of prompt form or few-shot form the authors employ.\n\nThe motivation of this paper is explicitly to make the LLMs remember the dependencies of responses and prevent making false positive answers. However, there are many fundamental errors in both model construction and experimental setups.\n\nIn method construction:\n\n1. The paper is very redundant to introduce the basic definition, and there is a lack of formal definition of basic memorization matrix in preliminaries. \n\n2. The introduction of the method is not clear, I think only two points need to be introduced clearly:\n\n2.1. How is the memorization matrix constructed between the answer and the truth to each query when the truth cannot be obtained during reasoning? How many times does each query need to be answered, and how is the sparsity of the matrix addressed when there are too many categories?\n\n2.2. Does the LLM actually capture dependencies between responses from the matrix? How does the LLM remember and avoid false positive responses? Are there any experiments to prove it?\n\n3. In Section 4.0, the paper introduces the learning conditions, it is very confusing how the sample features are introduced in the method, for example in Eq.3, how to find completely different features that can make the LLM generate the same prediction? I don't think any specific distinct feature alone can do that.\n\n4. Is it using the memorization matrix of few-shot hint samples to predict samples with no truth labels? What is the number of hint samples for each sample? **How to determine the LLM is not mimicking the memory matrix, but actually recording the response to each query.**\n\nIn related work:\nThe paper does not fully investigate the related work and has no logic in expression. The paper is based on LLM reasoning and constructs a strategy to improve prompt.\n\nIn the experiments:\n\n\n1. The paper **is lack of experimental setup**, and prompts should be introduced in detail, otherwise it is difficult to reproduce the work.\n\n2. There is too little experimental analysis in the paper, and the experimental Tables and Figures are not clearly introduced. Besides, the paper adopts three LLMs. Is there no difference in performance among the three LLMs? Meanwhile, the paper lacks interpretability experiments to validate the motivation.\n\n3. References to Figures1, 2 and 3 and Tables 1 and 2 are absent from the main paper. There are also several methods that are not detailed in baselines. What is used for correction? What is fot in the Tables? **I think the paper should at least be standardized and clear.**\n\n**There are multiple errors in the paper**, multiple quotes, punctuation, basic grammar errors, **the readability of the paper is poor**, I think the author should take the paper submission seriously." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 1 }, "primary_area": null, "questions": { "value": "Here is a revised version of your questions for the article:\n\n- Line 120: Why not simply use $(X, Y, \\mathcal{Y})$ or $(X, Y)$ here?\n\n- Line 132: Is the concept of a \"transition matrix\" derived from Markov models? If so, each row in M should sum to 1. If not, consider using a different term or providing clarification to prevent potential confusion.\n\n- If I understand correctly, Section 4 (Lines 184--241) essentially states that the smaller dataset has a similar distribution to the larger dataset, and thus the matrix M estimated from the smaller set can be applied to the larger set under certain assumptions. Why is such an in-depth discussion on conditional terms necessary if they are simply assumed valid? It seems the numbered equations and many unnumbered ones are not essential, as this is standard machine learning knowledge familiar to the intended audience.\n\n- Line 276: Certain GPT API versions can generate token probabilities alongside tokens. Therefore, the reasoning provided here is not strong enough to justify the innovation in this paper. Additionally, relevant studies should be cited in the related works.\n\n- Line 288: What is the difference between $D_{\\text{hint}}$ and $D_{\\text{small}}$ mentioned earlier? How many data points were used to compute matrix M in the experiments? It is mentioned later that $s=4$. If it accounts for 5% of the total data points, the total number of data points for each dataset should be around 80, which does not match the number in Table 2." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The strength lies in its approach to improving LLMs' reliability by addressing false positives through Memorisable Prompting (MP), which adds a layer of memory to the model's prompting strategy. This technique enhances model consistency by leveraging a memory bank that \"remembers\" past errors, allowing the LLM to self-correct based on learned dependencies. The approach is versatile, effectively improving accuracy across different domains and integrating well with existing prompting methods. Furthermore, the use of a memory masking matrix offers a structured way to manage and apply learned error patterns, adding a new dimension to LLM error management in a practical, adaptable manner." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a method to improve the accuracy and consistency of Large Language Models (LLMs) by enabling them to \"remember\" past errors. It uses small annotated datasets, or hint samples, to learn dependencies between predictions and actual labels, storing this knowledge in a memory bank to help LLMs avoid repeating false positives. By applying a memory masking matrix, MP enhances prediction accuracy across domains and integrates well with various prompting techniques, making it effective for error-prone, high-stakes applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The article is difficult to follow due to inconsistent terminology, unclear mathematical notation, grammatical errors, structural issues, and logical gaps, which detract from its clarity and coherence. Overall, the writing does not reflect careful attention to readability and precision.\n\n- The proposed approach appears limited in applicability, focusing primarily on classification tasks. However, this limitation is not addressed in the paper, leading to an incomplete understanding of the method's scope.\n\n- The narrative includes redundant explanations and lacks depth, limiting the reader’s engagement with the insights behind the proposed approach.\n\n- Baseline comparisons are insufficient; it remains unclear why traditional weak supervision methods, such as [1], combined with pre-trained language models were not included for a more comprehensive evaluation of the method.\n\n- The proposed approach shows minimal innovation compared to prior work using transition matrices for prediction calibration, which reduces the originality and significance of the contribution.\n\n[1] Ren, Wendi, et al. \"Denoising multi-source weak supervision for neural text classification.\" arXiv preprint arXiv:2010.04582 (2020)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024memorisable,\ntitle={Memorisable Prompting: Preventing {LLM}s Forgetting False Positive Alarm},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3viQDuclu0},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) are widely recognized for their superior performance across various domains. However, their tendency to generate inaccurate or misleading responses presents significant challenges, particularly in the natural language domain. This issue underscores the need to enhance both the explainability and reliability of LLMs. While recent advancements in prompting have focused on leveraging in-context learning—such as providing step-by-step explanations—these approaches often overlook the critical importance of understanding the response dependency of LLMs on specific datasets. This understanding is crucial for interpreting their outputs and improving their consistency. Moreover, if we can capture and encode these response dependencies, we can integrate them into LLMs as memorized knowledge to mitigate false positive predictions over time. In this paper, we tackle this challenge by introducing the Memorizable Prompting (MP) paradigm, which enables LLMs to retain and utilize information from past responses. Specifically, our approach leverages hint samples—a small set of annotated examples—to learn the response dependencies, defined as the relationship between LLM outputs and the ground-truth annotations for a given dataset. This equips LLMs with the ability to recall past false positives and use that knowledge for self-correction in future predictions. We have evaluated our method on a diverse set of domain-specific datasets, demonstrating its effectiveness across large-scale 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": [ "Prompt-based task", "Large language model", "Memorisable Prompting for Data Annotation" ] }, "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/af5cec55857e0a7ea1b04d7e94efbcc218e2bfa3.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": "Memorisable Prompting: Preventing LLMs Forgetting False Positive Alarm" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Considering CLIP's training process, is there a significant distinction in the representations output by CLIP when tracking-state-aware prompts, such as 'unoccluded' and 'occluded', are provided?\n2. Why does your method improve the ClA for novel categories compared to the baseline OVTrack but decrease scores for base categories? Does this suggest that your model introduces a conflict between tracking and classification? \n3. Could you include a visual representation and analysis of the prompt guided attention? This would more intuitively demonstrate its role and effect." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper offers a detailed explanation of the proposed method. Utilizing unlabeled data for self-supervision serves, to some extent, as an alternative to address the current scarcity of large vocabulary tracking data. \n2. The results show good performance in comparisons on the OVMOT benchmark." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Tracking-state-aware prompt guided attention, enabling the network to learn the detection of objects in different tracking states. A self-supervised approach is adopted to train tracking, leveraging large-scale, unlabeled video data across various categories. The experimental results on TAO datasets indicate that the proposed method achieves advanced performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The CIsA for basic categories in your method is lower than that of the baseline method OVTrack. Given that basic categories account for the majority of targets, the decrease on CIsA appears to reflect more than just typical fluctuation effects. Could this imply that your model's approach introduces a degree of conflict between tracking and classification based on the baseline? \n2. There is a lack of visualization and analysis for the prompt guided attention, which does not adequately demonstrate its direct impact." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 refer to the concerns and issues raised in the \"Weaknesses\"." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This manuscript is standardized and the writing is fluent, and the content is easy to understand.\n\n2. This manuscript proposes a new tracking-related prompt-guided attention for the localization and classification (detection) in the open vocabulary tracking problem. This method takes notice of the states of the time-varying objects during tracking, which is different from the open-\nvocabulary object detection from a single image.\n\n3. This manuscript develops a self-supervised object similarity learning strategy for the temporal association (tracking) in the OVMOT problem. This strategy, for the first time, makes full use of the raw video data without annotation for OVMOT training, thus addressing the problem of training data shortage and eliminating the heavy burden of annotation of OVMOT.\n\n4. Experimental results on the public benchmark demonstrate that the proposed VOVTrack achieves the best performance with the same training dataset (annotations), and comparable performance with the methods using a large dataset (CC3M) for training." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Open-vocabulary multi-object tracking (OVMOT), a significant challenge that involves detecting and tracking various object categories in videos, including both known (base classes) and unknown (novel classes) categories. The authors critique existing OVMOT methods for treating open-vocabulary object detection (OVD) and multi-object tracking (MOT) as separate modules, primarily focusing on image-based approaches. To address this, they present VOVTrack, which integrates object states relevant to MOT with video-centric training, approaching the challenge from a video object tracking perspective. VOVTrack features a prompt-guided attention mechanism that enhances the localization and classification of dynamic objects, and it employs a self-supervised object similarity learning technique for tracking using raw, unlabeled video data. Experimental results demonstrate that VOVTrack outperforms current methods, establishing it as a leading solution for open-vocabulary tracking tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experimental tables lack details on model complexity. It would be helpful to include a table or section comparing FLOPs, parameters, model size, and FPS across the different methods evaluated, including the baseline OVTrack method and other state-of-the-art approaches." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Some important details and explanations should be provided for clarity." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "A novel method that integrates object states based on prompt learning is proposed to combine OVD and MOT for open-vocabulary multi-object tracking.\n\nSome self-supervised losses are designed to learning better object associations.\n\nExperiments are conducted on public dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents a novel open-vocabulary multi-object tracking method that integrates object states based on prompt learning. Different from existing works, it combine OVD and MOT in a unified framework. Some self-supervised losses are designed to learning better object associations. Experiments on public dataset demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The following details are unclear. Are the designed prompts only used in training procedure? It would be better if visualized results are provided to validate the effectiveness of these prompts in handling challenging frames with occlusions or motion blur.\n\nThere is no comparison with works published in 2024, and the effectiveness of the proposed method is thus not fully validated. The relevent and recent trackers including both closed-set and open-vocabulary ones should be included in comparison.\n\nThe results suggest that the proposed method performs abnormal under ClsA metric in both result comparison and ablation study. Though it is better than all methods in most metrics, but it performs worse than OVTrack and OVTrack+RegionCLIP in a clear margin, and also worse than other methods in some cases. The reasons behind these results are not well explained.\n\nSome errors:\n\nLine 256: into ore framework model\n\nLine 456: Our method" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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 primary assertion of the paper is that existing OV tracking methodologies do not take into account the states of objects. However, the proposed approaches appear to be largely unrelated to tracking object states, and the introduced prompt, .e.g., occuded, and complete, seems overly naive in my view.\n2. The proposed classification method only focuses on the high-quality object. However, exclusively training the classifier on high-quality targets may result in the neglect of low-quality targets that are blurred or occluded. This seems inconsistent with the paper’s claim of addressing issues related to blurring and occlusion in object tracking. Furthermore, could this be the reason for the model’s lower classification performance on the base?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Competitive OV tracking performance on TAO.\n2. Self-supervised object similarity learning is compatible with unlabeled video data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes a series of improvements to the existing OV MOT framework (OVTrack), including a novel prompt-guided attention mechanism and a self-supervised object similarity learning method. With the support of additional video data, these enhancements achieve superior performance on the TAO dataset compared to OVTrack." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Novelty is limited: The overall architecture of the proposed OVOTrack is based on OVTrack. The specific implementation within the tracking-state aware prompt approach does not exhibit a clear association with the tracking state. Moreover, this method does not assess object quality specifically for tracking scenarios, as it remains overly limited to factors such as occlusion and blurriness. On the other hand, the self-supervised object similarity learning proposed in this work closely resembles QDTrack, with no significant differences observed.\n2. Writing requires further improvement: The overall writing is somewhat scattered and lacks coherence, with inconsistencies between the motivation of the paper and the proposed approach, raising concerns of possible over-claim. There is considerable redundancy in language, with some sections being overly simplistic while others are unnecessarily verbose. Additionally, certain assumptions and notations in the methodology are not rigorously standardized (e.g., line 218). Therefore, I believe there is substantial room for improvement in the writing. \n3. Experimental setting: The training process described in the paper involves four stages, which makes it overly complex and cumbersome. Compared to OVTrack, this work utilizes a substantial amount of additional TAO video data for training; however, the classification performance on the base categories in the TAO benchmark has declined. Moreover, the ablation study results indicate that the OVOTrack model consistently underperforms in classification metrics, yet no reasonable explanation is provided for this issue. The examples provided in the visualization results are overly simplistic, failing to showcase the model’s true capabilities." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. VOVTrack lacks of a clear definition, what's the full name?\n2. For Table 1, the propsed VOVtrack performs worse than OVTrack in terms of ClsA metric. What is the reason behind it?\n3. For Table 2, it demonstrate the effectiveness of Prompt-guided attention. But the baseline with the self-supervised association module fail to obtain obvious improvements, so that the contribution can not be effectively convinced." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper propose a new tracking-related prompt-guided attention for the localization and classification in OVMOT​.\n2. The self-supervised learning strategy leverages unlabeled video data for the temporal association, addresses the challenge of training data shortage.\n3. Extensive experimental results, demonstrating that VOVTracker outperforms existing methods on TAO benchmark." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a new tracking-related prompt-guided attention for the localization and classification, and develops a self-supervised object similarity learning strategy for the temporal association in OVMOT. Experiments demonstrate that the proposed method achieves state-of-the-art tracking performance compared to previous methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of novelty. The self-supervision is common and there are some considerable work in vision tasks. However, this paper just applies it to OVMOT, and does not compare the proposed method with previous methods. What's the main difference and contributions for OVMOT.\n2. While VOVtrack leverages unlabeled video data, its performance seems to depend on the quality and representativeness of the videos, which could affect its robustness across different real-world conditions​.\n2. The authors should conduct on the ablation sduty of different parameter settings. Such as parameters related to training and inference." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024vovtrack,\ntitle={{VOVT}rack: Exploring the Potentiality in Videos for Open-Vocabulary Object Tracking},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3vxfFFP3q5},\nnote={under review}\n}" }, "abstract": { "value": "Open-vocabulary multi-object tracking (OVMOT) represents a critical new challenge involving the detection and tracking of diverse object categories in videos, encompassing both seen categories (base classes) and unseen categories (novel classes). This issue amalgamates the complexities of open-vocabulary object detection (OVD) and multi-object tracking (MOT). Existing approaches to OVMOT often merge OVD and MOT methodologies as separate modules, predominantly focusing on the problem through an image-centric lens. In this paper, we propose OVTracker, a novel method that integrates object states relevant to MOT and video-centric training to address this challenge from a video object tracking standpoint. First, we consider the tracking-related state of the objects during tracking and propose a new prompt-guided attention mechanism for more accurate localization and classification (detection) of the time-varying objects. Subsequently,\nwe leverage raw video data without annotations by formulating a self-supervised object similarity learning technique to facilitate temporal object association (tracking). Experimental results underscore that OVTracker outperforms existing methods, establishing itself as a state-of-the-art solution for open-vocabulary tracking 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": [ "Object Tracking", "Open-Vocabulary" ] }, "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/b3790ab36130cf84b397cc30608e6e3e3e271afa.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/093f817c88c4c3d3606970dd055647c0d620bfd0.zip" }, "title": { "value": "VOVTrack: Exploring the Potentiality in Videos for Open-Vocabulary Object Tracking" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The experimental settings are somewhat vague, particularly regarding whether competitive approaches, such as FedAvg and FedProx, are trained from scratch or from the same pretrained backbone model as the MPFT. Could the authors specify this information in the paper, ideally in the experimental setup or implementation details section?\n\n2. The privacy protection measures appear informal. While Gaussian noise is added to prototypes (i.e., embeddings) as noted in Section 5.6, it is important to assess whether the noise level provides adequate privacy protection for sensitive attributes or labels. To strengthen this section, could the authors provide a more rigorous analysis of the privacy guarantees associated with their approach? For example, quantifying the level of privacy protection using established metrics or frameworks would be beneficial." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1, Experiments on two datasets indicate a significant performance boost from the proposal, increasing accuracy by 2%.\n\n2. The approach is intuitive, sharing prototypes using three sampling strategies.\n\n3. The proposal is communication-computationally efficient, as it involves a low number of interaction rounds and requires only prototypes for communication." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work studies domain adaptation in federated learning scenarios, employing prototype-based fine-tuning to leverage knowledge from other clients. The fine-tuning procedure requires only one round of communication between clients and the server, making it communication-computationally efficient." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The motivating scenario may be impractical, as the proposal assumes that clients already possess a well-pretrained model. This assumption can be particularly challenging in sensitive areas where federated learning is essential, such as healthcare or finance, where data privacy concerns limit access to robust pre-trained models. \n\nCan the authors discuss specific applications or scenarios within these sensitive domains where obtaining a pre-trained model might be more feasible? Additionally, it would be valuable to explore potential solutions or adaptations of their method for situations where a pre-trained model is not available." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "I suggest that the authors remove the overly powerful pre-trained feature extractor and instead use a simpler, more foundational model for feature extraction, such as ResNet or ConvNet. This would more accurately showcase the contribution of the prototype and methodology itself.\n\nThe authors should provide a validity proof to demonstrate that training the adapter on aggregated prototypes achieves comparable performance to training on aggregated raw client data or features—or, alternatively, establish an upper bound on the performance gap between these approaches.\n\nAlthough differential privacy (DP) experiments were conducted, a theoretical analysis is also necessary. Specifically, the authors should clarify the noise variance conditions required for their method to satisfy DP. In particular, the statement “Furthermore, we observe that specific noise configurations can reduce bias across heterogeneous datasets, enhancing the robustness of prototype data” requires a more robust theoretical explanation.\n\nI would encourage the authors to further explain, both theoretically and experimentally, the differences, advantages, and potential complementarities among various prototype selection mechanisms." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The main strength of the paper lies in its comprehensive experimentation, which provides a detailed view of MPFT’s performance across various scenarios, showing superior computational and communication efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes Multi-domain Prototype-based Federated Fine-Tuning (MPFT), a framework designed for Federated Domain Adaptation (FDA) by utilizing domain-specific prototypes. Instead of relying on traditional model aggregation techniques, which often falter due to data heterogeneity, MPFT transfers prototypes (compressed domain representations) to the server. This approach allows the server to learn a global adapter that improves both in-domain and out-of-domain performance. MPFT also incorporates differential privacy to protect prototypes from potential data leakage." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The Multi-domain Prototype-based Federated Fine-Tuning (MPFT) method proposed in this paper is overly simplistic, primarily relying on basic prototype sampling strategies—such as mean, cluster, and random sampling—to generate client prototypes. Additionally, because the model does not undergo training on each client’s data, it fundamentally contradicts the original design principles of federated learning (FL). The performance gains observed in the experiments likely stem from the strong pre-trained feature extractor, which does not require training on all client data but only a few representative prototypes to achieve reasonable performance. If the pre-trained feature extractor were removed and the model was trained solely on prototypes, the results would likely be poor. If only prototype fine-tuning is needed, then why even employ a federated learning framework? Thus, despite achieving certain results, this approach introduces no complex mechanisms or innovative architectures to enhance the FL system, lacking fresh or original design. It merely follows a standard process: feature extraction with a general pre-trained model, prototype selection, server training, and client fine-tuning, without presenting any real novelty.\n\nThe paper does not provide a thorough theoretical analysis of the different prototype sampling methods, nor does it uncover the fundamental differences between these sampling strategies in handling heterogeneous data distributions. Although the paper presents experimental results comparing mean, cluster, and random sampling methods, it lacks detailed explanations on the core differences among these strategies in terms of learning mechanisms, communication efficiency, and personalization effects. Consequently, the paper falls short in theoretical depth, failing to provide any deep insights into the impacts of these sampling strategies.\n\nThe theoretical analysis in this paper primarily focuses on convergence. However, training the adapter on aggregated prototype data is no different from centralized data training, so convergence is naturally expected. To provide valuable theoretical insights, the paper would need to show that the adapter’s performance on aggregated prototypes is comparable to that on aggregated raw client data or offer theoretical performance bounds. Furthermore, transmitting prototypes instead of models introduces a higher risk of privacy leakage, so the analysis should prioritize differential privacy (DP) guarantees rather than focusing solely on convergence. The current paper fails to address theoretical guarantees for privacy protection, which is particularly critical in privacy-sensitive federated learning contexts.\n\nAlthough the experimental section is comprehensive, covering multiple datasets, various sampling strategies, and different DP configurations, it does not clearly reveal the functional differences and applicability of each prototype sampling approach across different scenarios. While the experiments test different prototype configurations, they lack an in-depth discussion on how these configurations perform under different data distribution patterns and sample sparsity conditions. This results in a set of experiments that, while extensive, fails to provide a robust summary of the unique characteristics of the prototype methods, thus missing an opportunity to elevate the research value." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Would the authors explain how you chose the parameter $s$ in Section 5.6? \n\n2. Why is the class $k$ missing in the proof in Appendix B (while formulated before proof)?\n\n3. Can authors make more comparisons with recent (>=2023) works? For example,\nFedcp: Separating feature information for personalized federated learning via conditional policy. In KDD 2023.\nFedFed: Feature Distillation against Data Heterogeneity in Federated Learning. In NeurIPS 2023.\nFedgh: Heterogeneous federated learning with generalized global header. In MM 2023.\n\nIt would be better to compare aggregation-based improvements in recent works and the new mechanism proposed in this paper. Since the authors stressed the drawbacks of these works, it would be better to give detailed support or explanation." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ They propose an interesting idea -- MPFT as a one-round federated fine-tuning approach for multi-domain environments, demonstrating its performance improved over previous methods. \n+ A new metric is introduced to evaluate model adaptability, assessing both out-of-domain and in-domain accuracy to balance knowledge retention and domain adaptation.\n+ The writing and logic are clear and easy to follow. The problem formulation is clear and friendly to readers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Multi-domain Prototype-based Federated Fine-Tuning (MPFT), a framework designed to address data heterogeneity and data privacy in federated learning (FL) by using prototype-based training rather than traditional averaging methods. In MPFT, each client generates a set of representative data embeddings (prototypes) that capture essential domain-specific characteristics without transferring raw data. These prototypes are then aggregated at the server, allowing for a simulated centralized learning approach and enabling fine-tuning of a global adapter. This method aims to achieve performance on par with centralized learning while solving the challenges incurred by aggregating client models.\nThe efficiency of MPFT lies in that it requires only a single round of global communication, significantly reducing computational and communication costs compared to multi-round FL methods. Furthermore, by selectively sampling prototypes, the framework limits data transfer volumes. To ensure privacy, MPFT integrates differential privacy mechanisms, mitigating risks of data exposure and rendering prototype-based data reconstruction ineffective—even when the prototype encoder is known to potential attackers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proof of convergence in Section B lacks the differential-privacy-related analysis. A supplemental analysis of \"MPFT with DP\" should be added for theoretical completeness. Intuitively, the addition of differential privacy introduces a bounded randomness during the convergence process.\n\n- The selected experimental counterparts for comparison are mostly in 2017-2022, which are relatively not new. In Section 2, the authors think the \"averaging-based aggregation results in poor out-of-domain adaptation performance.\" It would be better to show the performance of out-of-domain adaptation from the recent advances atop model/parameter aggregation. Some recent works (with or without aggregation) relevant to data heterogeneity in federated learning are expected to compare, such as,\n\nFedcp: Separating feature information for personalized federated learning via conditional policy. In KDD 2023.\nFedFed: Feature Distillation against Data Heterogeneity in Federated Learning. In NeurIPS 2023.\nFedgh: Heterogeneous federated learning with generalized global header. In MM 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024enhancing,\ntitle={Enhancing Federated Domain Adaptation with Multi-Domain Prototype-Based Federated Fine-Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3wEGdrV5Cb},\nnote={under review}\n}" }, "abstract": { "value": "Federated Domain Adaptation (FDA) is a Federated Learning (FL) scenario where models are trained across multiple clients with unique data domains but a shared category space, without transmitting private data. The primary challenge in FDA is data heterogeneity, which causes significant divergences in gradient updates when using conventional averaging-based aggregation methods, reducing the efficacy of the global model. This further undermines both in-domain and out-of-domain performance (within the same federated system but outside the local client), which is critical in certain business applications. To address this, we propose a novel framework called \\textbf{M}ulti-domain \\textbf{P}rototype-based \\textbf{F}ederated Fine-\\textbf{T}uning (MPFT). MPFT fine-tunes a pre-trained model using multi-domain prototypes, i.e., several pretrained representations enriched with domain-specific information from category-specific local data. This enables supervised learning on the server to create a globally optimized adapter that is subsequently distributed to local clients, without the intrusion of data privacy. Empirical results show that MPFT significantly improves both in-domain and out-of-domain accuracy over conventional methods, enhancing knowledge preservation and adaptation in FDA. Notably, MPFT achieves convergence within a single communication round, greatly reducing computation and communication costs. To ensure privacy, MPFT applies differential privacy to protect the prototypes. Additionally, we develop a prototype-based feature space hijacking attack to evaluate robustness, confirming that raw data samples remain unrecoverable even after extensive training epochs. The complete implementation of MPFL is available at \\url{https://anonymous.4open.science/r/DomainFL/}." }, "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; Federated Domain Adaptation; Federated Fine-Tuning" ] }, "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/369006c3cf03bed575c9a067c733ba3efe1b3ae6.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": "Enhancing Federated Domain Adaptation with Multi-Domain Prototype-Based Federated Fine-Tuning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "- the last part of Definition 3.5 (\"Otherwise, ...\") seems to be repeated.\n\n- the abstract on OpenReview does not match the PDF abstract; could the authors clarify?\n\n- could the authors explain how worst-group accuracy is computed and how this leads to the conclusion that \"current language models struggle to properly utilize given premises for reasoning\"?\n\n- could the authors provide more detail on the different prompting strategies used in Section 4.1?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- the paper is well-written and includes extensive results across a diverse range of datasets and LLMs.\n\n- the analysis of language modelling bias and the language-thought gap offers an interesting perspective within the LLM community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper analyzes language modelling bias and the language-thought gap in the context of LLMs. It proposes a new prompting technique, LoT, to address these issues and evaluates its effectiveness across two bias benchmarks and eight general reasoning benchmarks, using six different LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- the construction of LoT does not appear fully derived from the analysis of language modelling bias and the language-thought gap (which contribute to a big part of this work). In other words, it seems that LoT is not a natural product/derivation from the bias analysis, e.g., how does LoT overcome the issue that \"one piece of information can have different expressions in language\"? LoT’s design choices, such as \"expanding thoughts,\" seem empirically beneficial but not systematically motivated by the initial analysis. Additionally, the claim in Section 5.2 that \"the expansion prompt may exacerbate language modelling biases\" potentially undermines the rationale behind this feature.\n\n- the performance of the \"echo only\" prompt, which frequently outperforms LoT as shown in Table 3, highlights the need for a deeper understanding of LoT's effectiveness (relating to the above concern). Although the work presents comprehensive testing of LoT across datasets and LLMs, readers would benefit from insights into the underlying mechanisms that make this prompting design work—or fail.\n\n- in Figure 5, the range of benchmarks and LLMs is commendable, but it is unclear why direct prompting is included in the comparison rather than other CoT-like techniques, which might provide a more meaningful comparison to LoT. Given that this study doesn’t focus on the benefit of having internal steps before LLMs generate outputs, the inclusion of direct prompting lacks relevance.\n\n- in cases where LoT performs worse than ablations or other baselines on general reasoning benchmarks, the authors provide conjectures to explain this. However, these conjectures are not substantiated by evidence (e.g., lines 428-431, 517-519).\n\n\n- the reference to the previously established CoT paradigm used in the experiments is missing.\n- the three types of bias used in the evaluation should be explicitly introduced in the main text rather than solely in the appendix.\n- providing practical examples of two-premise QA would help illustrate the generalizability of the training corpus used here and ground the analysis of the language-thought gap later in the main text.\n- in Figure 2, the arrows are stated to represent causal relations, yet it appears the arrows in the blue box denote topological order.\n- while formal propositions are generally beneficial for clarity, propositions 3.3 and 3.6 seem redundant as their informal explanations suffice, and they are not referenced further in the main text.\n- in Section 3.2, it is unclear how prompting LLMs to \"notice more details\" addresses the problem of \"ignorance of implicit premises.\" This approach may not necessarily lead to locating additional implicit premises, raising questions about the effectiveness of the LoT design.\n- the font size in Figure 5 is very small, impacting readability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- While somewhat inline with the paper's general theme, I think it should be noted in line 44 that the large language models using CoT merely strive to simulate system 2 way of thinking via the continuous application of their inherently system 1 capabilities and their method is not a true system 2 reasoning process.\n- I think a better explanation of equation 2 would be good to have. If I am not mistaken, this highlights the fact that if topological ordering of reasoning is not preserved by the language, the LLM will only learn a causal relationship between $L_1$ and $L_A$, relegating $C_2$ to a distributional shortcut. It would be adequate to briefly explain the corresponding behavior in the text as well.\n- In comparison to CoT, what do you think are the main differences of LoT that lead to its superior performance? \n- Regarding the extraction of implicit information as well as the exploration of various reasoning paths, the following papers may be of interest: \n\t- [ 1 ] aim to get better results via the exploration of contrastive reasoning paths in CoT. \n\t- [ 2 ] discuss various prompting methods including the distillation of explicit knowledge from implicit context or the utilization of implicit knowledge.\n\t- [ 3 ] aim to ameliorate affirmative bias of CoT via the exploration of counter-paths for each LLM response. Their overall bias and GPT-4 results are also consistent with those of this paper.\n- Consider writing the formula for the \"Bias Score\" rather than its natural language explanation for better readability.\n- Line 295, \"that relevant\" -> \"that is relevant\"\n- Line 359, \"dive the data\" -> \"divide the data\"?\n- Line 468, \" since the original evaluation results consider correct formats in the incorrect formats to be incorrect answers.\" -> \"Consider correct answers in the incorrect formats\"?\n- There is a discrepancy between the prompt name in the submitted version and the TLDR version, I believe CaT should be changed to LoT or vice versa.\n\n[ 1 ] Chia, Yew Ken, et al. “Contrastive Chain-of-Thought Prompting.” _arXiv.Org_, 15 Nov. 2023\n\n[ 2 ] Yu, Fei, et al. “Natural Language Reasoning, A Survey.” _arXiv.Org_, 26 Mar. 2023\n\n[ 3 ] Miandoab, Kaveh Eskandari, and Vasanth Sarathy. “‘Let’s Argue Both Sides’: Argument Generation Can Force Small Models to Utilize Previously Inaccessible Reasoning Capabilities.” _arXiv.Org_, 16 Oct. 2024" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The general concern of the paper is both valid and very interesting to explore in the literature. It is indeed true that given the autoregressive nature of Large Language Models, one can expect their inherent modeling biases to \"leak\" into the reasoning process, contaminating the possible results.\n- The paper is well-written and well-presented, with good formulation of assumptions as well as examples for the easier understanding of the readers.\n- The proposed method is relatively simple and effective, making its use-case accessible to users and across different models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper establishes the inherent gap between the language of communication, and the language of thought, showcasing that functionally equivalent language representation of a thought process might lead to biased behavior in Large Language Models during both training and inference due to their autoregressive objective definition.\nTo mitigate this shortcoming, the paper introduces LoT, a prompting method that aims to force the Language Model to echo and expand the facts contained in the input, converting the potentially unusable implicit knowledge, to usable explicit context. \nEvaluation of the above hypothesis shows that LLMs, when prompted via LoT, showcase superior behavior with respect to model fairness and bias, while gaining an overall boost in reasoning performance. \n\nOverall, I find the proposed prompting method interesting and potentially effective under reasoning and fairness intensive tasks. However, a number of concerns such as the applicability of method and its comprehensive analysis still remain as outlined below." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Despite the success of CoT-like prompting methods in the simulation of system 2 processes, I think we have to be careful when trying to achieve system 2 way of thinking via prompting, as LLMs, due to their training objective, will always carry their linguistic capabilities and biases to all prompting methods. Further architectural changes may even be necessary to reach true system 2 capabilities. \n- Regarding the inference time linguistic bias, I disagree that implicit representations of the context completely forbid the LLM from using them in its reasoning as the implicit representation is certain to encode parts of the explicit representation in itself as well. However, it is true that it can make the reasoning more difficult. I suggest changing the language in this section to reflect that.\n- Experimental results showcasing that the training, and inference bias, as mentioned in the paper, are the reasons behind the underperforming of the models in system 2 related tasks are sparse throughout the paper. Further evaluations and theoretical considerations should be made to further bolster the proposed conjecture.\n- The ablation study shows that the effect of each component can be inconsistent across models and tasks, therefore I suggest a deeper analysis of each component, possibly via the manual investigation of the model response change based on the inclusion of a component.\n- Following from the previous point, I still find it somewhat unclear how the proposed prompting method ameliorates the model bias, I suggest further evaluations to show the change in behavior.\n- I think that the relationship between the system 2 thinking of LLMs and the proposed method can be better explored via the construction of direct links between each other and showcasing how the prompting method addresses a problem in the large language model." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* LoT requires more inference time compute than plain CoT, given that it has to echo+expand. Have you considered doing a cost analysis to show how much gains one can have, vs how many more tokens are needed?\n* For domains where CoT is already good (e.g. math), is LoT still better or equal than using CoT? That would further show the potential of the technique. More in general, what are the potential drawbacks of using LoT?\n* Consider giving the phenomenon a more unique name than “bias”, as it is a very loaded term. In the first part of the paper it appears that we will focus on evaluating societal biases, and then the experimentation broadens and shows it is a general learning bias.\n* Consider presenting the results better in Figure 5, as it is hard to grasp the overall performance loss reduction vs general gains with LoT. It would be important to grasp when the technique can offer large improvement gains, on top of the bias reduction." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Simple prompting technique that can reduce language modeling bias (both societal and general learning biases) by echoing and expanding the required concepts before answering a complex question. I find particularly interesting that the technique shows that recent reports of performance degradation when using CoT may not be intrinsic to the step-by-step technique, but rather to the concrete execution shown in CoT.\n* The authors provide a useful intuition on why some biases may arise (Section 3), which is shared by many researchers but is important to spell it out like this work does. This intuition is useful for other applications besides the general motivation for a more structured CoT." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a prompting technique (Language of Thought, LoT) that provides a better structured step-by-step reasoning than the conventional Chain of Thought by asking the model to echo and expand the relevant information before answering. LoT alleviates the biases that are sometimes introduced by language modeling (societal or just general learning bias). LoT can also alleviate the regressions that CoT sometimes introduce in non-math domains. The paper also introduces some theory on language of thought bias, that is later used as intuition to support the design of LoT." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The theory presented is very useful to have in mind when designing prompts or generating synthetic data in general, but it is not really used anywhere in the paper except as general motivation, and it is not as formal as one would hope (e.g. propositions do not have a proof).\n* Consider rewriting the intro with more nuance, especially when describing psychological phenomena. E.g. System 1 and 2 should have more nuance, as these are useful theories, but they are not necessarily universally accepted in psychology. This psych discussion does not matter for CS research, as they are just useful analogies to our research, but the literature should be discussed well." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "This reviewer would have quite a few questions once the basic terms of the analysis are clarified.\n\nA basic one would be: why the insistence on causal relations between representations instead of deductive or inductively supported ones?\nPlease see my discussion above." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The authors are correct that permutations of sentences or premises for reasoning can lead to untoward results in LLM performance. The ideas behind the LoT prompting style have some promise." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper attempts to distinguish between language modeling and thought modeling. While LLMs right now model linguist to imitate human reasoning, the authors claim that there is a gap between language and though, which can introduce certain biases. They propose a new type of prompting that they call Language of Thought Prompting and provide various experiments comparing Language of Thought with CoT type prompts." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the ideas behind the new prompting style might be promising, this paper needs a radical overhaul to be acceptable. Too much in this paper is just hard to understand or vague. The leading question for the paper for example is already muddled: what does it mean to elicit the language of thought as humans??? Here is the quote: \n\"Do LLMs with CoT model and elicit the language of thoughts as humans?\"\nThe whole set involving a comparison up between language and thought in humans is somehow besides the point that the authors want to bring up. The motivation for the LoT style prompting seems to rely on is that different strings of linguistic tokens may express the same or at least very similar linguistic content to humans. This is not a language vs. thought issue but rather an issue of whether LLM objective functions as they stand, or even with CoT, capture real linguistic, semantic content. There is a lot of literature out there on insufficiencies of LLM in capturing semantic content; e.g. there's Bender, Emily M., and Alexander Koller. \"Climbing towards NLU: On meaning, form, and understanding in the age of data.\" Proceedings of the 58th annual meeting of the association for computational linguistics. 2020. But perhaps more relevant to the authors are two recent papers that look in detail at how LLMs fail to capture semantic meaning:\n \"Strong hallucinations from negation and how to fix them\" arXiv preprint arXiv:2402.10543 (2024);\n``Analyzing Semantic Faithfulness of Language Models via Input Intervention on Question Answering'' (Computational Linguistics 2024). \nSpeaking in terms of semantic contents instead of thoughts gives the authors a lot more ammunition to investigate where CoT approaches break down. Simply put, we know a lot more about the structure and features of semantic contents than we do about thoughts.\nIf the authors can show in detail that LoT methods capture semantic content better than CoT methods, that could be an important finding.\n\nThere's also a confusion between between logical and causal sequences. I quote:\n``Human conducts System 2 reasoning via the language of thoughts that organizes intermediate steps\nas a causal consequence of mental representations (Rescorla, 2024). For example, a human baby is\nable to abstract, construct, and reason over a causal map of the world in their minds.\" The important point it would seem to this reviewer in system 2 reasoning is that the intermediate steps follow each other in terms of logical or semantic consequence, not causal consequence. Or rather in a good system 2 reasoning system causal consequence and logical consequence merge. Why do I say that? Because you can have causal sequences of thoughts/representations in a psychopath that are completely crazy have no logical relation to each other and have nothing to do with type 2 reasoning. \n\n\nThe authors move quickly from evidence about thoughts without language to the thesis that \"As language is primarily a tool for communication instead of thinking and reasoning\". But there is nothing in the paper that warrants this assertion. And unfortunately,\nthis assertion is key to the paper and drives the move to find different prompts from standard CoT prompts.\n\nThe paper is woefully short on examples and often it's very difficult to understand what the authors want to say:\n\nFor example: \"Thoughts are the unobserved high-level random variables evaluated by\nbrains that drive us to generate language.\"\nIt's really hard to figure out what to do with this. And in addition it's a definition.\n\n\nanother example, 'When a premise is expressed in an implicit expression under a context, it is\nhard to notice and utilize it for downstream reasoning' what does it mean to have an implicit expression?\n\nHere's a sentence that just seems to be flat out false: \"For humans, since the language order does not determine the language\nmeaning when given proper conjunction words, one can easily change the order of presenting the\npremises in need.\" \nAs a counterexample, consider: \nJohn took off his shoes and went to bed\nvs. \nJohn went to bed and took off his shoes\n\nThe meaning conveyed by these two sentences is quite different. Changing the order of sentences often changes the meaning of a text.\nThis is part of the study of discourse structure and how it's formally interpreted. E.g. N. Asher & A. Lascarides, Logics of Conversation, \nCambridge University Press, 2003. And unfortunately this assumption seems to be key to distinguishing LoT from CoT\n\n\nA lot of the sentences in this paper aren't English or well formed in any language I know of . Eg. \"The Interplay between language and thoughts has intrigued a long historical discussion about\nthe role of language in human thinking in the literature (Rescorla, 2024; Fedorenko et al., 2024).\"\nIssues don't intrigue a historical discussion. In addition, if the history is long, why cite people from 2024? The authors might \nstart by citing Fodor Language of Thought (1975) but actually the issue already arises with early medieval thinkers like Saint Augustine and his concept of the \"verbum mentis\". See the Stanford Encyclopedia article on medieval semiotics.\n\n\nAgain: \"Consequently, modeling thoughts merely from the language\ncan easily integrate the language modeling biases into the learned model, such as the order (Wei et al.\"\nthis is largely unintelligible, and at this point language modeling biases haven't been defined.\n\n\nThe expanding thought part of the proposed prompt is too vague to be at all useful as it currently stands.\n\"instruct the model the expand those\" is not English or comprehensible. Please rephrase" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We demonstreate the gap of LLMs in modeling human thoughts for system 2 reasoning and propose Call-of-Thoughts to alleviate the gap." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024on,\ntitle={On the Language of Thoughts in Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3wrMRYuLlQ},\nnote={under review}\n}" }, "abstract": { "value": "System 2 reasoning is one of the defining characteristics of intelligence, which requires slow and logical thinking. Human conducts System 2 reasoning via the language of thoughts that organizes the reasoning process as *a causal sequence of mental language*, or thoughts. Recently, it has been observed that System 2 reasoning can be elicited from Large Language Models (LLMs) pre-trained on large-scale natural languages. However, in this work, we show that there is a significant gap between the modeling of languages and thoughts. As language is primarily a tool for humans to share knowledge and thinking, *modeling human language can easily integrate into language biases* that are not related to thoughts. Furthermore, we show that the biases may mislead the eliciting of “thoughts” in LLMs to focus only on a given part of the premise. To this end, we propose a new prompt technique termed **Ca**ll-of-**T**houghts ( CaT ) to alleviate the issue. Instead of directly eliciting the chain of thoughts from the potentially biased information, CaT instructs LLMs to focus and expand based on all the relevant information. We show that the simple strategy significantly reduces the language modeling biases in LLMs and improves the performance of LLMs across a variety of reasoning 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": [ "Language models", "system 2 reasoning", "language of thoughts" ] }, "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/1c2ed0377711d4baf2c5584ce59073e4f03dfb2f.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": "On the Language of Thoughts 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": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": { "value": "The submitted PDF is a placeholder and not a valid submission." }, "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": { "value": "Submission Desk Rejected by Program Chairs" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nnguyen2024enhancement,\ntitle={Enhancement of In-Context Reasoning in {LLM}s through Inductive Rule Learning},\nauthor={Tien-Dat Nguyen and Hai-Toan Nguyen and Nguyen Viet Ha},\nyear={2024},\nurl={https://openreview.net/forum?id=3x4vpeAclU}\n}" }, "abstract": { "value": "Currently, Large language models (LLMs) have achieved remarkable performance across various language tasks, largely due to their training on extensive datasets and their considerable model size. These models exhibit in-context learning abilities, which is to learn through few-shot learning. However, the underlying reasoning process remains ambiguous, it is unclear whether the model simply retrieves relevant information and instructions from its training data to generate similar responses, or whether it generalizes examples to form overarching rules, which are then applied to produce accurate answers. Another method for improving few-shot learning is Chain-of-Thought prompting that complement steps by steps instruction for LLMs, so they can follow this instruction to solve many reasoning tasks. Several approaches for evaluating the reasoning abilities of LLMs typically involve task-solving through code generation, which enables models to formalize problems and leverage a code compiler to solve them precisely. However, these methods are constrained to specific task types and are insufficient for a comprehensive assessment of the model's broader reasoning capabilities. Therefore, this paper proposes a method to enhance in-context learning capabilities through two main stages: generating general rules from the provided examples and utilizing LLMs to verify these general rules, thereby aiming to improve reliability and accuracy. At the same time, this approach seeks to investigate the inductive and deductive reasoning abilities, and can improve our understanding of the model’s reasoning by generating and applying general rules to provide transparent, clearly explained responses. The proposed method demonstrates competitive performance on the 1D-ARC benchmark and several traditional language tasks, suggesting its potential for more robust evaluation of LLM reasoning abilities." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Tien-Dat_Nguyen1", "~Hai-Toan_Nguyen1", "~Nguyen_Viet_Ha1" ] }, "authors": { "value": [ "Tien-Dat Nguyen", "Hai-Toan Nguyen", "Nguyen Viet Ha" ] }, "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", "Inductive Reasoning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "nguyen|enhancement_of_incontext_reasoning_in_llms_through_inductive_rule_learning" }, "pdf": { "value": "/pdf/6d6d76c526eaac210aa74569ad6fa951ecbd74d2.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": "Enhancement of In-Context Reasoning in LLMs through Inductive Rule Learning" }, "venue": { "value": "ICLR 2025 Conference Desk Rejected Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Desk_Rejected_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.\tThe remark in Section 3.1 suggests that increasingly complex tasks require stronger models. This is a commonly understood point and lacks novelty.\n2.\tIn Section 3.1, it is suggested that RAG is more suitable for tasks of moderate difficulty. However, according to Figure 2 and Table 1, there is no significant difference between LoRA and RAG in terms of performance.\n3.\tFive fine-tuning methods are compared in Section 3.1, but the first three are weaker than LoRA and RAG, offering limited practical insight.\n4.\tSection 3.2 suggests that smaller values for rank and alpha are more suitable for resource-constrained environments, but this finding also lacks innovation. Additionally, the models discussed are relatively small, making them inherently more compatible with LoRA in limited-resource scenarios, which somewhat disconnects the findings from real-world edge limitations.\n5.\tThe discussion on training duration in Section 3.2 does not specify which type of device is being considered. In the Appendix, the devices listed range from 4GB to 16GB of RAM, which would result in significantly different feasible training times.\n6.\tSection 3.4 proposes using only a limited amount of historical data in RAG, yet given the privatized edge LLM scenario suggested by the authors, it is realistic that users would only have access to a finite amount of data rather than unlimited data.\n7.\tThe performance loss due to model compression is a well-known trade-off rather than a novel finding specific to edge LLMs.\n8.\tAlthough the paper is framed as addressing the deployment of edge LLMs, it mainly focuses on the private fine-tuning of models. Important aspects of deployment, such as inference, are not covered." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This study is very comprehensive, employing a wide range of models and constructing various scenarios and methods for comparison." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors conducted extensive experiments and benchmarking to provide empirical guidelines for deploying large language models (LLMs) on edge devices, with a focus on fine-tuning models in private settings. However, the final conclusions of the paper align largely with common sense, and in some areas, the study lacks novelty." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although this paper provides an empirical guideline through extensive experimentation, many of the conclusions are quite intuitive, lacking some innovative findings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "The paper primarily evaluates popular pre-trained LLMs like Llama and OPT, with various modifications. There is little exploration of alternative architectures that could be inherently better suited for edge deployment." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Adequate experimental evaluation is carried out in this paper.\n2. The topic of deploying LLM at the edge is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on providing empirical guidance on LLM deployment on resource-constrained edge devices. The research focuses on how to optimize the design choices of LLMs in a resource-limited environment, balancing the computational demands with the capabilities of the devices." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper primarily restates existing strategies for model deployment and optimization, lacking substantial innovation. The guidelines and strategies discussed, such as model compression, parameter-efficient fine-tuning (PEFT), and retrieval-augmented generation (RAG), are already well-documented methods in the machine learning field. The paper offers an empirical evaluation rather than a novel methodological contribution.\n\n2. The experiments are largely confined to synthetic and benchmark datasets, which may not adequately represent the diversity of real-world scenarios where edge LLMs are deployed. This limits the applicability of the guidelines to practical use cases involving more dynamic environments." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Please see above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "This is an important and timely topic. If done correctly, such study can provide practical guidelines for researchers and developer." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies LLM personalization at resource-constrained edge devices, investigating the effect of design choices (e.g. what model to use, what personalization technique to apply) on the performance (e.g. accuracy). This is done by running a set of experiments to observe the effect of each choices. According to these observations, a set of guidelines are proposed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The study does not have the necessary depth. In particular, it reports only single-run experiments. It is however essential for an experimental study of this nature to perform multiple experiments per setting, enabling statistical comparison of different design choices. For instance, by providing means/medias and confidence intervals, one can assess if design choice A achieves a statically significant improvement over design choice B. I recommend the authors to follow approaches such as experimental design [1] to enhance the robustness of the study.\n \nBesides, the paper fails to provide clear guidelines on how select the design choices. For example, the text below is from Section 3.1.:\n\n“As task difficulty increases, such as with complex classification tasks and simple summarization tasks, the choice should gradually shift to RAG with the strongest model. Here, the strongest models are (quantized) LLMs that excel at general benchmarks and fit within the RAM constraint.”\n\nWhat are specific criteria on deciding what is a complex classification task? Does it depend on the number of classes? Or on the task? Is it possible to provide some quantitative measures on what is a complex classification task? Also, what does mean gradually? What I get from this guideline is some general roles, but it does not help me to make a clear decision.\n\nFinally, as I understand, the selected datasets for the fine-tuning process are available online, at Github. So, there is a possibility that the models which are studied in this paper have been already exposed to these datasets during the pre-training. This could change how we interpret the results, as fine-tuning a model over a subset of its training data is usually an easier task than fine-tuning over new (unseen) dataset. Please elaborate more on this aspect.\n\n[1]. R. A. Fisher et al. The Design of Experiments. Number 5th ed. Oliver and Boyd, London and Edinburgh, 1949." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Including some curiosity in weakness, I also raise the following questions.\n\n- Appendix A.1 shows different edge devices. I suggest the authors include and compare the computing resources, not only RAM, of the devices to give readers a better understanding of the range of resources considered in the study\n- In Table 3, the data samples per hour measured by the A10 GPU are illustrated. However, as stated earlier, processors like CPUs, GPUs, and mobile-targeted SoCs show significantly different computation times. Is the focus on computing resources solely on memory?\n- RAG and LoRA are quite different techniques, with one altering the model structure and the other not. I also think they serve different purposes. Can they be considered as alternatives to one another?\n- (minor) Rather than “performance,” please specify the exact metrics measured\n- Figure 9 shows five models, not 6, as mentioned in the caption.\n- The results suggest that insights vary significantly based on task type or complexity. Is there a way to categorize or quantify task characteristics or difficulty before running the tasks?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- As LLM deployment differs significantly on edge devices due to limited computing resources and the use of private data, the comprehensive evaluation and criteria for optimization techniques are crucial topics\n- The authors conduct comprehensive and well-defined experiments\n- The authors’ findings are novel, as they can guide future research directions" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses various design and deployment choices for running LLMs on edge devices. I enjoyed reading this paper, as the authors provide insights with clear and empirical experimental results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- In the introduction, many statements and findings seem to pertain to training (fine-tuning). Clearly stating the focus and deployment scenario would make the paper’s sections clearer (fine-tuning, inference, or both?)\n- Instead of focusing on individual techniques for LLM deployment, what about combining two or three methods? For example, RAG and LoRA could potentially be applied together\n- I understand the page limit and the authors’ efforts to address this issue, but many supporting results are in the Appendix, making it somewhat difficult to follow thoroughly.\n- The authors experiment with specified models on particular devices, considering memory constraints. What about testing similar models on different devices? Training a model for one hour on different devices might yield different insights\n- I believe many aspects, including the combination of design choices, remain unexplored in this study. Clearly specifying which aspects or potential experiment cases are covered and which are not would help readers better understand the study’s scope" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Provide Empirical guidelines for deploying and using LLMs on edge devices" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024empirical,\ntitle={Empirical Guidelines for Deploying {LLM}s onto Resource-constrained Edge Devices},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3xjc9PhEPd},\nnote={under review}\n}" }, "abstract": { "value": "The scaling laws have become the de facto guidelines for designing large language models (LLMs), but they were studied under the assumption of unlimited computing resources for both training and inference. As LLMs are increasingly used as personalized intelligent assistants, their customization (i.e., learning through fine-tuning) and deployment onto resource-constrained edge devices will become more and more prevalent. An urgent but open question is how a resource-constrained computing environment would affect the design choices for a personalized LLM. We study this problem empirically in this work. In particular, we consider the tradeoffs among a number of key design factors and their intertwined impacts on learning efficiency and accuracy. The factors include the learning methods for LLM customization, the amount of personalized data used for learning customization, the types and sizes of LLMs, the compression methods of LLMs, the amount of time afforded to learn, and the difficulty levels of the target use cases. Through extensive experimentation and benchmarking, we draw a number of surprisingly insightful guidelines for deploying LLMs onto resource-constrained devices. For example, an optimal choice between parameter learning and RAG may vary depending on the difficulty of the downstream task, the longer fine-tuning time does not necessarily help the model, and a compressed LLM may be a better choice than an uncompressed LLM to learn from limited personalized data." }, "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": [ "On-device Learning", "Edge Computing", "Efficient ML", "Large Language Models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/711b4df3bdb10abdaa5b085cebd86b48cc92f62d.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": "Empirical Guidelines for Deploying LLMs onto Resource-constrained Edge Devices" }, "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": "To the best of my knowledge, I see no ethical concerns regarding this work." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1) In line 271, you mention that $p(\\tilde{x_i}|X)$ is a Gaussian mixture distribution. Why did you choose this distribution? Did you evaluate the effect of this assumption on your results?\n \n 2) In your algorithm (see line 333), you split an email into sentences based on periods or commas. Have you examined whether this sentence-splitting method impacts your results? There are alternative ways to split text, such as by a specified number of characters.\n \n 3) Based on the main paper and appendices, it appears you developed a feed-forward neural network (see lines 879–881) for the selection model in your algorithm (see line 339). My understanding is that this neural network creates word embeddings based on the provided texts and then focuses on detecting phishing attacks. Why did you choose this model instead of a more advanced architecture, such as a transformer?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "Overall, this is a well-written and clearly explained work. The motivation and contributions are effectively communicated, as are the details of the algorithm, including the model architecture and data. According to the results presented in the tables, the AI2TALE model achieves state-of-the-art performance in detecting phishing attacks based on two measures—Label-Accuracy and Cognitive-True-Positive—which the authors believe are more appropriate for this task. The model also achieves state-of-the-art results with the well-known F1 score. The results are also validated by humans who found this model helpful. Another strength of this paper is that the results of this information-theory-based model are explainable, a claim that appears valid based on the presented figures. Finally, all code is reproducible and open-source." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work addresses the problem of phishing attacks and introduces an information theory-based model called AI2TALE to detect them without requiring ground truth while also providing explainable results. The authors validate this model on seven diverse real-world email datasets, demonstrating that the AI2TALE model achieves state-of-the-art results. The model is also tested with human participants, who found it helpful for detecting phishing attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have included a few comments regarding the presentation and also have some feedback on the experimental section.\n \n A) Presentation Comments\n \n A1) In lines 058 to 060, where you mention that early-stage AI approaches are among the most effective solutions for preventing and reducing negative effects, this seems like a strong statement. It would be beneficial to include a citation to support this claim.\n \n A2) In line 152, you introduce the term \"AI2TALE\" for the first time as the name of your model. It would be helpful to clarify what this acronym stands for and briefly explain your rationale behind the name selection.\n \n A3) In line 330, you refer to your method as “Algorithm 1.” Since this is the only algorithm presented in your paper, it is unnecessary to number it as \"1.\" Consider renaming it simply as “Algorithm” throughout the text.\n \n A4) In Section 4.1, “Studied Datasets” (lines 360-371), you list several datasets used in your research. Please add citations and links for each referenced dataset. Additionally, include citations and links in Section 6.2 of the appendices (lines 778-788).\n \n A5) In line 404, you mention that readers seeking more details can refer to the appendices. Please specify the exact section or appendix that contains the relevant information.\n\n A6) In Section 6.2 of the appendices (lines 789-799), this paragraph appears to be more relevant to the following section, 6.3, on data preprocessing and embeddings.\n\n A7) The link of your code, see line 897, should be in the main paper, such as in the introduction. \n\nA8) As mentioned in the author guidelines (see https://iclr.cc/Conferences/2025/AuthorGuide), you are encouraged to include a Reproducibility Statement at the end of the main text. This statement should detail the efforts made to ensure reproducibility and include any necessary information for those wishing to replicate your results.\n\n A9) You have referenced a good selection of papers with nice variation; however, I think a few relevant papers are missing. These include \"Feature-based Learning for Diverse and Privacy-Preserving Counterfactual Explanations\" by Vy Vo et al., \"The Anatomy of Deception: Measuring Technical and Human Factors of a Large-Scale Phishing Campaign\" by Anargyros Chrysanthou et al., \"Towards Modeling Uncertainties of Self-Explaining Neural Networks via Conformal Prediction\" by Wei Qian et al., and \"DIB-X: Formulating Explainability Principles for a Self-Explainable Model Through Information Theoretic Learning\" by Changkyu Choi et al. In the related work section of the appendices or the main paper, you could also include additional studies based on information theory that are not necessarily related to phishing attacks.\n \nB) Experimental Section Comments\n \n B1) In Table 1 (line 424), you present the results for each model. Since you used multiple datasets, it would be helpful to show the results for each dataset separately.\n \n B2) In Table 2 (line 443), it would be nice to include the second and third sentence top sentences of the model highlighting them with different colors and briefly discussing them.\n \n B3) In the human evaluation section (line 493), please include any available statistics on the participants, such as their expertise, gender, education level, etc.\n\n B4) In Section 6.7 of the appendices (lines 899–960), Table 3 presents results that, in my opinion, are quite significant and provide further support for the strength of your model. It would be better to incorporate the results from Table 3 into Table 1 and remove the \"average\" 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": { "value": "The paper presents results from a human study but do not provide operationalization details. Potential risk for leading questions/biased results." }, "flag_for_ethics_review": { "value": [ "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": 1 }, "primary_area": null, "questions": { "value": "Please see below some questions and comments to improve the manuscript:\n\n- Overall, the terminology used in this paper, e.g., phishing attack localization and phishing vulnerability are atypical for the cybersecurity community. Page 1, para 2 states that AI has been used for malware vulnerability detection. Malware detection and vulnerable code detection are two entirely different sub-domains. Moreover, page 3 states that post-hoc explanations do not offer a comprehensive understanding of the model's internal architecture or workings. Overexaggerated claim. Model-based XAI techniques like integrated gradients do exactly that! Deep models are also called \"self-explanatory\" methods in the manuscript. Again, popular literature dictates that deep learning is inherently a black box and thus cannot be self-explanatory/interpretable. The authors are advised to revise the manuscript and avoid redefining terms that already exist in the domain.\n\n- This work is not the first or the only work that labels an object as phishing and identifies which features led to the classification decisions. Could the authors present a comparative analysis against existing methods, e.g., [1-2], and justify why they were not mentioned in the manuscript? \n\n- Please add references for the datasets mentioned on page 7.\n\n- Please include the definition of the evaluation metrics (especially cognitive true positive) in the main body text (Section 4.2). Without the definition, it is impossible to understand the results. \n\n- The significance of the results in table 1 are hard to interpret because the authors do not provide any information regarding the dataset, e.g., number of phishing emails, sentences per email etc. Moreover, the improvement in the results is marginal (1-3%). Does the proposed approach provide benefits beyond the results in Table 1?\n\n- Please also report false positives and false negatives for each method in table 1. FPs are especially a big problem in cybersecurity tasks.\n\n- Please provide more details regarding the human study. 25 participants don't seem statistically significant for 10 emails, each with 5 options, as estimated by Cochran's formula. Moreover, it is unclear what questions are asked or how they are phrased exactly. For instance, if the participants are always shown the top-1 sentence and not given an option of top-3 sentences, this can very likely be interpreted as a rhetorical question, which leads people to respond in the affirmative most of the time. The authors are advised to increase the sample size to a statistically significant level, provide the exact phrasing of questions asked, and include control questions or alternative options (like top-3 sentences) to reduce bias." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "It is good to focus on explainable approaches so we can understand what ML models do. The method uses information theory approaches, which is indeed an innovative angle." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper builds a system that identifies both, whether an email text is a phishing attack and which top-1 sentence in the email has the highest contribution to the classification label. The method is based on information theory, which is used by the model to distribute importance weights to features. The paper claims that this is one of the first works in phishing localization and that XAI techniques have not been applied to phishing attacks in the literature. By comparing against 7 open source datasets and 5 popular baseline methods, the method shows an improvement of 1.5-3.5% over baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Firstly, the claim that this is one of the only works for interpretable phishing localization might be inflated. There is plenty of work in the domain of phishing attack detection that builds ML models and applies XAI techniques to understand why something was labelled as a phishing attempt, see e.g., [1-2] as just two random examples. These techniques even go beyond text and also tag why images or URLs were considered phishing. Thus, the novelty of this approach is severely lacking. Please provide a more detailed comparison of the proposed approach to existing XAI techniques in phishing detection, highlighting specific differences in methodology or capabilities.\n\nSecond, one of the most popular explanation techniques for text, especially for deep nets is the attention mechanism. The proposed approach seems to somewhat achieve a similar goal by learning one model that both classifies text as phishing and also discovers features (sentences) that led to the classification decision. Yet, the paper makes no mention of this popular technique, nor do they compare with it in terms of fidelity, speed, correctness, etc. Thus, there is insufficient comparison with existing methods. The authors are advised to include a comparison with attention-based methods, evaluating aspects like fidelity, speed, and correctness. \n\nThird, the writing can substantially be improved. There are several typos that hinder comprehension. A non-exhaustive list: page 2, second last para -> which information causes. page 4, first sentence -> given an email. page 5, line 255 -> we encourage. \n\n\n[1] Chai, Yidong, et al. \"An explainable multi-modal hierarchical attention model for developing phishing threat intelligence.\" IEEE Transactions on Dependable and Secure Computing 19.2 (2021): 790-803.\n[2] Lin, Yun, et al. \"Phishpedia: A hybrid deep learning based approach to visually identify phishing webpages.\" 30th USENIX Security Symposium (USENIX Security 21). 2021." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "I don't think any additional ethical review is required, but it could be good if the authors added a more elaborate discussion of the ethical implications of deploying this AI-based system in real-world applications, particularly concerning privacy concerns and the risk of misclassification." }, "flag_for_ethics_review": { "value": [ "No ethics review needed.", "Yes, Discrimination / bias / fairness concerns" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The paper compares the proposed method with state-of-the-art AI-based approaches. Did you consider including a more elaborate discussion of how traditional (non-AI) phishing detection methods, such as URL filtering and white and blacklisting IPs, sandboxing, etc., can be complemented by or replace improvements offered by AI2TALE?\n- Can you elaborate on what defense mechanisms organizations should prioritize to counter the increasingly sophisticated, region-specific threats posed by localizaion?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Strengths:\n- Original Contribution: The proposed method (enhances explainability) is a significant improvement in phishing defense and a good use of AI.\n- The use of information theory and the mutual information training principle to select relevant features is well-founded. The introduction of a selection network that utilizes latent variables to identify important sentences is also good.\n- The authors provide thorough evaluations across seven diverse datasets. The comparative performance metrics (Label-Accuracy and Cognitive-True-Positive) are appropriate and lend credibility to their claims of improved performance over baseline models.\n- The paper is well-organized, with clear sectioning and logical flow. The introduction succinctly frames the problem and the proposed solution, making it accessible even to readers less familiar with the technical intricacies.\n- The use of human evaluations to assess the interpretability of the selected sentences enhances the practical relevance of the findings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel approach to phishing attack localization using deep learning and information theory principles. The method aims to classify emails as phishing or benign and explain these classifications by identifying the most relevant information within the emails. In short, phishing attack localization aims to customize phishing attacks to align with a target region or group's specific cultural, linguistic, and contextual characteristics. The paper examines how localized phishing tactics, such as using native language, regional events, or imitating local institutions, enhance the effectiveness of phishing campaigns by increasing their credibility. The study compares localized versus generic phishing attacks and highlights key factors that improve success rates. It also discusses the growing role of AI in automating localization efforts, presenting a challenge for organizations to develop stronger, region-specific defense mechanisms to counter these advanced attacks. The authors have conducted extensive experiments on multiple datasets to demonstrate the efficacy of their approach. Overall, it’s an interesting paper with a sound contribution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses:\n- It could be good to add a more elaborate discussion of the ethical implications of deploying this AI-based system in real-world applications, particularly concerning privacy concerns and the risk of misclassification.\n- The authors acknowledge potential limitations regarding selecting non-relevant sentences. A more thorough discussion of these limitations and how they might affect real-world applications would be good. For example, the risk of misclassification and potential consequences of user confusion arising from irrelevant explanations.\n- The paper mentions the introduction of hyperparameters in the mutual information maximization and data-distribution mechanisms. It would be good to elaborate on the model’s sensitivity to these hyperparameters and their impact on the performance. It would also be beneficial for the authors to discuss how the introduced hyperparameters are likely to generalize to different contexts and model types. Understanding whether these hyperparameters remain effective across various datasets or when applied to different phishing detection frameworks.\n- The results section could benefit from more visualizations (figures or tables) to make the results more digestible and better illustrate the performance improvements." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We study an important problem of phishing attack localization aiming to tackle and improve the explainability (transparency) of email phishing detection. AI-based techniques for this problem have not yet been well studied." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024aitale,\ntitle={{AI}2{TALE}: An Innovative Information Theory-based Approach for Learning to Localize Phishing Attacks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3xpTXF5ALZ},\nnote={under review}\n}" }, "abstract": { "value": "Phishing attacks remain a significant challenge for detection, explanation, and defense, despite over a decade of research on both technical and non-technical solutions. AI-based phishing detection methods are among the most effective approaches for defeating phishing attacks, providing predictions on the vulnerability label (i.e., phishing or benign) of data. However, they often lack intrinsic explainability, failing to identify the specific information that triggers the classification. To this end, we propose an innovative deep learning-based approach for email (the most common phishing way) phishing attack localization. Our method aims to not only predict the vulnerability label of the email data but also provide the capability to automatically learn and figure out the most important and phishing-relevant information (i.e., sentences) in the phishing email data, offering useful and concise explanations for the identified vulnerability. \n\nThe extensive experiments on seven diverse real-world email datasets demonstrate the capability and effectiveness of our method in selecting crucial information, enabling accurate detection and offering useful and concise explanations (via the most important and phishing-relevant information triggering the classification) for the vulnerability of phishing emails. Notably, our approach outperforms state-of-the-art baselines by 1.5% to 3.5% on average in Label-Accuracy and Cognitive-True-Positive metrics under a weakly supervised setting, where only vulnerability labels are used without requiring ground truth phishing information." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Phishing Attacks", "Email Phishing Attack Localization", "Interpretability and Explainable AI", "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/98334b1cd3da556b99be966a527f6ba4e855eb45.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": "AI2TALE: An Innovative Information Theory-based Approach for Learning to Localize Phishing Attacks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 do you balance the four losses in the proposed method? Given the numerous modules and hyperparameters involved, does training the model from scratch carry a high risk of overfitting?\n2. Considering the frequent occurrence of the OOD generalization problem in brain network analysis, how could the proposed method be adapted or transferred to other models?\n3. Since the performance of fMRI-derived brain networks on the ADNI dataset is lower than that of structural MRI, do you believe it is appropriate or necessary to use it as a benchmark for the OOD generalization problem?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- **Originality**: This paper demonstrates a notable level of novelty, particularly in its combined approach of selecting critical node features and graph structures, along with the batch-level loss designed to identify key discriminative connections.\n- **Quality**: The methodology is thoroughly evaluated through comparisons with 16 existing methods across two datasets (ABIDE and ADNI), effectively highlighting its effectiveness and efficiency.\n- **Significance**: This research provides valuable insights into addressing the OOD problem in brain network analysis, contributing meaningfully to advancements in neuroscience." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work addresses the out-of-distribution (OOD) problem in brain network analysis. It introduces a framework called BrainOOD, which consists of a feature selector and a structure extractor. By filtering out noisy nodes and edges and enforcing the model to consistently select the same connections across all brain networks within each batch, the proposed method achieves strong performance on the ABIDE and ADNI datasets. Additionally, visualization results are provided to illustrate the method’s effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Contribution of the Benchmark**\n\n**The claim of introducing the first benchmark seems somewhat overstated.** The ABIDE and ADNI datasets have been long established in brain network analysis and are widely used for evaluating brain disorder diagnosis models. Simply partitioning these datasets to create an OOD scenario may not constitute a significant contribution.\n\n- **Alignment of Motivation, Method, and Analysis**\n\nThe motivation of this work is to address the OOD generalization problem. However, **it is not clearly explained how the proposed method specifically tackles this issue**. While reducing noisy nodes and structures could indeed improve brain disorder diagnosis performance, the methodology and interpretive analysis lack clarity on how this approach mitigates the OOD generalization problem. For instance, visualizing the top 10 connections with the highest scores on both the ABIDE ID and ABIDE OOD sets could help demonstrate the method’s generalizability more effectively.\n\n- **Paper Organization**\n\nThe organization of the paper could be improved for clarity. **It may not be necessary to dedicate extensive sections to GNN and brain network fundamentals.** Additionally, placing the related work section directly after the introduction or immediately before the conclusion could improve the flow and readability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 weakness above. \n\n- In addition, can you provide an ablation study on the feature selector and structure extractor by evaluating configurations such as $(X', A)$ and $(X, A')$? These results would help to clearly demonstrate the contribution of each module. Additionally, similar to the discussion on edge scores, the node mask should also be examined to strengthen the claim that the proposed method yields clinically relevant results.\n\n- While several GNNs and HPGNN are incorporated into the framework, certain aspects remain unclear. Specifically, what advantage does using HPGNN with multiple layers (hops) offer over simply multiplying the graph Laplacian matrix, especially if the goal is to capture deviations from local patterns? Furthermore, given your assertion that the brain structure matrix A contains noise, why did you choose to retain A rather than use A’ during feature selection?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper addresses a critical gap in brain network analysis by focusing on OOD generalization and interpretability, which are essential for deploying models in real-world settings. The work has high significance for the medical and neuroscience community. \n- It presents a framework that improves diagnostic tools for neurological disorders like AD and ASD, potentially leading to earlier and more accurate diagnoses. \n- The authors evaluate their method across two major datasets (ABIDE and ADNI) and compare it with 16 baselines including brain-specific networks, which adds credibility to their results. \n- The alignment of identified brain patterns with known neuroscience findings lends additional weight to the framework's interpretability. Also, ablation study demonstrates the needs of each loss types." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents BrainOOD, a framework designed to address the challenges of Out-of-Distribution (OOD) generalization in brain network analysis. Specifically, BrainOOD aims to enhance the performance and interpretability of Graph Neural Networks (GNNs) in diagnosing Alzheimer’s Disease (AD) and Autism Spectrum Disorder (ASD). The method incorporates a feature selector, structure extractor, and auxiliary losses, leveraging the Graph Information Bottleneck (GIB) framework to recover causal subgraphs. Through extensive experiments on those datasets, the framework demonstrates competitive performance, outperforming baseline models in OOD settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The technical contribution of this paper appears to be marginal despite addressing the OOD generalization problem and enhancing interpretability in brain network analysis. While the introduction of an OOD benchmark for brain networks is appreciated, it is unclear if this benchmark adds novel challenges beyond those already present in multi-site datasets like ABIDE and ADNI. Furthermore, many of the technical components, such as the auxiliary losses and discrete sampling strategy, are borrowed from existing work. Although the paper effectively motivates the need for the Graph Information Bottleneck (GIB) framework, the core technical innovations do not extend significantly beyond prior work.\n\n2) One of the primary technical contributions --- feature selection mechanism --- lacks clarity in its formulation. Specifically, the intuition behind $\\hat{X}$ derived from the covariance of $\\hat{H}$ and the use of the $tanh()$ as activation function is not well explained, leaving readers uncertain about the necessity of these design choices. \n\n3) The definition of the OOD problem itself also raises concerns. Table 2 indicates insignificant performance differences between in-distribution (ID) and OOD scenarios, even with the Empirical Risk Minimization (ERM) baseline, suggesting that the OOD scenario may not be as challenging as claimed. This raises the possibility that the proposed framework performs effectively only under moderate distribution shifts. Additionally, the paper would benefit from comparing the performance of other brain-specific models, such as BrainNetCNN or BrainGNN, under the same OOD conditions to better contextualize the reported improvements.\n\n4) Grammar should be double checked." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. For the adjacency matrix, were the top 20% connections identified based on correlation magnitude (including both positive and negative correlation)?\n2. The classification setting (6-class) on the ADNI dataset. It is confusing to have three classes related to MCI (MCI, EMCI, and LMCI), which may affect the evaluation results. EMCI and LMCI are used in ANDI GO/2, while MCI used in ADNI 1 is deemed LMCI. A 5-class (CN, SMC, EMCI, LMCI, AD) setting is more reasonable.\n3. It would be helpful to add more description about how the reconstruction loss can help select informative features.\n4. It is not clear how in-domain testing was performed. \n5. What are the differences between the 10-fold-CV and the overall test in Tabel 2 and 3?\n6. For evaluation, it is better to add some conventional ML methods (e.g., SVM) as baseline.\n7. What does the ID and OOD checkpoints mean in Fig.3? The edge score seems quite low (max value around 0.08, Fig.3 top left), how many edges were generally included in the extracted sub-graph?\n8. There are several other parameters in the framework (e.g., temperature in eq.11, number of sampling k for the final prediction). How do they affect the performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "It is novel to simultaneously identify informative features and extract causal subgraph for brain functional network based prediction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents BrainOOD, a novel GNN framework tailored for brain functional network analysis, which consists of feature selector and causal subgraph extractor for brain functional network to enhance the generalization to out-of-distribution dataset. The proposed framework has been evaluated on two multi-site datasets and demonstrated improved classification performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Several descriptions are not clear. Please refer to the Questions section for details.\n2. The classification setting (6-class) on the ADNI dataset. It is confusing to have three classes related to MCI (MCI, EMCI, and LMCI), which affects the evaluation results. EMCI and LMCI are used in ANDI GO/2, while MCI used in ADNI 1 is deemed LMCI. A 5-class (CN, SMC, EMCI, LMCI, AD) setting is more reasonable." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "BrainOOD boosts GNN generalization and interpretability for brain networks, outperforming 16 methods and introducing the first OOD benchmark." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024brainood,\ntitle={Brain{OOD}: Out-of-distribution Generalizable Brain Network Analysis},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3xqqYOKILp},\nnote={under review}\n}" }, "abstract": { "value": "In neuroscience, identifying distinct patterns linked to neurological disorders, such as Alzheimer’s and Autism, is critical for early diagnosis and effective intervention. Graph Neural Networks (GNNs) have shown promising in analyzing brain networks, but there are two major challenges in using GNNs: (1) distribution shifts in multi-site brain network data, leading to poor Out-of-Distribution (OOD) generalization, and (2) limited interpretability in identifying key brain regions critical to neurological disorders. Existing graph OOD methods, while effective in other domains, struggle with the unique characteristics of brain networks. To bridge these gaps, we introduce \\textit{BrainOOD}, a novel framework tailored for brain networks that enhances GNNs’ OOD generalization and interpretability. BrainOOD framework consists of a feature selector and a structure extractor, which incorporates various auxiliary losses including an improved Graph Information Bottleneck (GIB) objective to recover causal subgraphs. By aligning structure selection across brain networks and filtering noisy features, BrainOOD offers reliable interpretations of critical brain regions. Our approach outperforms 16 existing methods and improves generalization to OOD subjects by up to 8.5\\%. Case studies highlight the scientific validity of the patterns extracted, which aligns with the findings in known neuroscience literature. We also propose the first OOD brain network benchmark, which provides a foundation for future research in this 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": [ "Out-of-distribution Generalization", "Brain Network Analysis", "Graph Representation 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/c94ec9f94d0189748d4f620cb2dbc91600abe8c9.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/a34932cb521b3c4af5ff5094a65a62de9db64bf2.zip" }, "title": { "value": "BrainOOD: Out-of-distribution Generalizable Brain Network Analysis" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is easy to read and follow.\n2. The authors have a nice experimental framework that starts from manual annotations and builds up to a \"unsupervised approach\".\n3. The performance gains over previous methods are non trivial." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper improves upon methods that add a manual highlight over image to improve clip performance. The authors optimize a learnable visual prompt and show that they can start from manual annotations to eventually propose the final method which relies on image crops from a pretrained object detector. The authors show that their formulation improves upon previous methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors have not validated their initial hypothesis that different models will have different biases. Table 6 compares several CLIP models but I am more interested to know if there is any difference in this behavior based on pretraining objective and dataset. How would this compare for SigLIP, SILC, EVA-CLIP, MetaCLIP etc and models trained on OpenAI dataset, Webli, datacomp, etc.\n2. How does the method compare with different localization methods. The authors have used MAttNet. Is there data leakage in the localization network? Is it possible to use a generic model 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": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1.\tThis is in reference to the point 1 in the weaknesses. The apples-to-apples comparison (unsupervised regime) of the proposed approach with the existing ones can not cement the proposed approach’s betterness. Any comment on this would be good to have.\n2.\tComparison with the closest paper Xie et al. (2024a) is missing and in the same level playing ground (without using manual supervision to train) Xie et al. (2024a) works better. Any comment on this would be good to have.\n3.\tThis is related to the third point in the weaknesses. Is it possible to comment on the missing experiments mentioned there?\n4.\tThe strong unsupported claims as detailed in point 4 in ‘weaknesses’ needs to support.\n5.\tIn Table 1, the results are impressive. However, I would like to see, for completeness, what would be the performance of the proposed method if it is measured in the same level playing field i.e., it is unsupervised and uses ensemble of the same backbones as the rest of the competing methods.\n6.\tI am not comfortable with categorizing methods like RedCircle with sup=x. Theres a fine line between an 'unsupervised' approach and a zero-shot approach. RedCircle can best be described as zero-shot and not 'sup=x' (which means it gets training but no supervision, which is not true for redcircle as it does not get any training).\n7.\tIn figure 2, When both are vision encoders, what encoders are you using? Are the two vision encoders the same in such a case?\n8.\tThe next question is related to eqn. (2). What does this symbol a cross inside circle mean? What is the relation between the two $\\mathcal{F}(v)$'s and $I(v)$? What is the difference between upper case I and lower case i in eqn. (2)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThe paper presents a learnable way to predict both shape and color of a visual marker such that regions bounded by such markers in images can correspond to textual descriptions.\n2.\tThe discussion on the relevant works is good.\n3.\tThe use of image crops and synthetic captions for unsupervised training is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper works on a large Vision-Language Model (VLM), especially CLIP. It learns to generate markings on images so that a textual expression related to the image has high correspondence with the image region inside the marking. In short, the authors address the Referring Expression Comprehension (REC) problem using VLMs. While the problem of REC using VLMs like CLIP is not new, unlike previous studies, this paper does not assume the existence of predefined markers like red circle, square or arrow etc. Rather, the authors propose to learn the particular form of markers in both supervised and unsupervised manner. In supervised approach an infoNCE loss is minimized between annotated regions with expressions and the network generated regions (called highlights). In unsupervised approach, the authors use synthetic captions instead of manually annotated descriptions of the crops. Experiments show comparisons with related works that does not use supervision but ensemble of backbones in CLIP." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe novelty of the approach is limited. Let us take the example of the first work in this line which is RedCircle (Shtedritski et al., 2023). The authors there have done extensive analysis on drawing different types of visual prompting e.g., circle, square, arrow etc. with different colors on images and then processing it and the text query via a pretrained VLM. This paper concludes and shows that a red circle works best for REC with CLIP. One may argue that redcircle is dependent on manually tuning the markers for their shapes and colors. Manual tuning has already been replaced by learning to tune the visual markers in Xie et al. (2024a). Xie et al. also did not use supervision for training. The proposed work, on the other hand works good only when supervised training is used (compare row ‘Highlight (unsupervised)’ and ‘RedCircle’ rows as well as ‘Highlight (Supervised)’ and ‘RedCircle’ rows in table 2). The fact that ‘Highlight’ needs to use manual annotation from reference comprehension datasets to perform well, adds to novelty but takes away the major strength and flexibility of VLMs – i.e., zero-shot task transfer. The other incremental novelty of using synthetic texts does not perform well compared to the state-of-the arts.\n2.\tThe closest work to the proposed approach is Xie et al. (2024a) in flavor as this paper also learns to draw a visual marker in zero-shot manner. However, this has not been compared with in the zero-shot setting.\n3.\tThe proposed approach shows its usefulness only in one of the three tasks RedCircle (Shtedritski et al., 2023) evaluated themselves on. How does it perform on the other two tasks namely – ‘Naming Keypoints’ and ‘Keypoint Localization’ is missing.\n4.\tThere are a few strong statements that is not supported well. For example, Line 047 says - alt texts in pretraining image-text datasets summarise global image descriptions. This is in contrast to the common wisdom as alt-texts are usually very short and sometimes cryptic descriptions of a scene. Is there any reference or some examples that you can point to? Similarly, Line 60 says that whether a red circle is the prompt that best elicits these emergent behaviors in VLMs is unclear. However, in Shtedritski et al. (2023), there has been extensive studies on different types of markers e.g., circle, arrow, square, cross etc. or different colors of them. Table 2 in the main paper and table 10 and figure 10 in the supplementary clearly shows red circle works best.\n5.\tMinor typos: Line 63: ‘unfeasible’ -> ‘infeasible’; Line 261: ‘bounding’ -> ‘bounding box’." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "as shown in Weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "2. [Strengths]\n- a. It introduces Highlight, a method to automatically learn a visual prompt that highlights regions for a VLM.\n- b. The experiments are extensive, and the performance of the proposed method is promising.\n- c. The method is flexible, which can work in both supervised and unsupervised cases." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. [Summary]\nThis paper focuses on visual prompt learning for vision-language models. It introduces Highlight, a method to automatically learn a visual prompt that highlights regions for a VLM. The proposed method could work in both supervised and unsupervised manner. Experiments show that the proposed method achieves good performances." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "3. [Weaknesses] \n1. What is the different between this paper and the other popular visual prompt learning like [a, b]. [a, b] learn tokens as visual prompts. It would be better to discuss and clarify the similarity and difference between this paper and [a, b] and include the discussion in Section Related Work.\n[a] Visual Prompt Tuning.\n[b] Progressive visual prompt learning with contrastive feature re-formation.\n\n2. The core of this paper is to learn visual prompts for vision-language models. Have the authors tested the proposed method on other VLMs in addition to CLIP? Besides, the Section Related Work lacks a paragraph for introducing VLM. It would be better to add a new paragraph to introduce related work and recent advances in vision-language models, and could also cite the VLM survey [c] as the reference for the readers to know more about VLMs.\n[c] Vision-Language Models for Vision Tasks: A Survey\n3. Another question is that whether the learnt visual prompts are independent or dependent to input images? They are the same for all input images? Or works differently for different input images?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. I noticed the experiments were only conducted on CLIP models. Have you considered testing the method on other vision-language models? If not, what were the main challenges preventing such evaluations?\n\n2. From Figure 1, I can see that your method introduces some blur to the original images. Could you clarify:\n * Does this blur affect the model's performance?\n * Is this blur necessary for the method to work effectively?\n\n3. Your results show that learned prompts perform better than manual ones, but the mechanism behind this improvement is not clearly explained. Could you:\n * Provide more analysis on why learned prompts work better?\n * Share some failure cases to help understand the method's limitations?\n\n4. I'm interested in the robustness of your method:\n * How stable is the performance across different runs?\n * How sensitive is it to hyperparameter choices?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.The paper is clear and easy to follow.\n\n2.The research goal is meaningful. They propose to automatically learn visual prompts instead of designing them manually, which is a clever solution to a common problem in vision-language models.\n\n3. Their experimental results are impressive. The method works well across different versions of CLIP models and is more efficient than ensemble methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1.The authors present Highlight, a method that learns visual prompts for Vision-Language Models (VLMs) to improve their ability to localize specific image regions. The key innovation is automatically learning both the shape and color of visual markers in a differentiable manner, rather than relying on manually designed prompts like red circles. \n\n2.The method can be trained either with supervision from text-image region pairs, or without supervision using synthetic captions or images alone. \n\n3.The authors evaluate their approach extensively on RefCOCO datasets using different CLIP models, demonstrating that Highlight outperforms existing visual prompting methods and even ensemble approaches while being more computationally efficient." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.Robustness and Generalizability\n\n- The experiments are conducted exclusively on CLIP variants\n- I would like to see evaluations on other popular VLMs (e.g., Flamingo, GPT-4V)\n- Without broader validation, it's difficult to assess if this method is truly model-agnostic\n- I strongly suggest including at least 2-3 different families of VLMs in the evaluation\n\n2. Theoretical Understanding\n- The paper lacks insights into why learned prompts outperform manual ones\n- The visualizations don't effectively explain the mechanism's advantages\n- I recommend: \n(1) Providing case studies of success and failure cases\n(2) Including analysis of the learned prompt patterns\n\n3. Image Quality Impact\n- Looking at Figure 1, I'm concerned about the image blur introduced by the method\n- The authors should address: \n(1) How this blur affects the model's performance\n(2) If this is a limitation for certain types of images or tasks" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024highlight,\ntitle={Highlight: Learning Visual Prompts for Vision-Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3xxxoh92Mo},\nnote={under review}\n}" }, "abstract": { "value": "Large-scale Vision-Language Models, such as CLIP, demonstrate impressive capabilities and have multiple applications, from text-to-image generation to zero-shot classification. Recent work has suggested that visual prompts, such as a red circle, can steer the vision encoder to the circled region. While such vision prompts have now been used in various applications, they might be model-specific and depend on the model learning these behaviours from its training data. Discovering and evaluating various prompts might not be feasible given different models, tasks, and datasets. In this paper, we propose Highlight, a method to learn a visual prompt that highlights a region in an image or refines a manually engineered visual prompt. Using our framework, we can learn to highlight in a supervised way using a dataset of text-image region pairs or in an unsupervised way using synthetic captions or images only. Highlight outperforms other visual prompts, prompt learning approaches, and compute-intensive methods that use ensembles of multiple models and visual prompts." }, "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": [ "VLMs", "prompting", "visual prompting", "self-supervision" ] }, "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/ba2409ce1284cc058ce3575382d33e00fda787b3.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": "Highlight: Learning Visual Prompts 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": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "**Question 1.** Why did not SK RegMixup be applied to ViT in Table 4? Is there any potential challenges or limitations of applying SK RegMixup to ViT architectures?\n\n**Question 2.** [line 976-978] In the proof of the first part in Theorem 3.1., whatever the value of $\\lambda\\_{1}$ is, $\\lambda$ can be 0. When $\\lambda=0$, then $\\tilde{\\mathbf{x}}(\\lambda)=\\mathbf{x}\\_{l} \\in \\mathcal{M}\\_{j}$, not $\\mathcal{M}\\_{i}$, and so does symmetrical case. In my opinion, it would be $\\forall \\lambda \\geq \\lambda\\_{1},~\\tilde{\\mathbf{x}}(\\lambda)=\\lambda \\mathbf{x}\\_{k}+(1-\\lambda)\\mathbf{x}\\_{l} \\in \\mathcal{M}\\_{j}$. The authors are kindly requested to examine and fix them if there is an error.\n\n---\n\n**Things to improve the paper that did not impact the score:**\n\n* Figure 4 caption: there is no result about *Circles* toy datasets. The mention of it from the caption would be removed if it's not intended to be included\n\n* [line 442] The ECE and AECE of MIT-A in ImageNet-R would be in bold. The authors have to double-check their result highlighting in Table 4 to ensure consistency and accuracy.\n\n* In Table 4, is there any reason why ECE and AECE are exactly same in OOD settings? The authors are kindly requested to explain why ECE and AECE are identical in OOD settings or, if this is unexpected, verify whether there might be an error in the reporting or calculation of these metrics.\n\n* The experimental results show that the SK RegMixup is an effective method. However, Table 6 does not provide the efficiency of the SK RegMixup in terms of computational cost. It is recommended to include computational cost metrics for SK RegMixup in Table 6, similar to what they've provided for other methods.\n\n* [line 879] Typo: cccross" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "* The proposed approach was well motivated by targeting the appropriate research problem of calibration-driven Mixup methods in classification and regression: Manifold mismatch. Furthermore, this problem was theoretically defined and proved with the supplement materials.\n\n* The proposed method was in a clear and well-organized manner, and the proof was provided in a manner that supported its conclusions.\n\n* The research problem was validated through experimental results; a comparison was made with the baseline.\n\n* Comprehensive experiments in classification and regression, with a particular focus on performance and computational cost, demonstrated the efficacy of the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel Mixup framework that employs a Similarity Kernel (SK) called SK Mixup to achieve a stronger interpolation between similar points while reducing interpolation otherwise. As a motivation of this study, the authors defined the concept of manifold mismatch, which can negatively impact the calibration of confidence in Mixup. They conducted experimental validation to assess the impact of this phenomenon on the distance between points to mix. Following the presentation of SK Mixup, the effectiveness of the proposed approach in alleviating the manifold mismatch was demonstrated through extensive experiments in classification and regression. Nevertheless, the extension to more complex tasks and theoretical analysis on the regularization effect of SK Mixup still need to be undertaken -- This is why the authors did not receive the highest score in *Contribution*." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There is no weakness in the paper that would justify its rejection." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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": "- In Eq. (3), there is a -1 inside the exponential. Why not absorb it into $\\tau_{max}$? Are there any reasons why it is presented that way? \n- In the classification scenario where the data samples (but not the labels) are noisy and corrupted, thus blurring the true distance between the samples used in mixing, how would the proposed method perform?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Overall the paper is well written and is easy to follow\n- An innovative yet efficient method is proposed to improve mixup so that manifold mismatch likelihood is reduced, and better calibration and accuracies can be achieved\n- Adequate empirical results on both classification and regression tasks are provided to demonstrate the proposed method" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper first demonstrates that distance between the data used in mixup can impact the likelihood of manifold mismatch (a phenomenon where mixed samples lie outside of the class manifolds of the original data used for mixing). It then proposes an efficient framework to mitigate the occurrence of manifold mismatch. The key idea is to dynamically change the distributions of mixing coefficients via a similarity kernel that measures the distance between the mixed samples. Empirical results are provided to demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While Theorem 1 shows the existence of manifold mismatch, it does not tell us anything about the assumption (which is used in the paper) that the higher the distance between two points, the more likely their convex combination will fall outside of the original manifolds, and thus, the more likely a model would assign a different label than the original labels of the two points. Some theoretical results in establishing the validity of such assumption would further strengthen the paper. \n- There are some missing papers that could be worth mentioning in the related work. In particular, mixup related works such as [1,2,3]. \n- For the experiments, it is not clear how the proposed method would compare with the method proposed by [1,2,3], and also how would it perform for more complex datasets like ImageNet, ImageNet-C, ImageNet-R and ImageNet-P.\n\n---\n\n[1] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. Proceedings of the ICLR, 2020\n\n[2] Soon Hoe Lim, N. Benjamin Erichson, Francisco Utrera, Winnie Xu, and Michael W. Mahoney. Noisy feature mixup. Proceedings of the ICLR, 2022.\n\n[3] Erichson, N. Benjamin, Soon Hoe Lim, Francisco Utrera, Winnie Xu, Ziang Cao, and Michael W. Mahoney. Noisymix: Boosting robustness by combining data augmentations, stability training, and noise injections. Proceedings of the AISTATS 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": 1 }, "primary_area": null, "questions": { "value": "1. Line 011 or 012 on page 1. \"Along with improved performance\", what performance, please make it specified\n2. Line 012 or 013 on page 1. \"improving calibration and predictive confidence\", from my understanding they are the same thing, no need to use \"and\" here in my opinion.\n3. Line 014. Infact in this line the paper has mentioned \"calibration of confidence\", what is calibration of \"confidence\", again, in my opinion it's just calibration.\n4. Line 019 to 020. Again, please specify what \"performance\" it is when mentioning \"improve performance\".\n5. Line 054. Here the paper mentions \"label noise\", can the authors provide a brief description of it like the way they have described manifold intrusion in previous text?\n6. Line 140. Again, please specify the term \"performance\".\n7. Is there and related research works about calibration-driven Mixup methods in regression tasks?\n8. Line 161. If possible it's much better to turn this title to the next page, it would look better.\n9. Line 162 to 169. The notation preliminaries in this paragraph feel a bit too simple and careless. For example, when defining a dataset, one may want to first define a data space and label space, indicating that these spaces are subspaces of some vector spaces of certain dimensions, and then indicate that the training dataset is of some size with the examples being drawn from some data distribution over the data space, etcs.. Putting those details in one single line feels careless. Also, what is $M$? Many would of course just simply take it as the number of classes in the classification problems, but it also needs to be specified in the paper. \n10. Line 162 to 169. There is also some notations that are inconsistent. The paper first indicates that the labels are $M$-dim vectors, then what is the input and output dimension of the encoder $h_\\phi$? The way it is presented make the model output $f_\\theta(\\text{x})$ feels like a scalar, while it should be a vector the same dimension as the labels since $\\hat{y}:=f_\\theta(\\text{x})$.\n11. Line 193 to 194. \"bounded support $\\mathcal{M}_m\\subset\\mathcal{H}$\". Is this also how the manifolds are defined in this paper?\n12. Line 194 to 195. To assume that classes are separated, it means all samples belonging to \"different classes manifolds\" should disjoint, is that correct?\n13. Line 208. Why would the mixed points' belonging \"to no vlass manifold at all\" be a bad thing? In my opinion, when using Mixup for training, especially when the data dimension is high, it's actually mostly the case that the mixed point fall into the \"void\" areas in the entire data space.\n14. Line 212. \"the higher the distance between two points, the more likely their convex combination will fall outside of the original manifolds\". Here is a counter example. Suppose we have two classes of points in $\\mathbb{R}^2$. Suppose their manifolds are two separated line segments on the horizontal axis. If we combine the rightest point from the left manifold and the leftest point from the right manifold, the mixed points will almost all fall into the middle area which is outside both the original manifolds. If we combine the leftest points from both manifold, the fact is, a much bigger proportion of mixed points would fall inside the left manifold. But, the distance between the points in the second case is much larger than that in the first case.\n15. Line 215. \"Using a Resnet18 trained ...\", is it trained using ERM or Mixup?\n16. How to choose $\\tau_{std}$ and $\\tau_{max}$?\n17. How does SK Mixup help improve calibration? The derivation of the algorithm only suggests that SK Mixup can help mitigate manifold mismatch, but is manifold mismatch the necessary reason, or even the real reason, of bad calibration?\n18. In the process of trading off between diversity and uncertainty, how would calibration and generalization behave? Will they behave like a up-side-down U curve such that there is a sweet spot?\n19. The page number of page 7, and the header \"under review\" line in page 8, there are some strange hyperlinks around the text.\n20. Theorem 3.1 (i). \"$\\lambda\\in]\\lambda_1,\\lambda_2[$\", it should be $[\\lambda_1,\\lambda_2]$ right?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. It's a novel idea to summarize several observed phenomenons or properties of Mixup into one unified terminology: manifold mismatch\n2. The correlationship between data points pair distance and the occurence of manifold mismatch is carefully derived.\n3. The proposed algorithm is meant to balance the diversity and the uncertainty of the virtual examples in Mixup, which is meaningful and helpful" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper has come up with a concept called manifold mismatch, which is a phenomenon in Mixup that can harm the generalization and calibration performance of the trained models. The paper also show that such a manifold mismatch behavior is correlated to the distance between the data points being paired and mixed in Mixup. Then, by applying a similarity kernel, the paper proposes a variant training algorithm of Mixup that can dynamically adjust the mixing coefficient depending on the distance of the data points being mixed. Finally, the authors have empirically verified the effectiveness of the proposed algorithm on various models and tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Poor presentation\n2. Doesn't clarify the connection between manifold mismatch and calibration. \n3. The core claim that distance impacts manifold mismatch is too simple to be seen, and it's actually under-justified since there seems to be much more necessary conditions of manifold mismatch like data distribution, structure of learned features and manifolds, etc.. In fact, the idea of improving Mixup by forcing more mixtures between closer points is not novel, like k-mixup [1]\n4. The idea of dynamically adjust the mixing coeffecient is also not new, like AdaMixup [2].\n\n[1] https://arxiv.org/abs/2106.02933\n\n[2] https://arxiv.org/abs/1809.02499" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Instead of using a similarity kernel introduced in this paper, a simple implementation is to divide sample pairs based on their distances (like the case in Figure 2) and use two different Beta distributions to sample the interpolation weights. How will this baseline perform compared to the proposed method? Such comparison can further verify the motivation and support the proposed method. \n- While the authors have mentioned that non-linear interpolation methods have several disadvantages (larger computational cost, limited application) in section 2.1, some empirical comparison on their performance for image data sets in Table 2-4 should still be necessary. In addition, the authors may also include some of these interpolation methods in Table 6 to verify that their computational cost are muchh larger compared to linear mixup methods. \n- Moreover, can the proposed method be combined with these non-linear interpolation methods as well? I suppose such combination will be straight-forward, as we only need to change the sampling distribution of interpolation weights for different sample pairs. Some discussion (possibly with some empirical results) will be welcome here. \n- I am a bit puzzled by the performance of different methods on ImageNet-A in Table 4. None of them has an accuracy higher than 3%, which seems worse than most methods reported in [1]. Is this due to some discrepancies in experimental setup (e.g., less training epochs)? If so, the authors may need to clarify that such discrepancies are not in favor of their proposed method. \n- Also, how are the baseline methods chosen in Table 6? From my perspective, I suppose we need to compare the time cost of SK Mixup against Mixup or some other baseline methods with rather good performance. Nevertheless, it seems either of the above cases applies, and some explanation may be needed here. \n\n## References\n[1] On Feature Normalization and Data Augmentation. CVPR 2021" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed method is clearly introduced and easy to understand" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to improve different mixup methods by adjusting the probability distribution to sample interpolation weights according to the distance between a given sample pair. Using the proposed method, sample pairs that are far from each other are less likely to be mixed. Empirical results on different tasks and data sets demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The motivation of proposed method can be strengthened\n- Some baseline methods seem missing, which can make the empirical comparison not supportive enough" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We show that taking distance into account in mixup reduces occurence of mismatch between mixed labels and mixed samples, improving confidence, calibration and robustness." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tailoring,\ntitle={Tailoring Mixup to Data for Calibration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3ygfMPLv0P},\nnote={under review}\n}" }, "abstract": { "value": "Among all data augmentation techniques proposed so far, linear interpolation of training samples, also called Mixup, has found to be effective for a large panel of applications.\n Along with improved performance, Mixup is also a good technique for improving calibration and predictive confidence.\n However, mixing data carelessly can lead to manifold mismatch, i.e., synthetic data lying outside original class manifolds, which can deteriorate calibration of confidence.\n In this work, we show that the likelihood of manifold mismatch increases with the distance between data to mix.\n To this end, we propose to dynamically change the underlying distributions of interpolation coefficients depending on the similarity between samples to mix, and define a flexible framework to do so without losing in diversity. We provide extensive experiments for classification and regression tasks, showing that our proposed method improves performance and calibration of models, while being much more efficient." }, "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": [ "mixup", "calibration", "confidence", "robustness" ] }, "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/7540be5feee1e51a57ac516850038d494e500c62.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/c054a74fb185771b89711955ece573f6b749ad94.zip" }, "title": { "value": "Tailoring Mixup to Data for Calibration" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]