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": 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": "As written, still do not understand how data is exactly generated for the classification tasks. Can you try to explain it very concisely?\n\nAlso, what train/test splits did you use? In the different training stages (generative model training, classifier training etc.)" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* applies commonly used, yet fairly recentnovel deep learning methods from other fields to EEG, where they have not been used as much\n* supplies code\n* overall approach probably makes sense (even though have not fully understood it)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript describes a vector-quantization-based autoregressive transformer as a generative model for EEG. They use their generative model to generate synthetic data for EEG classification tasks. Results show that using the real data and the synthetic data with an auxiliary loss slightly outperforms using only the real data on motor imagery decoding tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I found the way how actually the synthetic data is generated hard to understand and may still not have fully understood it. This should be more clearly described early on I am still confused about it.\n\nSome of the writing I found vague and therefore hard to read, e.g. first sentences \"Large language models (LLMs) have been extensively utilized across various scenarios due to their powerful model characteristic: the generative models. These models are not restricted to producing specific forms of output; instead, they can generate output in any form\" I found these sentences mostly confusing.\n\nIt is unclear to me how the authors split the data into training and test, I think they did not follow official train/test splits of the datasets? Did not see this information, maybe I missed it.\n\nAlso, there are of course a lot of existing works on those data that should be compared to, and for this it is also necessary to check and align the train/test split with those works and to also mention their performances.\n\nAverage power spectra (e.g., for all trials fo one subject) for synthetic and generated data should be shown to assess quality of generated EEG." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "This paper will need a lot of work, to perform all the experiments and evaluate the synthetic EEG Data.\n\nCan the authors perform simple group-level beamformer source analysis for both real and synthetic data?\n\nAdditionally Time Frequency plots for beta power changes in case of motor imagery for both generated and real data?\n\nCan we get some loss plots for the training and fine-tuning?\n\nLine [318-320] \"We focus on three channels commonly used in motor imagery experiments, and two epochs are shown to allow us to confirm the robust performance of EEGTrans across different channels and multiple epochs.\" - which 3 channels - c3, c4, cz?\n\nLine [416]\"As shown in Table 2, only EEGTrans closely matches the ground truth (real data) with minor differences, retaining high sample entropy and thus indicating high complexity and variation. However, it does not retain high-frequency components, which is evident in the spectral entropy. \" - Can we have some statistical comparison?\n\nLine[463]\"Table 3 shows the classification accuracy achieved through various approaches: using only real data, only synthetic data, combining real and synthetic data, and incorporating real data with synthetic data along with auxiliary loss. For a comprehensive analysis, we employ five-fold cross-validation instead of the original train-test split used in the competition.\" - How did you do 5 fold cross- validation? What was your distribution of the test and validation set, or it was simple 80 -20? Since it's already a competition dataset, why did you choose to do cross-validation and not cross-session analysis?\n\nCan you also do inter-subject analysis?\n\nIs it justified to benchmark against cycle gan for non-stationary time series data?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper used Transformer architecture to generate synthetic EEG. The idea is not novel as there have been previous attempts on this, however the problem is not well studied. Thus utilising transformers seems a reasonable attempt at testing the capability. \nThey Validated on multiple datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors proposed EEGTrans (Encoder - Decoder) based architecture for generating synthetic EEG datasets. It is useful, as collecting EEG data is a human subject-dependent and time-consuming process. Authors evaluated on multiple motor-imagery datasets after downsampling them to 128 Hz." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There have been other methods where authors used GPT-based architecture to claim a foundation Model: \"Neuro-GPT: Towards A Foundation Model for EEG\". There is another paper \"A Time Series is Worth 64 Words: Long-term Forecasting with Transformers\" for time series data - The authors should have evaluated that.\n\nThere should be more details on architecture, and optimisation. The authors explained for EEGTrans but for CycleGAN approach, I couldn't find the information on layers/number of parameters/etc. There is some information in the appendix but still not complete.\nI would encourage authors to include maybe clear architecture diagram for both in the main text, so it is easier to compare and understand the implementation.\n\nAdditionally, There is an extreme weakness in the paper related to evaluating synthetic data. The authors didn't perform standard evaluations like PSD, STFT, activation in alpha and beta bands, or source analysis. Averaging all the subjects doesn't distinguish activations.\n\nThere are 5 datasets for evaluation: Only 1 dataset is there with 1 subject and 128 Hz sampling. Why use that dataset to bring the other 4 datasets with higher resolution and number of subjects to 128 Hz? Maybe let go of that dataset and deal with 250 Hz, as we have tried over the years to have higher sampling and temporal resolution for these datasets \n\nIn paper, I have seen repetitive occurrences stating the same stuff, authors can reduce that, to add more information in main text. ex:\n\"Visual inspection indicates that EEGTrans produces higher-quality synthetic data compared to CycleGAN. It is evident from the\nvisual comparison that EEGTrans’s generated data is significantly superior\". \n\nanother Line [493]The BCI Competition IV Dataset 2a was collected some time ago] - \"some time ago\" really? We can say, 2a was collected in 2008 and HGD was in 2017. However the may statement is speculative in nature, so authors can point towards noise/interference due to room/environment. \n\nLine [534] : \"This method produces high-quality synthetic data that enhances downstream classification tasks.\" - If you only intended for classification (only one downstream), just train a classification model utilizing transformers?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Does the next-token prediction converge well? I'm curious about how well the next-token prediction would work. Do you have a test set to verify the next-token prediction process? If there is the test set, how well can the model correctly predict the next token? Maybe reporting metrics like perplexity will help. \n2. Following Question 1., it's possible that the model is simply yielding EEG data from the trained dataset (with some kind of bias on frequency as described in 4.5). If so, the model will be downgraded to merely replicate an existing dataset and use it to other datasets.\n3. While the model is called EEGTrans, taken the name from transformer, I feel the RVQ plays the most important role for synthesizing realistic EEG data. So what if we use the discretization method such as RVQ, but do not use the auto-regressive manner to train a transformer model? For example, one can actually combine VQ with CycleGAN to also generate good-quality EEG data.\n4. I'm not sure whether it is my problem, but when I'm trying to visit the link to the code in the abstract, it tells me that \"The repository is not found.\"" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The general idea is taken from the language model, thus the overall methodology is reasonably backed up with empirical results, and is still novel in the EEG synth task. \n2. The experiments demonstrate their superior performance compared to a classic method CycleGAN. \n3. The overall presentation is good. The authors clearly conveys their ideas." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes the transformer-based EEG generative model, namely EEGTrans, to synthesize EEG data based on multiple datasets. The model used the RVQ autoencoder to train the discretize the EEG signal and build the codebook to represent EEG tokens. Then EEGTrans can model the generative task as a next-token prediction task in an auto-regressive manner just as what a language model is doing and then the RVQ decoder can generate the EEG signal similar to the real data. Experiments confirms the similarity by comparing the spectral entropy and sample entropy. Further experiments confirms that with these synthetic data as augmentation, a model can achieve better results on the target BCI dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "First of all, I have to mention that the key research problem of the EEG synthesis is not \"what is the best method to generate the synthetic EEG data\", but \"why do we need to generate the synthetic EEG data\". The paper answers this question in the most common way, which is to use these data as augmented data to improve the performance on a specific downstream task (and in this paper the BCI motor imagery task). From this point of view, the scope of the contribution of this paper will be limited to getting improved synthetic EEG data for data augmentation rather than extending our understanding for this area. \n\nThe above statement is actually not the weakness of this paper. I like paper that won't overclaim but focus more on concrete things. However, starting from it, there will be the following concerns about the paper. \n1. It's unclear to me whether the label needs to be generated together with the EEG data. If so, the usage of the synthetic dataset seems to be limited to highly related tasks and the claimed cross-dataset advantage will be weakened.\n2. If the synthetic data is used to boost the performance of downstream tasks, then how does it perform compared to more common but simple data augmentation methods like cropping and concatenating EEG fragments, jittering, etc.? It lacks such justification. \n3. If the synthetic data is claimed to integrate information from different datasets, then how should we compare this paradigm to a more common paradigm where model such as LaBraM will integrate information in the hidden representation space? This alternative paradigm can boost the performance of downstream tasks with unsupervised training on unlabeled EEG data, which can integrate information from more diverse dataset. \n\nThere are also some comments on the methodology: \n1. EEGTrans doesn't take the relationship among channels into consideration, and doesn't evaluate the cross-channel reconstruction quality.\n2. The role that the next-token prediction plays in the proposed method is unclear and unverified. See the questions for more details." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The EEG signals have high variations between subjects and between datasets, how are these addressed in this work? Is the generation subject-dependent?\n\n2. The generation quality of the proposed will also be largely dependent on the RVQ AE, which added additional complexity to train and tune, any comment on this?\n\n3. What are the real-world applications for the proposed method? It seems like in-order to generate synthetic data for a new target dataset, you still need to use real data as input. It is very rare that we only collect 'the first half of a trial' in BCI experiments.\n\n4. For auto-regressive generation, how does the 'prediction' performance decline with respect of length?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. the structure of the paper is good and clear\n2. the method section was clearly described and the figure 2 is informative.\n3. the source codes are provided to facilitate reproduction" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a transformer-based generative model, EEGTrans to autoregressively generate EEG data based on the input sequence. The proposed model was evaluated against CycleGAN and achieved more accurate 'prediction' or generation of future segments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The model was trained for next code prediction which means the proposed method is more of a EEG signal 'forecasting' or 'prediction' rather than generation for data augmentation purpose. Therefore, the introduction and problem setup is miss aligned.\n2. For synthetic data generation to facilitate training in a new dataset, both the reality and the diversity of generated signals are important. In a typical GAN or Diffusion setting, new signals can be generated or sampled from a random vector sampled from the normal distribution during inference to achieve diversity and variety. However, in the proposed method, it seems like there is no sampling process for diverse signal generation, so how can the proposed method augment the target dataset?\n3. There are GAN-based method for EEG that can directly generate the raw signal which should be added into comparison. For example, EEG-GAN.\n4.The language used in section 4.7 is very vague, for example, 'Dataset 2a was collected some time ago', 'the synthetic data likely does not retain subject-specific information', 'the classifier should yield very similar output'. These claims need to be justified better for example, providing a tSNE plot to show the generated data vs the real data, and the target classes vs the subject to demonstrate there is no subject-specific information. \n5. Why not use MSE as a performance metric since the method is 'predicting' the future segment?\n6. The use of GAN to replace the Transformer module as baseline for next coder prediction task should be justified better." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper introduces EEGTrans, a transformer-based generative model for sequentially generating synthetic EEG signals" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024eegtrans,\ntitle={{EEGT}rans: Transformer-Driven Generative Models for {EEG} Synthesis},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ydw2l8zgUB},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in Large Language Models (LLMs) have been significant, largely due to improvements in network architecture, particularly the transformer model. With access to large training datasets, LLMs can train in an unsupervised manner and still achieve impressive results in generating coherent output. This study introduces a transformer-based generative model, EEGTrans, designed for sequentially generating synthetic electroencephalogram (EEG) signals. Given the inherent noise in EEG data, we employ a quantized autoencoder that compresses these signals into discrete codes, effectively capturing their temporal features and enabling generalization across diverse datasets. The encoder of EEGTrans processes EEG signals as input, while its decoder autoregressively generates discrete codes. We evaluate our method in a motor imagery Brain-Computer Interface (BCI) application, where merging data across datasets is particularly challenging due to experimental differences. Our results demonstrate that the synthetic EEG data effectively captures temporal patterns while maintaining the complexity and power spectrum of the original signals. Moreover, classification results show that incorporating synthetic data improves performance and even surpasses that of models based on Generative Adversarial Networks. These findings highlight the potential of transformer-based generative models to generalize effectively across multiple datasets and produce high-quality synthetic EEG signals." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "LLM", "EEG", "BCI", "transformer" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/eedb4842487c689adce0b844e0cc888d50b6032a.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": "EEGTrans: Transformer-Driven Generative Models for EEG Synthesis" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The writing is in general good and clear. I liked the motivation of Bilevel optimization. \n\n- The paper performed extensive experiments on various datasets including some real-world examples. This is clearly a strength, though this could have been much more convincing if the paper is positioned to claim contributions on the empirical side, rather than the theoretical side." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates Bayesian Bilevel Optimization through a sampling-based approach. For general bilevel optimization problems, it provides a bound on the regret of the proposed algorithm and demonstrates its effectiveness across various synthetic and real-world problems." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper's introduction is somewhat misleading. Unlike typical optimization studies focused on the computational complexity of local-search algorithms, this paper addresses statistical complexity, assuming exhaustive search over a function hypothesis space. Consequently, it diverges from the usual challenges associated with noisy, constrained, and derivative-free settings in continuous stochastic optimization, making direct comparisons with gradient-based methods inappropriate.\n\nBelow are a few detailed comments:\n\n- Line 59-60: What is \"information flow\"? This is not a well-defined term. \n\n- Line 68-69: why functions are modelled using Gaussian Process? And this assumption appears abrupt. \n\n- Gaussian process: notation is messy and discussion is hard to follow.\n\n- At the beginning of Section 3, it is not clear what you exactly get from querying points.\n\n- Line 155 \"trusted sets\": in my first read, this was very confusing. I think the paper should have been much clearer about the interaction protocol, the fact that exhaustive search over hypothesis class is okay, and it is not going to be about convergence of local-search methods. \n\n- Line 301-302: What is \"information gain\"? Define it precisely.\n\nWhile the extensive experiments in Section 4 are impressive, it remains unclear how to evaluate this section if the main contribution is claimed on methodological rather than real-world applicability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Optimization from two constructed sets can be very hard; see Step 7 of Algorithm 1. How do you implement this step while maintaining two sets in Step 17 efficiently? I guess only approximated solutions can be given." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tBilevel Bayesian optimization is an important problem with many real-world applications in machine learning, economics, energy management, and investment.\n2.\tThe proposed algorithm also works for the decoupled setting where both level functions may come from different simulators or experiments.\n3.\tBoth theoretical and empirical studies are provided. Real-world experiments are reported to positively show the effectiveness of proposed algorithm.\n4.\tOverall the writing is good and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This submission proposes a novel algorithm for bilevel Bayesian optimization that optimizes both upper and lower level problems jointly in a sample-efficient manner. The key idea of this algorithm is using confidence bounds to construct trusted sets of feasible and lower level optimal solutions. Theoretical studies show that trusted sets guarantee points with instantaneous regret bounds and sublinear regret bound is proven. Empirical studies are also provided." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tI have concerns on novelty of this submission since all techniques in this paper are pretty standard. First, functions in both levels are modelled by Gaussian process and then trusted regions are defined using confidence bounds. Then theoretical results are derived from Srinivas et al., 2010. I cannot find challenges that are unique to the bilevel Bayesian optimization, or I might have missed something. Please clarify this point.\n2.\tIn Theorem 3.9, theoretical results are given based on $|\\mathcal{F}|,|\\mathcal{X}|,|\\mathcal{Z}|$, so all of them must be finite, which put more restrictions on problem setting." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Can you explain more about the motivations for studying the decoupled setting? And the reasons to consider $\\mathcal{F}$ and function query $h_t$ is the algorithm design and regret definition? For example, can we only focus on the upper and lower-level functions in the regret definition? \n- For regret bound in Theorem 3.9, the authors discussed the relationship to constrained Bayesian optimization in Nguyen et al. (2023), is there any existing bilevel BO regret result that can be compared with? \n- In experiments, BILBO is compared with TrustedRandom and Nested. As described in Appendix C.1, the Nested baseline used a nested approach with SLSDP. But the related works mentioned, e.g.  Kieffer et al. (2017); Dogan & Prestwich (2023) are not included. Can you compare it with the associated works directly? If not, can you explain the reasons that BILBO cannot compared with the related work? \n- In Figure 1a, can you explain why BILBO has a sudden drop at around 150 queries? And what is the regret of BILBO after 150 queries (cannot observe the line afterwards)? The uncertainty level looks huge as well, I am wondering what would happen if you increase the number of runs (or change random seeds/initialisations, etc). \n- In experimental results, we observe Nested is generally similar to or worse than TrustedRand (which is only a random sampling), can you explain the possible reason? Does it suggest Nested is not a strong baseline (as it is from a previous work)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper proposes a new algorithm, BILBO, specifically tailored for general bilevel optimization problems, where simultaneous optimization at both levels is essential.\n- The authors derive and prove a sublinear cumulative regret bound for the decoupled setting, enhancing the theoretical understanding of bilevel optimization.\n- Experiments are conducted on both synthetic and interesting real-world scenarios, demonstrating the practical applicability of BILBO.\n- The writing is generally clear and well-structured, making the paper accessible to readers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces BILBO, a novel bilevel Bayesian optimization framework that simultaneously optimizes functions at both upper and lower levels. The algorithm employs trusted sets based on confidence bounds at each level. The authors establish a sublinear cumulative regret bound in a decoupled setting and validate BILBO through experiments on simulated and real-world problems, using TrustedRand and Nested policies as baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the algorithm is well-explained, the motivations for certain design choices, particularly the decoupled setting and the specific function query mechanisms ($\\mathcal{F}$ and $h_t$), are not sufficiently discussed. A deeper exploration of these choices would help readers understand the broader implications and advantages of BILBO’s design.\n- The paper’s discussion of related work is somewhat limited. A more detailed comparison with existing bilevel optimization methods, particularly those offering theoretical guarantees or practical performance benchmarks, would provide a clearer picture of BILBO’s relative strengths and weaknesses.\n- The baselines used in the experiments—TrustedRandom and Nested—are not the strongest available. Including comparisons with state-of-the-art bilevel optimization algorithms, such as those from Kieffer et al. (2017) or Dogan & Prestwich (2023), would offer a more rigorous evaluation of BILBO’s performance.\n- No code is provided alongside the paper. This omission hampers the reproducibility of the results and limits the ability of the research community to build upon or validate the findings presented." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Have the authors considered a hybrid setting, where some parts of the bi-level problem are black-box while other parts are white-box with explicit expressions? I guess such problems also arise widely in practice." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The problem addressed in this work is significant, and the proposed algorithm introduces a novel approach. Most existing bi-level Bayesian optimization algorithms tackle the problem using a nested framework, where each upper-level query requires separately optimizing the lower-level problem to convergence. In contrast, the proposed algorithm optimizes both levels concurrently, greatly enhancing sample efficiency, as shown in the experiments.\n2. An infeasibility declaration method is also included.\n3. Additionally, the algorithm achieves a theoretical sublinear regret bound, extending the classic bound from single-level to bi-level problems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new algorithm designed to solve bi-level constrained Bayesian optimization problems. It demonstrates that the algorithm achieves a sublinear regret bound and provides experimental results on both synthetic and real-world datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This approach involves keeping track of a set of nearly optimal solutions for the lower-level problem. In low-dimensional cases, this can be managed through discretization. However, it likely becomes difficult to scale effectively to even moderately high-dimensional problems." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Are there any problems with extending the proposed method for an infinite input set? I believe that we can extend it by relying on the discretization arguments (as in [1]) under the four-times differentiability assumption of the kernel and the compactness of the input set.\n- Why does the author focus on the cumulative regret in Theorem 3.9? As with the existing works (e.g.,[2,3])\nthe upper bound of the simple regret (or the stopping time upper bound for finding $\\epsilon$-accurate solution) seems to be derived based on the pessimistic estimated solution. I believe that simple regret is a more suitable performance measure from the optimization perspective.\n\n[1] Srinivas, Niranjan, et al. \"Gaussian process optimization in the bandit setting: No regret and experimental design.\" arXiv preprint arXiv:0912.3995 (2009).\n\n[2] Bogunovic, Ilija, et al. \"Adversarially robust optimization with Gaussian processes.\" Advances in neural information processing systems 31 (2018).\n\n[3] Kirschner, Johannes, et al. \"Distributionally robust Bayesian optimization.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2020." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Overall, the writing of this paper is clear, well-structured, and easy to follow. \n- Comprehensive proofs are provided, and the proofs seem correct, as far as I can see.\n- As far as I know, the proposed algorithm is the first theoretically guaranteed approach for bilevel Bayesian optimization. \n- The practical behavior of the algorithm is well-described in Figures 1 and 2, which will be helpful for practitioners." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the black-box bilevel optimization problem based on the Gaussian process (GP) model.\nThe author proposes an algorithm that leverages the confidence set of the underlying objective function $f$ under the Bayesian assumption (i.e., $f$ is the sample path of GP.)\nThe proposed algorithm achieves $O(\\sqrt{\\gamma_T T})$ cumulative regret upper bound whose \ninstantaneous regrets are defined as the extension of existing work [1] for the formulation of the bilevel optimization. The numerical experiments are conducted, including two problems motivated by real-world problems.\n\n[1] Nguyen, Quoc Phong, et al. \"Optimistic Bayesian Optimization with Unknown Constraints.\" The Twelfth International Conference on Learning Representations. 2023." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concern is the lack of novelty. The confidence bound-based approach for \ncomplex structured optimization problems have already been extensively studied \nin the BO field (e.g., robust BO [1,2,3], constrained BO [4,5], and composite objectives [6, 7], etc.).\nSpecifically, it seems that the proposed algorithm can be constructed by combining the extended regret definition of [4] and the sampling for potential solutions based on a confidence set. As far as I see, the non-trivial points of the proposed algorithm construction are the following:\n\n1. The construction of $\\mathcal{P}_t$ with $\\overline{z}$ (Eq. 3.4, 3.5). Specifically, the natural construction of the potential solution set $\\mathcal{P}\\_t$ is based on maximum $\\max l\\_{f,t}(x, z)$; however, we cannot bound $r\\_{f}$ with such construction.\n2. Reassignment of $z_t$ based on Lemma 3.6.\n\nI believe that the other parts of the algorithm construction and analysis are not novel for the reader who studies related BO fields (e.g., robust BO, constrained BO, etc.) since they only use the basic, well-known result of the existing algorithm and are naturally derived. \nFurthermore, uncertainty-based input selection by fixing one input, such as the procedures of point 2, is also commonly leveraged to upper bound the instantaneous regret in the BO field (e.g., [2,3]). \n\nFor the above reasons, I slightly lean my score toward rejection.\n\n[1] Bogunovic, Ilija, et al. \"Adversarially robust optimization with Gaussian processes.\" Advances in neural information processing systems 31 (2018).\n\n[2] Kirschner, Johannes, et al. \"Distributionally robust Bayesian optimization.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2020.\n\n[3] Iwazaki, Shogo, Yu Inatsu, and Ichiro Takeuchi. \"Mean-variance analysis in Bayesian optimization under uncertainty.\" International Conference on Artificial Intelligence and Statistics. PMLR, 2021.\n\n[4] Nguyen, Quoc Phong, et al. \"Optimistic Bayesian Optimization with Unknown Constraints.\" The Twelfth International Conference on Learning Representations. 2023.\n\n[5] Xu, Wenjie, et al. \"Constrained efficient global optimization of expensive black-box functions.\" International Conference on Machine Learning. PMLR, 2023.\n\n[6] Li, Zihan, and Jonathan Scarlett. \"Regret bounds for noise-free cascaded kernelized bandits.\" arXiv preprint arXiv:2211.05430 (2022).\n\n[7] Xu, Wenjie, et al. \"Bayesian optimization of expensive nested grey-box functions.\" arXiv preprint arXiv:2306.05150 (2023).\n\n(Minor)\n- I believe that there is further room to enhance the paper's quality. For example:\n - The existing work [1] studies the particular case of the problem setting of this paper. The comparison with [1] will make the position and novelty of this paper clearer.\n - Finiteness assumptions of $\\mathcal{X}$ and $\\mathcal{Z}$ should be explicitly described in Section 2.\n - The references should be appropriately capitalized (e.g., bayesian -> Bayesian).\n - The notation $|\\mathcal{X}|$ usually denotes the cardinality of the set, not the dimension. I recommend using another notation.\n - Definition of the maximum information gain and its explicit upper bound for several commonly used kernels (SE or Mat\\'ern) is beneficial for the reader who are not familiar with theory.\n - The vector graphics figures (such as .pdf, .svg) are favorable.\n - The result of Section.4 seems yet unstable in $5$ trials.\n- (typo) Corollary 3.2: $[u\\_{h,t}(x, z), l\\_{h,t}(x, z)]$ -> $[l\\_{h,t}(x, z), u\\_{h,t}(x, z)]$\n- (typo) Line 107: The definition of the posterior mean misses the prior mean $m_h(\\cdot)$.\n- (typo) Line 108: \\sigma^2 I -> \\sigma^2 \\bm{I}" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024bilbo,\ntitle={{BILBO}: {BIL}evel Bayesian Optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ye1mxb79lw},\nnote={under review}\n}" }, "abstract": { "value": "Bilevel optimization, characterized by a two-level hierarchical optimization structure, is prevalent in real-world problems but poses significant challenges, especially in noisy, constrained, and derivative-free settings. To tackle these challenges, we present a novel algorithm for BILevel Bayesian Optimization (BILBO) that optimizes both upper- and lower-level problems jointly in a sample-efficient manner by using confidence bounds to construct trusted sets of feasible and lower-level optimal solutions. We show that sampling from our trusted sets guarantees points with instantaneous regret bounds. Moreover, BILBO selects only one function query per iteration, facilitating its use in decoupled settings where upper- and lower-level function evaluations may come from different simulators or experiments. We also show that this function query selection strategy leads to an instantaneous regret bound for the query point. The performance of BILBO is theoretically guaranteed with a sublinear regret bound and is empirically evaluated on several synthetic and real-world problems." }, "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": [ "bilevel", "Bayesian 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/83ffc7ed4a89f7c18c147755d24cff788136f272.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": "BILBO: BILevel Bayesian Optimization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Is the proposed method applicable to other types of data beyond images, such as text or time series data? If so, what modifications, if any, would be necessary?\n\nHow does the method scale with larger models and datasets? Are there computational bottlenecks, and if so, how can they be mitigated?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper presents a novel unsupervised framework for discovering latent space directions that are both interpretable and influential. This is an advancement over prior work that often requires annotated datasets or predefined concepts.\n\nThe method is theoretically grounded, extending previous models to a multi-label setting and introducing new loss functions such as the Uncertainty Region Alignment loss. The experimental results on synthetic and real-world data support the efficacy of the approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an unsupervised method for discovering interpretable and influential directions in the latent space of deep neural networks, specifically for image classifiers. The proposed approach leverages signal vectors and uncertainty region alignment to identify latent space directions that significantly influence model predictions while maintaining high interpretability. Unlike previous methods that require annotated concept datasets and predefined concepts, this method does not rely on prior knowledge and instead learns from the inherent structure of the feature space. The authors validate their approach on both synthetic data and real-world benchmarks, demonstrating that the discovered directions effectively fulfill critical debugging criteria and outperform supervised methods in certain aspects." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the experiments are promising, the evaluation is limited to a synthetic dataset and a single real-world dataset (Places365 with ResNet18). A broader range of datasets and models would strengthen the claims and demonstrate the generalizability of the method.\n\nThe paper could provide a more thorough comparison with existing methods, including additional baselines in both interpretability and influence metrics. This would help in positioning the proposed method within the existing literature more effectively." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- Are the learned directions always labeled by NetDissect? I would appreciate some clarity on how NetDissect is being applied to the proposed method and its necessity in generating explanations. If so, why is the proposed method preferable to simply using NetDissect on its own?\n\n- Please see above weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The problem of finding concepts or latent directions that are highly influential, which I take to mean causal, is an interesting and very active research area at the moment. Being able to use interpretability tools to exert influence over a model would be of high utility for future research and real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method to learn interpretable and influential directions in a trained model’s latent space, where these directions are modeled as linear “signal” classifiers. The authors propose a method called Uncertainty Region Alignment that aligns the subspaces where the model is uncertain and the subspaces where the learned “signal” classifiers are also uncertain, with the claim that this enhances the interpretability and influence of the learned directions. They present results from a small synthetic setting and from a slightly larger, more realistic setting on a single model and dataset (Resnet18, Places265), where they benchmark against IBD and find that their method produces better concept detectors." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- As it stands, it was difficult for me to understand the contribution of this work with respect to prior literature in this field. Further discussion of how this paper relates to the wide body of recent literature on generating concept-based explanations, including follow-up works to ACE/ICE/IBD such as CRAFT [1], work on Concept Bottleneck Model literature [2, 3, 4, 5], and recent work on sparse autoencoders and dictionary learning, would help clarify the gap in existing literature that this paper is trying to fill.\n- Furthermore, it seems like this paper is a close follow-up to the papers by Doumanoglou et al. (2023; 2024) cited in the paper, in that the majority of the method is the same with some minor adjustments. Can you provide further discussion of how this method and its contributions differs from the methods in those two papers?\n- The writing and presentation of this paper are not very clear to me - I struggled to understand the types of explanations given by the proposed method, and how those explanations could be used to gain a “deeper understanding of the model’s strategy, fostering trust, and enabling model correction and improvement” as stated by the authors in the abstract. Clarifying the intent of the proposed method and how the experiments validate its practical applications with respect to the stated goals from the abstract and introduction would significantly improve the paper.\n- Given that this paper is an explainability/interpretability paper, I believe it could strongly benefit from providing some example explanations yielded by the paper, and a more in-depth example or case study of how the explanations can be applied to the aforementioned end goals of model correction.\n\n[1] Fel, Thomas, et al. \"Craft: Concept recursive activation factorization for explainability.\" *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*. 2023.\n\n[2] Yuksekgonul, Mert, Maggie Wang, and James Zou. \"Post-hoc concept bottleneck models.\" *arXiv preprint arXiv:2205.15480* (2022).\n\n[3] Bhalla, Usha, et al. \"Interpreting clip with sparse linear concept embeddings (splice).\" *arXiv preprint arXiv:2402.10376* (2024).\n\n[4] Oikarinen, Tuomas, et al. \"Label-free concept bottleneck models.\" *arXiv preprint arXiv:2304.06129* (2023).\n\n[5] Gandelsman, Yossi, Alexei A. Efros, and Jacob Steinhardt. \"Interpreting CLIP's Image Representation via Text-Based Decomposition.\" *arXiv preprint arXiv:2310.05916* (2023)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The \"M\" term used in the maximum margin loss does not appear in any other loss equations (except perhaps in equation 5, where it seems to be a different \"M\"). Can you please clarify how the margin loss \"M\" is calculated?\n\n- Line 451 mentions \"Significant Direction Count (SDC) and Significant Class-Direction Pairs (SCDP). SDC represents the number of learned signal vectors that significantly influence at least one of the model’s classes, while SCDP counts the total number of class-direction pairs in which the learned signal vector significantly affects the class.\" -> what is \"significantly influenced\" here?\n\n- Conceptual clarification: Is the assumed data generation process in equation 1, identifiable? For example, if s_1 and s_2 are two signal directions, then is s_1 + s_2 also a signal direction? \n\n- Conceptual clarification: given a candidate direction \"v\", what is a simple test to determine whether it is a signal or a distractor direction?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ The topic of extracting a set of contributing semantic concepts underlying a representation is an important topic, that can potentially help efforts in debugging and controlling model behaviour. \n\n+ The proposed method is somewhat empirically successful, in that it is able to recover ground truth signals in synthetic settings; and recover signal vectors that score high on numerical interpretability metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to automatically learn the set of concept vectors underlying a given model representation in an unsupervised manner. Specifically, given a model representation and an unsupervised image dataset, the method aims to extract a set of \"signal\" vectors that explain the representation, and are semantically meaningful. Experimental results show that (1) in synthetic settings where the ground truth signals are known, the proposed method recovers the ground truth; and (2) when applied for an image classification setting, the extract signal vectors score high on numerical interpretability metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper is **hard to read**. At times, the issue is missing details, and at other times, it makes lots of references to text in the appendix and/or other papers. For instance, the method defines six heuristic loss functions, however it is unclear how they are combined and how the regularization parameters are set. The exact definitions and motivations for these are sometimes buried in the appendix. In addition, several terms defined in individual papers (Doumanoglou et al. (2023), Pfau et al. (2020), and Pahde et al. (2024)) are used without defining or explaining them. In particular, the contributions in these listed individual papers (S^1;S^2 interpretability metrics, RCAV, and Pattern CAV) are unexplained and used without context in the paper. I recommend that the authors rewrite these sections by clearly explaining the method, including being explicit about the exact loss functions used and the optimization problem being solved. In addition, it would help if terminology from other papers is first clearly defined before being used. \n\n**Missing Comparison with PCA / ICA / Dictionary learning**: One of the goals of this paper is to recover signal and distractor vectors that combine additively to generate the underlying inputs. However, this is precisely the setting in which classical methods such as PCA / ICA / dictionary learning are applicable (depending on the exact underlying assumptions re: orthogonality, etc), and the missing discussion and comparison with these methods in the context of this paper is a significant omission. An experiment comparing the proposed method with such classical baselines would help build the case for the proposed approach. In addition, it would be helpful to point out what is conceptually missing in such classical approaches that necessitates the present one.\n\nIt is **unclear whether recovered signals are interpretable**, in the sense that an unambiguous assignment exists from the signal vector to a semantic concept. In the interpretability literature, including network dissection (Bau et al., 2017), it has been discussed that not all nodes are interpretable in the sense that an assignment to a semantic concept is not always possible. If the authors can present visualizations of the set of recovered signal vectors and discuss what makes them interpretable, this would help claim the interpretability benefits of this approach. \n\n**Minor**: The paper claims that its method overcomes a common weakness of concept-based methods, i.e., the requirement of annotated concept datasets (line 88). However, even the proposed method uses the Broyden concept dataset to annotate concept labels (line 412), and without this, it is unclear how the method recovers interpretable signals." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "N/A" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The attempt to discover influential directions in a model's latent space without relying on labeled data is good. If it works, it will make the method more scalable compared to other techniques that require manual annotation or predefined concepts.\n\n2. The method has been tested not just on synthetic datasets but also on real-world tasks using state-of-the-art models like ResNet18. \n\n3. The writing can be improved by incorporating more motivations in introducing each loss." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a method to identify interpretable and influential directions in latent spaces of deep learning models. Since the latent space directions play a crucial role in understanding and correcting deep models, the goal here is to identify directions in the latent space that influence model predictions while being interpretable. Traditional approaches for finding these directions rely on supervised methods and predefined concepts. This paper introduces an unsupervised framework that combines signal vectors with uncertainty region alignment to discover interpretable latent directions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I commend the authors for developing an unsupervised method to identify interpretable directions, the approach faces a key challenge. Supervised methods generally ensure that these directions align closely with actual concepts, providing a clearer sense of interpretability. But unsupervised methods can not guarantee that. \n\nThe paper lacks experiments that measure cosine similarity with \"ground truth concept vectors\" on real datasets, such as Places365. This raises doubts about the authenticity of the learned vectors, as evidenced by the significant performance gap between unsupervised and supervised methods in Table 4. \n\nPutting together, if the learned \"concept vector\" doesn’t reliably represent the true concept, the claim of interpretability becomes less compelling." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 W1 to W4." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The presented method is very well theoretically motivated and is a valuable extension to existing (supervised) discovery methods like CAV.\n- The experiments have been carried out accurately and with care. However, I would classify both experiments more as proof-of-concepts than real-world experiments as the places365 dataset is very limited in complexity. However, this is totally valid in my opinion, as the focus of the paper is the theoretical contribution.\n- The ablation studies on the loss components are extensive and interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an unsupervised method for discovering interpretable and influential latent space directions in deep-learning models using signal vectors and uncertainty region alignment. The approach is validated on synthetic and proof-of-concept real-world data, showing its effectiveness in recovering key signal directions and improving model debugging." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**W1:** The paper is sometimes hard to follow and understand. While in the methods section, the equations help to understand what’s described in the text, the different experiments are sometimes hard to follow. Especially for a reader unfamiliar with latent space-based concept methods for interpretability the different methods and the interpretation of their results can be confusing. For example, the experiment setup in 4.1 and how it connects to section 2 could be better explained. The experiment setup in 4.2 refers to several methods from related work (e.g. Network Dissection, Broden dataset, RCAV (which was never introduced)), which readers are possibly not familiar with. \n\nTo improve clarity, I suggest adding an overview of key concepts at the start of the experiments section, explicitly explaining how Section 4.1 connects to Section 2, and providing brief explanations for unfamiliar methods (not limited to) like Network Dissection, the Broden dataset, and RCAV when they are first introduced.\n\n**W2:** The authors comment in Line 297 that experiments with the raw inclusion of the three loss terms do not converge, but do not further elaborate on how stable the final results are with their adaptations, i.e. if the learned directions are robust. \n\nPlease provide quantitative measures of the robustness in the learned directions, e.g. through the standard deviation or confidence intervals over several differently seeded runs for the tables in section 4 (if applicable).\n\n**W3:** Figures 1 and 2 are very helpful in communicating the concepts but are never referred to in the main manuscript. The paper would benefit from a general overview figure in the same style (in which these two figures would be a part), similar to a graphical abstract to communicate the different steps and concepts visually.\n\nPlease provide an overview figure on how the presented method is composed (and maybe why certain selections were made) in section 3. If it would make sense for you, also integrate Figure 2 into this overview figure.\n\n**W4:** The paper lacks a published code repository, making it difficult for others to easily apply the method to their own models.\n\nPlease provide a link to an anonymous repository with a (simple) implementation of the method. You can use \"Anonymus Github\" or GitLab with an anonymous email. \n\nI am open to raising my score if the identified weaknesses are either adequately addressed through revisions to the manuscript or convincingly argued to be non-issues." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "The proposed unsupervised method identifies a pair of latent space directions (filter and signal) with the first being able to answer questions of interpretability and the second to answer questions of concept influence on model's predictions" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning Interpretable and Influential Directions with Signal Vectors and Uncertainty Region Alignment},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yeEWZ8qvlS},\nnote={under review}\n}" }, "abstract": { "value": "Latent space directions have played a key role in understanding, debugging, and fixing deep learning models. Concepts are often encoded in distinct feature space directions, and evaluating impact of these directions on the model's predictions, highlights their importance in the decision-making process. Additionally, recent studies have shown that penalizing directions associated with spurious artifacts during training can force models to unlearn features irrelevant to their prediction task. Identifying these directions, therefore, provides numerous benefits, including a deeper understanding of the model's strategy, fostering trust, and enabling model correction and improvement. We introduce a novel unsupervised approach utilizing signal vectors and uncertainty region alignment to discover latent space directions that meet two key debugging criteria: significant influence on model predictions and high level of interpretability. To our knowledge, this method is the first of its kind to uncover such directions, leveraging the inherent structure of the feature space and the knowledge encoded in the deep network. We validate our approach using both synthetic and real-world benchmarks, demonstrating that the discovered directions effectively fulfill the critical debugging criteria." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "latent space", "interpretability", "concepts", "directions", "signals", "patterns", "distractors" ] }, "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/1bac30e18f7715bcb1333484a7bedb4d4b6efd9c.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": "Learning Interpretable and Influential Directions with Signal Vectors and Uncertainty Region Alignment" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I am wondering the results or analysis for more extreme expert reduction scenarios, such as reducing to 25% or 10% of the original experts. This would give insight into how the method performs under more aggressive compression." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. An output based similarity metric of expert clustering is proposed, which is more effective than previous works.\n2.The experimental results compared with the previous methods are very good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a retraining free experts merging approach which employs a hierarchical clustering strategy. The authors claim that using expert outputs as the similarity metric for clustering is more effective compared with using router logins or weights employed by prior works. The experimental results reveal the proposed approach gains more performance improvements compared with existing methods across various benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There is a lack of theoretical analysis on the performance of expert clustering using different similarity metrics." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 Weakness." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- HC-SMoE offers a practical solution for reducing parameters without the need for retraining, simplifying the implementation process.\n\n- The task-agnostic nature of HC-SMoE allows for broader applicability across different language tasks, enhancing its versatility.\n\n- The comprehensive experiments conducted on eight zero-shot language tasks provide strong empirical evidence of HC-SMoE's effectiveness in large-scale models like Qwen and Mixtral." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces Sparse Mixture-of-Experts (SMoE) models, which improve large language model performance without significantly increasing inference costs by activating only a subset of parameters. However, their high memory requirements hinder deployment. To address this, the authors propose Hierarchical Clustering for Sparsely activated Mixture of Experts (HC-SMoE), a task-agnostic framework that reduces SMoE parameters without retraining." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While this work demonstrates competitive accuracy, it lacks a comprehensive assessment of efficiency metrics, such as speedup and memory usage. Given that efficiency is a key contribution, this aspect of the experimental results is essential.\n\n- A theoretical analysis of the effectiveness of expert merging and HC-SMoE would enhance the understanding of the method's performance.\n\n- Although HC-SMoE is validated on eight zero-shot language tasks, its effectiveness may vary in more complex tasks or domains, potentially limiting its broader applicability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to my question 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 is simple but effective, \n* The paper is very easy to follow.\n* The experiments are comprehensive and the results are very promising." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a new expert merging framework, named Hierarchical Clustering for Sparsely activated Mixture of Experts (HC-SMoE), to reduce SMoE model parameters without retraining. The proposed method is simple but effective, and the experiments demonstrated the efficacy of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The motivation of using the \"Hierarchical\" clustering is not clear to me. I cannot intuitively get the idea of why hierarchical clustering is better than simple K-means clustering, although the results confirmed that K-means clustering is less effective. Besides, the paper proposed to use a \"hard\" hierarchical clustering, and I am wondering if it is more effective to use \"soft\" hierarchical clustering or simply \"soft\" clustering without hierarchies. \n* The choice of the calibration dataset. I did not see any ablation study about the choice of the calibration dataset, and I think the performance of the proposed method should highly depend on the calibration dataset. If the calibration dataset is not comprehensive enough, e.g., not covering enough domain specific data, the clustering may not be very informative, which may lead to poor performance. For example, if we want the LLM to perform well on a law-related or medical-related tasks, can you also rely on the same calibration dataset used in the experiments?\n* Some minor issues:\n * In Fig. 1, why you did not compare the methods on 14B model?\n * Section 3.2.1 presents the method of similarity metric but contains a lot of discussions about related work.\n * In line 299/300, is alpha_i fixed or not? If it is fixed, will it also suffer from the issue that you mentioned in line 199-203 about frequency-based method?\n * In Table 4, the best performance of 'ARC-c' should be the Average linkage using the Weight setting, right?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) Pruning experts in MoE models can indeed reduce the difficulty of deployment.\n2) The paper is easy to follow, and the ablation study is comprehensive.\n3) The experimental results on Qwen and Mixtral are convincing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Hierarchical Clustering for Sparsely Activated Mixture of Experts (HC-SMoE), a task-agnostic framework for merging experts within an SMoE model. HC-SMoE aims to reduce the model's parameters without requiring retraining. Experiment results on a series of benchmarks show its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) O-prune [1] requires enumerating all possible combinations of experts, resulting in significant time overhead. I would like to know how HC-SMoE compares to other approaches in terms of runtime and resource consumption.\n2) O-prune [1] also conducts experiments on domain-specific tasks (e.g., GSM8K, Math). I am interested in the performance of HC-SMoE on these datasets.\n\n[1] Lu, Xudong et al. “Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models.” Annual Meeting of the Association for Computational Linguistics (2024)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Our method (HC-SMoE) offers an efficient method for merging experts of large Sparse Activated Mixture of Experts (SMoE) models without retraining under task-agnostic settings." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024retrainingfree,\ntitle={Retraining-Free Merging of Sparse Mixture-of-Experts via Hierarchical Clustering},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yeeIGM3N6w},\nnote={under review}\n}" }, "abstract": { "value": "Sparse Mixture-of-Experts (SMoE) models represent a significant breakthrough in large language model development. These models enable performance improvements without a proportional increase in inference costs. By selectively activating a small set of parameters during task execution, SMoEs enhance model capacity. However, their deployment remains challenging due to the substantial memory footprint required to accommodate the growing number of experts. This constraint renders them less feasible in environments with limited hardware resources. To address this challenge, we propose Hierarchical Clustering for Sparsely activated Mixture of Experts (HC-SMoE), a task-agnostic expert merging framework that reduces SMoE model parameters without retraining. Unlike previous methods, HC-SMoE employs hierarchical clustering based on expert outputs. This approach ensures that the merging process remains unaffected by routing decisions. The output-based clustering strategy captures functional similarities between experts, offering an adaptable solution for models with numerous experts. We validate our approach through extensive experiments on eight zero-shot language tasks and demonstrate its effectiveness in large-scale SMoE models such as Qwen and Mixtral. Our comprehensive results demonstrate that HC-SMoE consistently achieves strong performance, which highlights its potential for real-world deployment." }, "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": [ "Sparse Mixture-of-Experts", "Merging", "Compression" ] }, "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/a23868b6c05aeef8a7957e0731f262ca0a2934ce.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": "Retraining-Free Merging of Sparse Mixture-of-Experts via Hierarchical Clustering" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How does CodeLutra perform on longer programs, e.g,, competitive coding? (This is nice to have; the material in the paper is enough for publication.)\n- What is the CodeLutra's performance vs program length on the current data sets?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- CodeLutra simple yet effective method, with clear articulation of how it differs from related work.\n\n- Impressive results: with only 500 samples, CodeLutra achieves GPT-4-level performance on a base model with just 8 billion parameters. For the Spider benchmark, it improves base model performance from 59.3 to 74.4 in just four iterations, surpassing GPT-4’s 74.4. On BIRD, it increases performance from 22.3 to 42.6 in four iterations, approaching GPT-4’s 46.3.\n\n- Comprehensive evaluation, covering three coding benchmarks (Spider, BIRD, and DS-1000) and three models (Llama-3-8B, Gemma-7B, and StarCode-7B), demonstrating the approach's generalizability.\n\n- Strong ablations that address key questions: (i) dual loss significantly boosts performance, raising it from 17.2 (DPO) to 76.6 on Spider; (ii) negative samples are crucial, as performance increases from 20 to over 40 with their inclusion, while positive samples alone yield minimal improvement." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents CodeLutra, a supervised fine-tuning (SFT) approach that demonstrates significant improvements on coding tasks. Specifically, CodeLutra achieves GPT-4-level performance fine-tuning an open-source 8-billion-parameter model using as few as 500 samples (with groundtruth solutions). The key idea behind CodeLutra is to use both positive and negative examples to fine-tune the model, creating a hybrid of SFT and DPO techniques. CodeLutra assumes a ground truth to generate these examples: if a sample code produces the same inputs/outputs as the ground truth, it is a positive example; otherwise, it is a negative example." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Current evaluation focuses on SQL queries and data science problems, which are relatively short (from a few lines of code to several 10s of lines of code). It would be interesting to see how this approach generalizes to longer programs.\n- Limited exploration of scenarios without ground truth. In such cases, CodeLutra relies on syntactic error detection, but the results are, as expected, less impressive." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- In line 150, \"if the model only predicts wrongly in the final token in a code snippet, the overall probability P (y|x) in the Equation 1 might still remain high as the preceding tokens are correct\". While the hypothesis makes sense, do you really observe this situation in real LLM and dataset? I doubt it.\n- Equation 6 is studied in a previous literature \"Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer\" with theoretical support, but not cited. This also limits the novelty contribution (at least on this part).\n- I'm confused with the experimental setup. What is the training dataset? It seems the experiment is using the test dataset to train. Could you clarify?\n- Could you explain the setting for SFT in Table 1? One baseline is the SFT model that only uses the groundtruth training solutions, or use the synthetically generated correct solutions. Which one are you using?\n- I don't think Table 2 is a right setting, where 17.2 and 12.4 is extremely low for DPO-only method. Normally we do DPO training on top of SFT model. The right setup should be training on top of SFT model. What is the gap between, SFT then DPO training and the SFT regularized preference training?\n- 500 samples might mean that 500 prompts or problems. What is the size of generated samples overall?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is well-written. The proposed method with training with correct and failed generations iteratively makes sense. Experiments show good improvement on benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method by iteratively generating successful and failed code and training with preference optimization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Some experimental setup is not clear enough, such as training data, SFT setting, and details of synthetically generated dataset. One of the contribution DPO and SFT loss is studied in previous literature. More experiments might be needed for comparing SFT then DPO with DPO+SFT loss." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "NA" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "NA" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces CODELUTRA, a framework designed to enhance the performance of LLMs in code generation tasks. However, the method is almost the same as an existing method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method closely resembles that presented in [1]. Applying the same approach to a different scenario does not warrant publication, especially since this new scenario is simpler and benefits from execution feedback. \n\n[1] Iterative Reasoning Preference Optimization. https://arxiv.org/abs/2404.19733" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "* Are the two answers in Figure 5 flipped? Given Currency is in table customers, I feel the first answer is correct, and the second is wrong." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Comprehensive evaluation, ablation, and analysis support the effectiveness of the proposed method. In particular, the necessity of negative training samples and of SFT loss are both well studied.\n* The paper is well written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes CODELUTRA, a preference-guided training framework to let code LLMs iteratively refine itself based on execution signals from its own generations. Specifically, given a task-specific training set, at each iteration, the model generates answers which are then evaluated by unit tests. Each correct answer is paired with an incorrect answer to form a preference pair. The preference dataset is then used for DPO training. To address the issue that DPO may reduce the generation probability of both correct and incorrect answers, supervised finetuning loss is added to DPO loss for joint training. Experiments show that CODELUTRA significantly improves performance on SQL and data science tasks, and is much more effective than DPO alone." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The technical novelty of this paper is somewhat limited. L233-246 claimed two major points of novelty: refinement from execution feedback and dual loss mechanism. First, using feedbacks from program execution to iteratively refine code LLMs is a direction that has been extensively studied (e.g., CodeRL [1], and NExT [2]). However, these works are not discussed in the related work section. Second, the dual loss objective (i.e. adding SFT loss in DPO training) was proposed in [3], known as RPO, which is not cited.\n* I find DS-1000 Pass@1 results in Table 1 are inconsistent with the public leaderboard (https://ds1000-code-gen.github.io/model_DS1000.html). In particular, pass@1 of Codestral-22B and Llama-3-70B-Chat is 51.2 and 48.6 respectively in the leaderboard, but 35.8 and 36.4 respectively as reported in the paper.\n\n\nReference:\n\n[1] Le, Hung, et al. \"Coderl: Mastering code generation through pretrained models and deep reinforcement learning.\" Advances in Neural Information Processing Systems 35 (2022): 21314-21328.\n\n[2] Ni, Ansong, et al. \"NExT: Teaching Large Language Models to Reason about Code Execution.\" Forty-first International Conference on Machine Learning.\n\n[3] Liu, Zhihan, et al. \"Provably mitigating overoptimization in rlhf: Your sft loss is implicitly an adversarial regularizer.\" arXiv preprint arXiv:2405.16436 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. See weakness 1, 2\n2. I am curious that if this method can lead to a model that is generalizable. For example, the authors split DS-1000 for training and evaluation. I wonder how the resulting model would perform on other similar datasets, e.g. MBPP?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-organized and easy to follow\n2. The proposed method can lead to a fine-tuned LLAMA3-8B model which has comparable performance to GPT-4.\n3. The authors conduct comprehensive ablation studies that the effect of every component involved in their method is clearly demonstrated.\n4. The method can still have good performance with limited annotations or training samples." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors proposed a new training framework called CODELUTRA which aims to fine-tune a small CodeLM to match or surpass the closed-source LLMs like GPT-4. CODELUTRA adopts an iterative method to learn by comparing the correct generation and the failed generation. At each iteration, CODELUTRA constructs the preference dataset by classifying the generation codes of the model from the last iteration and employs a dual-loss function that combines DPO with SFT for training. The authors show that their method can achieve a performance comparable to GPT-4 in the data query and data science tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Line 230 states that \"The refinement process continues until the improvement between consecutive iteration becomes marginal\". However, in the experiments, the authors seem to fix the iteration number to 4. In practice, how do you decide if the improvement between consecutive iterations is marginal?\n\n2. The baseline setup is not clear enough and may not be comprehensive.\na) For closed-source LLMs, it is unknown what prompting method is used. It is also not clearly stated what fine-tuning method is used. From the Appendix, I infer that the LoRA is used in CODELUTRA but is it also used in the fine-tuning baseline?\nb) In Table 1, since LLAMA-3 is used as the base model for CODELUTRA, the authors should apply more previous fine-tuning methods in the same setting and compare with them instead of comparing with different open-source CodeLLMs. For example, the related work section mentions other fine-tuning methods (Line 520), e.g. Self-debug and Codefort. The authors should apply them to fine-tune LLAMA3 and compare the results.\n\n3. Paper presentations can be further improved. Specifically, a) Line 142 $f$ is not defined. b) The notations in the legend of Figure 2 are not defined." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024codelutra,\ntitle={CodeLutra: Boosting {LLM} Code Generation via Preference-Guided Refinement},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yf30Al57nu},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) have significantly advanced code generation but often require substantial resources and tend to over-generalize, limiting their efficiency for specific tasks. Fine-tuning smaller, open-source LLMs presents a viable alternative; however, it typically lags behind cutting-edge models due to supervised fine-tuning's reliance solely on correct code examples, which restricts the model's ability to learn from its own mistakes and adapt to diverse programming challenges. To bridge this gap, we introduce CodeLutra, a novel framework that enhances low-performing LLMs by leveraging both successful and failed code generation attempts. Unlike conventional fine-tuning, CodeLutra employs an iterative preference learning mechanism to compare correct and incorrect solutions as well as maximize the likelihood of correct codes. Through continuous iterative refinement, CodeLutra enables smaller LLMs to match or surpass GPT-4’s performance in various code generation tasks without relying on vast external datasets or larger auxiliary models. On a challenging data analysis task, using just 500 samples improved Llama-3-8B's accuracy from 28.2\\% to 48.6\\%, approaching GPT-4's performance. These results highlight CodeLutra's potential to close the gap between open-source and closed-source models, making it a promising approach in the field of code generation." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "large language models; preference learning; code generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/48d18bdcda59e138b9ceb3f05bbc37c0cc2b29aa.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": "CodeLutra: Boosting LLM Code Generation via Preference-Guided Refinement" }, "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": "See weaknesses." }, "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 well-written and easy to follow.\n- It works on an important problem and provides critical insights: all protection methods today cannot protect artworks from diffusion-based mimicry. Though works like Glaze have been widely accepted by artists, they actually perform bad.\n- The proposed robust mimicry methods are simple but effective.\n- The authors did extensive experiments to support the claims." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors critically revisit the current efforts towards protecting artists' work from diffusion model mimicry. The author proposes that most protection nowadays cannot really protect artists, since the protection can by easily bypassed using some tricks of purifications e.g. Gaussian Noising, Diff-Pure and Up-scalers. Extensive experiments are done to support the claim in this paper. Overall, the paper is well-written, easy to follow and the proposed method is simple but effective." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Clarification:\n- The title of this paper is 'ADVERSARIAL PERTURBATIONS CANNOT RELIABLY PROTECT ARTISTS FROM GENERATIVE AI', while style-based mimicry is one part of diffusion-based mimicry, there are more basic mimicry e.g. inpainting, style-transfer by diffusion model, which are also tested in previous papers of protection e.g. Mist. Fine-tuning a diffusion model seems to have a more complicated mechanism compared with image-to-image applications of diffusion models. \n- I wonder does the proposed method also work for inpainting/image-to-image SDEdit, if it works, the proposed method becomes more general.\n\nMethods:\n- I noticed that the perturbation used in this paper is quite small, if the noise is scaled up, will the purification be worse\n- While Glaze and Mist are popular, there are many other protection methods that can be studied to get a safer conclusion. e.g. MetaCloak [3] and SDS [4].\n\nRelated Papers:\n[1, 2] are highly related to this paper, [1] also find that the current attacks are vulnerable to purifications, [2] proposed that latent diffusion models can be easily purified by pixel-space diffusion models. \n\n[1] Can Protective Perturbation Safeguard Personal Data from Being Exploited by Stable Diffusion?\n\n[2] Pixel is a Barrier: Diffusion Models Are More Adversarially Robust Than We Think.\n\n[3] MetaCloak: Preventing Unauthorized Subject-driven Text-to-image Diffusion-based Synthesis via Meta-learning\n\n[4] Toward effective protection against diffusion-based mimicry through score distillation" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. It would be interesting to see the results on Glaze 2.1 (the newest version) for completeness’s sake. Does this induce greater quality degradation compared to its previous versions?\n\n2. The “Best-of-4” method is not really a practical method since it depends on the human ratings and the same human ratings are used to evaluate the method. For fair comparison, the raters need to be split into validation and test raters, and the method selection should utilize only the validation raters and be evaluated by the test raters, with averaging across splits via cross-validation." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. While it was anticipated that prominent style protection techniques like Glaze might have weaknesses, the extent of their fragility is striking. The paper demonstrates that these methods fail even against rudimentary attacks, with Glaze unable to withstand a mere change in the fine-tuning script. This revelation underscores a concerning lack of “security mindset” within the style protection research community, which is particularly alarming given the recognition Glaze has received, including multiple awards from USENIX Security.\n\n2. The MTurk study is well constructed and the authors have taken pains to ensure that their work adheres to commonly accepted ethical standards." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper examines the effectiveness of existing style protection methods (such as Glaze, Mist, and Anti-DreamBooth) against both black-box and white-box attacks. The authors demonstrate that these protections can be easily circumvented using simple techniques like DiffPure, upscaling, and adding Gaussian noise, as well as more sophisticated methods like IMPRESS. The findings are supported by an MTurk study, where participants were asked to distinguish between images generated using style mimicry on unprotected vs protected artworks. The paper highlights the inherent asymmetry between attackers (style mimickers) and defenders (artists), cautioning against relying on adversarial perturbations for style protection due to the potential false sense of security they may provide." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The work has limited novelty since it was predictable that style protection would be vulnerable to DiffPure-like purification methods. It would be good if the purification methods evaluated in the paper could be packaged into a standard baseline, say on the lines of AutoAttack.\n\n2. In my opinion, the paper overstates the general case against style protection techniques based on adversarial perturbation. The presented argument makes an assumption that artists have the choice to release their artworks on the internet, and they may choose to withhold their artworks if they believe that AI models may be trained on their artworks. However, digital artists are extremely dependent on the internet to grow their customer pool and advertise their works, so this is likely not a feasible option. The appropriate counterfactual to artists using Glaze-like methods would be not using any protections at all. Also, since these methods seem to be improving against simple attacks at least, it may be enough to use them as a deterrent rather than as fool-rpoof security.\n\n\nI struggled to decide whether to rate this paper as borderline accept or clear accept due to the stated weaknesses. Ultimately, I decided to rate this paper as clear accept as the argument may indeed be persuasive to a small minority of artists who may decide to not publish any of their works if there is no secure style protection mechanism." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": 4 }, "primary_area": null, "questions": { "value": "See weaknesses above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The writing is clear and easy to follow.\n2. The topic is interesting and important." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a user study evaluating pre-processing techniques against current adversarial perturbations, finding that such methods can significantly reduce their effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weakness is the unreliable evaluation and whether the experimental results reflect real-world cases.\n\n1. The study heavily relies on MTurk workers, who may not be suitable for this task. They might lack the expertise to classify artistic styles or assess whether artwork quality has been degraded. Artists themselves, such as those mentioned in [1], should be included in the main evaluation. Additionally, these workers may not represent potential consumers of artists' works, contrary to what the authors propose (L 338). A more thorough study should report the workers' relevance to the art domain (e.g., how often they purchase art or visit museums).\n2. The effectiveness of the fine-tuning procedure is unclear. As shown in Fig. 3 (especially the bottom row) and Fig. 5, the generated images, even without protection, do not closely resemble the style of the training images. In contrast, previous work like Glaze [1] uses a fine-tuning setting with much stronger resemblance between generated images and training data (see Fig. 2 in [1]). The focus should be on whether adversarial perturbations effectively defend against mimicry once it has been successfully achieved. Even in Appendix K, where the limitations of using MTurk for evaluation are acknowledged, fewer than 50% of cases were rated as better or equal to “successful” mimicry.\n \n In an extreme case, if fine-tuning involved only one step, adversarial perturbations would likely fail to defend against mimicry. The core issue seems to be underestimating how closely the model need to fit training images, including adversarial perturbations, to learn the style, which may result in an underestimation of their impact.\n \n3. Even if the above weaknesses are overlooked, the paper provides no new insights on solving the problem of mimicry. Both pre-processing and adversarial perturbations degrade image quality, raising the question of whether removing adversarial perturbations is worth the cost, given that pre-processing might degrade the image quality in ways that hinder style recognition.\n4. The paper overlooks broader ethical implications, such as the removal of adversarial perturbations used as a watermark. Similar to traditional watermarks (such as a simple icon on the corner) that can be removed but are illegal to do so in many countries, stronger watermarks complicate removal and leave evidence. A discussion on these ethical issues would contribute more meaningfully to the community.\n5. The proposed methods lack novelty, largely offering improved hyper-parameter tuning of existing approaches.\n\n[1] Shan S, Cryan J, Wenger E, et al. Glaze: Protecting artists from style mimicry by {Text-to-Image} models[C]//32nd USENIX Security Symposium (USENIX Security 23). 2023: 2187-2204." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Overall, the technical details in the paper are clearly described, so I have no questions about the technical aspects. However, I do have some open questions for the authors that I would like to discuss:\n\n+ Q1: Practical and scalable evaluations? The core claim of this paper is to demonstrate the ineffectiveness of existing art mimicry protection methods. For this purpose, using a user study as an evaluation method is sufficient. However, for future papers proposing alternative protective solutions, what could be a scalable way to evaluate their effectiveness?\n\n+ Q2: Is it technically possible to protect artists from mimicry? The conclusions drawn from this work appear to be somewhat pessimistic. Given that generative AI has a strong capability to fit any content, it seems that the potential for mimicry might be inevitable. What could be potential solutions to mitigate this issue?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ S1: The motivation is clear and compelling, effectively serving the paper's purpose: to caution researchers about the limitations of using adversarial perturbations for protection against art mimicry.\n\n+ S2: The paper is well-organized and easy to follow. The experiments conducted are solid.\n\n+ S3: The conclusions drawn from this work are potentially significant for the community and could reshape the landscape of this research field. Researchers are encouraged to reconsider the current paradigms of artwork / copyright data protection in light of their practical effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the practicality of \"using adversarial perturbations to protect art works from mimicry\". The paper argues that, existing research works which uses adversarial noise for art copyright protection -- although published in top ML/Security conferences -- do not robustly achieve their claimed goals. Simple technical means are enough to bypass these existing protections. The work argues that researchers and practitioners should rethink the solution for art mimicry problem and develop alternative protections." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ Typos\n 1. Line 219, fare -> fail?\n 2. Figure 11, 13, 14, 15 seem to have the wrong label - now they are all labelled as \"gaussian noising\"" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We show that adversarial perturbations are not a reliable strategy to protect artists' images from being used to train generative models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024adversarial,\ntitle={Adversarial Perturbations Cannot Reliably Protect Artists From Generative {AI}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yfW1x7uBS5},\nnote={under review}\n}" }, "abstract": { "value": "Artists are increasingly concerned about advancements in image generation models that can closely replicate their unique artistic styles.\nIn response, several protection tools against style mimicry have been developed that incorporate small adversarial perturbations into artworks published online. In this work, we evaluate the effectiveness of popular protections---with millions of downloads---and show they only provide a false sense of security. We find that low-effort and \"off-the-shelf\" techniques, such as image upscaling, are sufficient to create robust mimicry methods that significantly degrade existing protections. Through a user study, we demonstrate that **all existing protections can be easily bypassed**, leaving artists vulnerable to style mimicry. We caution that tools based on adversarial perturbations cannot reliably protect artists from the misuse of generative AI, and urge the development of alternative protective solutions." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "security", "adversarial", "style mimicry", "generative ai" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/e93f7276629aae4a7b12fb8eec882fac1b217397.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": "Adversarial Perturbations Cannot Reliably Protect Artists From Generative AI" }, "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": "What is the actual dependency on k?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The algorithms presented in the paper are interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper the authors consider the problem of choosing at most k subsets from a stream such that the number of distinct items covered by the subsets is maximized. This is an interesting problem with applications in other areas - as demonstrated in the paper. The authors give a O~(d/\\epsilon^2) algorithm where d is the number of sets in the stream and \\epsilon is an approximation parameter." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is clear that the most important parameter is k. And when analysis the complexity of the algorithms the authors have avoided discussing the dependency of the space complexity on k. This makes it very hard to properly judge the true contribution of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 model is strict-turnstle model?\n2. Line 200: L_1 sketches are trivial. Do you mean L_0 Skteches?\n3. The description of the algorithm is confusing and not precise. For example\n a. Line 9. All rows are concatenated to obtain v. However, each entry in the matrix is 0-1. So $v$ is a binary vector? I assume $v$ contains row numbers/elements of the universe? (For example, if elements $i$ is in sets 3, 4 8. Then the vector $v$ contains the number $i$, at 3 positions.\n b. When you are keeping an L_0 sample of $v$, what do they contain? It seems to me that they contain a sample of rows (excluding all zero rows, and hashed into the same bucket) from $A'_m$? Is this correct?\n c. Line 23: I do not understand what it means to \"if $r$ has less then ..... edges among $L_0$ samplers\". Clarifying the above question will help understand this line.\n\n4. Claim 3.1: Consider the instance where $m = 1$. $A'_1$ contains approximately half the rows. Shouldn't this need $n/2$ memory? What am I Missing?\n5. Line 276-277: $k \\log d/\\epsilon^2$ is a fixed number. How can this be OPT?\n6. Claim 3.2: McGregor & Vu's proof relies on (set) insertion only model? Is it easy to see that it translates into a turnstile model?\n7. Line 294: What are $c_1, c_2, \\cdots$? They are not defined earlier\n8. As defined a linear sketch is a matrix drawn from a family of matrices. The algorithm is implicitly defining a family of matrices. Can you define these matrices more explicitly? For the sketch to be linear, the $L_0$ sampler needs to be linear. This should be clarified.\n\nI will revise my score after the discussion period." }, "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 algorithm for the maximum coverage problem in the turnstile model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work studies the maximum coverage problem in a streaming setting: Given $d$ sets over an universe $[n]$ and an integer $k$. Find $k$ sets whose union is maximized. The input (represented as $n \\times d$ matrix) arrives as a stream. Earlier works studied this problem in the insertion-only streaming model. This work studies the problem in the turnstile model, where deletions are allowed. The main contribution is the design of a sketch-based algorithm that use $\\tilde {O} (d/\\epsilon^3)$ space." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing could be improved to enhance the readability of the paper. I am not able to completely understand the proposed algorithm and verify the claims. Please see the Questions for details.\n2. There is a large body of work on streaming submodular maximization. A discussion on the relationships of those works to the current work is missing. For example, \n [1]. https://proceedings.neurips.cc/paper_files/paper/2020/file/6fbd841e2e4b2938351a4f9b68f12e6b-Paper.pdf\n [2]. https://dl.acm.org/doi/10.1145/3519935.3519951\n [3]. https://proceedings.neurips.cc/paper_files/paper/2020/file/9715d04413f296eaf3c30c47cec3daa6-Paper.pdf\n3. Experimental results are not evaluated on turnstile streams." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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.\tLine 209, $|x_i| \\ge \\epsilon^2 || x ||_p$. What is the value of $p$, or does this apply to arbitrary value of $p$?\n2.\tLine 222, $\\epsilon$ is missing in the $\\tilde{O}$ notation, whereas in line 224, the $\\epsilon$ is not omitted in the $\\tilde{O}$ notation.\n3.\tLine 226, $\\alpha - \\epsilon$ -- > $1 - 1 / e - \\epsilon$." }, "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 constructs a linear sketch that supports input updates for the maximum $k$ coverage problem.\n2. Experimental evaluations demonstrate a significant speedup compared to prior work.\n3. Overall, I find the paper to be well-written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the problem of constructing a linear sketch for the maximum $k$ coverage problem. Given a set of $n$ items and $d$ subsets, the objective of the maximum $k$ coverage problem is to select $k$ subsets that maximize the number of items covered. The problem can be represented using a matrix $A \\in \\{0, 1\\}^{n \\times d}$, and the goal of the linear sketch is to find a matrix $S$ such that $SA$ is significantly smaller than $A$ while still enabling an approximate solution to the original $k$ coverage problem using $SA$. \nSince the sketch is linear, it naturally extends to the turnstile model, where the entries of $A$ can be updated over time. \n\nThe paper also demonstrates the application of this sketching technique to the problem of fingerprinting for risk management, with empirical studies indicating substantial speed improvements over previous methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tSome explanation is missing for the algorithm. E.g., the paper claims that algorithm 1 construct a linear sketch. Does this imply that the $L_0$ sampler used in Algorithm 1 is also a linear sketch? The same question applies to the $L_1$ sketch. It would be helpful to explicitly clarify whether these sketches are linear, and if they are, to provide brief explanations or references that detail how they function as linear sketches. This additional context would aid readers in understanding the overall linearity of Algorithm 1." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Some questions are mentioned in the other parts of the review." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Extending the known results for the turnstile model is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides streaming algorithms for maximum coverage and fingerprinting for risk measurement problems. The streaming algorithms are in the turnstile model (input elements can inserted or deleted). Previous results on streaming algorithms were known only for the insertion-only model. Let us discuss the main results for both problems:\n\nMax-coverage:\nIn the maximum coverage problem, we are given subsets S1, ..., Sd of a universal set U (of size n) and an integer k, and the task is to find the k subsets that together cover the largest size subset of U. The problem is NP-hard. There is a poly-time (1-1/e)-approximation algorithm that is known to be tight. The input for this problem can be seen as a 0/1 matrix A of size nxd, where A(i, j)=1 iff i is in the subset S_j. In a previous work, McGregor and Vu (2018) gave a (1-1/e-\\eps)-approximation streaming algorithm in the insertion-only set-arrival model using O(d/\\eps^2) space. In the set-arrival model, the entire column of the matrix A is seen in one step. Bateni et al. (2017) gave a (1-1/e-\\eps)-approximation algorithm in the insertion-only edge-arrival model using O(d/\\eps^2) space. In the edge-arrival insertion-only model, a single matrix entry gets updated from 0 to 1. This paper explores the edge-arrival turnstile model, where in every step, one matrix entry may get updated from 0 to 1 or from 1 to 0.\n\nFingerprinting for Risk Management:\nIn targeted fingerprinting, the input is an n×d matrix A with n users and d features. The goal is to identify at most k features {f1,f2,...,fk} such that the number of users who share identical values at positions {f1,f2,...,fk} is minimized." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There is significant scope for improving the write-up. There is a lack of clarity in many of the statements that leaves the reader confused: \n- What is F_k in the abstract?\n- The introduction assumes the knowledge about the definition of a sketch. Writing 1-2 sentences defining a sketch before using it in the discussion would be good.\n- The abstract states the space usage to be O(d/\\eps^2), but the main theorem (Theorem 1) gives the space-bound as O(d/\\eps^3).\n- Remark 1 is unclear and confusing. It starts talking about 'sampling rates', l_0-sampler, etc. without discussing any random process or algorithm. I had no option but to move on without understanding Remark 1.\n- Lines (141-143): It is unclear what the estimation problem is. Are x_i values given in the streaming setting, or is it (i, \\pm 1)? Unless this is made clear, I am not sure how to interpret Theorem 3.\n- Understanding the sketch algorithm (Algorithm-1) is extremely challenging given that the format of the sketch is not defined. Is H_{\\leq d} a matrix or a subset of (element, subset) pairs? How is this sketch updated in the stream on an insertion and deletion? This does not come out clearly from the pseudocode. In line 14 of Algorithm 1, it is said that \"sketches and samplers handle updates.\" Which sketches are these?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024maximum,\ntitle={Maximum Coverage in Turnstile Streams with Applications to Fingerprinting Measures},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yfZJdCijo6},\nnote={under review}\n}" }, "abstract": { "value": "In the maximum coverage problem we aim to choose at most $k$ subsets such that the number of distinct items covered by the subsets is maximized. The input can be formalized by an $n \\times d$ matrix $A$ where there are $n$ items in the universe and $d$ input subsets. $A_{ij}$ is nonzero if item $i$ is in subset $j$ and is $0$ otherwise. To our knowledge, we are the first to create a linear sketch to solve maximum coverage which can lead to large runtime improvements and allow for implementation in distributed and streaming environments. We specifically focus on the application to turnstile streams which allows deletions. Here, the updates are of the form $(i,j,\\pm 1)$ which performs $A_{ij} = A_{ij} \\pm 1$. Previous work mainly considers the more restrictive set-arrival model where each update reveals an entire column of $A$ or the insertion-only model which does not allow deletions. We design an algorithm with an $\\tilde{O}(d/\\epsilon^3)$ space bound, which is nearly optimal for constant $k$. We then turn to fingerprinting for risk measurement where the aim is to monitor which $k$ columns of an input $n \\times d$ dataset poses the highest re-indentification risk. Our maximum coverage sketch directly enables a solution of targeted fingerprinting for risk measurement. Furthermore, we give an independent result of independent interest: a sketch related to the complement of $F_k$ for $k \\geq 2$. We use this sketch to create a streaming algorithm for general fingerprinting for risk management. Empirical evaluation confirms the practicality of our fingerprinting algorithms and shows a speedup of up to $210$x over prior work. We also demonstrate the use of our general fingerprinting algorithm as a dimensionality reduction technique, facilitating enhanced feature selection efficiency." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "maximum coverage", "turnstile streams", "sketching" ] }, "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/e548ead2a1696e9718b37570417c7c26c2dc5269.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/4880ef4eb1afc05d2fdf3b159e3caaaf6484831b.zip" }, "title": { "value": "Maximum Coverage in Turnstile Streams with Applications to Fingerprinting Measures" }, "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. Regarding function $f$ in line 135, is it a stationary function over the progression of severity -- that is, is it the same for every $Z_t$? If so, can you provide a justification for this assumption?\n2. There are hundreds, if not more, types of disparities among patients. Why do you believe that the three disparities used in your paper are the most significant or important?\n3. In the methodology section, the author mentions the need to pin a group $a_0$ first. How did you choose this particular group, and does selecting a different $a_0$ affect the performance of your model?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "This paper studies the important topic of predicting disease progression by capturing the disparities between patients." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a disease progression model that uses observed symptoms to model the progression of a patient's latent severity. Compared with previous research, it accounts for three types of health disparities: initial severity, disease progression rate, and visit frequency. The proposed method is identifiable and shows good performance on a private dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method appears to be a variant of a hidden Markov model (HMM). Instead of using transition probabilities in HMM, it employs simple functions to describe transitions between states and outcomes. This simplification might limit the model's ability to capture the complex dynamics of disease progression.\n2. The directed acyclic graph, the selection of functions between observed and hidden variables, and the specific types of disparities incorporated in Section 3 seem overly simplistic and heuristic. The paper lacks detailed reasoning or motivation for these design choices. Providing justifications or empirical evidence supporting these decisions would enhance the credibility of the model.\n3. The paper adopts a linear representation solely because \"it provides an interpretable characterization of the trajectory.\" However, real-world disease progression often involves intricate, nonlinear relationships between variables. Relying solely on a linear model may lead to suboptimal performance and may not capture the true underlying patterns, potentially undermining the trustworthiness of the explanations. Exploring nonlinear models could offer better performance and more reliable interpretations.\n4. The empirical results are based on a private dataset with a relatively small number of subjects (n=2,942) and only four features. This raises concerns about the model's generalizability to other datasets. I recommend validating the model on public EHR datasets such as MIMIC or UK Biobank to assess its broader applicability. Furthermore, the comparisons are limited to simple baselines like linear regression, quadratic regression, principal component analysis, and factor analysis. Evaluating the model against state-of-the-art methods, including neural networks, RNNs, and transformers, would provide a more comprehensive assessment.\n5. Insufficient Evidence of Superior Performance:\n - The model's explainability is only qualitatively assessed using medical knowledge. Incorporating quantitative evaluations or user studies could strengthen the claims about interpretability.\n - The model does not show (significant) improvements over the baselines in reconstruction and predictive tasks. Additionally, the absence of error bars or statistical significance tests makes it challenging to determine if the differences are meaningful. If the authors can include the code, my confidence in the results will be higher." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- can the authors add stronger baselines, both from classical machine learning and from the literature of disease progression with health disparities? With a better description of the feature engineering and alternative feature engineering strategies. As it is, the paper shows measurable health disparities of several types, but it is not clear that it is really important for the overall disease understanding\n- can the authors extend the ablation study to the real dataset? or to datasets simulated with alternative data generation processes?\n- can you describe more clearly the metrics used? there are several variables in Xt, how is the RMSE and MAPE aggregated over all of them?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper has several interesting ideas\n- the addressed problem is very important\n- the model is well explained, and the theoretical analysis seems strong\n- the authors provide an extensive empirical analysis, with interesting insights" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this work, the authors propose a Bayesian disease progression model that explicitly accounts for 3 types of disparities concerning health. The model contains several subgroup-specific parameters to account for inequalities in health. The authors provide a strong theoretical analysis of the model, as well as analyses of simulated data and a real application for heart failure patients." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "However, the paper suffers from several flaws. I am really willing to increase my grade if those points are addressed, but as it is, the contribution of the model compared to existing strategies, in terms of performance is really unclear.\n- the baselines seem quite weak. The authors report that several indicators are important, like the visit frequency. I am not sure to understand from the manuscript which features the baselines include, in particular do they include demographics information? and not sure either that PCA or FA are the best tools to do feature engineering. In particular all those approaches are linear, so they can not model interactions between demographics and other features. Maybe considering tree-based methods would be relevant here, with an appropriate feature engineering process. I know that prediction is not the end-goal of the model, but this constitutes the only \"measurable\" performance indicator.\n- the ablation studies are interesting, but they should also be conducted on the real-data application, to assess the variation in predictive performance, because it is not very convincing that data simulated with a model would be less accurately represented by a different model. \n- additional baselines to consider, both on simulated and real data it would be interesting to compare the performances with models that account for demographics characteristics differently (with one type of disparity for instance)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- For a continuous A, does there have to exist a reference value a_0 for which $\\mu(a_0) = 0, \\sigma(a_0) = 1$?\n- Is the second disparity mentioned a disparity because there might be worse clinical care for each ancestry group? Or is this supposed to capture biological differences? Wouldn't the point of any disease progression model with ancestry as a covariate be to infer disparities in disease progression rates?\n- What happens if some of the disparities are unmeasured? How well can the model infer the correct disease progression e.g. in simulations where some of the disparity variables are not shown to the model." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- Alleviating health disparities and fairness w.r.t. clinical algorithms is an important aspect of machine learning for health and the impact of disparities on disease progression modeling is important\n- Interpretability and estimating effects of disparities and covariates on disease progression is clinically meaningful" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a Bayesian hierarchical, linear model of disease progression that captures well known disparities in an interpretable way, while maintaining identifiability. The authors show on a simulated data set that the model can correctly identify the relevant parameters and that accounting for disparities is important to correctly estimate progression. They then demonstrate the utility on a real world data set, showing predictive ability, reconstruction compared to factor analysis and PCA and consistency with medical knowledge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the model is interesting from a clinical perspective, I am not sure this is the right venue for this publication due to limited technical novelty\n- Very limiting disease progression trajectory due to linear assumption in time. A patient can therefore not experience worsening and improvement on their disease trajectory\n- The proofs show that not taking into account disparities will bias the result, however other (non-linear) disease progression models can take \"baseline\" covariates like ancestry, effectively conditioning the probability of the latent progression $z$ on $A$ as well. Comparisons to models like these are necessary to underline the claims.\n- The authors need to compare the disease progression modeling with other state of the art methods of inferring latent disease trajectories. It is not clear how well more realistic disease trajectories are captured and how the model compares to SOTA models on this task.\n- For the synthetic data it would be helpful to also use data not sampled from the same architecture as the model under consideration, but more complex disease progressions to examine how well the model can infer these despite the relatively simple assumptions" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "None" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well written, easy to read and tackles an important problem: improving modelling when disparities mark model. The paper presents theoretical justification and thoroughly evaluates the proposed methodology on both synthetic and real-world data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a Bayesian modelling strategy to model disease evolution while accounting for three sources of biases present in medical data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The model makes assumptions upon the expression of disparities while being identifiable. The underlying process must verify the assumptions. It would be beneficial to discuss further how realistic and/or common these assumptions are. An analysis of a misspecified model when the underlying generating process does not meet these assumptions would be valuable." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose an interpretable Bayesian model that captures and accounts for three types of multiple health disparities, and we prove that failing to account for disparities leads to biased estimates of disease severity." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning Disease Progression Models That Capture Health Disparities},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yfkvUJEY6i},\nnote={under review}\n}" }, "abstract": { "value": "Disease progression models are widely used to inform the diagnosis and treatment of many progressive diseases. However, a significant limitation of existing models is that they do not account for health disparities that can bias the observed data. To address this, we develop an interpretable Bayesian disease progression model that captures three key health disparities: certain patient populations may (1) start receiving care only when their disease is more severe, (2) experience faster disease progression even while receiving care, or (3) receive follow-up care less frequently conditional on disease severity. We show theoretically and empirically that failing to account for disparities produces biased estimates of severity (underestimating severity for disadvantaged groups, for example). On a dataset of heart failure patients, we show that our model can identify groups that face each type of health disparity, and that accounting for these disparities meaningfully shifts which patients are considered high-risk." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "fairness", "equity", "bias", "health disparities", "disease progression", "bayesian 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/484cd25d66e920844aff5a395577a91717b88212.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/e8b7d2c86bcbf33ff6a30d04dbd994e456fa6cb2.pdf" }, "title": { "value": "Learning Disease Progression Models That Capture Health Disparities" }, "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) May include additional analysis or experiments on neural network models to demonstrate how Ensemble++ addresses the gradient coupling issue in that setting." }, "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 proposes the Ensemble++ architecture, which achieves scalability with O(d³log T) per-step computation. This closes a long-standing gap in scalable exploration theory.\n\n(2) The paper empirically validates the scalability and efficiency of Ensemble++ through experiments in bandit tasks, including language-input contextual bandits using a GPT backbone." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Scalable exploration in sequential decision-making is challenging, especially in high-dimensional environments. Ensemble sampling, an approximation of Thompson sampling, is widely used but can suffer from performance degradation due to ensemble coupling in shared-layer networks. The Ensemble++ architecture is proposed to overcome this limitation by introducing decoupled optimization and lifted index sampling. Empirical results show that Ensemble++ outperforms existing methods in regret minimization across various tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I appreciate the motivation and approach presented. However, I find that the methods, analysis, and experiments are inconsistent. The motivation of this method is primarily based on ensemble neural networks, which suffer from gradient coupling issues. Yet, the analysis and experiments focus on linear contextual bandits, where gradient descent is not commonly used. Additionally, an important related work, \"neural contextual bandits [1,2,3],\" has been overlooked. I believe that the analysis from neural contextual bandits could be adapted to this method.\n\n\n[1] Neural Contextual Bandits with UCB-based Exploration\n[2] Ee-net: Exploitation-exploration neural networks in contextual bandits.\n[3] Federated Neural Bandits" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* In the abstract (and elsewhere in the paper), how is ensemble sampling a computationally efficient approximation of Thompson sampling? It seems that it is in specific settings, e.g. the Gaussian setting in Lemma 3 in [1], and the linear setting in this paper where $P_\\xi$ is zero-mean and isotropic. If I am understanding correctly, then this limitation in the connection between ensembles and posteriors needs to be specified. \n* In line 195-196, it is mentioned that there are $M$ ensemble components, but $A$ is simply a matrix of size $d\\times M$; are these $M$ ensemble components separate in any way? \n* How is the variance of the $Z_{s,m}$ perturbations (139-140) chosen? Are these heuristics, or is there a principled reason for their existence? \n* The abstract / intro mention computational complexity but I don't see it derived anywhere. Is this supposed to be in the paper? \n\n[1] Ian Osband, John Aslanides, and Albin Cassirer. Randomized prior functions for deep reinforcement learning. Advances in Neural Information Processing Systems, 31, 2018." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The motivation seems strong, the method is largely straightforward and well-motivated (see questions for exceptions), the paper seems to be overall well-written and well-organized, and the experiments seem promising." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Scalable exploration is challenging in high-dimensional bandit settings. Ensembling is one (computationally expensive) way to try to approximate thompson sampling; ensembling can be made computationally cheaper by sharing one feature extractor between ensemble members, but can underestimate uncertainty due to the ensemble coupling issue, where ensemble members are too similar. \n\nAuthors propose Ensemble++, which uses stop gradients and a high-dimensional ensemble index. Empirically, authors find Ensemble++ outperforms existing methods in bandit settings. Authors analyze regret and computation for Ensemble++ in linear contextual bandit settings; Ensemble++ achieves the same regret bounds as exact TS, and has $\\tilde O(\\log T)$ per-step computation complexity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* **Insufficient discussion of related work** While there is a mention and empirical comparison to epinet in the paper, I felt there was insufficient discussion of how the proposed method is methodologically different from epinets, as they also use index functions.\n\n* **Unclear references to posterior approximation** There are references made to things approximating a posterior, but it was not clear what kind of an approximation that is (e.g. the way MCMC sampling is a posterior approximation, vs VI is a posterior approximation, vs a closed-form posterior for a different but similar problem is an approximation). \n\n\n * **Proposition 1** Proposition 1 is described as a closed-form update; while an earlier sentence mentions posterior approximation, from the Appendix, Proposition 1 is deriving the closed-form minimizers of the loss (2) with respect to $A$ and $\\mu$. Is this supposed to also be a posterior update? If so, what are the assumptions on the Bayesian model? \n\n * **Lemma 1** This result assumes $\\Sigma_t$ is the true posterior variance, but it is not clear why that should be the posterior variance. It seems like there are some assumptions and/or definitions missing. \n\n * **Experiments** There are no experiments that compare Ensemble++ vs other methods to a ground truth posterior, despite claims about Ensemble++ better approximating the posterior. (The first setting in 5.1 does appear to compare with ground-true TS, if I am understanding correctly, but this is the only example.)\n\n* **Sequential dependency** I still don't understand what exactly is the problem in regular ensembles and how it is mitigated in Ensemble++.\n\n* There are claims that the stop gradient prevents ensemble coupling; it would help if this could be empirically measured (beyond better bandit regret). \n\n* I think it could help for readability to have an algorithm box to summarize the method. \n\nAlso see questions for areas that were confusing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "More details should be provided on the stop-gradient operator in the network. Is it solely intended to block gradient flow? If so, how to do the ensemble process? I think the current explanation may not be sufficient for readers to fully understand this aspect." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(1) As demonstrated in the paper, Ensemble++ is computationally efficient, which is highly beneficial in practice. \n\n(2) Ensemble++ addresses the ensemble coupling issue by utilizing the stop-gradient operator." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an improved ensemble exploration method called Ensemble++, which incorporates decoupled optimization and lifted index sampling for more efficient exploration and uncertainty estimation. The architecture outperforms existing methods and is scalable in terms of computational cost." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In my understanding, the primary weakness of the paper lies in the theoretical analysis in Section 4. Specifically, the impact of the stop-gradient operator on regret performance is unclear, which raises concerns about whether the regret analysis fully supports the main idea. However, if I am missing something, please let me know, and I will reconsider my evaluation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 2 }, "primary_area": null, "questions": { "value": "1. The two key innovations in the proposed model are noted as (1) a variance-aware discretization method that prevents the exponential growth in ensemble size and (2) a reduction to sequential random projection techniques. However, the current presentation of the paper does not clearly explain how these innovations are applied. It is not clear how the two proposed changes improve upon the shortcomings of the existing approaches. Additionally, the solution approach lacks clarity without the algorithm.\n\n a) Can you provide a high-level description of how the variance-aware discretization method works and how it prevents exponential growth in ensemble size?\n\n b) Explain the key steps in applying sequential random projection techniques to their problem.\n\n c) Include a pseudocode description of the Ensemble++ algorithm to clarify the solution approach.\n\n d) Explain how the proposed modifications address the shortcomings of existing approaches like Ensemble+.\n\n2. The total computational complexity is $O(d^3 \\log T)$. However, it is not clear from the main paper how this result was concluded. Can you provide a proof sketch and key insights into how the specific approach resulted in this computational complexity and how it overcame the bottleneck in the existing approaches?\n\n3. The primary contribution is the computational advantage in high-dimensional settings. To validate the effectiveness of the proposed approach, experiments showcasing the variation of regret with respect to computational time will be beneficial. \n\n a) Can you include a plot showing regret vs. computation time for Ensemble++, Ensemble+, and the Langevin Monte Carlo approach?\n\n b) Provide a detailed comparison with the Langevin Monte Carlo approach in Xu et al., highlighting both performance and computational efficiency.\n\n c) Discuss any trade-offs between regret and computation time for each method." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper tackles an important issue: scalable exploration poses a significant challenge in sequential decision-making tasks, including reinforcement learning (RL) and contextual bandits, especially relevant in high-dimensional practical environments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Ensemble++, an ensemble sampling method that enhances the computational efficiency of Thompson sampling. Recently, Ensemble+ was developed to tackle this challenge; however, it faces an issue with ensemble coupling that adversely affects its performance. To address this problem, Xu et al. later introduced Langevin Monte Carlo Thompson Sampling, which, unfortunately, incurs high computational costs. The aim of this paper is to improve the approach to ensemble coupling while minimizing computational expenses. The proposed approach, Ensemble++, addresses this by implementing decoupled optimization and lifted index sampling, enhancing exploration and uncertainty estimation. Theoretically, the paper shows that it achieves the same regret bounds as exact Thompson sampling in linear contextual bandits, with a per-step computation complexity of $\\tilde{O}(\\log T)$.\nEmpirical results in the performance of Ensemble++ are presented." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper lacks details in many instances, which hinders understanding the contributions of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024scalable,\ntitle={Scalable Exploration via Ensemble++},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ygtmPu0xZy},\nnote={under review}\n}" }, "abstract": { "value": "Scalable exploration is a persistent challenge in sequential decision-making, especially in high-dimensional environments with neural networks. Ensemble sampling, a computationally efficient approximation of Thompson sampling, is widely used but suffers from performance degradation in shared-layer ensemble networks due to ensemble coupling. To overcome this limitation, we propose the Ensemble++ architecture, which introduces decoupled optimization and lifted index sampling for efficient exploration and uncertainty estimation. \nEmpirical results show that Ensemble++ outperforms existing methods in regret minimization while maintaining bounded per-step computation costs across a variety of tasks, including nonlinear bandits and language-based contextual bandits using a GPT backbone. Theoretically, we prove that Ensemble++ achieves the same regret bounds as exact Thompson sampling in linear contextual bandits, with $\\tilde{O}(\\log T)$ per-step computation complexity. This provides the first rigorous analysis demonstrating ensemble sampling as a scalable and effective approximation to Thompson sampling, closing a key theoretical gap in exploration efficiency." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Bandit", "Scalable Exploration", "Function Approximation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/01c64623f8a7ff1cacfa405bbb31c28c22dbcef9.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": "Scalable Exploration via Ensemble++" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ This work attempts to explore efficient knowledge transfer from super large teachers to tiny students for object detection, which is interesting to the community.\n\n+ The methed that integrates visual prompts with the knowledge distillation framework, bridging the gap between teacher and student models is somehow novel.\n\n+ The experiments across diverse knowledge distillation settings show some promising results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a knowledge distillation method to transfer knowledge from super large teacher object detectors to tiny student detectors. The proposed method incorporates three key distillation components: feature distillation, alongside external and internal prompt distillation mechanisms. The authors make use of the learnable prompts as bridges to mitigate the knowledge gap between teacher and student architectures by transfering the crucial information via the cross-attention modules. To validate their approach, experiments were conducted across various teacher-student detector pairs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "However, there are several weaknesses as follows:\n\n1. The presentation of the method section is not well-organized. The authors fail to provide a formulation of the total loss functions employed during distillation, and the optimization process of learnable components remains unclear. Moreover, the mathematical framework fails to establish clear connections between individual equations and their roles in the final losses. It would be better to show comprehensive formulations linking the learnable modules to their corresponding loss terms. \n\n2. The Figure 2 is notably limited to comparing feature maps between only the largest teacher and smallest student models, while omitting features from models of intermediate scales. Thus it is hard to understand the crucial relationship between model size differences and feature representation gaps.\n\n3. Tables 1, 2, 3 are missing essential metrics, such as the number of parameters and inference latency. The proposed KD method will introduce additional parameters and computational overhead during inference compared to other KD methods, which is potentially not fair. A thorough analysis of the additional computational burden and its impact on inference time should be discussed.\n\n4. The experiments are limited to the CNN-based detectors, which raises questions about the method's generalizability to transformer-based detectors. \n\n5. Moreover, the \"tiny\" students used in this paper are relatively large compared to some truly lightweight detectors (e.g., Tiny-YOLO, SSDLite, EfficientDet, TinyDet, etc.), leaving the method's effectiveness on extremely compressed models unexplored." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- In line 241~242, the author states that **the gradient of KD signals is prone to disappear in shallow stages, making it difficult to optimise effectively**. it is not a consensus as for me. It's better to cite some previous works or do some experiments to prove this statement.\n- What does $M$ denote in formula (3)?\n- How model select the top-N pixels in $Init$?\n- The inputs and outputs of most formulas in this paper use the same symbol, which makes it harder to understand. It's better to use different symbols to distinguish output from input." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This paper attempt to use prompt learning methods in object detection KD. The expeirments are sufficient to demonstrate the effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a simple prompt-based object detection distillation framework, termed DualPromptKD, which aims to improve knowledge transfer efficiency from both teacher and student perspectives. The author enables the student model to fully leverage proficient teacher knowledge by distilling teacher representations into compact external prompts. In terms of the limited learning ability of the student model, the author introduces lightweight internal prompts tailored to bolster the feature imitation capability for the target model. Comprehensive experiments on MS COCO dataset demonstrate the effectiveness of proposed DualPromptKD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Additional prompt learning modules and ConvLoRA bring extra parameters to the student, which goes against the purpose of compressing model of KD." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Overall, I think that investigating KD methods for lightweight object detection is meaningful. However, the current version can not fully support the claims. I expect more experiments in the rebuttal phase, including:\n\n- experimental comparions with most recent works to support the claims and demonste the superiority of this work;\n\n- more experiments with efficient detection pipelines, which can strengthen the core contribution of this work;\n\n- more experiments with general backbones to show the wide applications, which is a bonus of this work.\n\nI consider to improve my rating if these concerns can be addressed." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The topic is meaningful. The authors find the existing object detection KD methods are hard to apply on the light-weight setting.\n\n- Good performance. The authors provide extensive experimental results to show the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this work, the authors foucs on designing an effective knowledge distillation (KD) method for object detection. Specifically, they first showed that simply applying existing KD methods on (lightweight) object detection is not feasible, and proposed a new method, named DualPromptKD, for this topic. The authors proposed two designs in their model, i.e. external prompt KD and internal prompt KD. Finally, they conducted extensive experiments on the COCO dataset, which demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The authors claims they first apply existing KD methods [1,2,3,4] on the object detection task [5,6] but find limited improvement (e.g. line 14-17, line 50-53), which is the main motivation of this work. However, these baseline KD methods are somewhat outdated are not the SOTAs (line 52), and more recent works should be investigated. Also, the experimental results should compare the most recent KD methods.\n\n- This work mainly focus on KD for lightweight object detection. With this scope, I beleive the latency of the detectors should be listed. Besides, the authors mainly conducted the experiments with light-weight backbones, which is okay for this paper, but I still expect two further investigation: 1. Can the proposed KD method also work for the general backbones? 2. More experiments with efficient object detectors (such as EfficientDet, YOLOv4/7 etc.) should be conducted.\n\n\n[1] Focal and global knowledge distillation for detectors, CVPR'22\n\n[2] Pkd: General distillation framework for object detectors via pearson correlation coefficient, NeurIPS'22\n\n[3] Knowledge distillation from a stronger teacher, arXiv'22\n\n[4] Masked distillation with receptive tokens, arXiv'22\n\n[5] Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection, NeurIPS'20\n\n[6] Ghostnet: More features from cheap operations, CVPR'20" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 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. The writing is clear.\n\n2. \"Distillation for tiny models\" should be a good research question to discuss." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces DualPromptKD, a knowledge distillation framework designed to transfer knowledge from large teacher models to tiny student models for object detection. DualPromptKD employs a dual-prompt distillation strategy, including external prompts that capture teacher model knowledge and internal prompts that enhance the feature extraction capabilities of student model. Experiments on the COCO benchmark demonstrate the effectiveness of DualPromptKD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My main concern lies in the unfair comparison with existing methods. \n\nThis paper introduces prompts into the student model, which are additional parameters, while the methods compared in this paper do not add these extra parameters to the student model. Of course, the parameter number of prompt is small, as shown in Figure 4, usually about 20 M, but it is important to note that the GhostNet [1] network used in this paper also typically has only 20 M! This means that the student network parameters may have been doubled, so I think the current experimental comparison is very unfair.\n\n[1] Ghostnet: More features from cheap operations, CVPR 2020\n\n2. Lack of effective new insights.\n\nAs Lines 48-49 stated, the main difference between this paper's setting and the existing setting is \"much smaller and faster models.\" Therefore, the authors use the GhostNet network as the student network in their experiments. However, facing this new problem, this paper does not offer new insights. The insight in Lines 85-87 was already mentioned in [2]; the logic in Lines 89-90 is a bit strange - if distilling the output features on ResNet-50 does not constrain the consistency of shallow features, wouldn't the consistency of shallow features on small models be better constrained because there are fewer layers? The insight in Lines 91-95 lead to the design of prompts, but as previously stated, the effectiveness of prompts may lie in the introduction of additional parameters, rather than solving specific challenges when \"much smaller and faster models\" are used as student models. Therefore, I believe this paper lacks effective new insights, especially in-depth analysis of the specific challenges of the new scenario.\n\n[2] Pkd: General distillation framework for object detectors via pearson correlation coefficient, NeurIPS 2022\n\nDue to the unfairness of the experimental comparison and the lack of effective insights into the distillation task when using tiny models as student networks, I believe this paper does not meet the ICLR standard, and therefore I give a reject." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nzhao2024transfering,\ntitle={Transfering Knowledge into Efficient Tiny Models for Object Detection with Dual Prompt Distillation},\nauthor={Feng Zhao and Yukun Qi and Jiahao Chang and Lin Chen and Kun Li and Tianyou Song and Zehui Chen},\nyear={2024},\nurl={https://openreview.net/forum?id=yhKNCvYlCr}\n}" }, "abstract": { "value": "Knowledge Distillation (KD) has demonstrated significant benefits for learning compact models for object detection. Most current work focuses on general distillation settings, where student models are relatively large and learnable, then compete with the distillation performance. However, due to the model scale and inference speed, these models are seldom deployed in real-world applications. In this paper, we dive into a challenging but more applicable setting: how to distill rich teacher knowledge into tiny, faster models for object detection? We first show that simply applying previous KD strategies under such settings cannot achieve satisfying results, due to the extremely large model capacity gap between the teacher-student pairs. To this end, we propose a simple prompt-based object detection distillation framework, namely DualPromptKD, which aims to improve knowledge transfer efficiency from both teacher and student perspectives. Specifically, by distilling teacher representations into compact external prompts, we enable the student model to fully leverage proficient teacher knowledge even at inference time. In terms of the limited learning ability of the student model, we introduce lightweight internal prompts tailored to bolster the feature imitation capability for the target model. Extensive experimental results on the COCO benchmarks validate the effectiveness and generalization of our approach, including different image backbones and detector types. Notably, our DualPromptKD surpasses the previous best distillation strategies by more than 2.0 mAP under various experimental settings. The code will be available." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Feng_Zhao6", "~Yukun_Qi1", "~Jiahao_Chang2", "~Lin_Chen18", "~Kun_Li13", "~Tianyou_Song1", "~Zehui_Chen1" ] }, "authors": { "value": [ "Feng Zhao", "Yukun Qi", "Jiahao Chang", "Lin Chen", "Kun Li", "Tianyou Song", "Zehui Chen" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "knowledge distillation", "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": { "value": "zhao|transfering_knowledge_into_efficient_tiny_models_for_object_detection_with_dual_prompt_distillation" }, "pdf": { "value": "/pdf/15330ba2aeadb5d8d03e265eb15c3138ce479de6.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": "Transfering Knowledge into Efficient Tiny Models for Object Detection with Dual Prompt Distillation" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please list up and carefully describe any questions and suggestions for the authors. Think of the things where a response from the author can change your opinion, clarify a confusion or address a limitation. This is important for a productive rebuttal and discussion phase with the authors.\n\n\n\nBesides RMSE, it would be good to add other ablation study such as distribution analysis of the counterfactual prediction from utilizing the proposed method vs. baselines, which would provide more evidence to validate the the effectiveness of introducing the shared latent factor as illustrated in Figure 1\nIn table1 and table2, could the author elaborate on more details of the baseline THLTS(v)? Why author think this would be fair baseline to justify the rationality of learning shared part of latent factors compared to the proposed method\nIn Algorithm 1, what is the difference between forecast model pρ() and gρ()?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "A substantive assessment of the strengths of the paper, touching on each of the following dimensions: originality, quality, clarity, and significance. We encourage reviewers to be broad in their definitions of originality and significance. For example, originality may arise from a new definition or problem formulation, creative combinations of existing ideas, application to a new domain, or removing limitations from prior results. You can incorporate Markdown and Latex into your review. See https://openreview.net/faq.\n\nOriginality\nThe paper proposes a novel approach to capturing hidden heterogeneity in time series based counterfactual prediction, which is a significant domain problem in causal learning. The proposed Time-shared Heterogeneity Learning from Time Series method is a novel method that addresses this specific challenge by encoding the shared time-aware latent confounder and then utilizing them for counterfactual outcome forecasting.\n\nQuality\nThe paper provides a clear and well-structured presentation of the proposed method, including a detailed explanation of the shared latent confounder variable encoding process via VAE and how to adapt to time series data.\nThe experimental results basically demonstrate the effectiveness of the proposed method in improving the performance of mainstream models. \n\nClarity\nThe paper is well-written and easy to follow, with clear explanations of technical concepts and methods. The authors provide an informative context in each section that effectively organizes the story and summarizes the paper contributions.\n\nSignificance\nThe proposed THLTS method has the potential to improve the accuracy of counterfactual outcome in time-series data scenarios. The capture of hidden heterogeneity across time domains is a common challenge in many fields. The proposed method is flexible and can be easily inserted with arbitrary causal modeling framework, making it a valuable contribution to the field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a Time-shared Heterogeneity Learning from Time Series (THLTS) method which infers the shared part of latent factor across time steps with a variational auto-encoders (VAE), the method could capture the hidden heterogeneity by recovering the hidden factors and incorporate it into the outcome prediction. This method can be a flexible component to be easily inserted into arbitrary counterfactual outcome forecast models. The authors demonstrate the effectiveness of THLTS on (semi-)synthetic data in capturing shared patterns by combining several existing counterfactual outcome forecast methods to improve their performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Lack of Novelty in Methodology\nThe proposed THLTS method is based on the use of variational encoders (VAE) to improve counterfactual prediction in time-series data, which has been explored in other works such as\n1. https://doi.org/10.1145/3637528.3671950, \n2. https://doi.org/10.48550/arXiv.2310.18615 . \nWhile the shared latent factor encoding part is new, the underlying methodology is not entirely novel. To strengthen the contribution, the authors could provide a more detailed comparison with existing methods and highlight the specific advantages of their approach.\n\n\nLimited Experimental Evaluation\nThe paper only presents experimental results on (semi) synthetic datasets, which may not accurately reflect real-world scenarios. To demonstrate the practical applicability of the proposed method, it would be beneficial to include experiments on (or the connection to) real-world datasets ." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(1) New problem on TCP on state-space model\n\n(2) Design of novel de-correlation mechanism to reduce confounding bias." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper works with a time varying counterfactual prediction method using STATE-SPACE model. It introduces methods that de-correlate between current treatment and historical covariates. They claimed that their model is effective and lightweight. Finally, they performed experiments on several datasets to highlight the efficacy of their method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) My major concern is notation and presentation of the paper: The paper has too many overloading of notations-- for example, \"a\" or the actions are giving variable A_t but the system parameter is also A. This has been quite confusing to me for sometimes. \n\n(2) Re. experiments: I am not sure, results of Table 2 are statistically significant: I was looking for paired t test to see how well their method is effective with respect to baselines." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 sensitive is the CDSP mechanism to the choice of decorrelation threshold?\n2. Could the authors provide more insight into the computational complexity trade-offs between CDSP and traditional balancing methods?\n3. How does the method perform on extremely long sequences (e.g., >1000 timesteps)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Technical Innovation:\n\n-Novel combination of Mamba architecture with TCP\n\n-Well-designed CDSP mechanism that addresses known limitations\n\n-Efficient implementation with linear time complexity\n\n2. Practical Value:\n\n-Better handling of long sequences\n\n-Improved computational efficiency\n\n-Real-world applicability demonstrated on MIMIC-III dataset\n\n\n3.Experimental Design:\n\n-Comprehensive ablation studies\n\n-Multiple evaluation scenarios\n\n-Reasonable baseline comparisons" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Mamba-CDSP, a novel approach for time-varying counterfactual prediction (TCP) using state-space models. The key innovation lies in combining the Mamba architecture (a recent advance in state-space modeling) with a new Covariate-based Decorrelation towards Selective Parameters (CDSP) mechanism. The method addresses two major challenges in TCP: computational efficiency and the over-balancing problem. The authors demonstrate superior performance over existing methods like Causal Transformer and G-Net on both synthetic and real-world datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Theoretical Analysis:\n\n-Limited theoretical justification for why CDSP works better than traditional balancing\n\n-Could benefit from more formal analysis of the bias-variance trade-off\n\n2.Empirical Validation:\n\n-Could benefit from more diverse real-world datasets\n\n-Limited discussion of failure cases\n\n-More detailed hyperparameter sensitivity analysis needed" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Forecasing counterfactual prediction is highly applicable in real-world scenarios.\n2. The time-shared heterogeneity based learning method is easy to implement with VAE.\n3. This paper first utilizes longitudinal method to find the latent factor of each sample, which is intuitive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper tackles the challenge of forecasting counterfactual outcomes in longitudinal settings. Previous methods using LSTM networks and transformers often neglect hidden heterogeneity caused by unobserved factors, which complicates predictions. The authors propose the Time-shared Heterogeneity Learning from Time Series method, which captures shared hidden factors using variational encoders. This approach enhances any counterfactual forecasting method and demonstrates improved performance in experiments with synthetic datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Proposition 4.1, it would be helpful for the authors to explain more about when the prediction model $g$ is Lipschitz with respect to $e$, as this is critical for ensuring the model's effectiveness in identifying the latent factor.\n2. Since the latent factor is not directly observed, how can you guarantee that the latent factor identified by your method is the one you intend to find? It would be beneficial to provide some analysis regarding the identifiability of your method.\n3. Why did you choose VAE to implement your method? Could other structures, such as deterministic models, serve as the backbone? If so, is it possible to test different models as backbones in the experimental section?\n4. The compared baselines are not state-of-the-art methods. It would be better to select more recent methods as baselines to demonstrate the effectiveness of your approach, such as [1].\n\n\n\n[1] Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations. ICLR 2020." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Is CPSD on line 316 a typo?\n\n2. Which dataset was used for Table 3?\n\n3. How sensitive is the method to the choice of decorrelation hyperparameters?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Novel application of SSMs (specifically Mamba) to TCP, showing promising results in both effectiveness and efficiency. The paper leverages state-space models for counterfactual prediction, achieving significant improvements in both prediction accuracy and computational speed compared to existing methods.\n\n2. Well-motivated decorrelation approach that addresses key limitations of existing balancing methods. The proposed CDSP mechanism offers a novel solution to the over-balancing problem in sequential settings, effectively balancing between confounding bias correction and preservation of important covariate information.\n\n3. Comprehensive empirical evaluation across multiple datasets and settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Mamba-CDSP, a novel approach for time-varying counterfactual prediction (TCP) based on state-space models (SSMs). The key contribution is adapting the Mamba architecture with a covariate-based decorrelation mechanism to handle sequential confounding bias while preserving covariate information. The authors demonstrate superior performance compared to existing methods on synthetic and real datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited theoretical analysis of why covariance decorrelation works better than traditional balancing approaches.\n\n2. While performance improvements are shown, deeper analysis of where/why the improvements come from would strengthen the paper. For instance, Table 2 shows substantial gains from CDSP on the MIMIC-III real-world dataset, but this is puzzling since we cannot observe counterfactuals in this data and thus confounding bias should have minimal impact on evaluation. The authors should explain why CDSP shows such dramatic improvements if the test metrics don't actually measure counterfactual prediction ability. This suggests the gains might come from other aspects of the method beyond bias correction, which deserves further investigation.\n\n3. A more thorough literature review on temporal counterfactual estimation would enhance the paper by incorporating recent works like,\n - Chen et al, A Multi-Task Gaussian Process Model for Inferring Time-Varying Treatment Effects in Panel Data\n - Wu et al, Counterfactual Generative Models for Time-Varying Treatment\n - Wang et al, A Dual-module Framework for Counterfactual Estimation over Time\n - Berrevoets et al, Disentangled counterfactual recurrent networks for treatment effect inference over time\n\n4. The paper lacks sufficient implementation details for reproducibility. While the model architecture is described, key details such as hyperparameters (hidden dimensions, number of layers), the decorrelation coefficient, and dropout rates are not specified. These details are crucial for reproducing the reported results.\n\nThe reference for domain adversarial learning on line 269 is incorrect, for example Lim 2018 did not use domain adversarial learning strategy" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper presents a novel method for counterfactual prediction over time series with state-space models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024effective,\ntitle={Effective and Efficient Time-Varying Counterfactual Prediction with State-Space Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yheQRc5xWB},\nnote={under review}\n}" }, "abstract": { "value": "Time-varying counterfactual prediction (TCP) from observational data supports the answer of when and how to assign multiple sequential treatments, yielding importance in various applications. Despite the progress achieved by recent advances, e.g., LSTM or Transformer based causal approaches, their capability of capturing interactions in long sequences remains to be improved in both prediction performance and running efficiency. In parallel with the development of TCP, the success of the state-space models (SSMs) has achieved remarkable progress toward long-sequence modeling with saved running time. Consequently, studying how Mamba simultaneously benefits the effectiveness and efficiency of TCP becomes a compelling research direction. In this paper, we propose to exploit advantages of the SSMs to tackle the TCP task, by introducing a counterfactual Mamba model with Covariate-based Decorrelation towards Selective Parameters (Mamba-CDSP). Motivated by the over-balancing problem in TCP of the direct covariate balancing methods, we propose to de-correlate between the current treatment and the representation of historical covariates, treatments, and outcomes, which can mitigate the confounding bias while preserve more covariate information. In addition, we show that the overall de-correlation in TCP is equivalent to regularizing the selective parameters of Mamba over each time step, which leads our approach to be effective and lightweight. We conducted extensive experiments on both synthetic and real-world datasets, demonstrating that Mamba-CDSP not only outperforms baselines by a large margin, but also exhibits prominent running efficiency." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Time Series; State-space Models; Treatment Effect Estimation" ] }, "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/64411c7773960df01aedb09aa6a39b5a41b838f1.pdf" }, "presentation": null, "primary_area": { "value": "causal reasoning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Effective and Efficient Time-Varying Counterfactual Prediction with State-Space Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "None" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The paper presents the KL convergence results about the VP- and VE-types of diffusions models separately. Could you briefly explain how different types of forward processes affect your proof?\n- I find panel (b) of Figure 1 quite helpful as an illustration between theory and practice, but the paper also presents convergence results with respect to RK-2. What would the theoretical bounds of RK-2 look like on that graph?\n- Are equations 11 and 13 identical? If so, why?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper presents convergence results of second-order SDE solvers for diffusion models, which is very relevant to current research in diffusion modeling given the empirical usefulness of SDE-based simulation of the backward process, and the open question of suitable discretization techniques in this context. The paper gives a theoretical foundation on the application of high-order SDE solvers in diffusion modeling, which motivates further research on suitable solvers for diffusion generative modeling. \n\n- Within the scope of the paper, it presents a compelling argument in favor of SDE-DPM-2 over RK-2 or first-order discretization methods for the practical simulation of samples. I find the insight of \"not second-order solvers are equal\" overall interesting and helpful. \n- The paper also illustrates that the convergence bounds empirically with Gaussian mixture examples. \n- The theoretical results are quite general as they apply to both VP and VE diffusion models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Diffusion models (DMs) learn the score functions associated with a diffusion process, and use the learned scores to simulate an SDE corresponding to the backward process. While samples can be simulated by either an ODE or an SDE, SDE samplers are practically superior in terms of sample diversity and quality. This paper sets out to investigate second-order SDE solvers for the backward SDE, and concludes that second-order solver is preferable to the standard first-order discretization methods in terms of convergence with respect to the Kullback-Leibler divergence. \n\nThe paper mainly investigates two (approximate) second-order SDE solvers, SDE-DPM-2 and Runge-Kutta 2 methods, and compares the convergence results to first order SDE solvers such as EI. The paper presents theorems that suggest that SDE-DPM 2 is more preferable to RK-2 from the perspective of KL-divergence, mainly due to the added discretization error. \n\nWhile the paper mainly focuses on the VP-DMs based on the Ornstein-Uhlenbeck forward process, the main result also applies to the variance-exploding forward process as well, shedding light on the applicability of solvers on other forward processes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I have an overall positive outlook on the paper, I think the paper's overall organization seems confusing: the paper presents the main theorems and some empirical results, then jumps back to a sketch of the proof and how the theory works for the VE-type diffusion models. In my opinion, presenting the paper as theorems on SDE-DPM-2 and RK-2, proof sketch, discussion on VE and then experiments seems like more logical progression of the narrative. \n\nThere are a number minor issues in terms of the paper's presentation. Here is a list I have found: \n- Many discretization methods mentioned in the paper are known only as acronyms without mentioning what the acronyms are. \n- The mentions of $x_k$ in assumption should be $x_{t_k}$ in equations such as the one in Assumption 2, eqs. 11 and 13.\n- The use of partial derivatives w.r.t. $x_{t_k}$ seems confusing. I assume it means the Jacobian matrix. Perhaps the authors can explicitly denote a notation to describe the Jacobian matrix for clarity. \n- While it is useful to see that second-order SDE solvers makes improvements empirically, Table 1 presents quite little added information other than a somewhat vague empirical confirmation that SDE-DPM-2 does have empirical value, which has already been demonstrated by Lu et al. (2022b)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- I think the authors might have a mistake in assumption 4, because I don't see them using the operator $\\nabla^3$ anywhere. \n- Can the authors add error bars to the table of the FID scores? At present, I don't feel that these illustrate their point particularly well." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The authors address a question of significant interest in the diffusion model literature, namely which discretization schemes are most sample efficient at inference time.\n- The paper gives some additional theoretical support to the observation that higher order schemes can be important for sample complexity and differentiates between subtleties, such as the additional approximation in the linear term of the SDE. \n- Experiments show a modest improvement of CIFAR-10 FID with small numbers of sampling steps and improved convergence with discretization fineness using SDE-DPM-2 over RK-2." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the convergence properties of score-based diffusion models with a second-order discretization scheme called SDE-DPM-2, which improves the complexity over a first order exponential intergation scheme. Interestingly the result for SDE-DPM-2 is also stronger than the more widely used RK-2 scheme." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There is no comparison of the computational cost of RK-2 vs DPM-SDE-2 vs EI\n- I felt the authors should have more clearly delineated their contributions relative to Chen 2023, which they follow closely.\n- A number of the assumptions are quite strong. For example, the expectation of the second time derivative of the score is assumed to have a magnitude upper bounded by some time-independent constant. In practice, it is often the case that the score changes in quite a singular fashion near $t=0$." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "No question." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper analyzes the convergence of the higher-order discretization method (SDE-DPM-2). Under some smoothness condition as well as score estimation error and high oder estimation error, a sampling complexity at the order of O(1/epsilon) is established to ensure the KL divergence smaller than epsilon^2. In comparison, the complexity of second-order Runge–Kutta method (RK-2) scales as O(1/epsilon^2)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper analyzes the convergence of the higher-order discretization method (SDE-DPM-2). Under some smoothness condition as well as score estimation error and high oder estimation error, a sampling complexity at the order of O(1/epsilon) is established to ensure the KL divergence smaller than epsilon^2. In comparison, the complexity of second-order Runge–Kutta method (RK-2) scales as O(1/epsilon^2)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the following paper is posted after your submission, there maybe exist some conflict messages between your paper and this work: you said that RK-2 is less efficient, while this work claimed that RK-2 is provably fast.\nWu, Y., Chen, Y., and Wei, Y. Stochastic runge-kutta methods: Provable acceleration of diffusion models.\n\nLi et al. (2024) also provided a sampling complexity of O(1/epsilon) under KL divergence and a better complexity of O(1/sqrt(epsilon)) for TV, which may reduce the theoretical contribution of this work and was not discussed here. \n\nThere exists some other convergence analysis for high-order sampling of diffusion models. It seems that their rates are better than yours, but such comparisons are missed here.\nHuang, D. Z., Huang, J., and Lin, Z. Convergence analysis of probability flow ODE for score-based generative models.\nHuang, X., Zou, D., Dong, H., Zhang, Y., Ma, Y.-A., and Zhang, T. Reverse transition kernel: A flexible framework to accelerate diffusion inference." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- Assumptions 3 and 4 are both bounds for the third-order derivative of $\\log p_t$. However, I firmly believe that temporal derivatives can be represented as spatial derivatives, thereby revealing fundamental properties of the data distribution, as shown in Equation (22) in [2]. Could you please clarify why Assumptions 3 and 4 are considered separate?\n- If Assumption 2 is replaced with the corresponding assumption from [1], is the result for SDE-DPM-2 still valid? Is there any method to ensure the validity of this assumption during the training process?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper studies the SDE-DPM-2 scheme for the inference of diffusion models and improves the sample complexity from $O(1/\\epsilon^2)$ to $O(1/\\epsilon)$.\n- The mathematical proof looks sound to me.\n- Several experiments are conducted to validate the theoretical findings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates the convergence of the second-order discretization method (SDE-DPM-2). Given an $O(\\epsilon^2)$ $L^2$-accurate score estimation, the paper demonstrates that the sampling complexity of SDE-DPM-2 is $O(1/\\epsilon)$ instead of that of the exponential integrator scheme, which is $O(1/\\epsilon^2)$. Furthermore, the paper extends the analysis to the Runge-Kutta-2 (RK-2) method, proving that SDE-DPM-2 exhibits superior efficiency compared to RK-2." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The assumptions appear overly strong and artificial to me. Unlike the conventional assumption that the neural network score function $s(t, \\cdot)$ is approximately $\\epsilon^2$ close to the true score function $\\nabla \\log p_t$, Assumption 2 is, to my understanding, contingent upon the loss function employed in training diffusion models. Consequently, it is not feasible to guarantee or even evaluate this assumption for diffusion models.\n- I recommend redrawing Figure 1 in logarithmic scale to corroborate the theoretical findings.\n- The proof appears to follow the approach outlined in [1]. I believe it is possible to enhance the sample complexity in the data dimension from $O(d^{3/2})$ to $O(d)$ by drawing techniques inspired by the state-of-the-art results presented in [2].\n- I believe this paper lacks a comprehensive literature review. It fails to cite closely related empirical studies [3] and theoretical studies [4, 5], as well as the recent advancements in accelerating diffusion models, such as knowledge distillation [6], consistency models [7], adaptive stepsizes [8], parallel sampling [9], randomized midpoint [10], among others.\n\n[1] Chen, Hongrui, Holden Lee, and Jianfeng Lu. “Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions.” International Conference on Machine Learning. PMLR, 2023.\n\n[2] Benton, Joe, et al. “Nearly d-linear convergence bounds for diffusion models via stochastic localization.” (2024).\n\n[3] Dockhorn, Tim, Arash Vahdat, and Karsten Kreis. \"Genie: Higher-order denoising diffusion solvers.\" Advances in Neural Information Processing Systems 35 (2022): 30150-30166.\n\n[4] Wu, Yuchen, Yuxin Chen, and Yuting Wei. “Stochastic Runge-Kutta Methods: Provable Acceleration of Diffusion Models.” arXiv preprint arXiv:2410.04760 (2024).\n\n[5] Li, Xuechen, et al. “Stochastic Runge-Kutta Accelerates Langevin Monte Carlo and Beyond.” Advances in Neural Information Processing Systems 32 (2019).\n\n[6] Luhman, Eric, and Troy Luhman. “Knowledge Distillation in Iterative Generative Models for Improved Sampling Speed.” arXiv preprint arXiv:2101.02388 (2021).\n\n[7] Mei, Song, and Yuchen Wu. “Deep Networks as Denoising Algorithms: Sample-Efficient Learning of Diffusion Models in High-Dimensional Graphical Models.” arXiv preprint arXiv:2309.11420 (2023).\n\n[8] Jolicoeur-Martineau, Alexia, et al. “Gotta Go Fast When Generating Data with Score-Based Models.” arXiv preprint arXiv:2105.14080 (2021).\n\n[9] Chen, Haoxuan, et al. “Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity.” arXiv preprint arXiv:2405.15986 (2024).\n\n[10] Gupta, Shivam, Linda Cai, and Sitan Chen. \"Faster Diffusion-based Sampling with Randomized Midpoints: Sequential and Parallel.\" arXiv preprint arXiv:2406.00924 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Is it possible to get a better convergence rate for Runge Kutta methods?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The writing is very clear. It provides both theoretical and empirical comparisons with the most related papers. \n\n2. It proves a better convergence rate for a second-order sampling method. \n\n3. It also extends the setting to VE SDEs, showing that the analysis framework can be further generalized." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the convergence properties of a second-order discretization method, SDE-DPM-2, for diffusion models. The main result demonstrates that SDE-DPM-2 achieves an improved $O(1/\\epsilon)$ convergence rate to obtain an $O(\\epsilon^2)$ error in KL divergence, surpassing the performance of existing EI discretization methods. Additionally, using similar proof techniques, the paper shows that another widely used second-order method, Runge-Kutta, does not attain this level of convergence. Further analysis extends these results to the VE SDE, achieving a comparable convergence rate." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) The biggest weakness of this paper is the stringent assumptions. In Assumption 2, this paper assumes the Taylor expansion is accurate, while in most of the previous works for SDE analysis, only the value accuracy is needed. I have seen similar assumptions in [1], which assume the closeness of the Jacobian matrix with respect to $x$ in dealing with ODE analysis. They also showed that such an assumption is not required for SDE. However, Assumption 2 in this paper, though for SDE analysis, is even stronger than that, because it assumes that the time-derivative is also close.\n\nMoreover, Assumptions 3 and 4 are also very strong. When t is close to 0, the score function will get close to the gradient of the probability density function of the data distribution. Such boundness assumptions thus will require the smoothness of the data distribution. As is shown in Appendix B, it can only hold when the data distribution is a Gaussian mixture. This diverges from many useful data distributions, especially when the data is constrained on a low-dimension manifold. As a result, I think more discussions are required to verify the reasonability of these assumptions.\n\n(2) The writing of the paper is a little inconsistent. For example, in equation (6), the first-order derivative is approximated with the value of the score function, while in equation (13) it becomes the partial derivative concerning t and x. Moreover, the notation used here, defined in Line 204 is not standard and very confusing. In Line 201, it says “The difference between the EI and SDE-DPM-2 schemes lies in the approximation of the score function”, while in Line 401, it says “The key difference between EI and SDE-DPM-2 lies in the update scheme at each time interval” It is unclear whether they have the same meaning or not.\n\n(3) The description of the contribution is a little bit inaccurate. It claims that SDE-DPM-2 is more efficient than Runge Kutta. However, no guarantee has been given (Corollary 3.3 only shows that the method used in this paper cannot provide a better guarantee for Runge Kutta). It is possible that there exists an analysis of Runge Kutta that can achieve better results. As is shown in the experiment, the performance of Runge Kutta and SDE-DPM-2 is similar, both better than first-order methods. Thus, the claim seems a little strange to me. Moreover, this paper says that for VE SDE, the convergence is aligned with VP SDE. However, the remark under Corollary 5.1 shows that it only works when overlooking the initial error, which is the key difficulty of VE SDE. This point should also be emphasized in the introduction.\n\n(4) The paper is not self-contained. For example, the proof of Proposition 4.2, directly refers to Chen et al 2023a without any explanation. In my opinion, the argument here is far from trivial and should not be omitted.\n\n----\n[1] Li et al. 2024 Towards Non-Asymptotic Convergence for Diffusion-based Generative Models ICLR2024" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024the,\ntitle={The Convergence of Second-Order Sampling Methods for Diffusion Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yhmVrA8W0v},\nnote={under review}\n}" }, "abstract": { "value": "Diffusion models have achieved great success in generating samples from complex distributions, notably in the domains of images and videos. Beyond the experimental success, theoretical insights into their performance have been illuminated, particularly concerning the convergence of diffusion models when applied with discretization methods such as Euler-Maruyama (EM) and Exponential Integrator (EI). This paper embarks on analyzing the convergence of the higher-order discretization method (SDE-DPM-2) under $L^2$-accurate score estimate. Our findings reveal that to attain $\\tilde{O}(\\epsilon_0^2)$ Kullback-Leibler (KL) divergence between the target and the sampled distributions, the sampling complexity - or the required number of discretization steps - for SDE-DPM-2 is $\\tilde{O}(1/\\epsilon_0)$, which is better than the currently known sample complexity of EI given by $\\tilde{O}(1/\\epsilon_0^2)$. We further extend our analysis to the Runge-Kutta-2 (RK-2) method, which demands a sampling complexity of $\\tilde{O}(1/\\epsilon_0^2)$, indicating that SDE-DPM-2 is more efficient than RK-2. Our study also demonstrates that the convergence of SDE-DPM-2 under Variance Exploding (VE) SDEs aligns with that of Variance Preserving (VP) SDEs, highlighting the adaptability of SDE-DPM-2 across various diffusion models frameworks." }, "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", "reserve SDE" ] }, "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/05438fefd4ed15a3db0c998a83a272d5afc8da3f.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": "The Convergence of Second-Order Sampling Methods for Diffusion Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See Weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well written. The motivation is clear and the contribution BRO is well stated. \n- The BRO method leverage FFT for convolution in a similar fashion as the Caley approach\n- The authors have performed an extensive set of experiments to demonstrate their results (comparaison and ablation study)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose a new orthogonal parametrization, Block Reflector Orthogonal layer (BRO), that can be used in the context of Lipschitz neural networks and provide certified robustness against adversarial attacks. The authors state that the BRO method to construct orthogonal layers using low-rank parameterization is both time and memory efficient, while also being stable during training as it does need \nan iterative approximation algorithms. The authors also propose a theoretical analysis and develop a novel loss function, Logit Annealing loss, to increase certified accuracy. The authors perform an extensive set of experiments to demonstrate their finds." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- BRO is not compared to the Cayley approach in Figure 2. Does BRO offer better memory and runtime against Cayley? can Cayley orthogonal layers appear in Figure 2?\n- For the comparison on certified accuracy, there are a lot of moving parts in the experimental section and this tends to become confusing, I would suggest the authors to simplify the experiments and focus on the comparison with the state of the art, which is LiResNet (Hu et al. 2024). \n- It seems that the authors did not take the results of the latest version of Hu et al. 2024 (which came out in June 2024) as there seems to be a huge gap in the reported results. Can the authors comment on this?\n- Table 1 reports the first results of SLL, but there was an erratum in the latest version and the results have been revised (see Table 7 in the arxiv version of the paper). \n- It seems that BRO does not perform very well at large radius, is this due to the parameterization of the loss? \n- Hu et al. 2024 have shown that their approach scales to ImageNet, can the authors provide certified results for ImagNet?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed method is supported by good theoretical analysis.\n\n2. The BRO improves both efficiency and stability compared with previous work.\n\n3. The proposed Annealing Loss could serve as a scalable method to enhance the training for the Lipschitz neural network." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents a novel BRO layer for constructing Lipschitz neural networks that leverages low-rank parameterization, avoiding iterative approximations to enhance both memory and computational efficiency. The introduction of the Logit Annealing loss function addresses the complexity limitations of Lipschitz networks, contributing to improved learning of margin properties. This work promises good advancements in robust and efficient Lipschitz neural networks design." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **The scalability to other large models**: The proposed BRO layer is currently designed to function within specific neural network architectures, such as the BRONet Architecture, maybe with constrained parameters and fixed configurations. This limitation raises concerns about its practical applicability to large-scale foundation models. Ensuring the Lipschitz condition for these more complex and expansive models could be challenging. Nonetheless, the impact of this work would be significantly enhanced if the authors could empirically demonstrate that BRO effectively improves the robustness of more complex and diverse neural networks.\n\n2. According to lines 276-279, the BRO layer is not a universal approximation for orthogonal layers and the author empirically demonstrates BRO is competitive with that of LOT and SOC. A more detailed analysis would be beneficial to illustrate that the error introduced by this non-universal approximation is minimal.\n\n3. I am curious about whether the choice of activation function affects the Lipschitz constant." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Have you tested BRO's performance on larger datasets like ImageNet?\n- How sensitive is the method to the choice of rank in BRO and how should practitioners select it?\n- How does the LA loss perform compared to other margin-based losses beyond CE+CR?\n- What are the main failure cases or limitations of your approach?\n- Have you considered comparing against other certified defense methods beyond Lipschitz approaches?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Strong theoretical foundations with clear mathematical proofs for their proposed methods\n- Significant computational efficiency gains compared to existing approaches (SOC, LOT)\n- Comprehensive experiments across multiple datasets and architectures\n- Novel loss function with clear theoretical motivation and empirical benefits\n- State-of-the-art certified robustness with fewer parameters\n- Thorough ablation studies demonstrating impact of each component" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel approach called Block Reflector Orthogonal (BRO) layer for constructing Lipschitz neural networks with certified robustness guarantees against adversarial attacks. The key innovation is a new parameterization scheme that creates orthogonal layers without requiring iterative approximation algorithms, making it both computationally efficient and numerically stable compared to existing methods like SOC and LOT. The authors use BRO to develop BRONet, which achieves state-of-the-art certified robustness on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets. Additionally, they provide theoretical analysis showing that Lipschitz networks have inherent limitations in margin maximization due to limited model complexity, leading them to propose a new Logit Annealing (LA) loss function that employs an annealing mechanism to help models learn appropriate margins for most data points rather than overfitting to maximize margins for specific examples. Through extensive experiments, they demonstrate that their combined approach (BRO layer + LA loss) outperforms existing methods while requiring fewer parameters and computational resources." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- BRO layer is not a universal orthogonal parameterization (acknowledged by authors)\n- Limited experiments on larger datasets beyond Tiny-ImageNet\n- No comparison with empirical defenses or other certified robustness approaches beyond Lipschitz methods\n- Lack of investigation into potential failure cases or limitations of the LA loss\n- Some hyper-parameters (rank selection, LA loss parameters) require manual tuning" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Please see the weaknesses section for questions about the theory, notation and typos.\n\n- Can authors implement the convolution without the FFT? Just creating $W_{\\text{Conv}}$ and applying the standard Conv2D function from torch. Are the results the same as in Algorithm 1? Your outputs should be the same and its a good way to debug if anything is wrong.\n- Did authors try LiResNet + BRO only for table 2? That is, only changing the backbone without changing the loss.\n- The BRO layer is for sure orthogonal in the case of fully connected layers. Have authors tried comparing the different layers in the fully connected case? I.e. training fully connected classifiers in a small dataset (Cifar10 / MNIST). I don’t believe this experiment is super interesting but is the only case where orthogonality is clear at the moment." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The strengths of this work are mainly on the experimental side. Assuming the theoretical results hold:\n\n- More efficient and scalable models than the previous art due to exact 1-Lipschitz layers without the need of iterative approximations.\n- Improved certified accuracy. Authors demonstrate that both their BRO convolutional layer and LA loss help improving the certified accuracy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel 1-Lipschitz layer based on Block-reflector matrices (BRO convolution), which are orthogonal. Additionally, authors propose a Logit Annealing (LA) loss to optimize during training in order to favor certified accuracy. Their proposed recipe consistently improves certified accuracy in CIFAR10/100 and TinyImageNet and is shown to be more efficient than previously proposed orthogonal layers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weaknesses of the paper appear on the theoretical and implementation sides. In general, the paper is very hard to read and has many errors in key parts of the paper. This questions the validity of their 1-Lipschitz claims for the BRO convolution.\n\n- **Many typos and unclear notation:**\n\t- Everywhere: Please use $c_{\\text{in}}$ and not $c_{in}$ when subscripting or upperscripting with words.\n\t- Algorithm 1, line 1: I assume $c_{\\text{in}}$ and $c_{\\text{in}}’$ are the number of input channels for the current and next layers respectively. Then, why do you need $c_{\\text{out}}$? It is very confusing. I would just remove $c_{\\text{in}}’$ and use $c_{\\text{out}}$ instead.\n\t- Algorithm 1, line 4: $\\tilde{V}:= \\text{FFT}(V)$ should be $\\tilde{V}:= \\text{FFT}(V^{\\text{pad}})$.\n\t- Lines 205 and 214: Is $\\circledast$ the convolution operator? \n\t- Proposition 2: If $J \\in \\mathbb{C}^{m \\times m}$, then $J^* \\in \\mathbb{C}^{m \\times m}$ as well. How can you even perform $J\\tilde{V}J^{*}$ if $\\tilde{V} \\in \\mathbb{C}^{m\\times n}$? Does it have to be that $m=n$?\n\t- Proposition 2, equation 14: If assuming $n=m$, the proof is right, but equation 14 should be: $(\\tilde{V}^* \\tilde{V})^{-1} = J(J\\tilde{V}^*\\tilde{V}J^*)^{-1}J^*$\n\n- **Unclear if BRO convolutions are orthogonal:**\n\nFrom Proposition 1, it’s clear that in the dense case, the BRO layer is orthogonal. However, in the convolutional case there are many unclear aspects.\n\nIn lines 205 and 214, authors state that they construct their BRO convolution based on constructing the kernel:\n\t$$\n\t\tW_{\\text{Conv}} = I - 2V \\circledast (V^{\\top} \\circledast V)^{-1} \\circledast V^{\\top}\n\t$$\nAuthors argue that the circular convolution with this kernel is orthogonal, but do not provide any proofs. Then, given $\\tilde{V} = \\text{FFT}(V)$ authors conclude $\\tilde{W} = \\text{FFT}(W_{\\text{Conv}}) = I - 2\\tilde{V}(\\tilde{V}^*\\tilde{V})^{-1}\\tilde{V}^*$. It’s unclear how authors arrive to this expression as it would need to assume that: (i) $n=m$ to be able to do element wise products (which authors do not mark with $\\cdot$, leading to confusion) (ii) $\\text{FFT}(I)=I$ (iii) $\\text{FFT}((V^{\\top} \\circledast V)^{-1}) = \\text{FFT}((V^{\\top} \\circledast V))^{-1}$, which are false in general. \n\nMoreover, it is not clear what Proposition 2 implies regarding orthogonality, or how the case when $m=n$ is handled (lines 237-241 are very vague). It would be nice to include an explicit proof that the norm of the output of algorithm 1 is the same as the input. Exactness of the 1-Lipschitz result is important for ensuring the experimental results about the certified accuracy are valid. So, I believe more rigour should be put into proving the results claimed in this paper.\n\n- **Unclear motivation to use LA and differences with CR:**\n\nAuthors start motivating the need of their LA loss with a very convoluted argument about the minimal CR loss being constrained by the model complexity. Then, they simply present their loss without relating it with the theory they developed for the CR loss. I believe the analysis doesn’t add anything if it is not performed for the LA loss. While I see that LA is less strict I don’t see how LA solves the issue in Theorem 1. \n\nIt would be more beneficial to fully state the CR loss and compare it with LA. Right now, only the reference is provided. I also think the results in Figure 8 and section C.2 are more interesting than the current write-up of the main paper.\n\nAll in all, I believe authors should be more careful about their theoretical derivations and be very clear about why Algorithm 1 results in orthogonal convolutions. Without this being clear, the results are meaningless and I cannot propose the paper for acceptance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What factors contribute to the inconsistent performance gains observed in the proposed BRONet as the perturbation budget increases?\n2. Why were standard deviations or confidence intervals not included in the reported results? Given the marginal improvements, understanding the variability of the outcomes would be particularly valuable.\n3. Has the BRO method been evaluated on more complex datasets, such as full ImageNet, or applied to tasks beyond classification, to assess its scalability and generalization?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The BRO method effectively utilizes low-rank parameterization to construct orthogonal layers, resulting in significant improvements in both computational time and memory efficiency, which are critical for scaling neural networks.\n2. The paper provides a thorough comparison with state-of-the-art techniques, presenting results that are well-structured and articulate, allowing readers to easily grasp the contributions and effectiveness of the proposed method.\n3. The paper is well-structured and logically organized. The presentation of the results is clear and systematic." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the block reflector orthogonal (BRO) layer, which enhances the construction of Lipschitz neural networks. By employing low-rank parameterization and circumventing iterative approximations, the proposed approach achieves notable gains in both memory and time efficiency. Comprehensive evaluations against existing orthogonal layers reveal its superior robustness, underscoring its potential in advancing neural network performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The evaluation results in Table 1 indicate a degradation in the performance of the proposed BRONet with increasing perturbation budgets. A discussion on this inconsistency would enhance the paper's credibility and provide insight into the limitations of the model.\n2. Results in Table 3 show only marginal improvements over existing methods, raising concerns about the significance of these gains. Including standard deviations or confidence intervals would clarify the statistical significance of the results and help determine whether observed improvements are due to random variance in training.\n3. The experiments are conducted solely on CIFAR and Tiny-ImageNet datasets, which, while widely used, may not fully demonstrate the method's robustness and scalability. Including additional datasets, particularly larger or more complex ones (e.g., ImageNet or real-world benchmarks), would provide a more comprehensive evaluation of the method's efficacy." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a new orthogonal convolution and a novel loss function to enhance certified robustness." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024enhancing,\ntitle={Enhancing Certified Robustness via Block Reflector Orthogonal Layers},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yi3QcCGfP1},\nnote={under review}\n}" }, "abstract": { "value": "Lipschitz neural networks are well-known for providing certified robustness in deep learning. In this paper, we present a novel efficient Block Reflector Orthogonal layer that enables the construction of simple yet effective Lipschitz neural networks. \nIn addition, by theoretically analyzing the nature of Lipschitz neural networks, we introduce a new loss function that employs an annealing mechanism to improve margin for most data points.\nThis enables Lipschitz models to provide better certified robustness.\nBy employing our BRO layer and loss function, we design BRONet, which provides state-of-the-art certified robustness.\t\nExtensive experiments and empirical analysis on CIFAR-10, CIFAR-100, and Tiny-ImageNet validate that our method outperforms existing baselines." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Certified robustness", "Adversarial" ] }, "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/8de0265482b6bf213f9bf4c41983b4af6df5d33e.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/93b1b773838c55b3e5b3427eb3cb88fd6d88b65f.zip" }, "title": { "value": "Enhancing Certified Robustness via Block Reflector Orthogonal 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": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Nit: Please add a reference for the claim” We choose DiT as the backbone because it can be easily scaled up and is shown to have great\ngeneralization and expressiveness.” Line 210\n\nIn Table 3, it is not clear how many classes were predicted. This is important to assess the reported accuracies.\n\nCan you please comment on how would you set the number of model parameters for new unseen tasks?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Comprehensive Experimental Analysis: The paper includes a robust set of experiments, covering different prompt types, model architectures, dataset sizes, and scaling laws. These analyses provide a clear understanding of Tina’s capabilities and boundaries, and they validate (to some extent) the model’s effectiveness in generating personalized networks under varying conditions.\n2. Novel Approach to Model Personalization: The paper builds on the concept of train-once-for-all personalization, allowing a single pre-trained model (Tina) to generate personalized models dynamically based on text prompts. This can potentially eliminates the need for separate training per task, making the approach highly efficient and versatile." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new generative AI framework named Tina, which can generate \"personalized\" neural network models based on text prompts. This approach, called train-once-for-all personalization, enables a single model to generalize and create task-specific models on demand without the need to fine-tune the model on task related data. Tina leverages a diffusion transformer model conditioned on descriptions encoded with a CLIP model to understand and apply user-specific knowledge, even with a small training dataset. It demonstrates strong performance in generating models for both in-distribution and out-of-distribution tasks, supporting zero-shot and few-shot scenarios with images and adapting to different classification settings. The framework opens possibilities for text-to-model applications, expanding the range of personalization within neural network." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The approach is not scalable. The experiments are not showing a possibility to scale the approach for larger number of classes (limited to 10) or for more complex models. The paper presents what seems like a good proof of concept but it would require more work to demonstrate the effectiveness of the approach on larger more complex problems.\n2. The datasets used are too small and simple to validate the approach properly. \n3. One very important baseline that is missing is a direct fine tuning which should be an upper bound. The selected baselines are not representative enough to see what is the loss in performance to expect with Tina.\n4. The generic model in the experiments seems to be quite bad even on the in-distribution tasks. I would have expected it to perform better with improvements coming from Tina" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "see the Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. It’s an interesting idea of using text-conditioned diffusion models to generate neural network parameters based on varying requirements. \n\n2. Extensive experiments have been conducted to validate the effectiveness of the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on neural network parameter generation and utilizes diffusion models for text-to-model generation. With just one training session, the proposed method achieves outstanding results in both out-of-distribution and in-distribution model personalization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method is currently limited to personalizing models for image classification tasks. As a pilot study for generating neural networks with diffusion models, it does not fully support the title of \"train-once-for-all.\" Conducting more experiments on detection and segmentation would enhance the overall credibility of the study.\n\n2. The method can generate only a relatively small number of parameters—specifically, around 640 parameters in the classifier layers of ResNet-20. It still heavily relies on the feature extraction module of the generic model. Therefore, the significance of \"text-to-model\" is weakened if the partial model parameters are already provided.\n\n3. The ablation of text prompts indicates that the proposed method is sensitive to the input prompt. Could training with mixed prompts improve the stability?\n\n4. In traditional diffusion models, the inclusion of random noise could improve the diversity of the output. But it seems useless in the proposed method because the aim of Tina is to find the best classifier without considering diversity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 weakness" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The writing is clear and easy to follow.\n- The discussed topic and motivation are both innovative and significant." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Tina, a novel framework that leverages text-conditioned neural network diffusion for generating personalized models from textual prompts. It addresses the scenario of train-once-for-all personalization, aiming to create customized models for diverse end-users and tasks using text prompts. Tina is designed to generalize across in-distribution and out-of-distribution tasks, even with limited training data. The paper claims that Tina demonstrates an understanding of world knowledge by analyzing its capabilities under various conditions, including zero-shot/few-shot image prompts, different numbers of personalized classes, and predicting unseen entities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Though I'm not well-versed in the subject of this article, I'm still amazed by the \"text-to-model\" concept. I'm skeptical about the \"train-once-for-all\" approach since the author didn't provide any code or demos to back up the experimental results.\n\n- What kind of experimental settings did Tina use in text-to-model task—simple or challenging? What are the limits of Tina's capabilities?\n\n- I'm very curious whether the proposed Tina has theoretical support." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The main difficulty for this prediction is its limitation to relatively small neural network. Assuming that we want to predict a 1B parameter transformer network, how would you address it?" }, "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 has solid technical contribution.\n2. The proposed method is novel and clean.\n3. The experimental results are also strong." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigate the capability of GenAI for text-to-model generation, to see whether GenAI can comprehend hyperlevel knowledge embedded within AI itself parameters. The basic idea is to use diffusion transformers to generate parameter token by token. Each token is indeed a set of parameters in a specific layer. The model is trained with supervised learning approach to feed user-provided text description and then use diffusion model to synthesize the personalized network parameters. The results seem quite interesting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing of this paper need improvement. The introduction is quite obscure and high-level. It only shows some broad idea without elaborating the actual implementation much. I would suggest the authors to hint a bit in terms how they tokenize the parameters and use DDPM to predict the actual parameters, etc. This could help the authors gain more clear insights.\n2. The evaluated datasets are still a bit toy or simple. The whole paradigm still requires more thorough or large-scale experiments to validate." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nli2024texttomodel,\ntitle={Text-to-Model: Text-Conditioned Neural Network Diffusion for Train-Once-for-All Personalization},\nauthor={Zexi Li and Lingzhi Gao and Chao Wu},\nyear={2024},\nurl={https://openreview.net/forum?id=yiGSI7Ou3i}\n}" }, "abstract": { "value": "Generative artificial intelligence (GenAI) has made significant progress in understanding world knowledge and generating content from human languages across various modalities, like text-to-text large language models, text-to-image stable diffusion, and text-to-video Sora. While in this paper, we investigate the capability of GenAI for text-to-model generation, to see whether GenAI can comprehend hyper-level knowledge embedded within AI itself parameters. Specifically, we study a practical scenario termed train-once-for-all personalization, aiming to generate personalized models for diverse end-users and tasks using text prompts. Inspired by the recent emergence of neural network diffusion, we present Tina, a text-conditioned neural network diffusion for train-once-for-all personalization. Tina leverages a diffusion transformer model conditioned on task descriptions embedded using a CLIP model. Despite the astronomical number of potential personalized tasks (e.g., $1.73\\times10^{13}$), by our design, Tina demonstrates remarkable in-distribution and out-of-distribution generalization even trained on small datasets ($\\sim 1000$). We further verify whether and how \\Tina understands world knowledge by analyzing its capabilities under zero-shot/few-shot image prompts, different numbers of personalized classes, prompts of natural language descriptions, and predicting unseen entities." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Zexi_Li1", "~Lingzhi_Gao1", "~Chao_Wu1" ] }, "authors": { "value": [ "Zexi Li", "Lingzhi Gao", "Chao Wu" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "diffusion model", "parameter generation", "personalization" ] }, "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": "li|texttomodel_textconditioned_neural_network_diffusion_for_trainonceforall_personalization" }, "pdf": { "value": "/pdf/990e8544e0172b713c5c9296e849ad2c48a0c784.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": "Text-to-Model: Text-Conditioned Neural Network Diffusion for Train-Once-for-All Personalization" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "What is the starting point for AL and how to explain the initial differences?" }, "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 addresses an important problem of label imbalance and noises in AL.\n2. The paper improves upon existing methods like GALAXY." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an AL strategy that deals with label imbalance and noises. The proposed method uses separation thresholds similar to an existing method GALAXY, however could enable parallel annotations and be robust against label noises. The proposed method is compared with GALAXY and other AL baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presentation of the paper is poor. While there is additional content in the appendix, the paper does not utilize the main page limit well. The figure caption of Figure 3 overlaps with main text. The algorithm is not clearly presented with too much text and no numbers for equations. \n2. The results are presented poorly and unreliable. It is unclear where the AL starts and the different starting points of curves are confusing. It is also difficult to see the difference between variants of the proposed method and GALAXY in some 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": 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)\tAbout the construction of the imbalanced datasets, I am unclear about the rationale behind grouping certain classes into a larger class to create an imbalanced dataset, as this results in an inconsistent number of classes compared to the original dataset. A simpler approach, such as using a pre-existing imbalanced dataset like CIFAR-10-LT or selectively sampling data to form a minority class, might have been more straightforward. Could the authors explain the motivation behind this grouping strategy? \n\n2)\tThe methods used for comparison appear inconsistent across different datasets. Could the authors clarify the criteria for choosing comparison methods, and explain any reasons for these discrepancies?\n\n3)\tWhen plotting the learning curves, is the budget spent on B_{parallel} included in the overall labeling budget? Clarification on this point would help in interpreting the learning curves accurately?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The study tackles a complex and meaningful setting where both class imbalance and label noise are present, which is a valuable and practical area of focus for active learning. The idea of identifying and querying near the decision boundary, particularly for minority classes, is technically sound and shows promise for improving classification performance in imbalanced data contexts. The experiment includes recent active learning methods in its comparative analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an active learning algorithm designed to handle both class-imbalance and label noise in classification tasks. The proposed method reformulates the multi-class classification problem into multiple one-vs-rest tasks and employs the VReduce algorithm to estimate classification thresholds for each class. Data points near these thresholds are then selected for querying. Experiments are conducted to demonstrate the effectiveness of the approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1)\tMy primary concern is the limited technical contribution. The proposed method largely leverages previously established algorithms, and the theoretical contributions appear modest.\n\n2)\tAlthough the paper claims to address label noise, the algorithm itself does not explicitly manage or mitigate label noise. I think the noise may relate to the agnostic learning, but in section 5.2 in the experiment, the authors conduct experiments with different levels of label noise, which confuses me.\n\n3)\tEstimating the prediction threshold for minority classes could be challenging due to the limited sample sizes within these classes. This issue is not thoroughly addressed in the paper.\n\n4)\tThe paper would benefit from further proofreading to address minor errors and improve readability. For instance, lines L269 and L278 contain typographical or presentation mistakes that should be corrected." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See the above Weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The approach of combining class separation thresholds with one-dimensional active learning is innovative and provides a fresh perspective on tackling these issues.\n\n2. The extensive experiments conducted on imbalanced datasets provide a strong foundation for the claims made. The results indicating a 60% to 80% reduction in annotation budgets are compelling and demonstrate the algorithm's effectiveness compared to existing methods.\n\n3. The paper is well-organized, with clear explanations of the methodology and results, making it accessible to readers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel active learning strategy called DIRECT, which performs well under conditions of class imbalance and label noise. Specifically, the algorithm effectively identifies the optimal class separation threshold and adaptively selects samples for annotation. Experimental results demonstrate that DIRECT significantly improves labeling efficiency, providing an effective solution for scenarios affected by class imbalance and label noise." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is a formatting error at the bottom of page 5 that requires correction to improve the document’s presentation.\n\n2. The authors are encouraged to include additional visualizations to more effectively demonstrate the experimental results and clarify the method’s performances.\n\n3. The authors identified three limitations in the GALAXY method and addressed these by optimizing the separation threshold. It would be beneficial to provide a more detailed explanation of this approach to help readers better understand the proposed improvements.\n\n4. The authors demonstrated the robustness of their method under class imbalance and label noise conditions. Please clarify the sources of robustness in each scenario to highlight the model’s adaptability.\n\n5. In the experimental section, the authors only compare label noise levels at 0%, 10%, 15%, and 20%, leaving out 5%. It would strengthen the study’s comprehensiveness to either include results at 5% noise or explore higher noise levels to make the noise experiments more comparable and thorough." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The proposed approach relies on an \"optimal threshold\" derived from the training set, but this threshold is based on true labels, which are inaccessible during active learning. Given that this threshold may not generalize to the unlabeled dataset, how can we justify that the threshold remains optimal or even effective for active learning purposes, especially without analyzing data distribution shifts between labeled and unlabeled sets?" }, "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 a thorough summary of some classical methods in active learning within the related work section, which is informative for readers less familiar with the field.\n2.The paper aims to address the significant problem of active learning under conditions of label noise and class imbalance which are increasingly relevant in large-scale data applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a threshold-based active learning approach aimed at handling label noise and class imbalance. The main idea for handling class imbalance is to use a deep learning method to simplify and reformulate it into a one-dimensional active learning task with a threshold learner." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.I believe the method proposed by the authors contains fundamental and theoretical flaws. The validity of the “sample around the optimized threshold” approach itself is highly questionable. This sampling strategy is merely an intuition; even if the authors could accurately identify this threshold, they lack any proof that sampling around it is optimal. While the authors claim this method would encourage the model to select a balanced number of samples from each class, this strategy is not necessarily the best. From a learning theory perspective, as with all active learning methods, this sampling process distorts the training distribution, potentially undermining the learning guarantees. Furthermore, from a statistical learning perspective, if the majority class has a more complex distribution, naturally more data would be required to approximate the distribution.\n2.The so-called \"optimal threshold\" is based on true labels, which are inherently inaccessible in active learning, making accurate computation of this threshold infeasible in practice. Moreover, their hypothesis class is defined solely as a threshold classifier over the output of a deep learner. Without analyzing the behavior of this deep learning output and the underlying data distribution, any claims regarding the behavior of a hypothesis learned by empirical risk minimization (ERM) within this hypothesis class are, in my view, meaningless. Therefore, their proof is fundamentally flawed and fails to establish any meaningful guarantee for the proposed threshold and sampling strategy in active learning contexts. In fact, I think the authors’ proof in the appendix only establishes the equivalence between the ERM solution on the training data based on the learner’s output and an \"optimal threshold\" on this same training set. This is unrelated to what they actually need to prove—that this solution can serve as an optimal threshold for active learning on the unlabeled data.\n3. The authors claim to address label noise in addition to class imbalance; however, the paper lacks even a formal definition of label noise, and it remains unclear how label noise is actually addressed. Despite incorporating noisy scenarios in their experiments, the authors do not propose any specific strategies to mitigate or manage noisy labels, leaving it ambiguous how their method effectively handles this issue." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024direct,\ntitle={{DIRECT}: Deep Active Learning under Imbalance and Label Noise},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yiQCeXdPvs},\nnote={under review}\n}" }, "abstract": { "value": "Class imbalance is a prevalent issue in real world machine learning applications, often leading to poor performance in rare and minority classes. With an abundance of wild unlabeled data, active learning is perhaps the most effective technique in solving the problem at its root -- collecting a more balanced and informative set of labeled examples during annotation. Label noise is another common issue in data annotation jobs, which is especially challenging for active learning methods. In this work, we conduct the first study of active learning under both class imbalance and label noise. We propose a novel algorithm that robustly identifies the class separation threshold and annotates the most uncertain examples that are closest from it. Through a novel reduction to one-dimensional active learning, our algorithm DIRECT is able to leverage classic active learning theory and methods to address issues such as batch labeling and tolerance towards label noise. We present extensive experiments on imbalanced datasets with and without label noise. Our results demonstrate that DIRECT can save more than 60% of the annotation budget compared to state-of-art active learning algorithms and more than 80% of annotation budget compared to random sampling." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Deep Learning", "Active Learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/79f9aff73438cd6e08d054510745d2757fc7dcdc.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": "DIRECT: Deep Active Learning under Imbalance and Label Noise" }, "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": "- What is $R(\\cdot)$ in line 204?\n\n- Performance for ReverseGen in Tables 1 and 2 has not saturated with increasing iterations and it seems like you are under reporting results. What would happen if you ran this for 4 or 5 iterations? At what point would performance saturate? What if you generated 10000k instruction candidates and performed t=1 iteration versus 2000k instruction candidates with t=5 iterations?\n\n- When using harder and harder samples to train a model for example in active learning, or data selection, where data points are prioritized with a larger loss then this can cause a negative feedback loop with a catastrophic drop in performance from a target model. Did you observe similar artifacts in your experiments?\n\n- In Tables 1, 2, 3 you have a rows ‘without failure induction’ but you do not describe what this ablation is?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The idea of using the language model prediction errors to create a set of easy and hard examples to train a proposer model with DPO is novel and a nice idea.\n\n- Interesting results that show that harder to predict data points aka using a curriculum harder and harder questions is beneficial for some red teaming and honesty benchmarks. This is interesting since, in comparison other papers such as [1], show that hard samples actually hurts target model performance albeit in a different dataset domain.\n\n- Wide range of experiments: red teaming, honesty and mathematical reasoning to demonstrate that the method can generalize to multiple domains.\n\n[1] Evans, Talfan, et al. \"Bad students make great teachers: Active learning accelerates large-scale visual understanding.\" arXiv preprint arXiv:2312.05328 (2023)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose an iterative finetuning method for finetuning a target langague model by using synthetic data generation from a proposer language model which proposes harder and harder questions to a target language model. This is in effect a curriculum learning approach which trains a target model on harder and harder samples. The proposer model is also trained to propose harder and harder questions by using errors in the target model’s answers.\n\nA proposer language model generates few-shot candidate questions. Then the target model predicts answers to these questions. The answers are then compared to gpt-4o-mini's answers, which is used as a gold-standard. If the answers agree then this question is placed into the negative set {x^-}. If the answers from the target model does not agree with the gpt-4o-mini then the question is placed in the positive set {x^+}. These sets are then used by DPO is used to finetune the proposer model to produce harder and harder samples by leveraging the positive and negative sets. Finally, the proposer model generates synthetic questions which are deemed hard for the target model, labels are generated by the proposer model or gpt-4o-mini. The target model is then trained with SFT." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Weak results on mathematical reasoning which do not demonstrate considerable improvement in performance. Nor are many other similar iterative methods which have some sort of synthetic question-answer prioritization compared to [2, 3].\n\n- No ablation experiments, what is the performance with 1 iteration versus 5? What if you generate 1k or 10k samples to populate the positive and negative sets for DPO training?\n\n[2] Lee, Nicholas, et al. \"Llm2llm: Boosting llms with novel iterative data enhancement.\" arXiv preprint arXiv:2403.15042 (2024).\n[3] Jiang, Yuxin, et al. \"Lion: Adversarial distillation of proprietary large language models.\" arXiv preprint arXiv:2305.12870 (2023)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "None" }, "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 overall well written and easy to understand.\n* The proposed approach is novel. It performs RLHF to train the data generator, with the target model works as a preference provider." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a data synthesis approach by training a data generator and leverage the performance of the target model as a training signal. Specifically, the predictions of the target model are used to construct a preference dataset (target model's failure cases are preferred) for the training of the data generator which performs DPO on top of those preference data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The experiments miss 2 significant baselines. \n* To verify the effect of the proposed RLHF approach, there should a baseline finetuning the data generator (proposer LLM) with a collection of failed samples, and generate a dataset.\n* A strong LLM (i.e., gpt-4o) plays an important role in the proposed method when obtaining the oracle label, so there should another baseline directly prompting the gpt-4o multiple times to generate a synthetic dataset." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "Since the technical section and implementation details are somehow mixed, it is really hard to understand some of the details of each of the experiments. I have the following questions about the evaluation:\n\n- In what experiments, GPT-40-mini used? How was this decided? Is there an ablation study on its use?\n- For the honesty calibration and mathematical reasoning experiment, what number of \"proposer fine-tuning\" iterations are used in ReverseGen? \n- Why is Llama3 used in 4.4 but not in the previous experiments?\n\nCan you add a table that maps each of the variables in the technical section to the specific choice made in each experiment? This would make it easier for the reader to understand each experiment. For instance, the columns could be $M_{prop}, M_{tgt}$, number of examples, seed data, number of iterations, use of GPT-4o-mini, and other implementation details for each experiment." }, "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 new technique for synthesizing failure-inducing data for a target LLM \n\n- The technique is effective in 3 distinct domains of safety, honesty, and mathematical reasoning and shows improvement on SOTA on each. I appreciate the inclusion of Table 5 and Table 7 with examples in the evaluation section" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a technique called ReverseGen for generating failure-inducing examples for an LLM. The technique uses a proposer model that is fine-tuned using pairs of positive and negative examples. The evaluation shows that the generated data can be used to fine-tune and improve models in safety, honesty, and mathematical reasoning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The writing of the paper, especially the separation of the technical section and implementation details needs to be improved. Many of the evaluation choices seem a bit arbitrary and need to be organized better to understand how each of the applications of the proposed technique fits into a single framework (if they do). \n\n- Line 126: “ these studies tend to focus on particular tasks or inputs, and overlook the task generalization and the data salience on model enhancement. Our work differs from these prior works in that we focus on generating failure-guided data for target models”\nCan you be more specific in comparison with each of these prior works? It is unclear to me how ReverseGen differs from many of these works mentioned in the related works\n\n\n- Line 176: The term “solvable” and “unsolvable” is defined here and never mentioned here. The term “solvable” for positive examples is quite unclear. Use a more appropriate name for this.\n\n\n- Line 202: Section 3.2 is the technical section and it starts talking about “gpt-40-mini”. I would recommend authors to separate the implementation details from the technical section. “Gpt-40-mini” is a specific model and used to label the responses, define the model used for labeling as a parameter of the technique that’s instantiated as a specific model. \n\n## Minor:\n- Line 189: “We begin by warming up the proposer model $M_{prop}$ with the initial task-specific instruction set” - A bit unclear wording. Can be more technically precise, especially in the technical section of the paper\n\n- Line 190: “3-shot prompts” - this seems like implementation detail as well which would be more appropriate if it was in the evaluation section\n\n- Line 204: What’s R(.)? I don’t see it defined anywhere before\n\n- Line 220: “$M_{ref}$ is the reference model that remains unchanged”: this doesn’t really define what is $M_{ref}$\n\n- Line 227: typo - two dots “..”\n\n- Line 275: why were these hyperparameters chosen?\n\n- Line 284: “Responses from the target model are generated using greedy decoding for deterministic quality measurement, with the temperature parameter set to 0” - Doesn’t greedy already mean temperature doesn’t matter?\n\n- Line 287: “Instructions for the SFT data are produced by proposer models, while responses are derived from the target models using tailored prompts for safety tasks, generated by gpt-4o-mini for knowledge-intensive tasks.” -> the sentence is too long and hard to understand. What are “knowledge-intensive tasks” in this context?\n\n- Line 319: “well-safe”?\n\n- Line 429: “ReverseGen solely relies on synthetic data” - Doesn’t it use MMLU as the initial instruction seed?\n\n- Line 523: “This may result from inadequate assessment of REVERSEGEN’s proposed questions by the current benchmark or the need for more effective fine-tuning algorithms for difficult questions.” - a bit unclear" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 you provide theoretical or empirical justification for why failure-inducing data would be more valuable than standard training data for model improvement?\n* How does REVERSEGEN's performance evolve with increasing iterations? What determines convergence?\n* Could you analyze why failure induction appears less effective for mathematical reasoning (Table 6) compared to other tasks?\n* What are the computational requirements (tokens, API calls, training time) compared to baseline methods?\n* Could you provide more details about the reward mechanism design and validation process?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Novel approach to data synthesis through failure exploration\n* Comprehensive evaluation across three important tasks (safety, honesty, math reasoning)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces REVERSEGEN, a method for generating training data by identifying model weaknesses through failure-inducing exploration." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper lacks clear justification and motivation for why generating failure-guided data would improve model performance\n* No theoretical framework explaining why failure cases would be more valuable than standard training data\n* Table 6 shows similar result w and wo failure induction in math reasoning task, does this mean failure induction does not benefit math reasoning tasks?\n* No analysis of computational costs or token/API budget comparisons with baseline methods\n* Reward mechanism not clearly explained\n* Insufficient baseline comparisons, especially for mathematical reasoning task" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present a novel approach for automatically generating effective training samples from the target model's failure cases by optimizing another model to create samples via iterative preference learning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024forewarned,\ntitle={Forewarned is Forearmed: Harnessing {LLM}s for Data Synthesis via Failure-induced Exploration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yitH9xAHQs},\nnote={under review}\n}" }, "abstract": { "value": "Large language models (LLMs) have significantly benefited from training on diverse, high-quality task-specific data, leading to impressive performance across a range of downstream applications. Current methods often rely on human-annotated data or predefined task templates to direct powerful LLMs in synthesizing task-relevant data for effective model training. However, this dependence on manually designed components may constrain the scope of generated data, potentially overlooking critical edge cases or novel scenarios that could challenge the model. In this paper, we present a novel approach, \\name, designed to automatically generate effective training samples that expose the weaknesses of LLMs. Specifically, we introduce a dedicated proposer trained to produce queries that lead target models to generate unsatisfactory responses. These failure-inducing queries are then used to construct training data, helping to address the models' shortcomings and improve overall performance. Our approach is flexible and can be applied to models of various scales (3B, 7B, and 8B). We evaluate \\name on three key applications—safety, honesty, and math—demonstrating that our generated data is both highly effective and diverse. Models fine-tuned with \\name-generated data consistently outperform those trained on human-annotated or general model-generated data, offering a new perspective on data synthesis for task-specific LLM enhancement." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "data synthesis", "preference learning", "LLM 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/08b445879093527964bd89e7fbecbffdd8389f6b.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": "Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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 questions." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Abundant Experiments: The paper conducts many experiments to show the performances of the proposed and compared methods\n2. Clear Presentation: The paper's presentation is clear and easy to follow" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an improved method for LLM alignment called SAPO. The method targets on mitigating the need for paired preference data in the alignment stage, by introducing EMA model and data augmentation methods for creating dispreference data. Results show that SAPO achieves some improvements compared to selected baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Marginal and Inconsistent Improvement: Experiments show that the improvements from SAPO in many cases are marginal (e.g., DPO-based Llama-3-8B improves 0.8 and Mistral-7B improves 0.1 from the SPIN baseline). Sometimes, it is even worse than baseline methods (e.g., AlpacaEval 2.0).\n2. Potential Low Training Efficiency: Since the method involves sampling and building new dispreferred responses at each iteration, its training efficiency could be problematic compared to typical self-play method like SPIN which samples at the end of each phase. Can you provide detailed profiling of the time consumed for each stage in the iteration?\n3. Weak Baselines: I notice authors to use meta-llama/Meta-Llama-3-8B and mistralai/Mistral-7B-v0.1 (which are the base models, rather than instruct ones) as the baselines. However, these base models have not been aligned using SFT and RLHF to serve as proper baselines against those alignment methods.\n4. Intuition of generating B': I am concerned with the continuity of regenerating middle segment B' from B. How do you select the segment, by mere 256 tokens? In that case, how to ensure that the new B' is semantically continuous to the original C? Besides, it is doubtful that the regenerated B' is guaranteed to be worse than the original one. These questions make me unconfident of the reasonability of the method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Along with some points above, some additional questions are:\n\n**1. Clarification on segment-level augmentation?**\n\nAs the authors stated in Equations (4) and (5), segment B is selected as a target to refine/augment as a rejected response. Then, the segment C is appended to the refined B' for the next iteration. While refining B and C simultaneously as a continuation of A sounds intuitive, the rationale behind appending C to B' is quite unclear as some contextual mismatch could exist between B' and C. What are the intuitions behind the choice of this specific segmentation scheme?\n\n\n \n\n\n**2. Segment token length and general performance?**\n\nThe last paragraph of Section 5.3 ablates different segment token lengths and concludes that 256 was ideal. However, the impact of 256 tokens differs by the expected response length of the dataset, especially in the general chat dataset like Capybara. Interpreting the effect of segment token length by its ratio over the expected response length would more clearly demonstrate the effectiveness of SAPO.\n\n\n \n\n\n**3. Abnormal AlpacaEval generation length for Mistral-SFT in Table 2?**\n\nI noticed that the AlpacaEval 2.0 generation length for wandb/mistral-7b-zephyr-sft is excessively long, as are the DPO and SPIN-DPO trained versions. However, SAPO-DPO-Mistral-7B suddenly recovers to the normal range (referred to the overall generation lengths of the models registered in the official leaderboard). Some clarifications on the excessively long generation length of wandb/mistral-7b-zephyr-sft and some insights on how SAPO-DPO is the only method resolving such issue in the Table would be helpful." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. SAPO introduces a versatile LLM alignment framework that can be used in a single trajectory setting, widening the applicability of alignment methods to diverse tasks.\n2. SAPO demonstrates strong performance in general compared to the default settings of DPO and ORPO.\n3. The paper presents ablations over various experimental axes, providing a better understanding of SAPO's mechanism." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents SAPO, an LLM alignment framework applicable to pairwise preference-based methods like DPO and ORPO with a gradually updated reference model and self-augmented preference pairs. Mistral and Llama models trained with the SAPO framework outperformed conventional offline alignment schemes on both instruction-following benchmarks and leaderboard benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**1. Impact of EMA and segment-level augmentation in SAPO**\n\nSAPO comprises two main techniques that differ from the conventional offline methods (DPO, ORPO): EMA and segment-level response augmentation. While the method is widely tested over different models and methods, the impact of EMA strategies and hyperparameter choices has not been sufficiently studied. For example, the update coefficient $\\alpha$ and EMA update frequency were fixed to 0.5 and 2 throughout the experiments. Also, the impact of segment-level augmentation is partially studied, and only the high-level interpretations were presented in Section 5, lacking an in-depth understanding of how and why it makes SAPO a strong method (which is connected to the questions below). Thus, the necessity of EMA and segment-level augmentation in SAPO is left unclear.\n\n \n\n**2. Clarity in experiments and ablations**\n\nSome explanations are not clear enough to follow. For instance, in the first paragraph of Section 5.3, the experiments on \"on-policy sampling\" are not clear if means: (1) sampling from the trained policy in the middle of training [1] is assumed as rejected samples, (2) sampling multiple responses from the policy that the trained policy is initialized from and somehow labeled as chosen/rejected [2], or else. Also, it is unclear if \"epoch\" in the training details of SAPO and \"iteration\" in Algorithm 1 are equivalent or distinct values. Including these two examples, overall, the paper does not precisely state some terminologies. \n\n \n \n\n\n**References**\n\n[1] Direct Language Model Alignment from Online AI Feedback (Guo et al., 2024)\n\n[2] SimPO: Simple Preference Optimization with a Reference-Free Reward (Meng et al., 2024)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How do the authors address the potential discontinuities caused by segment-level supervision, and what impact do they anticipate these discontinuities might have on the overall model performance and generalizability?\n\n2. In the ablation study (Table 3), the segment-level augmentation shows varying effects on different benchmarks. Have the authors explored any adjustments to this augmentation strategy to mitigate these inconsistencies, or are there alternative augmentation methods they would consider?\n\n3. What is the rationale behind starting with epoch 3 rather than epoch 1 in Figure 2? \n\n4. Can the authors include results for SPIN in Figure 2 to strengthen the comparative analysis and provide more context regarding the effectiveness of their method?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The motivation to integrate EMA and replay buffer into the self-play pipeline is well-grounded. Experimental results across various benchmarks demonstrate the proposed method's effectiveness over contrastive and self-play baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce off-policy reinforcement learning (RL) techniques, including Exponential Moving Average (EMA) and replay buffer, into the self-augmented preference optimization, where only a dataset of desired responses is needed, while negative samples are generated by the model itself during optimization. The proposed approach aims to reuse timely feedback to train the model, thereby mitigating the delays often encountered in the traditional self-play training paradigm, which has separate sampling and training phases. Additionally, the authors propose a segment-level data augmentation strategy, where a segment of the full response is regenerated by the EMA model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. One crucial aspect of the self-play policy optimization paradigm is to construct negative samples that are generalizable. However, the segment-level supervision employed by the authors could lead to noticeable discontinuities due to the concatenation of the regenerated segment B' with the original segment C. This may introduce unintended bias in the preference learning stage, which could affect the model's ability to generalize.\n\n2. The paper lacks a direct comparison between the effect of curriculum learning and the approach of SPIN that alternates sampling and training phases. For instance, including a comparison with the SPIN method in Figure 2 could provide clearer insights into the effectiveness of this approach.\n\n3. The ablation study in Table 3 presents inconsistent results across different benchmarks. The proposed segment-level augmentation does not consistently outperform direct sampling, at least based on the current experimental results. Additional studies or further refinement of the augmentation technique could help clarify its benefits, e.g., investigating how varying the segment length affects performance across different benchmarks, or exploring alternative segmentation strategies that might lead to more consistent 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": 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": "Did the author conduct any experiments to show that the generated $y^{-}$ is indeed worse than $y^{+}$? For example, you might use some off-the-shelf reward model to compute the average rewards of $y^{+}$ and $y^{-}$." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The strengths of this paper are listed as follows\n\n1. The paper proposes a very novel and interesting way to construct negative responses from high-quality dataset\n\n2. The experiments are conducted over a wide range of models and evaluation benchmarks, which makes it solid and comprehensive\n\n3. The idea of leveraging of EMA and replay buffer is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a new method, SAPO, for aligning language models with human preferences. Previous methods either requires labeled pairwise data or an external reward signal provider. SAPO overcome this by constructing negative responses from SFT dataset and then adapt DPO / ORPO to train the model. Experiments are conducted to empirically verify the performance of SAPO" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My first concern pertains to the high-level intuition behind the methodology. SAPO requires to generate negative samples to pair up those targets in SFT dataset. However, the motivation behind this is unclear. Specifically, SAPO generates negative responses by replacing segments of the SFT response with the model's own generation, which produces off-policy negative responses. The the DPO / ORPO loss will push the model to further penalize these responses. My concern is that, given that these responses is already unlikely generated by the model, what is the necessity of penalizing them further? What specific benefit does penalizing such off-policy data offer, especially when their likelihood of occurrence is already minimal?\n\nAdditionally, while SAPO introduces several improvements over SPIN—including a new method for negative sample generation, as well as the use of EMA and a replay buffer—these enhancements can be integrated into SPIN. However, whether SPIN can benefits from these techniques and how SPIN with these compares to SAPO, both at a high level and empirically, is not discussed sufficiently.\n\nFinally, since SAPO's objective is ultimately to fit the SFT dataset responses, I believe that SFT itself should also be included as a baseline for comparison." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Our paper introduces Self-Augmented Preference Optimization (SAPO), a dynamic, scalable training paradigm that outperforms traditional methods by autonomously generating negative responses and integrating real-time data updates." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024selfaugmented,\ntitle={Self-Augmented Preference Optimization: Off-Policy Paradigms for Language Model Alignment},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yizEOJVFFd},\nnote={under review}\n}" }, "abstract": { "value": "Traditional language model alignment methods, such as Direct Preference Optimization (DPO), are limited by their dependence on static, pre-collected paired preference data, which restricts their adaptability and practical applicability. To address this limitation, we introduce Self-Augmented Preference Optimization (SAPO), an effective and scalable training paradigm without the need of existing paired data. Built upon the self-play concept that autonomously generate negative responses, we further involve the off-policy learning pipeline to improve the data exploration and exploitation. Specifically, we employ an Exponential Moving Average (EMA) model along with a replay buffer to enable dynamic updates of response segments, effectively integrating real-time feedback with historical data insights. Our comprehensive evaluations of the LLaMA3-8B and Mistral-7B models across benchmarks—including the Open LLM Leaderboard, IFEval, AlpacaEval 2.0, and MT-Bench—demonstrate that SAPO matches or surpasses established offline contrastive baselines, such as DPO and Odds Ratio Preference Optimization (ORPO), and outperforms offline self-play methods like SPIN." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large Language Model", "Fine-tuning", "Self-play" ] }, "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/5e9b373f0bb3e7ab24dea515874441bacf7dc126.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": "Self-Augmented Preference Optimization: Off-Policy Paradigms for Language Model Alignment" }, "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) The server updating procedures of Algorithm 1 can be further explained. Will the prediction on every testing example be executed on the server? Is the the known confidence of the input $u^i(x)$ required to provide the ensemble prediction?\n\n(2) What are the implications of $\\alpha$ in Definition 3.5 and Definition 3.6? How will it affect the design of practical algorithm, e.g., FedAdav without $\\alpha$?\n\n(3) In Table 1, FedAdav has very large standard deviation on MNIST (#C=2). This phenomenon can be further explained." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**Originality:** It proved the learnability of one-shot FL ensembles with OSL. Specifically, this theoretical analysis highlighted the importance of OOD detection. By combining multiple OSL methods, the proposed FedAdav algorithm achieved superior performance than baselines.\n\n**Quality:** Theorem 3.7 showed the impact of OOD detection function in proving the learnability of one-shot FL ensembles. Experimental results confirmed the effectiveness of the combination of multiple OSL methods in one-shot FL problems.\n\n**Clarity:** The paper was well-written. The motivation of the proposed theory and algorithm is clearly illustrated.\n\n**Significance:** It provides theoretical supports for analyzing one-shot FL ensembles with OSL." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studied open-set learning (OSL) for one-shot federated learning under label shifts. It proved the learnability of one-shot federated learning ensembles with open-set learning. Then it introduced a FedAdav algorithm to combine multiple OSL approaches. Experimental results showed that FedAdav could outperform SOTA baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) The connection between the theoretical analysis and the proposed FedAdav algorithm is weak. Theorem 3.7 shows that OSL is key to improving the performance of one-shot FL. However, it generally works well for all OSL methods, including baselines FedOV and RotPred. It is unclear what properties a OOD detection approach require to enhance one-shot FL. \n\n(2) The proposed FedAdav is simple combination of FedOV and RotPred. As illustrated in section 4.1, both approaches have some limitations. It is not explained why the simple combination of FedOV and RotPred can solve these limitations. Moreover, it is confusing why the combination of FedOV and RotPred enables better learnability of one-shot FL ensembles in Theorem 3.7.\n\n(3) The parameter sensitivity of FedAdav is not analyzed. There are two key hyperparameters: $T_{check}$ and $\\tau$. Both affects how FedOV and RotPred loss function are performed in the proposed 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": 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.FedAdav simply combines the loss functions from FedOV and RotPred. What is the purpose of doing so? How does this enhance the method’s capability?\n2.The trade-off parameters in FedAdav are derived through experimental results. How can you ensure that the same parameter settings would work effectively if the dataset or model changes?\n3. The experiments only use very simple CNN models, such as LeNet, on simple datasets. Can such a simple setup properly evaluate the impact of OSL in one-shot FL? Why did the authors not use more sophisticated models, such as Transformer-based models, on more complex datasets?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper includes a comprehensive theoretical analysis proving the utility of OSL in the context of one-shot FL.\nExtensive experimental comparisons with multiple baselines are provided." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the effectiveness of open-set learning (OSL) in improving one-shot federated learning (FL), especially when facing label skews. The authors provide a theoretical proof of OSL's benefits and propose a new method, FedAdav, combining features from previous works." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The proposed method lacks innovation, as it is simply a combination of the loss functions from two existing works (FedOV and RotPred).\nThe experiments use simple models and datasets, making it difficult to effectively validate the proposed method's robustness and general applicability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 carefully address W1." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper introduces a novel approach, FedAdav, that combines multiple OSL methods, enhancing the handling of label skew in one-shot FL. This combination of OSL signals is innovative for managing highly imbalanced class distributions across federated clients.\n\n2. The theoretical analysis is rigorous, with proofs supporting the benefits of using OSL for addressing label skew in one-shot FL. \n\n3. The paper is well-organized, presenting both theoretical foundations and empirical results to support the proposed approach. The steps in the algorithm and experiment setups are clearly described." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the challenge of label skew in one-shot federated learning (FL), where clients communicate with the server only once, and class distributions across clients are imbalanced. Existing open-set learning (OSL) methods, like FedOV, help by identifying unknown samples but are limited in flexibility. The paper proposes an adaptive algorithm, FedAdav, combining multiple OSL signals to improve accuracy under label skew. The theoretical contribution proves the learnability of one-shot FL with OSL, and extensive experiments show FedAdav's effectiveness in enhancing performance over other state-of-the-art (SOTA) OSL methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My biggest concern is about the setting of this paper, which only considers the one-shot FL. In a more general case, we usually have multiple communication rounds and the model will be aggregated in each round. However, the paper only focuses on one-shot FL, where clients communicate only once and the models are not aggregated. Instead, only the model predictions are aggregated. This limitation may restrict the effectiveness (at least from the theoretical part) in the general multi-round communication settings.\n\n2. Limited Real-World Data Testing: While the experiments use standard datasets like MNIST and CIFAR, these do not fully represent the complexities of label skew in real-world FL applications, such as varying regional disease prevalence in healthcare. Adding real-world data (if any) could strengthen the practical relevance of FedAdav." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 further insights into the scenarios where FedAdav might underperform compared to other methods? The paper could benefit from a more detailed discussion on the limitations of the proposed method, including scenarios where FedAdav might not perform as expected." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper theoretically proves that closed-set learning cannot effectively address label skews, whereas integrating OSL into FL could ensure the learnability of one-shot federated learning, which is the main contribution and a significant exploration.\n- The theoretical analysis and empirical results are of reasonable quality, providing support for the proposed method.\n- The paper is generally well-structured, but some sections could benefit from more detailed explanations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the issue of label skews in one-shot federated learning (FL) by integrating open-set learning (OSL) techniques. The authors provide a theoretical analysis proving the learnability of one-shot FL ensembles with OSL algorithms and propose FedAdav, an adaptive algorithm that combines multiple OSL signals to improve ensemble accuracy under label skews. Extensive experiments demonstrate that FedAdav outperforms state-of-the-art OSL algorithms in severe label skew conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper makes a reasonable contribution to the field of federated learning, yet there are areas that could be improved.\n- Although the theoretical part seems solid, the proposed algorithm appears somewhat incremental, which shares similarities with existing OSL methods. I suggest the authors to explore the algorithm's performance in other federated learning challenges such as data heterogeneity with different Dirichlet distributions to demonstrate its broader application potential.\n- The empirical results are somewhat limited and could be strengthened by additional experiments on more diverse datasets or real-world applications. I recommend the authors conduct experiments in practical application areas such as natural language processing to validate the algorithm's effectiveness in different domains. Additionally, consider testing on more challenging datasets, such as large-scale image datasets and multilingual text datasets, to demonstrate the algorithm's robustness and generalization capabilities.\n- The presentation of the theoretical proofs could be made more accessible to readers who are not experts in the field. I suggest the authors add more intuitive explanations at key steps, such as using charts to illustrate the proposed algorithm." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We study the theory and experiments on open-set learning for label skews in one-shot federated ensemble." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024openset,\ntitle={Open-Set Learning for Addressing Label Skews in One-Shot Federated Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yj6P8OdWyj},\nnote={under review}\n}" }, "abstract": { "value": "Federated learning (FL) is crucial for collaborative model training, yet it faces significant challenges from data heterogeneity, particularly label skews across clients, where some classes may be underrepresented or absent entirely. In one-shot FL, where clients only communicate with the server once, this problem becomes even more challenging. Recent solutions propose incorporating open-set learning (OSL) to tackle this issue by detecting unknown samples during inference, but current methods like FedOV lack adaptability to varying client data distributions. In this paper, we provide a theoretical analysis proving that improving OSL algorithms can effectively address label skews in one-shot FL, since one-shot FL is learnable through good OSL algorithms regardless of label skews. We also empirically evaluate state-of-the-art OSL algorithms and identify their limitations. Based on these insights, we propose FedAdav, an adaptive algorithm that combines OSL signals to significantly improve ensemble accuracy in one-shot FL under label skews. Through extensive experiments, we demonstrate that exploring better OSL is key to overcoming label skew challenges in federated learning." }, "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", "open-set 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/6ac6434d8b21591937ca84fcd668caf89564d366.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/3a19bad288d903f0e961395da58314ecff7f71d2.zip" }, "title": { "value": "Open-Set Learning for Addressing Label Skews in One-Shot Federated 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": 3 }, "primary_area": null, "questions": { "value": "Can the method proposed in this paper be applied to general audio such as sound and music?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper introduces the first unified pre-training framework (UniWav) for speech representation learning and generation.\nUniWav can compete with different foundation models with low bitrate speech tokenization and high-quality resynthesis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "UniWav is an encoder-decoder framework designed to unify pre-training representation learning and generative tasks. With the appropriate design choices for pre-training, UniWav can jointly learn a representation encoder and generative audio decoder that can be applied to both types of tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The results of the method in this paper do not show an advantage over existing methods in speech recognition and speech generation. \n2. In the speech tokenization section, there is a lack of experiments related to the modeling performance of the speech language model (LLM-TTS) as mentioned in SpeechTokenizer[1]. Such experiments could effectively evaluate the potential of the tokenizer proposed in this paper when applied to autoregressive speech synthesis.\n[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": 3 }, "primary_area": null, "questions": { "value": "1. How does the experimental section provide evidence for the claim that simultaneous pre-training on speech representation learning and generation enhances performance compared to pre-training on only one task, and what justifies the assertion of reduced overhead and cost?\n\n2. Could you include comparisons with models like Natural Speech 3, Funcodec, Encodec, DAC, and DinoSR to improve the evaluation of the proposed methods in speech generation and tokenization tasks?\n\n3. Could you incorporate additional subjective and objective metrics, such as Mean Opinion Score (MOS) and Mel/STFT distance, provide a more comprehensive assessment of model performance?" }, "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 conducts experiments across multiple tasks like speech recognition, text-to-speech, and tokenization. It includes analyses, such as ablation studies and mutual information metrics.\n2. The paper is well-organized and clearly presents technical details, such as the encoder-decoder structure and the Flow Matching method, making it easy to follow. Visual aids and concise explanations further contribute to the clarity of the complex concepts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes UniWav, a unified pre-training framework for speech representation learning and generation. Traditionally, pre-training models for speech have been specialized either for discriminative tasks, like speech recognition, or for generative tasks, such as text-to-speech. UniWav aims to bridge this gap by integrating both functions into a single encoder-decoder model. The encoder is responsible for learning robust speech representations, while the decoder generates speech through Flow Matching." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Firstly, I have some doubts regarding the motivation and novelty validation in this paper. The introduction states that ideally, speech representation learning and generation can mutually enhance each other and that this approach can reduce the overhead and cost of pre-training. However, I did not find evidence for these conclusions in the experimental section. Specifically, it seems that the paper does not address whether pre-training on just one task (either speech representation learning or generation) would yield better performance on downstream tasks compared to simultaneous pre-training on both tasks. Additionally, the model's performance on downstream tasks does not appear to be very strong; for instance, its performance on the speech recognition task is worse than that of other baselines. So the experiments can not show that the overhead and cost can be reduced. Because there is neither proof that pre-training on both tasks together is better, nor evidence that there are significantly better results on a single downstream task. If these advantages cannot be clearly demonstrated, then why combine the two tasks? Merely putting speech representation learning and speech generation together without thoroughly explaining the rationale and benefits of this approach significantly limits the contributions of the paper. Overall, the results seem to contradict the stated motivation and novelty, or these points have not been well validated. \n\n2. Secondly, the selection of baselines in the paper is quite limited, especially for the speech generation and tokenization tasks. Speech generation could be compared with Natural Speech 3 [1], while the speech tokenization task could benefit from comparisons with the latest models like Funcodec [2], Encodec [3], and DAC [4]. Furthermore, the paper should include comparisons with the experimental results of DinoSR for both the speech recognition and tokenization tasks, as the encoder component of the paper is primarily based on the DinoSR model.\n\n3. Lastly, the metrics used for evaluating downstream tasks are insufficient. For example, in the speech generation task, subjective evaluations such as Mean Opinion Score (MOS) should be included, as this is a critical metric. For speech tokenization, additional metrics related to speech reconstruction, such as Mel/STFT distance, could be incorporated.\n\n[1] Ju, Zeqian, et al. \"Naturalspeech 3: Zero-shot speech synthesis with factorized codec and diffusion models.\" arXiv preprint arXiv:2403.03100 (2024).\n\n[2] Du, Zhihao, et al. \"Funcodec: A fundamental, reproducible and integrable open-source toolkit for neural speech codec.\" ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024.\n\n[3] Défossez, Alexandre, et al. \"High fidelity neural audio compression.\" arXiv preprint arXiv:2210.13438 (2022).\n\n[4] Kumar, Rithesh, et al. \"High-fidelity audio compression with improved rvqgan.\" Advances in Neural Information Processing Systems 36 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. Based on Figure 2, have the authors considered only conditioning the decoder on the first, e.g., 12 layers of the encoder? \n2. Can the authors clarify how speaker similarity is computed? Is it the average cosine similarity between the WavLM speaker embedding of pairs of ground-truth and synthesized utterances? If so, I think it would make sense to share the standard deviation as well. \n3. Can the authors comment on their batch size and how they are sampled (e.g., like wav2vec 2.0)? This would help with reproducibility, and figuring out how much data is seen throughout pre-training, and where UniWav lies in Figure 2 of the DinoSR paper (trade-off between performance and data efficiency)\n4. Line 233, \"sinusoidal pos. enc. is appended to the input.\" Is this appended to the feature dimension or the time dimension? Isn't it normally the case that positional embeddings are summed with the input features? Can the authors comments on this design decision?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "### Originality\n\nThe method proposes a new speech SSL method, combining existing methods DinoSR (encoder-only) and SpeechFlow (encoder-decoder) . This method has strong performance on generative and discriminative tasks compared to foundation models which are generative-only or discriminative-only. \n\n### Quality\n\nThe method is evaluated on multiple speech technology tasks, and a small ablation study is performed for further insights. \n\n### Clarity\n\nThe paper is well-written, easy to follow, and appropriately places itself into existing literature.\n\n### Significance\n\nThis work will definitely spark future work in the speech community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes a self-supervised speech representation learning objective which combines 1) masked prediction with online clustering from an EMA teacher model (DinoSR) to train a transformer-encoder network and 2) reconstructing noise-inserted input data based on the encoder representations with Flow Matching to train a transformer-decoder network. The aim of this method is to unify the creation of foundation models for discriminative tasks (such as ASR) and generative tasks (such as TTS). \n\nThe method is evaluated by pre-training on 60k hours of LibriLight, and fine-tuning for speech recognition, speech synthesis, and speech tokenization and resynthesis, on Librispeech. \n\nFor speech recognition, they show limited degradation of performance compared to SSL methods like HuBERT, WavLM, and data2vec. \nFor speech synthesis, they show performance matching contemporary models like VoiceBox and SpeechFlow. \nFor speech tokenization and reconstruction, performance exceeds SpeechTokenizer and HuBERT+HifiGAN. \n\nThe method is also ablated on the encoder and decoder depth, which shows that a 12-layer encoder does not benefit from adding the decoder objective, while a 24-layer encoder does benefit. Moreover, it is shown through a mutual-information analysis that their encoder has different characteristics on how speaker and speech information is processed compared to HuBERT." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "### domain\nThis works only evaluates on Librispeech, i.e., the audio book domain, which has very homogenetic speaker conditions. The observation that ASR performance closely matches HuBERT, WavLM, etc, does not take into account the robustness these models have to other domains. The authors could discuss this limitation or add experiments with the SUPERB benchmark.\n\n### modelling \nAs the proposed method is an extension of DinoSR, it would be nice to see UniWav w/o decoder, as in Table 3 for 200k steps, in Table 1 with 600k steps of pre-training. Moreover, Speechflow also uses an encoder-decoder model. This encoder could in theory be fine-tuned for ASR. The original work did not do this. Can the authors comment on expected results when the SpeechFlow encoder would be fine-tuned for ASR, and how this would compare to UniWav? Could this be discussed in a future work section?\n\n### speaker recognition\n\nOne of the central claims is that UniWav bridges the inherent orthogonality between speech recognition, which normalizes over speaker and environment information, and speech synthesis, which requires speaker and environment information. This is touched upon slightly by the mutual information analysis, where UniWav is seen to lose speaker information through the encoder layer. It would seem to me that UniWav will not be better for the speaker recognition task than, e.g., WavLM, based on the speculation that most environment and speaker information is stored in the decoder. I think evaluating UniWav on the (SUPERB) speaker recognition and speaker verification task would strengthen the claim significantly. I think for now, this limitation should be made more explicit in the limitation section, or if possible, the authors could perform additional experiments on SUPERB.\n\n### Minor comments\n\n1. line 215: Encodec has not been introduced yet, so cite it here. I find this paragraph confusing due to missing context on how the network uses Encodec as input features (line 239) instead of Mel spectrograms as suggested in line 108. \n2. line 236: \\proposed~is\n3. line 242: in~\\ref \n4. line 358: kmeans instead of k-means\n5. line 370: we follow...run k-means, use e.g., by first identifying .. on which to run" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "On line 108, the authors claim that the model uses mel spectrograms as input, but on lines 239–240, they mention using Encodec features instead." }, "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 to jointly pre-train for discriminative and generative tasks. The paper shows SOTA, if not comparable, results on the generative tasks and lag a little on the discriminative task.\nThey show very good results when tokenizing speech using UniWav, compared to other works in the literature. \nThe paper is well written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces UniWav, a new pre-training framework focused on both speech representation learning for discriminative task and for generation task. Using an encoder-decoder architecture, UniWav learns speech representations and generates speech together, letting it perform well in both discriminative tasks and generative tasks. UniWav aims to unify the approach, simplifying the speech processing pipeline and reducing need for many specialized pre-trained models. Experiments in speech recognition, text-to-speech synthesis, and speech tokenization show UniWav's effectiveness, achieving results competitive with other top models for specific tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors propose UniWav, a pre-training method for jointly learning representations for both discriminative and generative tasks in speech processing. This method uses an encoder-decoder architecture to learn effective representations as part of the training process.\n\nFor the discriminative task of speech recognition, the authors fine-tune the encoder on 100-hour and 960-hour splits, showing competitive results compared to other methods. However, they do not provide results for low-resource scenarios (e.g., 10 hours or 1 hour of data) or results using a frozen encoder (SUPERB benchmark). These two setups are essential to support their claim that the learned encoder representations are effective for speech recognition tasks. Because with large finetuning data, effect of pre-training is less. \n\nFigure 2 shows that the encoder’s full capacity is not optimized for learning features specific to speech recognition. UniWav achieves the highest mutual information (MI) on layer 10 out of 24, which then gradually decreases. This pattern suggests that the model divides its capacity between the discriminative and generative tasks, similar to observations in WavLM and prior works. This could explain UniWav’s strong performance in speech generation tasks, as a significant portion of the model’s capacity is allocated to optimize for generation. This trend is consistent across the paper’s reported results.\n\nRegarding speech tokenization, it’s unclear if the comparison is entirely fair. For speech tokenization, the input to the SpeechTokenizer is raw audio, while for UniWav, the input is Encodec encoder features, which are further transformed by UniWav’s 10-layer encoder. It is as if UniWav is heavily overparameterized and most of the parameters are used for generation 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": 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 \"For speech recognition with a shallow encoder, we found introducing the decoder degrades WER regardless of the size. Interestingly, an opposite trend is observed when the encoder size is doubled\"? This is something requires more analysis and explanation.\n\nIn section 3.1, \"We extract the prequantized latent feature from a 16kHz EnCodec (D´efossez et al., 2022) encoder, pre-trained on the same dataset, to serve as the input for our model.\" This is contradictory to the claim in early sections that the input to your model is mel spectrum? In addition, this makes the comparison with other approaches unfair since you used another model as the pre-processing module (i.e., total model size is actually increased significantly)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The main contribution of this work is the combination of DinoSR and a flow matching model to train a unified representation model. While not considered significant, this special setup is somewhat novel.\n\nThe paper contains sufficient experimental results to support their claims although some setups are questionable (in detail below). \n\nThe paper clearly indicated the limitations of the work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework named UniWav. It is an encoder-decoder framework aimed at unifying pre-training representation learning and generative tasks. The authors claimed that this is the first such framework and achieves comparable performance to different existing foundation models, each trained on a specific task, on speech recognition, text-to-speech, and speech tokenization tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The claim their work is \"the first unified pre-training framework for speech representation learning and generation\" is arguable. For example, the paper \"Speechgptgen: Scaling chain-of-information speech generation.\" introduced an approach to train a discrete representation that works for both understanding and generation. The paper \"Moshi: a speech-text foundation model for real-time dialogue\" further improved the approach. \n\nThe presentation needs to be improved. \n1. the DinoSR part in Figure 1 is confusing since the softmax layer is not clearly shown since the encoder, according to the text description, does not include that layer. In contrast, the figure in the original DinoSR paper is very clear. \n2. The majority of Section 2.1 is describing DinoSR. However, notation is not clearly explained sometimes. for example, what is \"k\" in s^k_v right below eq. 3? it's unclear why Euclidean distance (instead of COS distance) is used given that Cos distance is usually more preferred in higher-dim spaces. \n3. notation \"z\" is overloaded in eq 9.\n4. the footnote under Table 1 is not clear. alignment is only used for generation tasks?\n5. eq 11 is wrong? argmin returns an index instead of a representation.\n\nWhen researchers train a representation model they usually keep the encoder fixed when using them in downstream tasks such as ASR. however, in this work the encoder is finetuned. This causes the claims weaker. Similarlly if flow matching is used in the generation part the quality of the generated speech will of course become better. However, this gain comes from the flow matching model instead of the way the representation is learned. Some clarification here is needed." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "First unified pre-training method for speech representation learning and generation" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024uniwav,\ntitle={UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yj9lLwMjnE},\nnote={under review}\n}" }, "abstract": { "value": "Pre-training and representation learning have been playing an increasingly important role in modern speech processing. Nevertheless, different applications have been relying on different foundation models, since predominant pre-training techniques are either designed for discriminative tasks or generative tasks. In this work, we make the first attempt at building a unified pre-training framework for both types of tasks in speech. We show that with the appropriate design choices for pre-training, one can jointly learn a representation encoder and generative audio decoder that can be applied to both types of tasks. We propose UniWav, an encoder-decoder framework designed to unify pre-training representation learning and generative tasks. On speech recognition, text-to-speech, and speech tokenization, \\proposed{} achieves comparable performance to different existing foundation models, each trained on a specific task. Our findings suggest that a single general-purpose foundation model for speech can be built to replace different foundation models, reducing the overhead and cost of pre-training." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "speech foundation model", "generative pre-training", "self-supervised learning", "speech generation", "speech tokenization" ] }, "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/043d484a53d1bb42e7cdc6c81e226e7fc074cecc.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": "UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Given the model's limitations with non-rigid motions, did the authors explore any alternative solutions, such as enforcing additional temporal consistency constraints or incorporating motion priors for articulated objects?\n\nWhile the quantitative results are promising, did the authors consider conducting a user study to assess perceived motion realism, as subjective assessments might capture nuances that FID/FVD cannot?\n\nCould the authors elaborate on how sensitive the model's performance is to the choice of the 180-degree rotation in the self-attention map? Did they experiment with other configurations for reversing the temporal interaction?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Strengths\nThe paper’s fine-tuning approach makes effective use of a pre-trained model (SVD) to generate backward motion without requiring extensive additional data or full retraining. This demonstrates an efficient approach to model adaptation.\nBy developing forward-backward motion consistency through temporal self-attention, the method generates smooth and coherent transitions, especially in scenarios with long differences between keyframes. \nThe paper provides good experimental results, using both qualitative comparisons and metrics like FID and FVD to validate performance improvements over established baselines (FILM, TRF, etc.).\nAblations explore the impact of various components and the paper transparently discusses limitations, providing clarity on the model's boundaries, especially with non-rigid motion types." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Generative Inbetweening, a method for creating intermediate frames between two keyframes by adapting a pre-trained image-to-video diffusion model. This model adapts Stable Video Diffusion with dual-directional diffusion: generating video frames that interpolate both forwards and backwards in time. \nThis approach achieves motion-coherent inbetween frames through a technique that involves reversing the temporal self-attention maps within the U-Net model to generate backward motion from the endpoint keyframe, then combining this with forward-motion frames to produce smooth video sequences.\nEvaluations on the Davis and Pexels datasets show the method’s performance against the existing techniques, including TRF and FILM, in terms of frame coherence and motion fidelity for larger motion gaps." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the paper includes comparisons with baseline models, it lacks an in-depth discussion on the unique metrics or benchmarks used to capture differences between models, particularly in subjective aspects like motion realism. Including a more detailed discussion on why certain metrics (e.g., FID or FVD) were selected over others could clarify the relevance of the performance gains.\n\nThe model relies heavily on SVD’s motion priors, which, as the authors note, can struggle with non-rigid or complex kinematic movements. \nWhile the paper acknowledges this, further discussion on how future models might address such limitations, possibly by incorporating other motion datasets or additional temporal constraints, would add depth to the future directions.\n\nAlthough the fine-tuning approach is a strength, it may be challenging for readers unfamiliar with diffusion models to follow the model adaptation process fully. More visual aids or pseudocode detailing the fine-tuning and dual-directional sampling steps would enhance clarity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. As noted in the weaknesses, I observed that in the cases provided, the camera and object movements between keyframes are slight. Can this method still perform well when there is a significant difference between the given keyframes? Additionally, when the keyframe difference is large, the backward generation may be unable to reuse the rotated attention matrix from the forward generation, potentially causing large discrepancies in frames generated at the same time step. In such cases, can fusion still work effectively?\n2. This generation pipeline seems to require a substantial number of corresponding points between keyframes. Beyond the issue of low overlap mentioned earlier, I’m also curious whether the method could still generate smooth transitions if, for example, one object in the keyframes—such as a fish in the ocean—undergoes a mirrored flip, meaning every point has a mapped counterpart but with an orientation change.\n3. The paper adopts simple averaging for intermediate frame fusion (line 281), but intuitively, frames generated closer to the initial keyframe might exhibit higher quality. Why not use weighted averaging instead? For example, linearly blending frames based on their proportional distance from each keyframe might yield smoother transitions and higher quality." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "This paper fills a gap in the field of large-scale video generation, specifically keyframe interpolation, at a related lower cost. As summarized earlier, this paper presents a novel pipeline to generate synchronized frames and targeted frame fusion techniques to achieve smooth transitional videos." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on the keyframe interpolation problem, which has been overlooked in existing large-scale video generation models. The article proposes a solution to this task by treating keyframe interpolation as a forward video generation from the first frame and a backward generation from the last frame, followed by a coherent fusion of the generated frames. Based on this, the paper reuses existing large-scale image-to-video models to obtain a video generation model for backward motion by reversing temporal interactions. Additionally, it uses sampling techniques to blend paired frames generated by the forward and backward temporal directions with synchronized paths, producing intermediate frames." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The keyframes shown in the paper have relatively small motion ranges and require extensive pixel mapping; otherwise, obvious artifacts occur (as mentioned in the limitations), making this approach unsuitable for large-scale object or camera movements." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 kindly refer to the Weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is clear and easy to understand, with well-presented motivation and methodology.\n- The proposed method is novel, straightforward, and effective, demonstrating improvements over the selected baseline interpolation methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author proposes a novel method for distant keyframe interpolation, leveraging pretrained image-to-video diffusion models. This approach generates intermediate frames by predicting and fusing forward and backward video sequences, conditioned respectively on the given start and end frames. The author introduces a lightweight fine-tuning technique to tackle the key challenge of predicting backward video sequences from the end frame. Additionally, a dual-directional diffusion sampling strategy is employed to effectively fuse noise in both forward and backward directions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Further ablation studies on the proposed method could explore:\n1. **Training Dataset Scale**: In the paper, the model is fine-tuned with only 100 videos. It would be interesting to investigate how the scale of the training dataset affects the model’s performance.\n2. **Fine-tuning Modules**: The paper fine-tunes only the value and output projection matrices in the self-attention layers of the backward framework. Since there might be a gap for the forward motion in the context of the image-to-video task and the interpolation task, it would be worth exploring whether the interpolation performance could be further improved by fine-tuning both the forward and backward framework matrices while preserving the attention map rotation operation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- There are three publicly available weights for Stable Video Diffusion (img2vid, img2vid-xt, img2vid-xt-1-1). Which of these weights did the authors use?\n\n- Stable Video Diffusion applies different classifier-free guidance scales to each frame. Did the authors use the same approach in this paper?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This method is both parameter-efficient and data-efficient, making it highly effective even with limited resources.\n\n- It leverages an open-source model, which enhances its accessibility and contributes to the broader video interpolation research community.\n\n- It demonstrates superior qualitative and quantitative performance compared to FILM, a well-known method for large motion interpolation, as well as TRF, which also uses Stable Video Diffusion.\n\n- In Section 5, the qualitative results are thoroughly explained, clearly highlighting the strengths of this approach in various aspects." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the keyframe interpolation problem by leveraging the large-scale image-to-video diffusion model, Stable Video Diffusion, to generate frames between a pair of keyframes. \n\nUnlike traditional image-to-video models that generate frames in a forward-moving manner, this paper proposes finetuning the model for backward-moving videos and utilizing both the original and finetuned models together during inference. \n\nTo leverage the knowledge from the forward-moving model, only the value and output projection matrices of the 3D self-attention layers are trained, and the attention maps from the forward-moving videos are rotated by 180 degrees and inserted into the finetuned backward-moving model. \n\nDuring inference, the attention maps generated by the forward-moving model are rotated and applied to the finetuned backward-moving model, and the predictions from both models are fused. \n\nThis approach demonstrates superior performance over FILM and TRF on the Davis and Pexels datasets, despite being trained on only 100 videos." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Video interpolation performance can vary significantly based on FPS and the magnitude of motion, but the paper does not provide any analysis of these factors. Besides the motion bucket ID mentioned in the paper, Stable Video Diffusion also takes FPS as a condition. The paper would benefit from demonstrating whether the method still outperforms FILM and TRF when varying the motion bucket ID and FPS during finetuning and inference.\n\n- The proposed method requires both the base forward-moving model and the finetuned backward-moving model during both training and inference, making it more computationally intensive compared to a baseline of \n fine-tuning on video interpolation dataset." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024generative,\ntitle={Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ykD8a9gJvy},\nnote={under review}\n}" }, "abstract": { "value": "We present a method for generating video sequences with coherent motion between a pair of input keyframes. We adapt a pretrained large-scale image-to-video diffusion model (originally trained to generate videos moving forward in time from a single input image) for keyframe interpolation, i.e., to produce a video between two input frames. We accomplish this adaptation through a lightweight fine-tuning technique that produces a version of the model that instead predicts videos moving backwards in time from a single input image. This model (along with the original forward-moving model) is subsequently used in a dual-directional diffusion sampling process that combines the overlapping model estimates starting from each of the two keyframes. Our experiments shows that our method outperforms both existing diffusion-based methods and traditional frame interpolation techniques." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "generative keyframe interpolation", "image-to-video diffusion 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/e1142e3c8ea4f762b7354c1eae008048722f605c.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/8c91d7dd13923cc570777edb10c41c551c8600e2.zip" }, "title": { "value": "Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation" }, "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. Given the grid search used to determine optimal hyperparameters, could a deeper analysis reveal why certain values work best for measuring sample difficulty? Specifically, how do these parameters influence the balance between easy and hard examples selected for the coreset, and could this inform a more consistent method for tuning them?\n\n2. ELFS currently uses SwAV and DINO as feature extractors for clustering. Would more powerful encoders, such as CLIP, improve the quality of pseudo-labels or provide more stable performance across datasets? Additionally, what effect might these alternative encoders have on the distribution of selected hard and easy examples?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. ELFS effectively addresses the limitations of previous label-free coreset selection approaches, providing a feasible solution that leverages deep clustering for pseudo-labeling.\n\n2. By employing double-end pruning, ELFS improves the selection of informative samples, achieving consistent performance improvements over baselines, even in challenging scenarios.\n\n3. The evaluation across multiple datasets and pruning rates, along with an ablation study, showcases ELFS's flexibility and robustness, which may benefit a range of vision tasks.\n\n4. The authors show that including more challenging samples enhances model performance, with ELFS effectively prioritizing hard examples through double-end pruning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents ELFS (Effective Label-Free Coreset Selection), a method designed to improve label-free coreset selection by estimating data difficulty scores without requiring ground truth labels. The authors tackle challenges in label-free selection by employing pseudo-labels from deep clustering to approximate training dynamics and mitigate distribution shifts with a double-end pruning technique. ELFS shows superior performance over existing label-free methods across various vision benchmarks (e.g., CIFAR10, CIFAR100, and ImageNet-1K) and achieves results close to those of supervised selection methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiments involve numerous hyperparameters, optimized through grid search. A more in-depth analysis of the underlying reasons behind these optimal values would strengthen the understanding of how different parameters affect the measurement of sample difficulty, offering clearer insights into the importance of hard examples.\n\n2. The approach heavily relies on feature extractors like SwAV and DINO for clustering. It remains unclear if using more advanced encoders, such as CLIP, could further improve performance or stability, suggesting potential limits in ELFS's generalizability with different encoders." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 consider responding to the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. It is an elegant and effective idea to estimate the data difficulty score through deep clustering. This handles the challenge to measure the prediction uncertainty and sample difficulty without any human labels.\n\n2. The proposed method is evaluated on multiple classification benchmark, showing notable performance gain compared with state-of-the-arts.The design of each module is well justified through ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new label-free coreset selection algorithm called ELFS to relieve the costly human annotation efforts. ELFS utilizes the deep clustering to generate pseudo-labels and estimate data difficulty scores. Afterwards, a double-end pruning method is introduced to mitigate the bias of data difficulty scores. Experiments show that ELFS can surpass previous label-free coreset selection baselines on several benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My major concern lies in the selection of hyper-parameter $\\beta$. I can understand they require some grid search for hyper-parameters. However, according to Fig. 5, the optimal value is different for multiple datasets or sampling ratios, which is quite inefficient. For example, if there is a large dataset with millions of images, it is infeasible to do grid search on it.\n\n2. Based on Tab. 7, it is quite strange that ResNet50 cannot outperform ResNet18 on the selected subset. I assume it reasons from the simplicity of CIFAR10. Maybe the authors can do the transferability experiments on complex datasets like ImageNet since it is a main difference between corset selection and active learning. \n\n3. For Sec. 4.1, I assume the formulation of label-free coreset selection is already covered in previous work. It may be moved to Sec. 3 for clarity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "ELFS presents a compelling label-free coreset selection method that reduces the need for extensive and costly labeled datasets while achieving accuracy close to supervised methods. By effectively utilizing pseudo-labels, ELFS not only significantly outperforms other label-free baselines but also exhibits strong performance despite the inherent inaccuracies and noise associated with pseudo-labels. Moreover, the method demonstrates robustness and versatility, showing good transferability across different datasets and model architectures, thereby enhancing its applicability in diverse machine learning tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel method called ELFS (Effective Label-Free Coreset Selection) for selecting coresets without relying on labeled data. This approach uses pseudo-labels derived from deep clustering to approximate training dynamics, enabling the estimation of data difficulty scores. These scores help identify coresets that can be labeled for training high-performance models while minimizing human annotation costs. ELFS addresses the significant performance gap typically found in label-free coreset selection by introducing a double-end pruning technique to manage the distribution shift caused by pseudo-label inaccuracies. This method shows notable improvements in various vision benchmarks over existing label-free methods, demonstrating its ability to approximate the effectiveness of supervised coreset selection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The ELFS method is quite effective, but it mainly builds on familiar techniques like pseudo-labeling and coreset selection. This might make it seem less novel or groundbreaking to those familiar with the field. Despite this, it does a great job using these methods to ensure high accuracy and reliability.\n\nMoreover, to really show how well ELFS works and to expand its use, it would be beneficial to test it on a wider variety of datasets. This includes tackling larger and more complex datasets such as ImageNet, as well as datasets with uneven distributions or long tails. Testing ELFS in these contexts would help validate its effectiveness across different challenges and environments.\n\nPotential Application Areas for ELFS: Beyond vision tasks, are there other types of data or tasks where ELFS could be effectively applied? Exploring its adaptability to different domains like text, audio, or even structured data could open up new applications.\n\nExplanation of Hard and Easy Examples in Section 4.4.2: Could a visual representation or graph be used to clarify the difference between hard and easy examples as discussed in the section? Visual aids could help illustrate how ELFS handles these types of data, enhancing understanding of its approach.\n\nAnalysis of Data Distribution in Table 1: Is it possible to analyze further how the data distribution of the coreset selected by Random compares to that selected by ELFS? Understanding the differences in selection criteria and resulting coreset characteristics could provide deeper insights into the strengths and limitations of ELFS compared to simpler random sampling methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Please explain why double-end pruning helps the performance.\n2. Do you fine-tune the model with the coreset? Or do you train the model from scratch?\n3. Do you use only the coreset to train the model? It would be better to show the result of using the coreset as the labelled set and the rest data as the unlabelled set to train a model with a semi-supervised learning algorithm such as SemiReward[1]. If, with the help of semi-supervised learning, a randomly sampled labelled set achieves good performance, and the labelled set selected by your model yields similar performance, then the benefits of using a coreset to train the model need to be clarified.\n```\n[1] SemiReward: A General Reward Model for Semi-supervised Learning, Siyuan Li and Weiyang Jin and Zedong Wang and Fang Wu and Zicheng Liu and Cheng Tan and Stan Z. Li, ICLR 2024\n```" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The motivation of this paper is solid, and the topic of this paper exactly matches ICLR.\n2. The writing of introduction clearly delivered the motivation and idea.\n3. The experiment result looks good. It's interesting that many methods even cannot beat random sampling as suggested in Tab.1.\n4. The ablation study is extensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new policy to sample a core subset for deep models. It introduces a deep clustering with the pseudo-labelling to estimate the score for each sample. Meanwhile, they try to fix the bias issue of pseudo-labelling. Experiments demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some sentences are redundant, such as these two questions proposed in the paper.\n2. It would be better to move sec 4.1 to sec 3 to give readers an overview of the problem you are solving.\n3. My **main concern** is that more benchmarks in different distributions should be evaluated. As described in the paper, this method relies on a pretrained vision encoder to get the visual features for each sample. Then, a deep clustering algorithm is introduced to get the pseudo labels and scores. However, the evaluated datasets in this paper are too easy for pretrained vision encoders. I believe that much of the data in the evaluation datasets is included during pretraining. If we use a dataset in a different distribution, such as a medical image dataset, without a good visual feature, will this method still work?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a novel label-free coreset selection methods (ELFS) that outperforms existing baselines on four vision datasets." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024labelfree,\ntitle={Label-Free Coreset Selection with Proxy Training Dynamics},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yklJpvB7Dq},\nnote={under review}\n}" }, "abstract": { "value": "High-quality human-annotated data is crucial for modern deep learning pipelines, yet the human annotation process is both costly and time-consuming. Given a constrained human labeling budget, selecting an informative and representative data subset for labeling can significantly reduce human annotation effort. Well-performing state-of-the-art (SOTA) coreset selection methods require ground truth labels over the whole dataset, failing to reduce the human labeling burden. Meanwhile, SOTA label-free coreset selection methods deliver inferior performance due to poor geometry-based difficulty scores. In this paper, we introduce ELFS (Effective Label-Free Coreset Selection), a novel label-free coreset selection method. ELFS significantly improves label-free coreset selection by addressing two challenges: 1) ELFS utilizes deep clustering to estimate training dynamics-based data difficulty scores without ground truth labels; 2) Pseudo-labels introduce a distribution shift in the data difficulty scores, and we propose a simple but effective double-end pruning method to mitigate bias on calculated scores. We evaluate ELFS on four vision benchmarks and show that, given the same vision encoder, ELFS consistently outperforms SOTA label-free baselines. For instance, when using SwAV as the encoder, ELFS outperforms D2 by up to 10.2% in accuracy on ImageNet-1K." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Coreset Selection", "Data pruning", "Label free coreset selection" ] }, "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/f458be4e5699f8fca6e3d45dbb4a8dfec290677c.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/65b14e08567aabdff96068d0173a7651ac7ff51c.zip" }, "title": { "value": "Label-Free Coreset Selection with Proxy Training Dynamics" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- **Q1:** Use of a `pretrained diffusion model` is mentionned. I can see this type of models for images, but for complex data with varying domain of definition, like in many scientific applications, the diffusion model needs to be retrained for each type of data no? IT is not completely clear how reusable these models are outside computational imaging.\n\n- **Q2:** The prediction-correction scheme strongly relate to the usual Plug-and-Play methods classicaly used in inverse problem resolution. The proposed scheme is related to the Forward-Backward Splitting (FBS) method, which is typically used in the DiffPIR [B]. From what I understand, equation (9) in this paper corresponds to equation (13) in [A]. Could you elaborate on the main difference of between the proposed method and the DiffPIR framework? (except the derivation free aspect). Other missing references are [B], a survey on using diffusion models for inverse problems, and [C], which also consider diffusion prior for inverse problems, with gradient based method but applied to the Black-hole imaging problem. Adding comparison with their method would be very useful.\n\n- **Q3:** To make is possible to evaluate how good the results are compared to methods that are able to access the gradient of the forward operator, it is necessary to add a few methods that have access to the forward operator's gradients. Indeed, even though ODE solvers are not always natively differentiable, there are more and more works that consider making them differentiable, for instance using `jax` for Navier stokes using pseudo-spectral solver [here](https://github.com/google/jax-cfd). The interesting question is: should we spend some time making them differentiable or do we not gain much by doing so. Therefor, quantifying how much is lost on simple cases as the ones presented here is necessary to make the case of derivative free methods. In particular, adding the results of the DPS method and the DiffPIR method would be very useful at least for the imaging task. Note that these models are both implemeted in the [`deepinv`](https://deepinv.github.io) library (see [here](https://deepinv.github.io/deepinv/deepinv.sampling.html)). Also, adding the DPS base line and PnP-DM form [C] for black hole imaging would better illustrate how much we loose by note considering the gradient of this differentiable operator. Adding them for Navier-Stokes would also be very interesting but probably more challenging.\n\n- **Q4**: The value of $J$ and $Q$ in the experiments are not reported. Could the author provide them? From equation (16), I understand that we need to compute the forward operator $J$ or $Q$ times at each iteration. From the NAvier stokes experiment, assuming that the procedure is run for 1k steps, I guess $J = 295$ and $Q=2048$? How were these value chosen? What are their impact on the results? Also, would having access to the gradient mean going roughly 100 times faster than EnKG? (The computational cost of computing the gradient through autodiff is approximately x2/3 times the cost of evaluating the forward operator). Note that the metrics chosen (# of evaluation of Fwd/DM, Seq) are not clear. Better definition of what they represent mean would be useful. Adding the total runtime of the method would help a lot to assess the computational cost.\n\n- **Q5:** In the black-hole experiment, how many simulation are used to train the diffusion model?\n\n### Minor comments, nitpicks and typos\n\n- Missing ref: \n- l.066: \"One more challenging\" -> \"On\"\n- l.069: \"More computationally efficient\" -> than what?\n- l.078: \"often Gaussian\" -> They are almost never Gaussian, but they are modeled as such and this gives reasonable results.\n- l.141: \"and Nelder-Mead simplex methods\" -> Extra `,`.\n- l.198: The drop of the subscript $x$ for the gradient is confusing\n- Eq (7) -> The notation $\\Delta x_i$ is not defined. As it is not used anywhere else, I would recomment using $\\|x_{i+1} - x_i\\|$ instead.\n- Eq (12) -> $x_i'$ should have a super script $(j)$. It is also not completely consistent for $x_{i+1}$.\n- l.302: \"instead\"\n- l.316: `our approach outperforms the standard strong DPS baseline` -> The DPS result are not present in the table, so I think they are missing?\n- l.335: What is the `EDM` framework?\n\n\n- l.847: The change from \"l\" to \"j\" should be made explicit as it is not immediately clear and can appear as a typo. \n\n### References\n\n[A] : Daras, Giannis, et al. [A survey on diffusion models for inverse problems.](https://arxiv.org/pdf/2410.00083) arXiv preprint, 2024. \n[B] : Zhu, Yuanzhi, et al. [Denoising diffusion models for plug-and-play image restoration.](https://arxiv.org/pdf/2305.08995) Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. \n[C] : Wu, Zihui, et al. [Principled Probabilistic Imaging using Diffusion Models as Plug-and-Play Priors.](https://arxiv.org/pdf/2405.18782) arXiv preprint, 2024." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Interesting approach to solve inverse problems.\n- Derivative-free approaches can be useful in many cases and have received much less attention than gradient-based methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new approach to solve inverse problems using a derivative-free optimization method based on the Ensemble Kalman Filter. The core idea is to approximate the data-fidelity term gradient with a statistical\nlinearization from the ensemble Kalman methods. The method is applied to three types of inverse problems: computational imaging problems, the Navier-Stokes equation, and the black-hole imaging problem. The method is compared other derivative-free baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Positioning relative to the state-of-the-art is not clear, in particular with respect to the proposed framework, which seems to be a variant of the existing ones (see **Q2**).\n- The evaluation is not completely satisfactory (see **Q3, Q4**)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1) How does EnKG perform when applied to larger-scale problems or high-dimensional data? Are there specific limitations to its computational efficiency?\n2) Is there any dependence on the performance and the pertrained model? How sensitive is the method to the quality and type of pre-trained diffusion model used? What are the implications if a suitable model is not available?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1)\tEnKG operates without gradients, needing only black-box access to forward models, making it highly applicable to complex inverse problems with unknown or undefined derivatives. The proposed PC framework generalizes existing methods, enabling adaptability across various inverse problems without retraining.\n2)\tThe current work demonstrates strong performance, notably in complex tasks like the Navier-Stokes equation, outperforming gradient-based solutions\n3)\tProvides deeper understanding and new interpretations of diffusion-based approaches, contributing to the field of inverse problem-solving." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Ensemble Kalman Diffusion Guidance (EnKG), a novel derivative-free method for solving inverse problems using pre-trained diffusion models. Traditional approaches often require detailed knowledge of the forward model, such as derivatives, which limits their applicability. EnKG overcomes this by relying solely on black-box evaluations of the forward model and a pre-trained diffusion model.\nKey contributions are twofold: 1) EnKG operates solely with black-box access to forward model evaluations and a pre-trained diffusion model, making it particularly useful in scenarios where derivative information is inaccessible. 2) The authors introduce a prediction-correction (PC) framework that utilizes the empirical covariance matrix of ensemble particles during the correction step. This innovation allows EnKG to effectively bypass reliance on gradients, enhancing its applicability in non-linear inverse problems.\nThe paper demonstrates the effectiveness of EnKG across various inverse problems, including applications in fluid dynamics. These examples highlight the method's capability to handle complex, non-linear scenarios that are common in scientific research.\nIn summary, this work expands the toolkit for addressing inverse problems in machine learning by introducing a flexible and robust approach that maintains the generative power of diffusion models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The method may face challenges when scaling to very large models or high-dimensional data, as ensemble-based approaches can become computationally expensive. A further insight along this line would be useful. Also, it relies on pre-trained diffusion models, which might limit effectiveness if high-quality models are not available for certain tasks.\n2) The empirical validation focuses on specific problem sets; broader testing across diverse applications would strengthen the generalizability claim. \n3) A further analysis on the algorithm complexity would be beneficial as the combination of the prediction and correction steps might introduce additional computational and implementation complexity due to ensemble covariance estimation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- Line 316: Is the DPS baseline included in the table?\n- Black-Hole Imaging Problem: In the black-hole imaging problem, how is $G(\\phi(x_i, t_i))$ computed if $G$ is unknown and $\\phi(x_i, t_i)$ differs from the observed data? Could you provide a step-by-step explanation of how the black-box forward model is handled in the case? Additionally, please clarify if any assumptions are made about $G$ or generated samples in this scenario.\n- Proof of Lemma 1: The proof shows a monotonic decrease in $\\text{tr}(C_{xx}^{(i)})$. Why does this quantity converge to zero? From another perspective, a vanishing covariance implies the trajectories converge to a single point. Does this implication contradict the ill-posed nature of inverse problems, which typically have many possible solutions?\n\nPossible Errata\n- Lines 225 and 284: Remove \\Gamma .\n- Line 731: Remove the extraneous ‘(‘.\n- Line 847: Replace with $\\approx$." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tNovel Approach: The paper introduces statistical linearization within ensemble Kalman methods to diffusion-based inverse problems, a novel concept in this context.\n2.\tInnovative Guidance Term Formulation: The authors present a unique formulation for the guidance term, with a clever trick that replaces the derivative of the forward model with covariance from forward evaluations.\n3.\tComprehensive Validation: The effectiveness of EnKG is demonstrated across three different scenarios: (1) cases where the forward model is known and differentiable, (2) cases where the forward model is known but differentiating it is impractical (e.g., PDE-based models), and (3) cases where the forward model is a black box, with observations as the only available information." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Diffusion models have been used to address inverse problems, with numerous diffusion-based solvers that avoid retraining existing diffusion models. These approaches typically rely on pseudo-inverses or derivatives tied to the forward model. This paper introduces a novel diffusion-based inverse solver designed for cases where the forward model is unknown.\n\nThe authors propose Ensemble Kalman Diffusion Guidance (EnKG), a derivative-free method that utilizes only forward model evaluations along with a pre-trained diffusion model.\nIn the proposed method, guidance term is computed as follows:\n1. Particles are initialized to compute covariance in following steps.\n2. During the diffusion trajectory, the particles are pushed by ODE solver.\n3. Then, synthesized samples are applied to the forward model and compute covariance of them.\n4. Diffusion trajectory is updated by the formula given the EnKG.\nThe proposed method replace the derivative of the forward model by computation of covariance and ODE solving.\n\nThe empirical results demonstrate the effectiveness of EnKG across diverse inverse problems, including cases where the measurement operator is treated as a black box. Specifically, the method is applied to image inversion problems with explicit forward models, Navier-Stokes inverse problems where the forward model is computed by solving PDEs, and black hole imaging, where the forward model is a black box." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited Discussion of Related Work: The paper lacks depth in interpreting and explaining related works.\n- The motivation behind the weighting matrix $w_i C_{xx}^{(i)}$ is unclear. Could you provide further explanation on the intuition and reasoning behind the choice of weighting matrix?\n- Although the Kalman method is a core component, the paper does not provide a thorough explanation of its role and mechanics in this context. Specifically, which parts of the method are directly applied from existing literature and which represent novel contributions of this paper? For example, the introduction of weighting matrix, the derivation using local linearity of the operator in the proof, and the convergence claim. Related to above questions, please clarify the intuition and motivation provided by the literature versus that introduced by the authors.\n2. Overstated Contribution: The claimed contributions seem somewhat overstated. The concept of Predictor Corrector interpretation in guidance-based methods is not entirely new." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Plot (b) in Figure 3 is difficult to interpret. Could you clarify what the values 0.2, 0.4, ..., 0.8 represent? Additionally, what do \"Seq # DM\" and \"Seq # DM grad\" refer to in this context?\n- Plot (c) reports the runtime of the proposed algorithm, with an average of approximately 140 minutes. This is a considerable computational cost (about 2 hours) for solving one inverse problem.\n * How does this runtime compare to other algorithms (aside from EKI) ?\n * Can you comment on the practical applicability of the method given this runtime?\n * Considering the high computational cost, how does the algorithm perform relative to methods that fine-tune or train smaller network components for the guidance term, as seen in [1, 2, 3, 4]?\n\n\n---\n.. [1] Black, Kevin, et al. \"Training diffusion models with reinforcement learning.\" arXiv preprint arXiv:2305.13301 (2023).\n\n.. [2] Uehara, M., Zhao, Y., Black, K., Hajiramezanali, E., Scalia, G., Diamant, N. L., ... & Levine, S. (2024). Fine-tuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194.\n\n.. [3] Fan, Ying, et al. \"Reinforcement learning for fine-tuning text-to-image diffusion models.\" Advances in Neural Information Processing Systems 36 (2024).\n\n.. [4] Denker, Alexander, et al. \"DEFT: Efficient Finetuning of Conditional Diffusion Models by Learning the Generalised $ h $-transform.\" arXiv preprint arXiv:2406.01781 (2024)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Introduction of an algorithm that solves inverse problems with diffusion models prior that only requires point-wise access to the forward model" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors address inverse problems using diffusion models as priors within a restrictive setup where the forward model is accessible only point-wise.\nThey employ a Predictor-Correct framework, where in the prediction stage, samples are drawn from the prior using the ODE describing the diffusion model.\nIn the correction stage, a MAP problem involving the forward model at the current diffusion step is solved through gradient-based updates.\nUnlike previous approaches, the authors approximate the intractable term in this MAP formulation by evaluating it on samples generated via the ODE.\nSince only point-wise access to the forward model is available, the authors estimate the gradient through statistical linearization and ensemble Kalman methods.\nThis approach maintains a set of particles and uses them, along with their centroids, to approximate the gradient.\nThe authors validate their algorithm on three different inverse problems." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Methodological concerns**\n\n- The authors’ motivation for their \"Derivative-free correction step\" in Lines 246–263. The matrix $C_{xx}$ appears without adequate justification, and in Equation (12), the authors invert $C_{xx}$ even though it is a singular matrix, as the number of samples used to compute it is less than the dimensionality of the problem. Consequently, $C_{xx}$ cannot be treated as a preconditioning matrix in this context.\n- In the appendix, the authors base their proofs on Equation (20), which is introduced without sufficient explanation. This equation appears to correspond to the ensemble update, which assumes an estimation of the gradient, the very objective of the lemmas and propositions that follow. This circular reasoning raises concerns about the validity of the proofs.\n\n\n**Technical concerns**\n\n- In Lines 125–127 (paragraph following Equation (4)), the authors suggest that the guidance term depends on the noise scheduler $\\dot{\\sigma} \\sigma$. However, this dependence result from their specific formulation. As verification, the authors can review Equation (5) in [1] or write DPS's algorithm in terms of the score. The guidance term is not scaled by the noise scheduler.\n- The statement in Lines 194–195 is misplaced. Specifically, $\\log \\hat{p}(y | x_{i+1})$ is a composition of the simulated ODE and the forward model, making it highly non-convex. Therefore, the hypothesis of convexity is unrealistic. Besides, this term varies at each diffusion step, as it is composed with the ODE at different time steps, which diverges from the requirements outlined in [2], Chapter 4. Hence, the iterative updates may not converge to a true MAP estimate within this setups.\n\n\n**Errors and clarifications**\n\n- Equations (12)–(14) lack clarity. The variable $x_{i+1}'$ is undefined, and while the argmin is specified with respect to $x_{i+1}^{(j)}$, this variable does not appear in the equations.\n- In Lines 897–903, the gradients of $p$ are missing a logarithmic term.\n- The first part of Assumption (3) regarding $C_{xx}$ is redundant. Since $C_{xx}$ is a positive semi-definite matrix, its trace is non-negative, being the sum of positive eigenvalues. Therefore, if the trace of $C_{xx}$ is zero, $C_{xx}$ must be the zero matrix.\n- In Lines 316–317, \"DPG\" should replace \"DPS,\" as DPS is not included in the experiments.\n- The term \"guidance\" is repeated twice in Line 52.\n\n\n---\n\n.. [1] Chung, Hyungjin, et al. \"Diffusion posterior sampling for general noisy inverse problems.\" arXiv preprint arXiv:2209.14687 (2022).\n\n.. [2] Parikh, Neal, and Stephen Boyd. \"Proximal algorithms.\" Foundations and trends® in Optimization 1.3 (2014): 127-239." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024ensemble,\ntitle={Ensemble Kalman Diffusion Guidance: A Derivative-free Method for Inverse Problems},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ykt6I21YQZ},\nnote={under review}\n}" }, "abstract": { "value": "When solving inverse problems, it is increasingly popular to use pre-trained diffusion models as plug-and-play priors. This framework can accommodate different forward models without re-training while preserving the generative capability of diffusion models. Despite their success in many imaging inverse problems, most existing methods rely on privileged information such as derivative, pseudo-inverse, or full knowledge about the forward model. This reliance poses a substantial limitation that restricts their use in a wide range of problems where such information is unavailable, such as many scientific applications. To address this, we propose Ensemble Kalman Diffusion Guidance (EnKG) for diffusion models, a derivative-free approach that can solve inverse problems by only accessing forward model evaluations and a pre-trained diffusion model. We study the empirical effectiveness of our method across various inverse problems, including scientific settings such as inferring fluid flows and astronomical objects, which are highly non-linear inverse problems that often only permit black-box access to the forward model." }, "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": [ "inverse problem", "diffusion model", "derivative-free" ] }, "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/3da29777b80ef4b3784cee06c0329a5d8f04d3ed.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/e9918658d4fe1d1e0440c2f7e877f1f68437545d.zip" }, "title": { "value": "Ensemble Kalman Diffusion Guidance: A Derivative-free Method for Inverse Problems" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The only two clarification issues I had are:\n- Details regarding human annotator guidelines, annotator recruitment, and compensation. This would help in better understanding the reliability of the human annotations.\n- More specifics regarding licensing (i.e., a table with licensing terms for each dataset being used)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "A strong benchmark paper should satisfy some of the following dimensions (along with some commentary): \n- the task is useful\nDifficult document retrieval is a long- and widely-studied problem. It is both more important in the era of LLMs due to increased reasoning capabilities, but potentially less important as more information is encoded in the parameters (modulo time-sensitive, etc.). Thus, additional motivation could help justify the significance of the work. \n\n- the dataset is large enough and non-trivial to construct\nThis is also mixed; the dataset isn't particularly large and is ostensibly of varying quality between Stack Exchange, Coding, and Math. That being said, it is clearly more complex than many existing benchmarks for at least a subset of the questions. For some questions, the quality is seemingly higher (i.e., more human validation) than existing datasets.\n\n- there are sufficient details regarding the construction of the benchmark\nIncluding the appendices, there are a lot of details -- to the point where I am confident I could replicate most of the results. However, the amount and clarity of the procedure for different data sets (Stack Exchange, Coding, and Math) isn't as detailed for all cases. Also, it isn't clear in general what the human annotation guidelines were, how annotators were recruited, and how they were compensated (unless it is just the authors and volunteers). However, the details are solid overall.\n\n- the tools provided reduce friction for new people to work on this\nCode is provided and was used to run several experiments. I didn't dig through the code and thus do not know how easily it is to conduct experiments. However, I am reasonable confident it is sufficient.\n\n- the baseline models tested on the benchmark are non-trivial\nThe authors conduct several experiments over several different retrieval engines including state-of-the-art systems on related datasets.\n\n- the benchmark answers new questions or enables new solutions\nThe authors did conduct experiments beyond just IR performance and were able to address some of these questions using this dataset. The discussion in these sections could be strengthened, but it is solid in this regard overall.\n\nEvaluating the paper with respect to the stated dimensions,\nOriginality: There are multiple 'hard QA/IR' datasets, but the emphasis here is on IR for reasoning-heavy scenarios -- which is timely and a useful contribution. Many have likely considered such datasets, but the execution here is better than a first attempt.\nQuality: Overall, the work is well-motivated, well-executed, and sufficiently rigorous. My primary concern in this regard is variance in quality between different benchmark types (QA, Math, Coding) and that this is a relatively small dataset.\nClarity: Overall, the paper is easy to understand and has sufficient details, especially when considering the appendices. The figures are helpful. My two suggestions in this regard are a Table comparing Bright to the most related datasets and more discussion regarding the empirical results including specific references to cells in the tables (i.e., I didn't always know which cells I was looking at when validating quantification claims).\nSignificance: I am fairly certain that at least part of this benchmark will be used, but not sure if all parts will be used. Additionally, it would have more potential impact if it was a larger dataset (or there was clear evidence that it covers some expected 'production' distribution)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose Bright, an information retrieval benchmark that focuses on questions that require 'reasoning-intensive' retrieval to provide relevant content. Whereas most existing benchmarks focus on factoid questions where there is significant semantic overlap between the query and the retrieved content (or some 'translation', but not 'reasoning'). Specifically, Bright is actually 12 datasets including 7 datasets from StackExchange covering {Biology, Earth Science, Economics, Psychology, Robotics, Stack Overflow, Sustainable Living}, 2 coding data settings from {LeetCode (python), Pony}, and 3 math reasoning datasets {TheoremQA-Question Retrieval, TheoremQA-Theorem Retrieval, Art of Problem Solving}. Details are provided regarding the procedures to collect each dataset. In short, for StackExchange, human annotators browse recent posts and select a post with at least 5 upvotes and contain at least one URL link -- which are human validated to produce questions, answers, and relevant documents. Negative documents are collecting via identifying semantically similar but irrelevant documents (i.e., negatives) via a Google search powered method. These are human validated for unanimity. Pony coding adapts a code generation dataset to retrieving pages from manuals to cover syntax and LeetCode via a fairly straightforward crawling procedure. Math reasoning adapts TheoremQA to retrieve math queries that either use the same theorem as the query's solution (question retrieval), theorems from ProofWiki (theorem retrieval), or Math Olympiad problems (Art of Problem Solving) matching other problems that use the same theorems. Experiments are conducted based on 13 different retrieval engines including sparse retrieval, open-source dense retrieval, and proprietary models. The important findings is that nDCG@10 has variance amongst different systems (i.e., improvements can be made) while being relatively low as compared to other benchmarks (i.e., it is difficult). Additional experiments show that querying with LLM reasoning (i.e., chain-of-thought) improves performance (i.e., reasoning is needed for retrieval, irrespective of underlying retrieval method), retrieval improves RAG-based results (i.e., retrieval is an important problem). They also demonstrate that reranking with increasingly powerful LLMs improves retrieval performance, Bright appears robust with respect to data leakage in pre-training (i.e., pre-training doesn't cover reasoning requirements as much as most tasks), and that long-content retrieval (e.g., legal, medical) is more difficult." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "On the other hand, below are some of my concerns regarding this paper (some previously mentioned):\n- I would like to see a statistics-level comparison between Bright and competing datasets in a table (\"Benchmarking retrieval\" (line 104) and \"Benchmarking reasoning\" (line131) in Related Work section)\n- I would like clarification regarding the annotation guidelines, recruiting experts, compensation for lines 248-253.\n- To get the details for coding and math, one pretty much has to read the appendices and from what I can tell, the significance and quality of the Stack Exchange questions is the strongest aspects of the paper.\n- The appendices are more detailed (to the point where they actually seem different from the text). However, still no details regarding annotation guidelines, recruiting experts, and compensation (unless the authors did all of this)\n- The dataset seems relatively small; if I am incorrect, I would recommend a table contrasting this with other datasets (along with other aspects).\n- For reasoning steps, a bit more from the appendix (e.g., StackExchange vs. coding vs. math stratification) would be helpful in the main text with discussion.\n- In general, it isn't clear that ordering matters for RAG settings, so NDCG-based results may not be that useful as in IR settings. I also would recommend rank-biased precision and recall (i.e., evaluations similar to 'needle-in-a-haystack' settings.\n- As implied in other areas, there are a lot of results, so more specific interpretation would be helpful (but I am aware of the page limit).\n- While the authors claim that there are not licensing issues, I wasn't able to verify this. Obviously, if there are licensing issues (for academic research within commercial organizations?)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "### Questions \nPerhaps I missed something, but do you think any of the following analysis will be helpful for future works - \n- A quantitative analysis that shows the rank of the gold document for the best retriever (QWEN).\n- A qualitative analysis that examines why strong LLMs (e.g., GPT-4 in Table 3) fail to correctly rank relevant documents. The gold document can be added as an oracle to when it was not retrieved. If not, do you think that a different analysis regarding model errors can be helpful? \n- The ratio of evaluator failures described in lines 454-458 (although this is minor).\n\nAgain, maybe I am missing something, but did you also experiment with the StackExchange answers as positive passages?\n\n### Suggestions \n- The abstract mentions that “The leading model on the MTEB leaderboard which achieves a score of 59.0 nDCG@10,1 produces a score of nDCG@10 of 18.0 on **BRIGHT**”. Consider adding which model this is, because the MTEB leaderboard is changing constantly.\n\n- The example for “Level 1: Keyword-based Retrieval” only states that the part of highway 401 is *one of the widest*, while the question asks for the widest highway. I understand this is from the NQ dataset, but a positive passage that directly answers the query might be simpler for the reader.\n\n- The appendix is sometimes referenced as Appendix (e.g., Appendix C) and sometimes as § (e.g., §B.3). Consider using one for consistency (this is of course very minor).\n\n- The subsets of **BRIGHT** are relatively small, with the smallest including only 78 examples. Adding error bars or variance for the results, either in the paper or in a leaderboard, could help future researchers focusing on a specific subset." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The data collection pipeline is thorough and includes verification by two domain-expert PhD students. \n- Lots has been done to ensure diversity by focusing on several different StackExchange domains, two coding tasks, and an additional effort for including theorem-based questions.\n- The experiments are extensive and cover 13 different retrievers, in addition to two re-rankers.\n- The paper is well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents **BRIGHT**, a benchmark for evaluating retrievers on reasoning intensive tasks. **BRIGHT** contains 1,398 realistic tasks over 12 domains, including Biology, Psychology, Coding, and Math. By experimenting with 13 retrievers, the paper shows that **BRIGHT** is challenging for current systems - the best model reaches an nDCG@10 score of 22.1. Moreover, enhancing queries with reasoning traces improves retrieval performance, retrieval-augmentation with current retrievers improves QA accuracy, re-ranking with LLMs can increase retrieval accuracy, and pre-training on documents in **BRIGHT** does not further increase performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It could have been helpful to add a quantitative analysis in the analysis section. For example, an analysis that examines when models err could be useful for future research (see the Questions section for further discussion and some suggestions).\n\n- There are a few details in the appendix that are not referenced from the main paper (e.g., the Limitations section), and I found the appendix a bit hard to follow. Consider verifying all main sections are referenced, or alternatively adding a small Table of Contents in the beginning of the Appendix." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "- Have you tried finetuning retrievers on this benchmark?\n- In figure 1, level 2 queries (NQ, MSMARCO) are outdated. It would be better to compare BRIGHT’s “intensive” reasoning to recent retrieval benchmarks, such as RAR-b or subset of BeIR and MTEB that goes beyond semantic matching." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The benchmark is challenging and has the potential to drive future research toward developing retrievers that handle difficult queries more effectively.\n- The dataset is human-collected, ensuring authenticity rather than relying on artificially generated data. \n- I liked the comprehensive appendix, which provided valuable insights into the annotation process and the dataset's structure." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents BRIGHT, a challenging retrieval benchmark designed to require deep, beyond-surface-form matching between queries and documents. Human annotators constructed a dataset of 1,398 instances spanning diverse domains. Current retrieval models perform poorly on this benchmark, highlighting its difficulty. The authors also propose several techniques aimed at improving retrieval performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Detailed analysis of results is lacking and some (RAG, reranking) are not surprising.\n- In line 413, how can LLM-reasoning queries enhance the reasoning capabilities of retrievers? If the primary effect is an increase in lexical similarity (BM25's strong performance), should models be specifically trained to leverage this feature to perform well on BRIGHT? Additionally, the results for Coding and Theorem-based datasets (Table 38) appear inconsistent.\n- In line 428, regarding retrieval-augmented generation, the QA performance seems to be already quite high when paired with current retrievers that are reported to perform poorly." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "N/A" }, "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 focuses on creating a new dataset to benchmark reasoning-intensive queries, which is a very important type of queries for RAG systems and search engines, and there does not exist such a benchmarking dataset so far.\n\n2. The dataset covers a variety of domains, including economics, psychology, mathematics and coding, which is quite comprehensive.\n\n3. The retrieval process in the experiments incorporated explicit reasoning about the query and achieved an up to 12.2 points improvement, which demonstrated that reasoning is indeed a bottleneck for such types of queries, which aligns very well with the motivation of the datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper argues that for retrieval tasks, many real-world queries require in-depth reasoning to identify relevant documents, which goes beyond traditional keyword or semantic-based matching focused by existing benchmarks.\nTo benchmark retrievers on such reasoning-intensive queries, this paper created a new dataset, Bright, which consists of 1398 queries spanning diverse domains. \nThe benchmarking on this new dataset demonstrated that state-of-the-art retrieval models perform poorly on this dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is no unified definition of either relevance or required reasoning across data from different domains in this dataset. Though the author wrote different \"relevance\" in different subsections, they are actually not definition of relevance but how the corpus is created.\nIt is acceptable as a dataset, however, there is a lack of deep scientific understanding about what exact (reasoning) ability is required for a retriever model to success on this dataset. Instead what we can learn from this dataset is that the retriever model may need to overfit some ad-hoc collection process.\n\n2. Due to a lack of unified relevance definition, the dataset is more like a collection of different benchmarking applications where different relevant (can be either closely or loosely) information to the query can be helpful. Therefore, rather than using it as a retrieval benchmark to evaluate retrievers, it is more suitable to use it to evaluate the application. For these applications, the retrieval results are just intermediate results and can be difficult to judge whether they can actually help the final 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": 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. (Lines 248-253): Why not use all three annotators independently? It seems you biased the expert annotators with annotations performed by a non-expert\n2. Section 5.2 claims that continued pre-training does not help. However, you mention in Appendix A.3 that the average number went to 21.0! Why was 20.4 picked as the final number in Table 5. It is possible that continued pre-training in this fashion (only showing new documents from Stack Exchange) is making the model lose its generalization abilities. There are two options to do this ablation study -- train a model with a replay buffer, which samples data from the original training data along with stack exchange. Train a model where all data is used together (original training data and stack exchange)\n3. How does an LLM (such as GPT-4o) work on the downstream task. That is give it a question, and evaluate its performance on answer quality (without giving any document)" }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. A high quality, relevant and challenging benchmark for information retrieval tasks\n2. Comprehensive evaluation on a wide variety of models\n3. Love the section on how chain of thought reasoning helps improve the models -- especially BM25!\n4. Downstream task evaluation is also very helpful -- confirms the documents are relevant" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "## Summary\nThe paper presents a challenging and novel retrieval benchmark which requires reasoning to answer queries\n\n## Contributions\n1. The first benchmark to focus solely on reasoning intensive retrieval. The benchmark is constructed from diverse domains and carefully curated to remove possible noise\n2. Thorough experiments are done with respect to a variety of retrieval models -- sparse, open source small models, open source large models and even proprietary models\n3. The authors list the data creation and curation process in detail. This would further help in creating similar such benchmarks" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I could not think of any questions I had which were not answered in the paper. I have some observations which are mentioned in the questions" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024bright,\ntitle={{BRIGHT}: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ykuc5q381b},\nnote={under review}\n}" }, "abstract": { "value": "Existing retrieval benchmarks primarily consist of information-seeking queries (e.g., aggregated questions from search engines) where keyword or semantic-based retrieval is usually sufficient. However, many complex real-world queries require in-depth reasoning to identify relevant documents that go beyond surface form matching. For example, finding documentation for a coding question requires understanding the logic and syntax of the functions involved. To better benchmark retrieval on such challenging queries, we introduce BRIGHT, the first text retrieval benchmark that requires intensive reasoning to retrieve relevant documents. Our dataset consists of 1,398 real-world queries spanning diverse domains such as economics, psychology, mathematics, coding, and more. These queries are drawn from naturally occurring or carefully curated human data. Extensive evaluation reveals that even state-of-the-art retrieval models perform poorly on BRIGHT. The leading model on the MTEB leaderboard (Muennighoff et al., 2023), which achieves a score of 59.0 nDCG@10,1 produces a score of nDCG@10 of 18.0 on BRIGHT. We show that incorporating explicit reasoning about the query improves retrieval performance by up to 12.2 points. Moreover, incorporating retrieved documents from the top-performing retriever boosts question answering performance by over 6.6 points. We believe that BRIGHT paves the way for future research on retrieval systems in more realistic and challenging 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": [ "Retrieval benchmark", "Reasoning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f8d579bcbed2e911c91ea37cb767128ab6930a75.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/b21e98fbc277aaa47c3b6a11f9e18b5e1e83c1c4.zip" }, "title": { "value": "BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "We thank the reviewers for reviewing our paper. After careful consideration, we think our paper is inappropriate for ICLR, and we decided to withdraw our paper." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In lines 260–261, the authors mention using Mask R-CNN to extract a motion mask. Given that Mask R-CNN is primarily an object detection and segmentation method, how is it specifically utilized to obtain the static mask in this context?\n2. In section 3.2, the authors propose to use the scale and shift estimated from the static part for the dynamic part. If a particular frame contains an excessively large moving part (which are common in object-centroid dataset), resulting in a small static part and insufficient static depth, how does this affect the accuracy of the estimated scale and shift and how to handle this? Given that the same scale and shift are applied to the dynamic part, the motion could complicate optimization. Without accurate scale and shift estimates, how can the authors ensure successful optimization in such scenarios?\n3. It is evident that the initial module is crucial, as it provides both point clouds and the initial pose, with the core of this module being optimization based on view consistency. However, the depth information provided for optimization is inconsistent in scale and shift, and the pose estimation begins from scratch, espectically under synamic condition. How easy is it to optimize under these conditions? Is there a risk of falling into local optima during the optimization process?\n4. In lines 294–296, the authors mention using 5% and 10% thresholds. Why were these specific percentages chosen? They appear intuitive. Additionally, what criteria are used to sort the depth values within a frame to derive the 5% and 10% results?\n5. How is the proposed CSM network trained? Where does the ground truth for the Gaussian splats (GS) come from? \n6. How is the canonical space for each object defined? Can this be applied to each object individually? Additionally, is the static part also transformed into the canonical space? The authors mention using CSM to transfer points to the canonical space, but how are these points converted back to global space before rendering?" }, "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 understand.\n2. the results show the effectiveness of proposed pipeline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a dynamic Gaussian reconstruction method. The proposed pipeline consists of two steps: the initial module and the reconstruction module, which handle static and dynamic components separately. The static initial module initializes point clouds and estimates the pose based on static elements, while the dynamic module focuses solely on initializing dynamic point clouds. In the reconstruction phase, the static reconstruction leverages pose estimation and the point clouds, whereas the dynamic part employs a deformation network for dynamic reconstruction and pose estimation. Finally, both static and dynamic components are combined for the final rasterization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In the Methods section, the authors should provide succinct descriptions of the methods employed, such as COGS and monocular depth, rather than solely citing them. Including specific aspects—such as key features/primary steps/crucial settings. This would enhance clarity and allow readers to understand the context and relevance of these techniques within the proposed framework.\n2. In line 292, the authors claim that the dynamic point cloud obtained from the process is unstable due to adjustments made with the static scale and shift values. If the results of the dynamic part are indeed affected by an overall incorrect scale and shift, then simply removing outliers (top 5% and bottom 10%) may not adequately address this issue. While outlier removal can eliminate extreme values, it does not rectify the underlying problem of an incorrect overall scale and shift. The authors should discuss any additional steps taken to correct scale and shift for stability or acknowledge this as a limitation if unresolved.\n3. In line 299, how are the parameters (r and s) for the Gaussian representation obtained? The authors only describe the point cloud in the previous sections. To improve completeness, the authors should provide the mathematical formulation or algorithm used to derive these parameters from the point cloud data.\n4. Figure 3 and the lower part of Figure 1 share the same framework, which is redundant. I recommend removing the lower portion of Figure 1 to streamline the presentation.\n5. Considering NeRF is also a good 3D representation, the authors should also compare to the state-of-the-art NeRF-based dynamic methods, such as DynNeRF[1], CTNeRF [2] , DynPoint [3], MonoNeRF [4]… .\n6. I have some doubts regarding the novelty of this work. Previous dynamic reconstruction efforts typically involve first segmenting the scene into dynamic and static components, using depth prior knowledge to assist in the process. This paper appears to be more of an engineering implementation, resembling a combination of existing approaches rather than presenting novel research contributions. To clarify, the authors could explicitly articulate the unique technical contributions of their approach and explain how it advances the field beyond engineering implementation.\n\n[1] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712–5721, 2021.\n[2] Xingyu Miao, Yang Bai, Haoran Duan, Yawen Huang, Fan Wan, Yang Long, and Yefeng Zheng.\nCtnerf: Cross-time transformer for dynamic neural radiance field from monocular video. arXiv preprint arXiv:2401.04861, 2024.\n[3] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36, 2024.\n[4] Fengrui Tian, Shaoyi Du, and Yueqi Duan. Mononerf: Learning a generalizable dynamic radiance field from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17903–17913, 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the above section for my questions and suggestions." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. A novel approach to handling dynamic scenes without requiring traditional SfM-based initialization, and proposes to use the separation of static and dynamic components for independent optimization.\n2. Enhance the quality of complex spatiotemporal scenes through a combination of high-dimensional feature loss and annealing frequency loss.\n3. Well-structured presentation with clear pipeline illustrations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents IF-MODGS, a novel approach for reconstructing and rendering dynamic scenes using only monocular camera input without requiring pre-computed camera poses or point clouds. The key contributions include a pipeline that separates static and dynamic regions to estimate camera poses from static backgrounds and generate point clouds for dynamic objects; a Canonical Space Mapper (CSM) that defines a canonical space and applies deformation to link it with different viewpoints and timestamps; enhanced quality in complex spatio-temporal scenes through a combination of high-dimensional feature loss and annealing frequency loss." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The reviewer appreciates the authors' efforts in addressing dynamic scene reconstruction without relying on SfM points and camera parameters as known priors. However, there is considerable room for improvement in the experimental design and comparison of results:\n\n1. First, it appears that Table 1 and Table 2 are redundant, as both seem to display quantitative results on the NVIDIA dataset. Out of all 7 samples in this dataset, the proposed method outperforms the baselines in only 3. Additionally, the quantitative results for the UCSD dataset are missing from the paper.\n\n2. Similar to RoDynRF, which employs two separate neural radiance fields to represent static and dynamic regions of a scene, the proposed method uses two distinct sets of 3D Gaussians for modeling these regions. The reviewer wonders about the performance of the baseline if the same motion mask were applied, optimizing static and dynamic scenes independently.\n\n3. How does the computational efficiency of the proposed method compare to the baselines in terms of training time and testing speed? This aspect is highlighted as one of the main advantages of using GS.\n\n4. The proposed method depends on the motion mask predicted by Mask R-CNN. How well would the method perform if the motion mask were imperfect? How robust is the motion mask prediction 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": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Why was the method not tested on the HyperNeRF dataset?\n- In Table 1, what does “w/o initial” mean? Does it omit an initialization module?\n- Could you explain the motivation behind introducing CSM? I am unclear on why two warping fields (CSM and HexPlane) are necessary. A clearer explanation and an ablation analysis in this regard would be helpful.\n- How do you define the static region based on the Mask R-CNN output? For example, in the upper row of Fig. 4, the window could potentially be dynamic as well." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The first dynamic 3DGS approach that operates without pose initialization.\n- Superior performance compared to baseline methods.\n- The paper is clearly written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for dynamic 3D reconstruction without pose initialization using 3D Gaussian Splatting (3DGS). By leveraging color, depth prediction, and motion mask images as inputs, the method decomposes the scene into static and dynamic parts, achieving competitive results on benchmark datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The method is overly engineered, utilizing multiple priors (depth maps, motion masks), loss functions, and multi-stage processing. This complexity can obscure the core idea and contribution of the work. Specifically, the method includes separate \"initialization\" and \"reconstruction\" modules, despite claiming to be an initialization-free approach. Why can’t these modules be unified?\n- Limited novelty: The approach consists of existing modules. The static initialization component is essentially the same as COLMAP-free 3DGS, while the warp field parameterization resembles 4DGS. The method simply breaks down the problem into smaller parts, supported by monocular predictions.\n- The method heavily depends on Mask R-CNN’s motion mask, making it susceptible to errors in network predictions. An ablation analysis assessing the sensitivity to motion mask accuracy from different mask inputs would be valuable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The separation of processing static and dynamic regions in a scene is indeed a reasonable approach. However, it's important to note that many existing methods in SLAM have adopted similar strategies for handling static and dynamic elements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors present IF-MoDGS, a novel approach for scene reconstruction and novel view synthesis (NVS) that eliminates the need for precomputed camera poses and point clouds from Structure-from-Motion (SfM). The method divides the scene into static and dynamic regions, using the static background to estimate camera poses and a specialized dynamic module to handle moving objects. To enhance spatio-temporal consistency, the authors introduce a high-dimensional feature loss and an annealing frequency loss, which improve rendering quality across complex dynamic scenes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are some concerns about the paper:\n\n1. I agree that SfM often struggles to extract accurate camera poses and obtain sparse point clouds in certain scenarios. However, the proposed method has only been tested on the NVIDIA and UCSD datasets, which utilize multi-view camera setups with significant camera angles. In such settings, SfM methods generally perform well in estimating camera poses. If the authors wish to emphasize their contribution to camera pose estimation, it would be beneficial to test the method on casually captured monocular videos, such as those from the DAVIS dataset, where camera angle changes are minimal and pose estimation is more challenging for SfM.\n\n2. What is the accuracy of camera pose estimation compared to the ground truth? How effective and reliable is the proposed pose estimation method?\n\n3. The provided renderd videos exhibit noticeable visual artifacts, such as the car with weird moving speed. Additionally, the videos are rendered from a fixed camera position and angle. How does the rendering performance vary with different viewpoints and positions? Could the authors provide more analysis on this?\n\n4. More importantly, the videos focus solely on the authors' method. Could the authors provide additional video comparisons with baseline methods?\n\n5. Given that the proposed method uses masks, do the baseline methods also utilize masks? How can fair comparisons be ensured when additional information is involved? Furthermore, how critical is the accuracy of the masks? If the mask is inaccurate at the boundaries between static and dynamic regions, what issues might arise?\n\n6. Why are the results for D3DGS so poor? This seems unusual. Could the authors provide more details on the implementation of each baseline method listed in Table 2?\n\n7. Please add citations for the referenced methods in Table 1." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Monocular Dynamic Scene Rendering and Camera Pose Estimation using 4D Gaussian Splatting" }, "_bibtex": { "value": "@misc{\nyeomsuwoong2024ifmodgs,\ntitle={{IF}-{MODGS} : {INITIAL} {FREE} {MONOCULAR} {DYNAMIC} {GAUSSIAN} {SPLATTING}},\nauthor={Yeomsuwoong and Jimin Roh and Eunho Shin and Kyeongbo Kong and Joonsoo Kim and Songju Na and Suk-Ju Kang},\nyear={2024},\nurl={https://openreview.net/forum?id=ylgg2RE7ub}\n}" }, "abstract": { "value": "In the field of scene reconstruction with moving objects, recent studies have utilized 3D Gaussian Splatting (3DGS) for spatial representation. This method typically relies on camera poses and point clouds obtained through the Structure-from-Motion (SfM) algorithm. However, in scenes captured with monocular viewpoints and containing moving objects in each frame, the SfM algorithm struggles to obtain accurate camera poses and points clouds. As a result, it often either removes point clouds of dynamic objects or fails to find camera poses for each frame, thereby leading to sub-optimal rendering of dynamic scenes. We propose a novel approach, Initial-Free Monocular Dynamic Gaussian Splatting (IF-MoDGS) which does not require precomputed camera poses and point clouds in dynamic scenes with moving objects. Our approach estimates camera poses using the static background, separated from dynamic objects by a motion mask, and generates point clouds specifically for the dynamic objects. To handle dynamic objects, we define a canonical space and apply deformation to link it with each viewpoint and timestamp. Then, to improve quality in complex spatio-temporal scenes, we utilize a high-dimensional feature loss and an annealing frequency loss. Extensive experimental results demonstrate that our method can effectively render dynamic scenes without relying on precomputed camera poses and point clouds, achieving the state-of-the-art performance in dynamic scene rendering tasks using a monocular camera. Our project will be available at:https://anonymous.4open.science/w/IF-MODGS-67F5/" }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Yeomsuwoong1", "~Jimin_Roh1", "~Eunho_Shin1", "~Kyeongbo_Kong1", "~Joonsoo_Kim2", "~Songju_Na3", "~Suk-Ju_Kang1" ] }, "authors": { "value": [ "Yeomsuwoong", "Jimin Roh", "Eunho Shin", "Kyeongbo Kong", "Joonsoo Kim", "Songju Na", "Suk-Ju Kang" ] }, "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": [ "novel view synthesis", "4D rendering", "camera pose estimation", "3D reconstruction" ] }, "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": "yeomsuwoong|ifmodgs_initial_free_monocular_dynamic_gaussian_splatting" }, "pdf": { "value": "/pdf/e4b6f188d0ef636dc4f4a82ef2018ad047123f8c.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": "IF-MODGS : INITIAL FREE MONOCULAR DYNAMIC GAUSSIAN SPLATTING" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "> Although the paper provides an approach with computational efficiency, the LP studied in this paper differs from and looks more accessible than the prior work (Osogami et al., 2023).\n\nThe LP studied in Osogami et al. (2023) is a special case of the LP studied in this paper. Specifically, letting $\\theta\\equiv 0$ and $\\rho=0$ in our LP gives the LP in Osogami et al. (2023). Notice that even the LP in Osogami et al. (2023) is unlikely to admit analytical solutions in general, and this is the key technical challenge. Our Lemma 1 gives a sufficient condition that allows us to analytically solve the LP under consideration. Lemma 2 shows that this sufficient condition is necessary whenever types are independent, so our analytical solution is optimal for a wide range of interesting cases in mechanism design.\n\n> The theoretical results' organization is not easy to follow.\n\nWe appreciate your suggestion. We agree that the presentation can be improved. However, some of the intermediate results are of independent interest and are worth being formally stated. For example, Lemma 1 contains essential information that is not contained in Corollary 1." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you very much for your review.\n\n> I would love to hear the author's opinion on the novelty of the BME design in this work.\n\nBME has not been studied as a problem of MAB (i.e. from the perspective of sample complexity), although related problem of estimating the best mean from given sample (i.e. bias correction) has been widely studied in machine learning (as we discuss in Line 161). As is acknowledged in your review, a key novelty is in the connection from mechanism design to multi-armed bandits, where we reveal that the sample complexity of BME is a relevant problem. Although we only discuss its relevance to mechanism design, BME is a fundamental problem that may find a wide range of applications, such as those that require estimating the worst case expected cost (Worst Mean Estimation). For this new problem of BME, we show that a simple approach can achieve the best possible sample complexity. Although the approach is simple, it is nontrivial that this simple approach matches the lower bound (as we discuss in paragraphs staring at 154, 365, and 404). As you suggest in your review, there may be more efficient approaches that improve constant factors, have better instance-dependent complexity, etc., and we expect that our results will provide a basis for such extensions." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you very much for your review.\n\n> [W1] In this regard, in Section 5, the authors are essentially stating the fact \"LP has a solution only when the feasible region of the constraints is non-empty\", which is really trivial.\n\nAlthough it is trivial that every LP has a solution only when the feasible region of the constraints is non-empty, this does not mean that one can analytically derive optimal solutions to all LPs. Our Lemma 1 gives a sufficient condition that allows us to analytically solve the LP under consideration. Lemma 2 shows that this sufficient condition is necessary whenever types are independent, so our analytical solution is optimal for a wide range of interesting cases in mechanism design. Note also that the (special case of) LP under consideration has been solved numerically in the prior work of mechanism design (e.g. Osogami [2023]), which also indicates the nontriviality of our results in Section 5.\n\n> [W2] Similar to the first point, the method described by the authors in the Section 6 is essentially just the basic mean estimation of each arm's reward in stochastic MAB.\n\nThe suggested approach of estimating the mean reward of each arm with $O((1/\\varepsilon^2) \\log(1/\\delta))$ samples can only guarantee that the best mean is estimated within error \\varepsilon with probability at least $(1-\\delta)^K$. To provide a $(\\varepsilon,\\delta)$-PAC guarantee, one would need $\\Omega((1/\\varepsilon^2) \\log(K/\\delta))$ samples from each arm, resulting in the suboptimal sample complexity of $\\Omega((K/\\varepsilon^2) \\log(K/\\delta))$. The novelty of our MAB results lies in proving that $O((K/\\varepsilon^2) \\log(1/\\delta))$ samples are sufficient for best mean estimation, and that this is the best possible sample complexity (Theorem 1).\n\n> [W3] However, in Section 3, the authors do not introduce any information regarding MAB.\n\nThe background on MAB is not needed until Section 6, and we are concerned that some of the readers would not remember what has been stated in Section 3 when they read Section 6. However, we would appreciate the reviewer's guidance on what specific information about MAB would be most helpful to include in Section 3.\n\n> [W4] These two statements seem to conflict.\n\nThe two statements do not conflict. As is stated in Section 3, the types ($t_1, t_2, ..., t_N$) are generated from a fixed distribution $P$. For this fixed distribution $P$, we consider a conditional distribution $P(\\cdot\\mid t_n)$ and take i.i.d. sample from $P(\\cdot\\mid t_n)$ in Section 6.\n\n> [Q2] Prior to line 346, the paper does not mention MAB at all. Are the authors assuming that $t_n\\in[K]$ for all $n\\in[N]$ here?\n\nYes. Note that we assume a finite number of players $N$ and a finite size of each type space $K$ (line 185). In Section 6, we state our results on general MAB problems (using the language of MAB), which are then connected to mechanism design at the end of Section 6 (Theorem 2 and Proposition 3). Specifically, for each player $n\\in[N]$, the type space $\\mathcal{T}_n$ corresponds to the set of arms $[K]$." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. See weakness.\n\n2. Prior to line 346, the paper does not mention MAB at all. Are the authors assuming that $ t_n \\in [K] $ for all $n\\in [N]$ here?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well written and the theoretical results appear to be correct.\n2. The paper improves the previous results in Osogami [2023].\n3. The paper proposes numerical experiments to show the advantages of their designs.\n\n\nOsogami [2023]: Takayuki Osogami, Segev Wasserkrug, and Elisheva S. Shamash. Learning efficient truthful mechanisms for trading networks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies mechanism design problem under multi-armed bandit framework. The authors analytically derive a class of optimal solutions for such an LP that gives mechanisms achieving standard properties of efficiency, incentive compatibility, strong budget balance (SBB), and individual rationality (IR), where SBB and IR are satisfied in expectation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I hold reservations about the contributions in the paper. In Section 3, the authors introduce four properties that the mechanism needs to satisfy: Dominant Strategy Incentive Compatibility (DSIC), Decision Efficiency (DE), $\\theta$-IR, and $\\beta$-WBB/SBB. Such properties should be the key challenges in the mechanism design. However, as the authors stated, directly using the VCG mechanism can satisfy the first two properties. Furthermore, regarding the other two properties, they can be represented as two linear constraints of the optimization problem. In this regard, in Section 5, the authors are essentially stating the fact \"LP has a solution only when the feasible region of the constraints is non-empty\", which is really trivial. In summary, I am not convinced that the method proposed in this paper is innovative or makes sense.\n\n2. Similar to the first point, the method described by the authors in the Section 6 is essentially just the basic mean estimation of each arm's reward in stochastic MAB. The novelty of the proposed method should be further clarified.\n\n3. The title of the paper is \"Mechanism design with multi-armed bandit\". However, in Section 3, the authors do not introduce any information regarding MAB.\n\n4. In Section 3, the authors assume that the types are generated from a fixed distribution. However, in Section 6, the authors state that the algorithm can access to an arbitrary size of the sample that is independent and identically distributed (i.i.d.) according to $P(\\cdots|t_n)$ for any $t_n$. These two statements seem to conflict." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- I would love to hear the author's opinion on the novelty of the BME design in this work. I understand that it serves as a tool for the overall mechanism design; thus it is acceptable if the novelty of this part is limited (in that case, I might need to rely on other reviewers to get an assessment for the novelty in mechanism design)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The automated mechanism design is an interesting problem. While I do not have exact background in this direction, I believe the efforts provided in this work are of relevance and importance to the community.\n\n- The connection from mechanism design to multi-armed bandits is inspiring. With my background in MAB, I largely appreciate such intersection that leverages MAB techniques to faciliate other domains.\n\n- The overall presentation and writing is clear. It has been a smooth reviewing experience for me." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work studies automated mechanism design. First, a class of optimal solutions is derived that requires an exponentially smaller number of essential variables than the previous version of linear programming. To resolve the computational issue, a connection is drawn towards best mean reward identification in MAB. Then, provably efficient design to perform best mean reward identification is provided, which is further plugged back in the original mechanism design problem." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- As I do not have a strong background in mechanism design, I would leave the further judgement of the significance and novelty of this part to other reviewers.\n\n- For the MAB part, while the connection is interesting, I found the adopted technique is a bit straightforward. In particular, while best mean identification (BMI) and best arm identification (BAI) have their differences (e.g., the example in line 380), the upper bound is obtained in Theorem 1 is from an algorithm that perform BAI first while following up with additional samples to do BMI. I, in general, have doubts that this can be done in a more efficient way." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "### Minor Comments\n\n- Line 170, notation $\\mathcal N=[1,N]$ is confusing; how about $\\{1,2,\\dots,N\\}$?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The numerical simulation section is designed to verify several theoretical results, which are good paper complements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies how to solve an LP for mechanism design. It first formulates this LP, which can satisfy four conditions, and then illustrates that the solution of this LP enjoys an exponentially smaller variable size. Then, to approximate the solution, the paper proposes to use the MAB algorithm and shows that this approximation is asymptotic optimal. Numerical simulations are also reported." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Unclear contribution. Although the paper provides an approach with computational efficiency, the LP studied in this paper differs from and looks more accessible than the prior work (Osogami et al., 2023). So, it is hard to evaluate this paper's contribution from the aspects of significance and methodology. It would be helpful if the author could discuss the technical challenges they encountered in this paper. \n2. The theoretical results' organization is not easy to follow. This is a theoretical paper, providing a lot of lemmas and corollaries in Sections 5 and 6, where the essential parts are. However, the authors should put more effort into revising the presentations in these two sections. For example, in Section 5, the Lemmas 1 and 2 composes the Corollary 1. Why not directly give Corollary 1 and move Lemmas 1 and 2 to the appendix? This could help the reader quickly understand the meat of this paper.\nAnother example is that Corollaries 3, 4, and 5 are all on different conditions; why not just have one corollary with three bullets? For Section 6, Lemmas 4 and 5 are components to support Theorem 1. Why not use a proof sketch to posit Lemmas 4 and 5 so that readers familiar with these materials can directly skip them?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We analytically derive the optimal solution to a mechanism design problem and evaluate the solution via a bandit algorithm with theoretical guarantee." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024mechanism,\ntitle={Mechanism design with multi-armed bandit},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ylhKbwJrjC},\nnote={under review}\n}" }, "abstract": { "value": "A popular approach of automated mechanism design is to formulate a linear program (LP) whose solution gives a mechanism with desired properties. We analytically derive a class of optimal solutions for such an LP that gives mechanisms achieving standard properties of efficiency, incentive compatibility, strong budget balance (SBB), and individual rationality (IR), where SBB and IR are satisfied in expectation. Notably, our solutions are represented by an exponentially smaller number of essential variables than the original variables of LP. Our solutions, however, involve a term whose exact evaluation requires solving a certain optimization problem exponentially many times as the number of players grows. We thus evaluate this term by modeling it as the problem of estimating the mean reward of the best arm in multi-armed bandit (MAB), propose a Probably and Approximately Correct estimator, and prove its asymptotic optimality by establishing a lower bound on its sample complexity. This MAB approach reduces the number of times the optimization problem is solved from exponential to linear. Numerical experiments show that the proposed approach finds mechanisms that are guaranteed to achieve desired properties with high probability for environments with up to 128 players, which substantially improves upon the prior work." }, "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": [ "mechanism design", "incentive compatibility", "efficiency", "individual rationality", "budget balance", "multi-armed bandit", "probably approximately correct" ] }, "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/839b4b38995f2d874fb0800392878cbad62efc75.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/9ec9bed47b10aa51e10aadb63b1806ec08e42564.zip" }, "title": { "value": "Mechanism design with multi-armed bandit" }, "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": "* LLaMA-VID [1] can encode images and videos using only two tokens, whereas the proposed method requires a minimum of 64 tokens and does not support video input. What advantages does the proposed method offer over LLaMA-VID?\n\n[1] Li, Y., Wang, C., & Jia, J. (2025). Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision (pp. 323-340). Springer, Cham." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Despite reducing computational demands, the method maintains 98.1% of the original performance, indicating that it is highly efficient without compromising on accuracy.\n* By reducing the number of visual tokens processed by the model, the method is more scalable and flexible than original MLLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To reduce the computational costs associated with MLLMs, the authors propose a method that leverages pre-trained vision encoders to group similar image segments into semantically related concepts without the need for segmentation masks. Besides, the method employs isolated attention to preserve the integrity of original image representations while allowing semantic tokens to aggregate image features into meaningful regions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* My main concern is the novelty of the proposed method. The use of clustering algorithms or q-formers to reduce the number of vision tokens fed into LLMs has been examined in several previous works, including Chat-UniVi [1]. Additionally, the concept of isolated attention is not novel. I recommend that the authors provide a more in-depth analysis of the proposed method to strengthen the paper.\n* To demonstrate the generalizability of the proposed method, I suggest that the authors validate it across a broader range of MLLM architectures or base LLMs.\n\n[1] Jin, P., Takanobu, R., Zhang, W., Cao, X., & Yuan, L. (2024). Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 13700-13710)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "I don't think there is any necessity for ethics reviews for this paper." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The biggest concern should be the motivation of the two designs, and the current experiments cannot support to prove the effectiveness of them. I think the authors should well figure out these points before claiming the necessity of introducing this method." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The overall design is technically sound, which can be easily implemented. \n2. This paper offers a clear background on why we need vision token compression in vision-language models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces VisToG, a vision token compression method for efficient inference in large vision-language models. The basic idea is to group the similar vision tokens with the token similarity computation, and leverage the vision encoder to initialize the group tokens. The training configurations adopt the LLaVA-v1.5 style. The experiments cover the comparison with previous methods on vision token compression and some ablation studies, where VisToG achieves promising lower inference cost." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The overall pipeline is highly similar with Q-Former based vision token compression methods. After going though the paper, I feel the only differences are two particular designs: \n- The first one is encouraging the vision encoder to initialize the query tokens (i.e., the group tokens in this paper) by learning some tokens to abstract the vision information in patch tokens of the vision encoder; \n- The second one is using Gumbel-SoftMax based operation to calculate the Q-K similarity, resulting in the one-hot selection on vision tokens rather than the weighted sum of vision tokens. Also, it only requires a single grouping layer. \n\nHowever, the motivation of the above two designs remains unclear. I cannot find any ablation results in the paper to validate the effectiveness of these designs. Especially, I am wondering the model performance if using common cross attention rather than Gumbel-SoftMax based vision token selection. To be honest, it is really hard to comprehend the motivation of the second design. \n\n2. As shown in Table 1, the proposed method does not appear fully comparable to the vanilla LLaVA-v1.5 baseline, especially with a notable drop of approximately 5 points on TextVQA. In my experiences, a 5% decline in multi-modal benchmark scores often signals a disproportionately larger impact on a model's multi-modal capabilities. Therefore, it is unconvincing to claim that the model ‘retains xx% performance’ just based on the percentage drop in benchmark scores. \n\n3. And, It is still hard to judge the superiority of VisToG compared with Q-Former. Maybe the authors could conduct a fair comparison on the LLaVA-v1.5 backbone, with the same data and training settings. That will be much better. \n\n4. Please show the full results of table 2 rather than the average results. \n\n5. Not good presentation with some typos like \"8×NVIDIA100-40G\" in Line 314." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Have you tried using a visual token compression method similar to QwenVL on LLava1.5?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The method has a certain level of innovation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for compressing visual tokens in the multimodal large model by adding additional learnable tokens in the vision transformer. It designs a attention mask of the vision transformer to keep original output of vision transformer and compresses visual tokens into learnable tokens using a method similar to cross attention. This approach compresses the size of the visual tokens while maintaining a certain level of model performance reduction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "More comparative experiments on visual feature compression under the same experimental settings are needed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024efficient,\ntitle={Efficient Multi-modal Large Language Models via Visual Token Grouping},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ym1dS37mZE},\nnote={under review}\n}" }, "abstract": { "value": "The development of Multi-modal Large Language Models (MLLMs) has significantly advanced various downstream applications, including visual question answering and image captioning. However, the substantial computational costs associated with processing high-resolution images and videos pose a barrier to their broader adoption. To address this challenge, compressing vision tokens in MLLMs has emerged as a promising approach to reduce inference costs. In this paper, we introduce \\methodname, a novel grouping mechanism that leverages the capabilities of pretrained vision encoders to group similar image segments without the need for segmentation masks. With the isolated attention we adopt, \\methodname can identify and eliminate redundant visual tokens, which effectively reduces computational demands. Extensive experiments demonstrate that the effectiveness of\\methodname , maintains over 98.1% of the original performance while achieving a reduction of over 27% in TFLOPS." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large Language Model", "Multi-modal 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/a50cb35cb716f9a2b69fcb151cedb690e35e2d41.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": "Efficient Multi-modal Large Language Models via Visual Token Grouping" }, "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": "The authors are encouraged to provide more details concerning the above-mentioned weaknesses. For example,\n1. What is the significance of the PCQM4Mv2 dataset in the pretraining of the NoiseVAE model?\n - How does the dataset help in learning atom-specific noise distributions?\n1. How are the noise distributions generated by DenoiseVAE used in the downstream molecular property prediction tasks?\n - What is the architecture of the property predictor(s) used in the downstream tasks? What type of model is used?\n - How are the noise distributions used in the property prediction tasks, in addition to the input 3D molecular geometry?\n1. What are the correct metrics/criteria for the results in Tables 1 and 2?\n\nAdditionally, the authors may consider addressing the following questions:\n1. The paper provides proof that the DenoiseVAE provides a higher theoretical evidence lower bound (ELBO) guarantee for the real conformation distribution of isoenergetic molecules. Is there a more direct way to show the correctness of the noise distributions generated by DenoiseVAE with the ground truth by first-principles calculations?\n - For example, in Figure 4, the authors showed the generated noise distributions for two molecules. How close are these distributions when benchmarked against first-principles simulations?\n1. How does the DenoiseVAE model help with conformational sampling in quantum chemical simulations?\n - The DenoiseVAE model generates noise distributions that are specific to a 3D molecular geometry (to the best of my understanding). By sampling from these noise distributions, are the sampled conformations more likely to be energetically favorable/stable than randomly sampling from a preset Gaussian distribution?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "### Originality\n*My expertise in related literature is limited.*\n1. The NoiseVAE model leverages a VAE for a more robust atom-specific and physics-related generation of noise distribution.\n\n### Quality\n1. The experiments/results are very comprehensive.\n1. The shown results on the downstream tasks suggest the robustness of the generated noise in providing quantum chemical information for accurately predicting molecular properties.\n\n### Clarity\n1. The pretraining objective and method for NoiseVAE are well introduced with details.\n\n### Significance\n1. Exploring the conformational space efficiently is of significance in quantum chemical simulations. NoiseVAE provides a way to efficiently sample conformations around a given geometry." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces NoiseVAE, a novel molecular pre-training method, for learning atom-specific noise distributions in different molecules. The variational autoencoder (VAE) based model consists of a Noise Generator and a Denoising Module that are jointly trained. The pretraining objective is to learn atom-specific noise distributions that align better with quantum mechanical theories than empirically derived strategies. The results show that the pretraining DenoiseVAE can generate noises specific to atomistic environments and improve the prediction accuracy on molecular properties such as the HOMO-LUMO gap and potential energy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I appreciate the amount of information and results presented in the paper, I find the experiments and results a bit difficult to follow.\n\n- The authors showed proof that the proposed DenoiseVAE provides a higher theoretical evidence lower bound (ELBO) guarantee for the real conformation distribution of isoenergetic molecules.\n\n- Subsequently, to show the effectiveness of the DenoiseVAE, the authors conducted downstream tasks with the pre-trained DenoiseVAE for predicting molecular properties.\n\n- However, the authors did not provide a clear explanation of how DenoiseVAE was used in the downstream molecular property prediction tasks. To the best of my understanding, the DenoiseVAE was used to generate noise distributions for the input molecules. However, how were these noise distributions leveraged was not clear to me.\n\n- The tables with results are also a bit misleading. For example, tables 1 and 2 mentioned \"force prediction\", but the 12 properties in Table 1 were not force-related and the values in Table 2 seemed more energy-related than force-related.\n\n- In addition, the paper lacks details on the training process of the NoiseVAE model. Specifically, the pretraining is performed with the PCQM4Mv2 dataset, but the paper only briefly mentions that the dataset has 3.4 million organic molecules." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Major:\nIt’s a bit unclear how the downstream predictive tasks are performed after pretraining. Specifically:\n\n(1) Is the force prediction achieved by a task-specific prediction head? If so, what is its architecture and how is it trained (e.g. end-to-end fine tuning)?\n\n(2) What are the features used for the prediction (e.g. pooled EGNN outputs or latent embedding)?\n\nMinor:\n(1) Following the previous questions, as the model is considered a VAE, how is the latent space obtained?\n\n(2) Is it possible to visualize the learned energy landscape (e.g. for MD17 samples)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. This work addresses an important problem in molecular representation learning. The learnable and molecule-specific per-atom noise distribution is a reasonable setup and aligns better with the physical intuitions.\n2. The authors provide comprehensive theoretical analysis of the rationale of their method.\n3. The proposed method shows consistent improvement over the baselines in the majority of the tasks, while also being more parameter efficient.\n4. The method has good interpretability and the learned noise patterns are physically relevant." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The This work proposes a 3D molecule pretraining method with learnable atom-specific noise generation, as opposed to the common practice of hand-crafted and shared noise. The pre-trained representation shows higher performance than baselines in most of the downstream energy prediction tasks. Furthermore, the authors show the learned noise intensities correspond with the atomic reactivity and rigidity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The results are overall solid. However, some additional experimental and implementation details about the downstream tasks should be provided to help better understand and assess the results. See Questions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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 can different formulations of noise distributions affect the result? Currently, the noises are Gaussians with diagonal covariance, what if the covariance is non-zero at off-diagonal positions? E.g., the neighboring atoms connected by a bond will have non-zero covariance.\n\n2. How well will the model perform if the training and testing are on molecules of different sizes? Say the model is trained on small molecules but is tested for large molecular structures such as proteins." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The use of adaptive noise generation is novel. Previous works use hand-crafted or uniform noise across molecules, but DenoiseVAE uses an automated, atom-specific noise generation method with stochastic parameterization, which respects the unique structural and chemical properties of each molecule.\n\n2. The paper proposes theoretical derivations based on PES which aligns with quantum chemistry principles. The derivation for Theorem 1 based on ELBO offers theoretical insights into the effectiveness of the proposed method.\n\n3. Experimental results are promising and convincing: the proposed method achieved state-of-the-art performance across all benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents DenoiseVAE for 3D molecular pre-training that adapts to the anisotropic nature of molecular force fields. The authors proposed Noise Generator, which learns molecule-specific noise distributions for each atom, allowing DenoiseVAE to generate conformations that align closely with a molecule's potential energy surface. The network architecture is a variational autoencoder, where noisy molecular conformations are denoised using a denoising module. The authors claim that this method leads to more accurate representations for downstream molecular tasks. Experiments on molecular property prediction tasks including QM9 and MD17 are conducted to demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The use of DenoiseVAE introduces additional computational complexity. The training of the VAE with molecule-specific noise sampling could be computationally expensive or impossible for large datasets or complicated molecules. A more thorough analysis of training time and resource requirements could be more helpful.\n\n2. It is unclear to me how the proposed DenoiseVAE can be adapted across different dataset scales. It would be great if the authors could offer a series of studies on the performance of DenoiseVAE on different sizes of datasets.\n\n3. I can imagine the training of DenoiseVAE can be very unstable and will be sensitive to initialization and optimization settings. It would be great if the authors could show convergence of loss curves across all datasets." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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>\n\n - The prior distribution constrains the atom coordinate distribution to follow a predefined Gaussian form. But what is the range of these coordinate distributions? Are they confined within a narrow range close to the prior distribution, or are they spread across a finite range? If so, what is that range?\n\n - The article addresses 3D coordinate noise by adding Gaussian noise based on a learnable distribution, but what about rotations?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "<Strengths>\n\n - DenoiseVAE employs a learnable noise generation strategy that allows for adaptive, atom-specific noise distributions across different molecules.\n\n - The authors introduce a variational approach to jointly optimize the Denoising Module and Noise Generator of DenoiseVAE. Here, the denoising objective encourages generated noise to adhere more closely to physical constraints.\n\n - A KL divergence-based regularization term is applied to prevent low noise intensity and increase the diversity of sampled conformations.\n\n - Optimizing pre-training objective is proven to be equivalent to maximizing the evidence lower bound (ELBO) of the log-likelihood.\n\n - DenoiseVAE outperforms existing denoising methods across various datasets for both molecular and complex property predictions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "<Summary>\n\n - The authors propose a denoising-based 3D molecular pre-training method, called DenoiseVAE, which employs a learnable noise generation strategy instead of existing hand-crafted strategies. This allows for adaptive, atom-specific noise distributions across different molecules.\n\n - The authors introduce a variational approach to jointly optimize the Denoising Module and Noise Generator of DenoiseVAE. Here, the denoising objective encourages generated noise to adhere more closely to physical constraints, enabling the recovery of equilibrium conformations. Additionally, a KL divergence-based regularization term is applied to prevent low noise intensity and increase the diversity of sampled conformations.\n\n - Theoretically, the authors demonstrate that optimizing their pre-training objective is equivalent to maximizing the evidence lower bound (ELBO) of the log-likelihood.\n\n - Extensive experiments reveal that DenoiseVAE outperforms existing denoising methods across various datasets for both molecular and complex property predictions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "<Limitations>\n\n - The Boltzmann distribution is commonly used in the classical regime. However, for precise coordinate computation of interacting atoms, a quantum approach is necessary. This involves, for example, treating the energy function as a quantized operator and finding energy eigenvalues for a given basis. However, this article does not carefully consider such approaches.\n\n - This approach essentially adopts a force field method, which may be inadequate for precise atomic-level computations.\n\n - Additionally, the energy is modeled as a simple harmonic potential, lacking any exchange correlation or other accurate potential forms." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a novel denoising method for 3D molecular pre-training." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024denoisevae,\ntitle={Denoise{VAE}: Learning Molecule-Adaptive Noise Distributions for Denoising-based 3D Molecular Pre-training},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ym7pr83XQr},\nnote={under review}\n}" }, "abstract": { "value": "Denoising learning of 3D molecules learns molecular representations by imposing noises into the equilibrium conformation and predicting the added noises to recover the equilibrium conformation, which essentially captures the information of molecular force fields. Due to the specificity of Potential Energy Surfaces, the probabilities of physically reasonable noises for each atom in different molecules are different. However, existing methods apply the shared heuristic hand-crafted noise sampling strategy to all molecules, resulting in inaccurate force field learning. In this paper, we propose a novel 3D molecular pre-training method, namely DenoiseVAE, which employs a Noise Generator to acquire atom-specific noise distributions for different molecules. It utilizes the stochastic reparameterization technique to sample noisy conformations from the generated distributions, which are inputted into a Denoising Module for denoising. The Noise Generator and the Denoising Module are jointly learned in a manner conforming with the paradigm of Variational Auto Encoder. Consequently, the sampled noisy conformations can be more diverse, adaptive, and informative, and thus DenoiseVAE can learn representations that better reveal the molecular force fields. Extensive experiments show that DenoiseVAE outperforms the current state-of-the-art methods on various molecular property prediction tasks, demonstrating the effectiveness of it." }, "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": [ "3D Molecular pre-training via denoising", "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/f25583612c43755ed41b08af71bc6064f585b2d5.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": "DenoiseVAE: Learning Molecule-Adaptive Noise Distributions for Denoising-based 3D Molecular Pre-training" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please check my previous comments." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The topic of backdoor attacks and defenses is important.\n- The writing in the paper is easy to follow.\n- The experimental setups are comprehensive, and the results look good with appropriate replications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of backdoor attacks and defenses. The authors begin by listing three requirements that a successful backdoor attack should satisfy. Next, they point out the 'high binding' effects and propose a new attack algorithm named Key-Locks to generate backdoor data. Experiments are conducted to evaluate the effectiveness of the proposed algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have two major concerns:\n\n1. First, I suspect that there are already previous theoretical results [1, 2] that can explain your proposed \"high binding\" phenomena and why your method works well. Considering that the literature on backdoors in computer vision is well developed, I think all the points listed in your paper have been somehow mentioned (although in different forms) in previous work; however, there seems to be a lack of discussion on this.\n\n2. Second, more defenses are needed. I think your proposed method can potentially be defended against detection-based algorithms (originally designed for OOD detection). I would like to see how your proposed methods perform under the two more recent strong defenses [4, 5].\n\nRefs:\n[1] Manoj et al., \"Excess Capacity and Backdoor Poisoning\"\n[2] Wang et al., \"Demystifying Poisoning Backdoor Attacks from a Statistical Perspective\"\n[3] Guo et al., \"SCALE-UP: An Efficient Black-box Input-level Backdoor Detection via Analyzing Scaled Prediction Consistency\"\n[4] Xian et al., \"A Unified Detection Framework for Inference-Stage Backdoor Defenses\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "This paper introduces an interesting backdoor attack mechanism using keys and locks. The authors highlight that the vulnerability of existing backdoors stems from their strong binding to model parameters, making them susceptible to defenses based on normal model updates. To enhance robustness, the paper proposes a new backdoor attack. The core technique is similar to adversarial attacks, where a keyhole is generated and embedded using gradient descent towards the target class. During inference, a trigger (the key) is generated using a similar algorithm. The paper includes extensive evaluations, comprising various analyses and ablation studies, to demonstrate the effectiveness of the proposed attack, which outperforms existing methods.\n\nHowever, I have several concerns regarding the design and evaluation:\n\n(1) The attack appears to focus on reducing the class distance between the target class and other classes, as indicated by the poisoning loss in Equations (3) and (4). Without a fixed backdoor generation function, the model learns perturbations from any class towards the target class in each iteration. Once convergence is achieved, generating an effective trigger using a similar function, as described in Equations (5) and (6), becomes straightforward. While this is intriguing, it raises certain issues.\n\nFirst, the necessity of the poisoning process is questionable. Why not directly generate natural backdoor sample, e.g., [1]? This would simplify the process while maintaining a reasonable ASR.\nSecond, the corrupted target class might be easily detectable using existing trigger inversion methods, such as [2] and [3], since generating adversarial perturbations to flip the label to the target class becomes easier than for other classes.\n\n(2) The paper lacks sufficient support to validate its hypotheses, particularly those in Lines 211-215. Recent work [4] models backdoor learning as a special case of continual learning, identifying the orthogonal nature between backdoor behaviors and normal classification. There appears to be a connection between this claim and the authors' hypothesis, which should be explored and discussed.\n\n(3) The baseline attacks and defenses used in the paper are not actually up-to-date. For instance, state-of-the-art attacks such as [5,6,7,8] demonstrate robustness against various defense methods. It is important to clarify how the proposed attack differs from these approaches. Additionally, FST [9], a latest defense effective against a wide range of attacks, should be included in the comparison.\n\n(4) The triggers illustrated in Figures 19-30 appear misaligned with the original images, even if the L-2 distance metrics are favorable. These measurements may not align with human perception. Moreover, adversarial perturbations might be challenging to implement practically, requiring careful tuning of pixel values. They may also be vulnerable to input purification techniques, such as those used in [10], where diffusion models are leveraged to reconstruct inputs and reduce trigger effectiveness. It is suggested to provide a discussion about this issue.\n\n(5) The paper's structure could be improved. Several important details are relegated to the appendix without proper referencing in the main text. For example, Line 237 mentions Algorithm 1 without a clear link and description to it. Properly integrating and referencing content from the appendix would enhance the paper's clarity and flow.\n\n--------------------------\nReference:\n\n[1] Tao, Guanhong, et al. \"Backdoor vulnerabilities in normally trained deep learning models.\" arXiv preprint 2022.\n\n[2] Wang, Bolun, et al. \"Neural cleanse: Identifying and mitigating backdoor attacks in neural networks.\" IEEE S&P 2019.\n\n[3] Wang, Zhenting, et al. \"Unicorn: A unified backdoor trigger inversion framework.\" ICLR 2023.\n\n[4] Zhang, Kaiyuan, et al. \"Exploring the Orthogonality and Linearity of Backdoor Attacks.\" IEEE S&P 2024.\n\n[5] Zeng, Yi, et al. \"Narcissus: A practical clean-label backdoor attack with limited information.\" CCS 2023.\n\n[6] Qi, Xiangyu, et al. \"Revisiting the assumption of latent separability for backdoor defenses.\" ICLR 2023.\n\n[7] Cheng, Siyuan, et al. \"Lotus: Evasive and resilient backdoor attacks through sub-partitioning.\" CVPR 2024.\n\n[8] Huynh, Tran, et al. \"COMBAT: Alternated Training for Effective Clean-Label Backdoor Attacks.\" AAAI 2024.\n\n[9] Min, Rui, et al. \"Towards stable backdoor purification through feature shift tuning.\" NeurIPS 2024.\n\n[10] Shi, Yucheng, et al. \"Black-box backdoor defense via zero-shot image purification.\" NeurIPS 2023." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Proposes a novel and effective backdoor attack strategy.\n\n2. Provides in-depth analysis of the backdoor properties contributing to vulnerability against defenses.\n\n3. Includes extensive evaluations across various models, datasets, and ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel backdoor attack mechanism based on keys and locks. The primary concept is to reduce the high binding of backdoors to model parameters, thereby enhancing robustness against defenses. Extensive experiments demonstrate the effectiveness of this approach, showing its superiority over existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Insufficient support for design choices and key hypotheses.\n\n2. Potential vulnerability to detection by existing trigger inversion methods.\n\n3. Lack of latest baseline attacks and defenses in the evaluation.\n\n4. Concerns regarding the practicality of the trigger pattern.\n\n5. Some 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": 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. Why was clipping excluded during training? Has the potential impact of significant perturbations during the embedding locks of K&L been evaluated against in-training defenses such as Anti-Backdoor Learning?\n2. Could you clarify how K&L differs from adversarial training?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ Analysis of the concept of high binding between backdoor triggers and model parameters.\n+ Extensive experiments using various datasets and neural network architectures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the Key-Locks (K&L) backdoor attack algorithm, designed to bypass existing defenses. Unlike traditional backdoor attacks that show high binding with model parameters, making them vulnerable to defenses, the K&L algorithm decouples the backdoor process from model parameters. This decoupling allows the method to adjust the unlocking levels of backdoor activation, enhancing its robustness against a diverse set of defenses. The paper also introduces the Accuracy-ASR Curve (AAC) as a new metric for evaluating backdoor attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- First of all, the paper requires improvements in editorial quality. There are instances where terms more characteristic of language models are used, such as \"assaults\" instead of the more appropriate \"attacks.\" A thorough human proofreading is recommended to ensure precise usage of terminologies and enhance overall clarity. In addition, the paper excessively relies on the appendix, making it difficult to follow without constant cross-referencing.\n\n- In Algorithm 1 for embedding locks, the backdoor samples are iteratively modified at each training epoch, where the learning rate $\\eta$ is added or subtracted to the backdoor inputs based on the gradient's sign in the direction of the backdoor target. Unlike the inference stage, clipping is not applied during training. This lack of clipping over multiple iterations could result in less stealthy and significantly perturbed images. Such perturbations may potentially lead to distinguishable loss patterns compared to benign inputs. Such characteristics could be susceptible to in-training defenses, such as Anti-Backdoor Learning, which, notably, has not been considered in the evaluation.\n\n- The paper attempts to distinguish K&L from adversarial attacks in Section 3.4. The K&L algorithm generates adversarial perturbations similar to the FGSM method using gradient signs and incorporates these examples during training. The claim is that this approach reduces the binding between the backdoor trigger and model parameters. However, as K&L uses adversarial perturbations as backdoor triggers, this approach shares similarities with adversarial training, which enhances the robustness of models against adversarial attacks. A more precise explanation would help to understand why this method reduces robustness to adversarial perturbation instead of improving it.\n\n- The paper does not fully discuss the rationale behind introducing the new metric, the Accuracy-ASR Curve (AAC), alongside existing backdoor evaluation metrics. It would be beneficial to elaborate on AAC's added value to backdoor attack evaluations. Additionally, the meanings of AAC1, AAC3, and AAC5 are not clear without a proper definition of AAC, making it difficult to understand their relevance in the evaluation. Also, the definition of AAV is not sufficiently explained.\n\n- The backdoor sample similarity rates presented in Table 2 could be potentially misleading, given that they are based on only two input examples, as illustrated in Figure 2. To ensure a robust and reliable evaluation, it would be more appropriate to compute the similarity attribution using multiple images across multiple runs to achieve statistical significance.\n\n- All the results presented in the paper must be evaluated across multiple runs to ensure the statistical significance of the reported values." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 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": "NA" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Large set of experiments: The authors re-implement eight defense methods and six attack backdoors across four datasets. Through these experiments, they demonstrate that the proposed approach can outperform the six existing attacks in bypassing defenses." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an adaptive attack designed to bypass existing backdoor defenses. However, the paper is underprepared, and several concepts are not well explained." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tPresentation needs improvement: Key concepts such as \"high binding,\" \"key,\" and \"lock\" are not explained clearly. For instance, what does \"high binding\" represent? On line 75, the authors state that it refers to the tight coupling between a backdoor and a specific trigger. However, on line 181, it appears to mean something different, namely, the binding between the backdoor and model parameters.\n2.\tUnclear limitations of existing backdoor attacks: It is not clear why current defenses are successful against these attacks. Since \"high binding\" is ambiguous, it is difficult to understand why existing attacks fail. Additionally, according to Table 1, some existing attacks can also bypass these defenses. The authors should clarify why attacks sometimes succeed and other times fail in bypassing defenses.\n3.\tDistinction from adversarial examples: In the KL backdoor attack, the poisoned test sample is generated via gradient descent, similar to adversarial examples. The authors should explain how their approach differs from adversarial examples. Moreover, I suggest that the authors conduct further experiments to distinguish the effects of backdoor attacks from those of adversarial examples. For example, when the poisoning ratio is zero, they could apply the KL attack during testing to check the ASR (Attack Success Rate). If the ASR is not too low, this would suggest that adversarial examples, rather than backdoor attacks (which rely on poisoning training data), are the primary factor leading to successful backdoor injection.\n4.\tSummary of backdoor attack requirements: Summarizing the three requirements for a backdoor attack should not be considered a primary contribution, as many other papers have already discussed this." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024kl,\ntitle={K\\&L: Penetrating Backdoor Defense with Key and Locks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ymqLAmqYHW},\nnote={under review}\n}" }, "abstract": { "value": "Backdoor attacks in machine learning create hidden vulnerability by manipulating the model behaviour with specific triggers. Such attacks often remain unnoticed as the model operates as expected for normal input. Thus, it is imperative to understand the intricate mechanism of backdoor attacks. To address this challenge, in this work, we introduce three key requirements that a backdoor attack must meet. Moreover, we note that current backdoor attack algorithms, whether employing fixed or input-dependent triggers, exhibit a high binding with model parameters, rendering them easier to defend against. To tackle this issue, we propose the Key-Locks algorithm, which separates the backdoor attack process into embedding locks and employing a key for unlocking. This method enables the adjustment of unlocking levels to counteract diverse defense mechanisms. Extensive experiments are conducted to evaluate the effective of our proposed algorithm. Our code is available at: https://anonymous.4open.science/r/KeyLocks-FD85" }, "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": [ "backdoor attack", "backdoor defense", "AI security" ] }, "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/1fbd959d796ccf58481f79b45530f9b428f82437.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": "K&L: Penetrating Backdoor Defense with Key and Locks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see Weakesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The problem of automatic benchmark generation in a guided manner is an important one. While LMs have been used as judges to automatically evaluate other LM's answers, this work proposes using LMs to also generate questions.\n\n2. The problem is formalized and packaged in an elegant and extensible way in the AutoBencher framework, and two important instances of the framework are studied. The two-step division (first generate topics and then generate datasets per topic) is especially novel and effective.\n\n3. The proposed approach has two primary novel aspects: the use of a tool (such as a Wikipedia knowledge database or a Python library for mathematical calculations) to avoid the evaluating LM from generating answers that could be incorrect, and an algorithm to generate the benchmark in a guided rather than brute-force manner.\n\n4. The empirical evaluation shows the promise of the proposed approach in generating datasets that are more novel and difficult than even hand-crafted ones like MMLU." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents AutoBencher, a declarative framework for automatic benchmark construction, and use it to scalably discover novel insights and vulnerabilities of existing language models. Specifically, it is instantiated in two scenarios: 1) capability evaluation and 2) safety evaluation. In each scenario, a set of desirable characteristics is first formally defined. For capability evaluation, these are salience, difficulty, separability, and novelty. For safety evaluation, they are harmfulness and attack success rate. Then, a language model is used to automatically construct descriptions of topics along with datasets in those topics, where a dataset is a set of (question, answer) pairs. Empirical results using GPT-4 as the evaluator show that the created datasets are on average 27% more novel and 22% more difficult than existing benchmarks. AutoBencher also helps identify specific gaps not captured by existing benchmarks: e.g., Gemini-Pro has knowledge gaps on Permian Extinction and Fordism while GPT-4o fails to decline harmful requests about cryptocurrency scams." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. AutoBencher is currently stand-alone. The paper would be stronger if it integrated AutoBencher into existing popular evaluation frameworks like Stanford's HELM or HuggingFace's Open LLM. Adoption of AutoBencher by one of these frameworks would make a more convincing case for its usefulness and viability.\n\n2. There seems to be a mismatch between the capabilities of the evaluating LM and the evaluated LMs: the former has access to tools whereas the latter does not. The paper does not make a convincing case why future LMs should be expected to have such capabilities without using tools themselves.\n\n3. The evaluation is rather limited. I would expect a benchmark/evaluation focussed paper to be more comprehensive and derive more insights than those currently presented. For instance, AutoBencher currently only generates one-turn (question, answer) pairs; it would be interesting to see it extend to multi-turn data, chain-of-thought data, etc. Another direction to extend it could be in the domain of multi-modality.\n\n4. I found the safety evaluation less convincing than the capability evaluation. While the paper does use recent baselines such as XSTest and HarmBench, it would be more convincing if the paper would report on how AutoBencher could be integrated into a mainstream framework for safety evaluation and discuss challenges that were overcome in such an integration (this is related to item 1 above but is more specific to safety evaluation)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- In Section 3.1 on separability, are you taking the mean over $acc(\\cdot) \\in V_c$? If so, this should be clarified. Additionally, I am not convinced that high deviation alone ensures that rankings between any two models, $LM_i$ and $LM_j$, with $i \\ne j$, are robust to noise.\n- Using vector notation could help make the equations clearer.\n- In Section 3.1 on novelty, it seems that there is a coefficient for each model being evaluated. Could you elaborate on why there are coefficients for each model rather than for each dataset?\n- How are the salience and harmfulness binary variables obtained in practice?\n- I am unclear about the procedure for qualitative assessment in Section 6.3. What does \"high quality\" mean, and how many question-answer pairs were sampled? What do the \"ranks\" in this section refer to?\n- What language model was used to generate Table 1? Adding this information to the caption would improve clarity.\n- Are there any ablation experiments on the following: the language model used, the criteria for evaluating dataset descriptions, and the impact of including privileged information?\n- Are the human benchmarks also grouped into distinctive categories and descriptions similar to those proposed by the language model?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper conducts experiments across a wide range of knowledge domains.\n- The presentation is generally clear, with well-structured sections and a logical flow.\n- Equations and figures are used effectively to enhance clarity.\n- The automatic generation of benchmarks is a topic of strong interest to the community.\n- The proposed method demonstrates significant performance gains over human-generated benchmarks, particularly in terms of novelty, difficulty, and separability metrics as defined in the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a method, Autobencher, to automatically construct benchmarks for evaluating language model (LM) capabilities and safety. Specifically, based on predefined criteria, the authors prompt an LM to propose dataset descriptions and generate questions, which a candidate LM then attempts to answer. The dataset is subsequently scored according to these criteria, enabling the selection of high-quality dataset descriptions. The authors aim to demonstrate that Autobencher generates dataset topics/descriptions that are both more novel and more challenging than existing benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It is not entirely clear to me what methodological novelties the paper introduces in dataset construction, aside from its use of retrieval during generation. Optimizing for various criteria objectives is not a particularly significant contribution.\n- The introduction and abstract could benefit from revisions to be more specific and descriptive.\n- The method and criteria rely on accuracy scores computed across both existing models and preexisting datasets, which can be computationally intensive, especially if generating multiple datasets across a large set of models.\n- Additionally, language models (LMs) can exhibit high variance across different runs. A robust scoring criterion should account for this issue by incorporating variance between different runs, both for generating questions and answers. Introducing a statistical method to handle this variance would strengthen the contribution.\n- Ensuring that the dataset description is clear and unharmful does not guarantee that all generated questions are equally clear and unharmful. For instance, a dataset labeled as \"Nuclear technology\" might contain questions of varying levels of harmfulness.\n- There are concerns regarding the practicality of the method. While the paper claims high automation, many of the criteria still seem to require potentially time-intensive manual crafting. Furthermore, LM generations can be compute-intensive; a discussion on the computational resources required for these generations would be beneficial.\n- Additional ablation studies could help demonstrate the impact of various components of the method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "There are ethical issues with the topic under study as it enables attack vectors on LLMs; however, the authors take due diligence to properly discuss their study process and acknowledge the broader impact." }, "flag_for_ethics_review": { "value": [ "No 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. To the knowledge of this reviewer, it should be possible to have a hyper-/hyponymy graph of Wikipedia topics. Did you consider using such a graph, or a similar topic relationship graph to further guide the adaptive search?\n2. The optimisation problem for case (1) is multi-objective, is there a reason why multi-objective optimisation was not employed and instead the loss is linearised by adding hyperparameters? (evolutionary/population-based methods?)\n3. Related to W1, how is the math/science salience set constructed? Unless I missed it in the paper, the inclusion-exclusion criteria were not clear or indeed the exact methodology. Did you perform, for example, open card sorting or a similar methodology?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "+ On novelty, the approach generates human interpretable topics where model ranks exhibit surprising results.\n+ On safety, the qualitative examples align with the experience of this reviewer when trying to manually jail-break LLMs: pose the question as a hypothetical or philosophical debate; this vector being auto-discovered is encouraging.\n+ Well-executed research methodology with manual validation of the discovered benchmarks" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose to use an LM-guided adaptive search to construct benchmarks that optimise (1) difficulty (benchmark headroom), separability (existing model variance), and novelty (rank correlation with other benchmarks) and (2) attack success rate (when trying to extract harmful information from models). They demonstrate the need for this by showing that existing benchmarks exhibit low values in novelty (especially) even for benchmarks such as MMLU where there is sufficient headroom. Further, they use privileged information to enable a weaker LM to assess a stronger LM, use translation tools to enable multi-lingual data and employ source code/python to evaluate numerical and symbolic expressions. When optimising, the authors employ an adaptive search strategy that uses the history of explored subjects to guide the new candidates restricted to a salience set. To assess the quality of the generated data, the authors employ Mechanical Turk and find sufficiently low error rates with high salience on the questions. The value of adaptive search is demonstrated via an ablation on the use of the history vector. Both regimes demonstrate improvement over HumanEval on the desired metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The approach to the math and science categories suggests an open vocabulary problem that is not clearly tackled. For other categories, this is tackled via Wikipedia and a popularity metric.\n- The translation induces an issue for low-resource language: such languages are both less likely to be tackled by LLMs and by translation tools, creating a catch-22. (I feel it would be sufficient to acknowledge the issue as a limitation since tackling the issue requires significant manual effort, future work can consider specialised models for specific language pairs.)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "Yes, Discrimination / bias / fairness concerns" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1- How does AutoBencher differ from previous work on automatic benchmarking, such as that in [2]?\n\n2- Why is a comparison with the baseline included?\n\n\n\n\n\n[1] Taskbench: Benchmarking large language models for task automation." }, "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 very easy to understand.\n\n2- The paper provides an in-depth empirical evaluation, uncovering vulnerabilities and assessing model performance. It also includes results from human evaluations to further validate the quality and relevance of the generated benchmark.\n\n3- AutoBencher automatically constructs datasets, making it highly scalable and reducing reliance on costly human labor. This is useful for evaluating LLMs across various domains." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces AutoBencher, a declarative framework that automatically generates benchmarks to evaluate language model performance by defining specific desiderata, such as difficulty, novelty, and safety. The proposed approach, AutoBencher, leverages language models to iteratively construct benchmarks that reveal weaknesses and safety vulnerabilities in LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1- The details of the qualitative analysis are missing.\n\n2- AutoBencher uses GPT-4 as the evaluator, which may introduce potential bias in datasets that could favor LLMs from the same family (e.g., OpenAI's LLMs)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present AutoBencher, a declarative approach to constructing new datasets, revealing model weakness and safety vulnerabilities." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024autobencher,\ntitle={AutoBencher: Towards Declarative Benchmark Construction},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ymt4crbbXh},\nnote={under review}\n}" }, "abstract": { "value": "We present AutoBencher, a declarative framework for automatic benchmark construction, and use it to scalably discover novel insights and vulnerabilities of existing language models. Concretely, given a few desiderata of benchmarks (e.g., question difficulty, topic salience), we operationalize each desideratum and cast benchmark creation as an optimization problem. Specifically, we experiment with two settings with different optimization objectives: (i) for capability evaluation, we declare the goal of finding a salient, difficult dataset that induces novel performance patterns; (ii) for safety evaluation, we declare the goal of finding a dataset of unsafe prompts that existing LMs fail to decline. To tackle this type of optimization problem, we propose to use a language model to automatically construct datasets and iteratively revise the dataset to optimize for the declared desiderata. We use AutoBencher (powered by GPT-4) to create datasets for math, multilinguality, knowledge, and safety. The scalability of AutoBencher allows it to test fine-grained categories and tail knowledge, creating datasets that are on average 27% more novel and 22% more difficult than existing benchmarks. AutoBencher also helps identify specific gaps not captured by existing benchmarks: e.g., Gemini-Pro has knowledge gaps on Permian Extinction and Fordism while GPT-4o fails to decline harmful requests about cryptocurrency scams." }, "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": [ "automatic evaluation", "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/2b55ac880bfa78148082c3e5dccd7a61a651bcf8.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": "AutoBencher: Towards Declarative Benchmark Construction" }, "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 might the ASR be affected if the safety filter remained active? Including this would provide a clearer picture of BoN’s practical effectiveness and real-world implications. Would the authors consider conducting additional experiments with the safety filter enabled to better align the study with real-world conditions?\n\n- Audio quality after attack modifications is an important factor for stealthy and practical attacks. Could the authors provide details on the perceptual quality of the modified audio samples, perhaps using a metric like the ViSQOL score?\n\n- The term \"ASR\" is commonly used to denote \"Automatic Speech Recognition\" in the speech research community. Could the authors consider using an alternative term?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Proposed Best-of-N (BoN) Jailbreaking is novel, specifically tailored for attacking ALMs. BoN is unique in its application of audio augmentations to create high-entropy inputs, exploiting the ALM's sensitivity to continuous input variations. The combination of BoN with other jailbreak methods, such as the PrePAIR technique, showcases an innovative blend of audio augmentations and iterative refinement for more effective attacks.\n\n- The discovery and application of power-law scaling to predict ASR with increased samples indicate a high-quality analysis, providing valuable insights into the scalability and potential impact of the proposed BoN method.\n\n- The paper is well-structured and effectively explains complex ideas, making the novel BoN method accessible. Visualizations such as graphs that show the ASR progression with sample size and the power-law behavior, enhance clarity by illustrating key results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces \"Best-of-N (BoN) Jailbreaking,\" a black-box algorithm designed to exploit weaknesses in Audio Language Models (ALMs) by extracting harmful information through audio-based attacks. The BoN method uses repeated audio augmentations, such as changes in pitch, speed, and background noise, to malicious prompts until it elicits harmful responses from the target ALM. The study shows that BoN achieves a high attack success rate (ASR), with results exceeding 60% ASR on several top ALMs, including a preview version of GPT-4o’s voice mode. Additionally, the authors discover power laws that allow them to predict ASR based on the number of samples, helping forecast the efficiency of BoN jailbreaking. The approach becomes even more effective when combined with other jailbreak techniques, reaching up to 98% ASR in some models. This paper highlights the difficulty of securing ALMs due to their sensitivity to continuous input variations, proposing BoN as a scalable and effective method for targeting ALM vulnerabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- A significant limitation of this study is the decision to turn off Gemini’s safety filter. In real-world applications, LLMs and ALMs rely on both alignment techniques and safety filters for safeguarding against misuse. By disabling these filters, the study's attack success rate may be artificially inflated, making the findings less practical and relevant in environments where safety filters are essential. Including experiments with the safety filter enabled would provide a more realistic assessment of the BoN method's effectiveness and its relevance to real-world deployments.\n\n- The study does not evaluate the audio quality of the modified samples post-attack, which is an important aspect for assessing the stealthiness of these attacks. A low audio quality in the altered samples could make the attacks easily detectable or unnatural. A quality metric, such as the ViSQOL score, would allow for a quantitative comparison between the original and post-attack audio samples. Without such an analysis, it is unclear if the BoN attacks are feasible in scenarios where high-quality audio is essential for covert operations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "See above" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "This paper successfully introduces a new black-box jailbreaking algorithm called Best-of-N (BoN) that effectively extracts harmful information from Audio Language Models (ALMs) by exploiting their sensitivity to variations in audio input. The audio domain is in general underexplored so I appreciate this effort.\n\nThe paper looks into different aspects such as combinations with other attacks. In addition, the paper provides detailed insights into the workings of BoN jailbreaking, including analysis of the transferability and semantic meaningfulness of the augmentations used. The research highlights the challenges of safeguarding ALMs with stochastic outputs and continuous input spaces." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This research paper explores the vulnerabilities of Audio Language Models (ALMs) to audio-based jailbreaking attacks. The authors introduce a novel algorithm called Best-of-N (BoN) Jailbreaking, which leverages random audio augmentations to elicit harmful responses from ALMs. They demonstrate the effectiveness of BoN jailbreaking on several state-of-the-art ALMs, including Gemini and GPT-4o. The authors also uncover power laws that predict the effectiveness of BoN as a function of the number of sampled augmentations. Finally, they investigate the composition of BoN with other jailbreaking techniques, demonstrating that combining these methods can lead to more potent attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concerns regarding this paper are that the methodology of the attack is not adequately described, and the evaluation begins without sufficiently introducing the approach and settings.\n\nFor example, to understand the experiments, the reader needs to know what is defined as \"harmful information.\"\n\nIt is difficult to assess the novelty of this work. Although the authors propose an attack against an alternative domain, the method used is unclear. Additionally, we do not gain many modality-specific insights and could potentially derive the same findings from other, text-only models. It would be beneficial if the paper included some audio-specific insights.\n\nFor instance, what is the signal-to-noise ratio (SNR) of the input? Would a human notice the attacks? Does the attack also work if it is played over the air?\n\nIn the introduction, the paper describes the \"robustness\" of the models (third paragraph). However, the method used is not described, and the findings there are not particularly useful.\n\nIt would also be appreciated if the authors uploaded a few audio samples for listening." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. See weakness 1. Are there any more insights and discussion?\n\n2. See weakness 2. Do the PrePAIR have any relationship with the audio part in ALMs?\n\n3. See weakness 3." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The research question, studying the jailbreak vulnerabilities of ALMs, is relatively interesting.\n\n2. The experimental part is comprehensive and complete. The appendix is detailed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces BoN, a new jailbreaking technique for ALMs. Based on various audio augmentation methods, BoN bypasses the safeguards of SOTA ALMs through repeated sampling with higher ASR. Authors also show that BoN jailbreaking can be composed with other jailbreak techniques." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As stated in line 157-159, PAIR and TAP are classical jailbreak techniques, and both of them achieve good performance on jailbreaking text-only LLMs. However, when applying single audio augmentation and TTS engine, the ASR on ALMs decrease to below 5%. This result looks counterfactual and lacks of detailed discussion. The authors should provide a more detailed analysis of why this decrease occurs, or discuss the potential reasons when transferred from text to audio.\n\n2. Technical contribution is limited. (1) Lack of detailed description of the proposed algorithm PrePAIR. (2) From the limited description of PrePAIR, while universal, this algorithm is not highly relevant to the previous sections and the “audio” part in language models, which makes the motivation unclear. The authors should clarify the connection between PrePAIR and the audio aspects of language model, and explain how this algorithm relates to the overall motivation of the paper. \n\n3. As a supplementary section of the pre-experiments, Section 3 (especially Section 3.1 & 3.2) does not give clear and valuable insights, which also makes the pre-experiment part lengthy. The authors should highlight the key findings more clearly, which would help this reviewer and other readers understand the value of this section.\n\n4. The presentation of the paper needs to be improved. For example, the duplicated part in upper left figure 1." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce BoN Jailbreaking: a composable, and highly effective black-box algorithm for attacking state-of-the-art ALMs." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024attacking,\ntitle={Attacking Audio Language Models with Best-of-N Jailbreaking},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yougZBoUY3},\nnote={under review}\n}" }, "abstract": { "value": "In this work, we investigate the susceptibility of Audio Language Models (ALMs) to audio-based jailbreaks and introduce Best-of-N (BoN) Jailbreaking, a black-box jailbreaking algorithm to extract harmful information from ALMs. To craft jailbreak inputs, our approach samples audio augmentations and applies them to malicious prompts. We repeat this process until we find a set of augmentations that elicits a harmful response from the target ALM. Empirically, we find that applying BoN with 7000 sampled augmentations achieves an attack success rate (ASR) of over 60% on all models tested, including the preview model for the released GPT-4o. Furthermore, we uncover power laws that accurately predict the ASR of BoN jailbreaking as a function of the number of samples. These power laws allow us to forecast the effectiveness of BoN jailbreaking as a function of the number of sampled augmentations over an order of magnitude. Finally, we show that BoN jailbreaking can be composed with other black-box attack algorithms for even more effective attacks—combining BoN with an optimized prefix attack achieves 98% ASR on Gemini Pro and Flash. Overall, by exploiting stochastic sampling and sensitivity to variations in a high-dimensional input space, we propose a scalable, composable, and highly effective black-box algorithm for attacking state-of-the-art ALMs." }, "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", "jailbreaks", "audio language model", "speech language model", "multimodal", "adversarial attack", "audio jailbreak", "safety", "trustworthy", "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/c5a6020ec0ed18c4f098ef984b84144796cf0d4d.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/840e5a8474267456ae2457ecdb23e3a3b3ddd198.zip" }, "title": { "value": "Attacking Audio Language Models with Best-of-N Jailbreaking" }, "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)\tThe motivation example in Figure 1 is a bit misleading. The authors suggest that the question “Who is the father of the artist who painted Head I?” has three hops of reasoning, whereas it only requires two hops, i.e. finding the artist in 1st hop and then identifying the author in the 2nd hop. This example can be a bit confusing to the reader on what the overall motivation of the approach is.\n\n2)\tIt is very surprising to see almost similar performance for models of highly different capabilities such as GPT-3.5-turbo and GPT-4o. The authors should provide more insight/analysis on why this is happening? Also, what is the performance when the LLM used is an open-source model such Llama-3 8B.\n\n3)\tIt would be interesting if the authors could show performance separately based on the number of reasoning hops present in the question. Also, does the approach show any benefits over RAPTOR for single hop/simple queries?\n\n4)\tThe authors should also evaluate more recent multi-hop QA datasets, such as FanOutQA or FreshQA. All the datasets considered are pre-2022, raising concerns about leakage into LLM pretraining data. \n\n5)\tThe authors should also consider adding a few qualitative analysis examples that demonstrate how and “why” (i.e. which part of the method helps) SiReRAG improved over RAPTOR due to incorporating the relatedness." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1)\tThe approach shows considerable performance improvements over RAPTOR on a variety of multi-hop QA datasets.\n\n2)\tThe experimental ablation settings are thorough and show the benefit of different design choices made by the authors for the SiReRAG approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces SiReRAG, which proposes to consider both similar and relatedness information when creating retrieval indices that address queries with multi-hop reasoning. The paper gives motivations to demonstrate the bottleneck of solely modelling relatedness or similar information only. For similar information, the paper constructs a similarity tree based on recursive summarization, while for relatedness, SiReRAG extracts propositions and entities from text, and groups them via shared entities to construct a relatedness tree. Experimental results show SiReRAG considerably improves over other baselines like HippoRAG, GraphRAG and RAPTOR when evaluated on various multi-hop QA datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1)\tThe paper does read like incremental work over RAPTOR, and it is hard to be convinced that the paper has enough novelty for acceptance at ICLR.\n\n2)\tSome important baselines such as the closed book approach, i.e. directly getting the final answer from the LLM without any retrieval, or iterative retrieval, such as Self-Ask [1] or DSPy [2] are missing. Also, the authors should include these baselines when doing the inference latency comparison.\n\n3)\tThe writing needs to be improved in Sections 1 and 3 of the paper since it’s not easy to grasp the main intuitions or motivations of SiReRAG. \n\n[1] Measuring and Narrowing the Compositionality Gap in Language Models\n\n[2] DSPY: COMPILING DECLARATIVE LANGUAGE MODEL CALLS INTO SELF-IMPROVING PIPELINES" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The reasoning behind why HippoRAG does significantly better for 2Wiki dataset is not very clear in my opinion: I think what might be happening here is the following: when the RAPTOR text clustering pipeline is applied to the extracted entity-relatedness propositions, it treats the full propositions as text, completely ignoring the underlying structure of the proposition/fact (which is usually of the form Subject/Predicate/Object). That destroys certain information from the triple. My guess is that if we modify the RAPTOR encoder to encoder these features separately (e.g. with SPO markers tokens or with different encoders altogether), then we could see that we're able to encode the triple structure better, and we might be able to recover that difference.\n\n- It would be very informative to share more information about the average number of candidates considered by the method compared to the baseline methods per query. This would show the computational increase compared to the baseline." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality: The work presented in SiReRAG is original in the sense it combines ideas from two existing methods/philosophies about indexing, similarity and entity-relatedness into one solution.\n\nQuality: The thought process and reasoning of the paper is mostly intuitive and simple (merge two ideas that have been shown to improve performance), the experimental section supports this reasoning and is somewhat comprehensive with high quality and well-studied datasets. \n\nClarity: The paper is mostly well-written and easy to follow, with certain exceptions are included in the Weaknesses sections. The claims of the authors about coverage of relatedness or similarity only in table 1 serves as a good benchmark and motivation for the work.\n\nSignificance: I consider this method to be an incremental addition on top of RAPTOR, The main metrics reported show significant improvement over the baseline methods that use only one signal." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents SiReRAG: a new indexing and retrieval technique that takes into account both \"semantic similarity\" and \"entity relatedness\" for answering complex multi-hop queries.\n\nThe suggested technique is built mostly on top of another technique called RAPTOR (which indexes text chunks on similarity only), but augments it by: first, extracting entity-level facts or propositions s (which are single facts/statements about entities), using LLMs like chatGPT or LLama-instruct, then applying the RAPTOR pipeline to extracted propositions to create relatedness trees akin to the original RAPTOR similarity trees. The final method merges the output from the two trees (one from original RAPTOR built on text chunks, and another from RAPTOR built on extracted and aggregated propositions).\n\nThe paper includes comprehensive evaluations and shows significant improvement over baseline methods that use only similarity or relatedness features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper is not very clear regarding how to merge the result from the two trees (similarity and relatedness trees), but I might have missed this.\n\n- The metric \"time per retrieval pool size\" which is used to quantify the time of the SiReRAG vs other baselines method seems very contrived and not at all relevant, since it seems to say more about the number of candidates (the denominator) than about the time spent answering a query. Which means that the number of candidates generated by this method can be an order of magnitude more than the baseline methods in some cases, which can be prohibitive in many use-cases.\n\n- The main body of the paper has only two examples, without much detail: I recommend adding a full simple example for how the relatedness tree would look for a simple paragraph, as it's more difficult to visualize the entity-related tree as opposed to the similarity/summary between text chunks.\n\n- The suggested method ends up performing badly on the more structured datasets (2Wiki): this indicates that the way it's encoding the relatedness is not representative enough to capture the structure of the graph/triples/facts." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. It is unclear whether you re-ran the baseline or obtained the results from the original papers. I could not find the exact numbers reported elsewhere in Table 4.\n2. What was the retrieval size of the baselines?\n3. Overall, I found that TPRS was not meaningful. Wasn't it simply a result of whether the documents were short or long?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The paper's main idea is well-motivated and supported by its preliminary experiments (Table 1). I found that its arguments were well-articulated and supported by existing literature, yet the hard evidence provided in Table 1 was a helpful addition.\n2. The paper acknowledged an alternative approach in some steps and some results to reject the alternatives (Sections 4.1 and 4.3).\n3. The method proposed in the paper consistently improved over three multi-hop reasoning datasets (Tables 4 and 6). In addition, ablation analysis also justified the need for the relatedness tree (Table 5).\n4. The paper included an inference time experiment, showing that inference time increased with pool size." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the retrieval component of the retrieval-augmented generation. It argues that existing RAG methods tend to focus exclusively on semantic similarity or relational links, leading to suboptimal performance in complex, multi-hop reasoning tasks. The paper motivates this with a coverage study to show that similarity or relatedness can only return a few correct entity connections, and the return results are half overlapped. The paper proposes SIRERAG, which is designed to optimize information retrieval by indexing data based on similarity and relatedness. The index is based on a similarity and relatedness tree using a a method similar to RAPTOR's.\n\nExperiments with three datasets (MuSiQue, 2WikiMultiHopQA, and HotpotQA) showed improvements in EM and F1 scores compared to previous indexing methods, which include similarity or relatedness. The gain in performance had a slight negative impact on the inference time compared to a similarity-only method due to a larger retrieval pool size. Further analyses showed that the relatedness tree was indeed helpful." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although not related directly to the paper, the paper should still acknowledge the state-of-the-art multi-hop reasoning method such as [Open-RAG](https://openragmoe.github.io/).\n2. Section 4.3, which described an alternative method rather than the main method, did not explain flattened indexing well. Figure 2 showed an \"+\" sign but was not technical enough to confirm what was being done.\n3. A few prompts were not provided, such as the summary prompt (maybe similar to RAPTOR?), the topic extraction prompt (Section 3), and the hierarchy prompt (Section 4.1).\n4. Some missing details of the baseline might be crucial to interpret the results (see Questions)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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 experimental results on other multi-hop reasoning tasks.\n2. The performance using other relatedness-based trees (e.g., entity-based, or entity pair-based)" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "In general, I think it is helpful to leverage multiple types of relevance for better retrieval results, and the experimental results also verify its effectiveness. It is also helpful to investigate and identify that similarity-only is not enough for complex QA tasks (although this is not a new discovery as many previous studies in traditional IR studies)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Indexing is critical for IR or RAG systems. In this paper, the authors propose a RAG indexing method which considers both semantic similarity and relatedness for better retrieval, named SIRERAG. Specifically, for similarity-based indexing this paper constructs a similarity tree via recursive summarization, and for relatedness this paper constructs a relatedness tree via entity and proposition extraction and grouping. Experimental results show some performance improvement on multi-hop QA datasets over previous similarity-based baselines and relatedness-based baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Because the similarity-based indexing component just follows RAPTOR (Sarthi et al., 2024), I think the main contribution of this paper maybe the relatedness-based indexing component. Unfortunately, I found the relatedness-based indexing algorithm is ad hoc. Firstly, the relatedness are modeled using entity-specific propositions, which I think is over-specialized to multi-hop QA tasks(where the hop is just entity-entity associations and this is why the proposed method can achieve performance improvements). However, I think entity-specific propositions may not a good decision for many other complex reasoning tasks, and the authors should explain and verify the effectiveness and generality in more tasks. Secondly, I found there are many heuristic decisions, such as how to filter proportions, how to resolve entity references, etc.\n2. The similarity and relatedness trees are constructed and used independently, which I think is straightforward, it is important to consider the interaction between different relevance scores.\n3. There are many multi-hop QA baselines, such as iterative RAG-based, agent-based, etc. The authors should compare them for more convincing experimental results.\n4. The writings of this paper should be improved." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce an innovative RAG indexing approach that considers both similarity and relatedness when organizing data with strong performance." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024sirerag,\ntitle={SiRe{RAG}: Indexing Similar and Related Information for Multihop Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yp95goUAT1},\nnote={under review}\n}" }, "abstract": { "value": "Indexing is an important step towards strong performance in retrieval-augmented generation (RAG) systems. However, existing methods organize data based on either semantic similarity or related information, but not both. As our analysis reveals, modeling only one perspective leads to suboptimal performance on complex tasks requiring multi-hop reasoning. In this paper, we propose SiReRAG, a novel RAG indexing approach that explicitly considers both similar and related information. On the similarity side, we follow existing work and explore some variances to construct a similarity tree based on recursive summarization. On the relatedness side, SiReRAG extracts propositions and entities from texts, groups propositions via shared entities, and generates recursive summaries to construct a relatedness tree. We index and flatten both similarity and relatedness trees into a unified retrieval pool, demonstrating that SiReRAG consistently outperforms state-of-the-art indexing methods on three multi-hop datasets, with an average 4.6% improvement in F1 scores. SiReRAG also enhances existing embedding and reranking methods, with an average improvement of 7.8% and 4% in F1 scores." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Retrieval-augmented generation (RAG)", "RAG indexing", "Multi-hop question answering" ] }, "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/5ab04091aef7e78e96d6255bd9eacf96bed592a5.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": "SiReRAG: Indexing Similar and Related Information for Multihop 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": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Questions\nMajor questions:\n* What is the distribution used to initialize the weights of the RNNs? How does the initial weight distribution affect the results? In particular, does the DSA distance still decrease with task complexity when initializing the RNNs with small weights close to zero?\n* Why use a Frobenius invariant norm to measure weight degeneracy instead of an orthogonal invariant Frobenius norm as measured in DSA?\n\nAdditional questions:\n* In general, how does the degeneracy evolve during training? Does it increase or decrease with training? This might depend on the weight initialization.\n* It would be interesting to look at other more commonly used similarity measures than DSA such as CKA or Procrustes distance." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The study addresses an important question on diversity and degeneracy in RNNs.\n* The authors investigate this question across multiple levels: behavior, dynamics, and connectivity, and across multiple settings: different loss functions, network sizes, and weight regularization methods.\n* The paper is clearly structured and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Summary\nThe authors study the variability in RNNs across four tasks: a flip-flop task, a delay discrimination task, a sine wave generation task, and a path integration task. For each of these tasks, the authors propose a way to vary the task complexity, as measured by the entropy of the inputs and outputs. They found that the average DSA pairwise distance between 50 RNNs trained to reach a certain performance threshold, decreases as the task complexity increases. On the other hand, the degeneracy of the RNN weights, measured with a permutation invariant Frobenius norm, was found to increase with task complexity. RNNs trained on more complex tasks also showed decreased variability in their outputs when evaluated on the original tasks with longer time durations. Additionally, the authors showed that the weights and dynamics degeneracy varies depending on the task loss, the network size, and the weight regularization method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses\nIt seems that the central claim of the paper is: “while harder tasks lead to more consistent neural dynamics across task-trained networks, their underlying recurrent weight matrices are more degenerate/variable.” (Section 4). However, this statement may be overly simplistic and may significantly depend on factors that were not considered in the proposed work:\n(a) Dependence on the initial weight distribution of the RNNs\n(b) Dependence on the specific distance metrics used to quantify dynamics and weights degeneracy\n(c) Dependence on the specific notion of complexity\n\n(a) Prior work has shown that the initial weight distribution can have a large influence on the training dynamics and the solutions found by the RNNs. However, there is little information on how the RNN weights were initialized and there is no analysis on how the main claims may depend on their initial distribution. In particular, the decrease in dynamics degeneracy with task complexity may not hold for RNNs initialized with small weights with values close to zero.\n\n(b) The authors measured weights degeneracy with a stricter Frobenius norm than the dynamical degeneracy. The norm used for DSA is invariant under orthogonal transformation, which is less strict than the permutation invariant norm used to quantify the weights degeneracy. The authors found that dynamics degeneracy decreases with task complexity but weights degeneracy increases. How much does this result depend on the specific norm and its invariance class used to measure degeneracy? Would the weights degeneracy decrease with task complexity when considering a similar matrix norm than for the dynamics degeneracy, i.e. orthogonal transformation invariant Frobenius norm?\n\n(c) The authors quantify task complexity as the entropy of the task inputs and outputs. However, this captures only a specific aspect of the task complexity. For example, as stated in section 4: “Moreover, as we modified the task characteristics to study its effect on degeneracy, we observed that degeneracy remains invariant under certain transformations of the task, e.g., changing the delay duration in Delayed Discrimination or altering the environment size in Path Integration tasks.”. It would be really interesting to see these results because increasing the duration of the delay period or increasing the size of the environment can be seen as ways to make the tasks more complex, in the sense that it would probably take longer for the RNNs to learn the tasks. The amount of training required to learn the task is another metric that can be used to measure task complexity. It would be interesting to study how it relates to the entropy based task complexity considered by the authors.\n\nMeasuring task complexity as the entropy of the task inputs and outputs implies that it is invariant under shuffling of the timesteps, which completely changes the task structure. Does the relationship between degeneracy and task complexity still hold when considering time shuffled variants of the tasks?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "How does your permutation-independent distance contrast with the distance provided in Generalized Shape Metrics on Neural Representations paper Eq. 6?\n\nIt is written that \"within each task, larger marker size and less opacity indicate task variant with higher complexity.\" Did the authors mean larger markers size and _more_ opacity? If not, I don't think your figure is interpretable.\n\nFor out of distribution tasks: What was the reasoning behind showing the CV but not mean error? Can you also show mean error and standard deviation separately? \n\nWhat is the reasoning behind doubling the delay period or entire trial for OOD part?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors tried their method in various tasks and settings. \n\nThe authors goal of measuring and controlling the degeneracy across different scales is very important and I think an important question.\n\nThe figures are very explanatory and the paper is written well, easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides ways of how to measure the behavior, dynamics, and weight-space degeneracies in RNNs, and then they provide methods to control these degeneracies. They show their findings using various tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "\"We chose this measure because it directly reflects the amount of information present in the task’s inputs and outputs, which the network must process and represent through its hidden state (Tishby et al., 2000).\" In the experiments done in paper, the dimensions of the RNNs are most of the time higher than the dimensions of your input and output. I do not understand how does this method (of measuring the entropy of input and output _separately_) works, in the case when the goal of an RNN is to reproduce its input? Say one presented a highly complex input signal. And since the goal is to reproduce, then the output signal is also highly complex. But, the RNN only needs to implement an identity function, but the method suggested would classify this as a difficult task. I think this clearly demonstrates the problem of taking into account the input and output signal separately. At the end of the day, the goal of an RNN is to _map_ its inputs to outputs, so to the best of my understanding, there lies a significant problem in this approach. Since authors base all of their claims to this methods, I am highly confused and would like to have a discussion about this. \n\nControlling methods are only tried in one task but the results are written in general form. The claims in the paper, I believe could only be made if the authors have tried these methods on different tasks.\n\nIn the task settings in the paper I think it is true that increasing the number of channels makes the task more difficult, but I don't think it is always the case. Consider the delayed match-to-sample task. One can either have two different input channels to represent your signals or one. I think in this case representing with two channels makes the task easier. Does this show that this method is _not_ task agnostic?\n\nFig 4A, third column does not support your claim but you did not discuss it.\n\nAuxiliary loss: I do not believe the defined auxiliary loss is in fact _auxiliary_ in the delayed discrimination task. When the task of a RNN is to output f1-f2 together with sign(f1-f2), I believe the main task becomes the former one." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. As hinted at in the weaknesses, it's unclear if this measure relates to the true underlying generative process. Can they authors provide additional details on the generative processes' of these tasks (i.e where the uncertainty is, for instance in the delay period, or the probability of the bit-flip) and instead provide increased/decreased complexity of tasks based on this, and if the central claims of the paper still hold?\n2. As detailed in 3.1, entropy calculated is discrete, however some tasks are continuous, yet no detail is provided on how binning can affect the final measure of H_{task}. Can the authors provide some results on this, and how this might change the 3 clusters of tasks they observe?\n3. Can the authors provide any justification for why a joint/conditional/marginal entropy on input & output wasn't used? \n4. Fig 2: while the general trend of more PCs are needed for more complex tasks, how can this trend be reconciled across tasks? For instance DD (high input, low output) needs more PCs than the most \"complex\" n-bit task. Similarly, the opposite is observed for path integration (fewer PCs) and n-bit which have similar complexity measures. \n5. Fig 2: For panel E can the authors use either % or decimals consistently for variance?\n6. Fig 3: Can the authors provide some justification or discussion for the trends observed, as the current writing goes it is unclear why such a trend makes intuitive sense. For instance, why does the output task complexity provide this trend v/s the input task complexity?\n7. Fig 3: While the trend stated holds for same tasks with higher complexity, can the authors provide clarification on why between tasks this isn't observed? For instance sin-N and n-bff have similar output complexity measures but vary on DSA values. Can the authors provide some insight/clarification on this?\n8. It appears the authors use different hidden dimensionalities in RNNs between tasks (i.e 64 v/s 128) while showing how their measure of task-complexity relates with degeneracy. As discussed in 3.3 this influences the capacity of the network. Can the authors provide a justification for why this was done and how the trends remain the same/change if all networks have the same dimensionality?" }, "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 provides a unified approach to understanding the wide range of solutions in task-trained RNNs. The authors present multiple reasons/possibilities for these solutions (i.e due to behavioral, weight space or neural dynamic variability), and quantify its relationship with the amount of information present in each task. The information theoretic approach provides a way to categorize more complex v/s less complex tasks, which is then used to observe trends with other covariates. The authors further increase same task complexity by increasing the amount of information associated with the task (by adding independent channels). Their results suggest a similar trend in complexity with weight space degeneracy, and an inverse trend for neural dynamics and generalization. Lastly, they use these insights to increase/decrease the solution-space associated with tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an approach to explore & control the solution space of task-trained vanilla RNNs. They provide an information theoretic approach to measure task complexity and provide empirical results on its correlation with behavioral, weight space and neural dynamic degeneracy. Lastly, the also provide ways in which such degeneracy of solutions can be controlled. Their results are performed on four neuroscience inspired tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have some main concerns/weakness which I will outline here with specifics under questions.\n\n1. The authors present an interesting approach to quantify task-complexity using entropy. However, as detailed in the paper, this is very closely tied to the probability encoding the underlying generative process for each task. Specifically, while adding channels increases the information, a more robust measure would be changing the probability of the generative process (for instance the probability of a bit flip in a single channel). While in principle adding channels increases information associated with tasks, it's unclear if this is representative of the \"randomness of the task\".\n2. The authors discuss ways to control solutions associated with task-trained RNNs. Other than increasing task-complexity and noting it's trend associated with weight and dynamical degeneracy (presented in this paper), the other four methods have been previously observed. Further, the authors do not provide any theoretical or discussion regarding how this relates to the method presented in this paper. Can the authors expand on the novelty of this 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": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please see the questions in the Weaknesses before." }, "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 structured and clear. \n- Jointly considering several degeneracy metrics (in particular DSA) and tasks is useful for the community. \n- The tasks considered are diverse and rooted in the literature, giving a considerable degree of generality to the results.\n- The related works section is thorough and bridges to the feedforward literature" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper considers the effect of task complexity on “solution degeneracy”. The latter is characterized through examining how three measures behave as a function of task complexity: 1\\) Similarity of the dynamical system between trained networks, 2a) difference in weights from initialization and 2b) pairwise similarity of weights between trained networks, and 3\\) difference in generalization performance between trained networks. \n\n### Recommendation\nOverall, I think the paper is a valuable survey of the influence of task complexity on degeneracy and am looking forward to see it published. The main issue I see is with how expected the results are, depending on the way that it is regularized: An underconstrained system (as per less input-output channels) will be more variable due to unconstrained directions in the dynamics. If regularization in training, as is common practice, would turn out to remove for example DSA variability, I would not find the results sufficiently novel for ICLR. Until this point is clarified, I cannot recommend the paper to the conference." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(more important points first)\n\n- **Regularization.** I could not understand from the paper whether a weight penalty was present when training on the task loss with BPTT. Weight decay for Adam is not specified, I assume it is the default, in case of which there would be some penalty. Was this the case, and if not, how would the measures change if the penalty got increased? For a stronger penalty, I would expect that the DSA metric shows less variation, as presumably noise directions that give rise to dynamical variability would then be penalized to decay. This would make the paper's finding much less interesting to me, because we already know that controlling task-unrelated dimensions is useful. I wonder for example whether Fig. 7 would be induced by strong regularization, biasing the network towards the presumably simpler line attractor solution.\n- **Expectedness.** As the authors state in the discussion, most of the results can be understood with the following argument: As task entropy increases, the network parameters are increasingly constrained. Hence, there are less parameters that are unconstrained and give rise to degeneracy: Effectively, the network has now to solve $N$ tasks in parallel, but has the same representational capacity. \n This makes me wonder how the metrics change when task complexity is not scaled up by replicating the same task over channels, but changing the complexity of the function that needs to be learned? As a very simple and probably too naive example, how would the network behave if task complexity was controlled by the number of Fourier modes in the target signal? It seems that the paper identifies a specific kind of task complexity whose generality is questionable: The proposed measure is not rooted in the literature apart from a loose connection to information theory.\n- **Weight degeneracy.** I find the metric not well motivated. Why should I expect a pair of networks that is non-degenerate (i.e., similar in some way) to have small $d\\_{PIF}$? Would I not rather consider an (orthogonal) similarity measure on matrices, for example $d\\_{sim} \\= min\\_O ||W\\_1 \\- O^{-1} W\\_2 O||\\_F$, where $O$ is an orthogonal or general invertible matrix. This is exactly the $d\\_{procrustes}$ the authors discuss for DSA. I expect that using this metric on the weights would essentially make it the DSA metric. It would be especially interesting if that changes Table 1\\. \n- Do the authors have a hypothesis why change weight degeneracy depends on control paradigm?\n- **Relevance.** Why should we care about whether a networks solution is degenerate or not? I understand that it is interesting descriptively to know a connection between task parameters and degeneracy metrics, but I am not sure where I would need this knowledge. The paper would benefit from discussing this. \n- **Controllability.** It is hard to generalize from the findings for controlling degeneracy from the DD task only. For example, it is hard to say what an auxiliary loss term would look like for the other tasks, possibly involving some amount of engineering. \n- **Generality.** Do the authors suspect that there a tasks where the observed trends do not hold?\n- No code is supplied, limiting transparency and reproducibility.\n\nAgain, I like the paper and think it is a valuable contribution, but these points make me question its novelty and generality." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the Weaknesses part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper proposes a unified approach to analyze the degeneracy of solutions obtained by RNNs on several tasks from several perspectives.\n\n2. This paper incorporates certain measures from information theory to quantify task complexity. The authors conclude that an increase in task complexity will lead to a reduction in degeneracy in neural dynamics and generalization behavior. However, it will simultaneously increase the degeneracy in weight space.\n\n3. Based on their analysis, the authors propose several strategies for controlling solution degeneracy in practical applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a unified approach to quantifying and controlling the degeneracy of solutions obtained by RNNs on several tasks by analyzing their degeneracy from three levels: behavior, neural dynamics, and weight space. They introduce some measures from information theory to quantify task complexity and conclude that increasing task complexity will reduce degeneracy in neural dynamics and generalization behavior, but increase the degeneracy in weight space. Based on these discoveries, they propose several strategies to control solution degeneracy in practice." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The results in this paper are empirical and lack of theoretical analysis. For instance, it would be better if the authors could add more discussion on the reason why increasing task complexity will increase the degeneracy in weight space. Maybe some theoretical analysis of some toy examples will help. \n\n2. The strategies proposed for controlling solution degeneracy in practical applications are too general and may not be very useful in practice. For example, increasing task complexity reduces dynamical degeneracy and increases weight degeneracy, it is not clear which strategies we should use for specific tasks. More discussions on this aspect are welcome." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024measuring,\ntitle={Measuring and Controlling Solution Degeneracy across Task-Trained Recurrent Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ypBYdetYd9},\nnote={under review}\n}" }, "abstract": { "value": "Task-trained recurrent neural networks (RNNs) are versatile models of dynamical processes widely used in machine learning and neuroscience. While RNNs are easily trained to perform a wide range of tasks, the nature and extent of the degeneracy in the resultant solutions (i.e., the variability across trained RNNs) remain poorly understood. Here, we provide a unified framework for analyzing degeneracy across three levels: behavior, neural dynamics, and weight space. We analyzed RNNs trained on diverse tasks across machine learning and neuroscience domains, including N-bit flip-flop, sine wave generation, delayed discrimination, and path integration. \nOur key finding is that the variability across RNN solutions, quantified on the basis of neural dynamics and trained weights, depends primarily on network capacity and task characteristics such as complexity. We introduce information-theoretic measures to quantify task complexity and demonstrate that increasing task complexity consistently reduces degeneracy in neural dynamics and generalization behavior while increasing degeneracy in weight space. These relationships hold across diverse tasks and can be used to control the degeneracy of the solution space of task-trained RNNs. Furthermore, we provide several strategies to control solution degeneracy, enabling task-trained RNNs to learn more consistent or diverse solutions as needed. We envision that these insights will lead to more reliable machine learning models and could inspire strategies to better understand and control degeneracy observed in neuroscience experiments." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Recurrent Neural Network", "Dynamical System", "Neural Computation", "Computational Neuroscience" ] }, "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/3d8fd8097a4afd121525a164ce773b1a7597e589.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": "Measuring and Controlling Solution Degeneracy across Task-Trained Recurrent Neural Networks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Is the task or audio-visual video segmentation supervised or unsupervised? (e.g. do have ground truth video segmentation pixel maps for training or not?)\n - There is mention of a semi-supervised approach in the related works, is this paper following the same approach? Please consider adding such detail.\n- In eq. 1, what prompt is x? Why is a prompt needed at all? How do you mathematically define an audio frame? Is it a spectrogram?\n- Is the training of Keyframe Mask Generation supervised, unsupervised or semi-supervised? I am reading that it is fine-tuned on keyframes extracted from the training dataset by RCPG, fine-tuned to do what exactly?\n - Please consider adding a step-by-step description of the training process.\n- What do the scores in the Table 1 correspond to?\n- Section 4.3: what is the \"cosine similarity\" used for? Please consider more details about its use. \n- What happens when sounding event are overlapping? Please discuss how your method handles overlapping sound events, or if this is a limitation of your approach." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The core novelty of this paper is the proposal to rely on audio segmentation to decompose the audio-visual video segmentation task into sub-problems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focusses on the task of audiovisual video segmentation: given a sounding video, generate pixel-level maps of sound-producing objects that align with the ongoing audio.\n\nExisting methods suffer from poor temporal alignement.\n\nTo tackle this issues the authors introduce a two-steps framework:\n- LLM-assisted audio event segmentation.\n- Segment-based downstream video segmentation.\n\nThe paper also proposes a new dataset and benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although the high-level task description is clear, the paper lacks a deeper explanation of the workflow with simple terms (it is unclear whether the task is supervised, unsupervised or semi-supervised). The introduction dives into convoluted acronyms (e.g. Retrieval-augmented Control Points Generation Module (RCPG)) took quickly without telling what they are actually meant to achieve (detecting keyframes + reference masks). This makes the paper hard to understand.\n - I would suggest adding an overview of the task setup in the introduction (there is no description of the Figure 1 anywhere in the paper).\n- RCPG is meant to perform audio segmentation, it should be treated as such and compared with other audio segmentation methods.\n- Results are presented before metrics and data.\n- The training protocol for the Keyframe Processor is unclear." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 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 intends to solve an important problem in AVS.\n\n2. The visualization and representation are relatively clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "An AVS model to solve the temporal misalignment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Temporal misalignment: It seems an important problem to me, but the author should have expressed this specific problem more clearly. How many failure cases are caused by temporal misalignment? Are there any ratios? Is there any quantitative analysis beyond the qualitative cases? Additionally, the impact of temporal misalignment in Figure 5 is not that clear. In my perspective, most of the cases in Figure 5 are caused by simply incapable segmentation networks.\n\n2. Missing important references and comparison: The major CV conferences, including CVPR, ICCV, and ECCV, have already published numerous papers on supervised AVS. However, the author does not compare any of these top works in Table 1. Here is a list of papers: [1-8].\n\n3. Propagation process: Ablation on Audio-insert Propagator only tests with 4 layers and 1 layer. The natural process involves testing more settings, selecting the best, and reporting the peak performance. Would it achieve better results if using 8 layers?\n\n4. Low amount of testing set: MOC (Multiple-sound Source Conversion) only contains 17 cases. Can these 17 cases serve as solid proof of the performance? I think it requires more data.\n\n5. Accumulation error of key frame result: The model appears to rely on the segmentation result of the first frame. What if it is incorrect?\n\n6. Temporal misalignment: In previous models like TPAVI and AVSegformer, there is no specific temporal information included in the model. It would be better to compare it with other models with temporal perception.\n\n7. Some important AVS works in the related work: Works [10-12], including unsupervised/weak-supervised AVS and open-vocabulary AVS, need to be discussed in the related work section.\n\n[1] (Cited but not compared) Chen, Y., Liu, Y., Wang, H., Liu, F., Wang, C., Frazer, H., & Carneiro, G. (2024). Unraveling Instance Associations: A Closer Look for Audio-Visual Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 26497-26507). \n\n[2] Chen, Y., Wang, C., Liu, Y., Wang, H., & Carneiro, G. (2024). CPM: Class-conditional Prompting Machine for Audio-visual Segmentation. arXiv preprint arXiv:2407.05358.\n\n[3] Ma, J., Sun, P., Wang, Y., & Hu, D. (2024). Stepping stones: A progressive training strategy for audio-visual semantic segmentation. arXiv preprint arXiv:2407.11820.\n\n[4] Hao, D., Mao, Y., He, B., Han, X., Dai, Y., & Zhong, Y. (2024, March). Improving audio-visual segmentation with bidirectional generation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 3, pp. 2067-2075).\n\n[5] Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., ... & Gao, P. (2024, March). Referred by multi-modality: A unified temporal transformer for video object segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 6, pp. 6449-6457).\n\n[6] (Cited but not compared) Yang, Q., Nie, X., Li, T., Gao, P., Guo, Y., Zhen, C., ... & Xiang, S. (2024). Cooperation Does Matter: Exploring Multi-Order Bilateral Relations for Audio-Visual Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 27134-27143).\n\n[7] Sun, P., Zhang, H., & Hu, D. (2024). Unveiling and Mitigating Bias in Audio Visual Segmentation. arXiv preprint arXiv:2407.16638.\n\n[8] Nguyen, K. B., & Park, C. J. (2024). SAVE: Segment Audio-Visual Easy way using Segment Anything Model. arXiv preprint arXiv:2407.02004.\n\n[9] Li, J., Yu, S., Wang, Y., Wang, L., & Lu, H. (2024, October). SelM: Selective Mechanism based Audio-Visual Segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia (pp. 3926-3935).\n\n[10] Liu, J., Liu, Y., Zhang, F., Ju, C., Zhang, Y., & Wang, Y. (2024). Audio-Visual Segmentation via Unlabeled Frame Exploitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 26328-26339).\n\n[11] Liu, J., Wang, Y., Ju, C., Ma, C., Zhang, Y., & Xie, W. (2024). Annotation-free audio-visual segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 5604-5614).\n\n[12] Guo, R., Qu, L., Niu, D., Qi, Y., Yue, W., Shi, J., ... & Ying, X. (2024). Open-Vocabulary Audio-Visual Semantic Segmentation. arXiv preprint arXiv:2407.21721.\n\nI will consider raising my score if the authors can address the questions above." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. I'd be happy to increase my rating if the authors address the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The use of a retrieval-augmented LLM approach to identify audio control points and address temporal misalignment is novel within the AVVS domain.\n2. The proposed Co-Prop framework addresses a pain point in AVVS, which is useful for content creation in complex audio-visual environments like AR or video editing.\n3. Experimental results show improved alignment rates, especially in scenarios with multiple sound sources, demonstrating the method’s efficacy in AVVS.\n4. The proposed method is well presented, with comprehensive diagrams and examples illustrating the misalignment issues in prior models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents the Co-Prop, a new framework that addresses temporal misalignment in audio-visual video segmentation by enhancing alignment between audio cues and visual segmentation outputs. The Co-Prop framework consists of two core modules: (1) Retrieval-Augmented Control Points Generation Module, which anchors key transition points in the audio, and (2) Audio-Insert Propagator, which propagates the segmentation frame-by-frame, integrating audio information to improve synchronization and reduce memory load. Evaluations on multiple datasets demonstrate better performance in alignment rates and segmentation precision than baseline models, particularly on multi-source audio benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I am concerned that the paper does not explore how segmentation results would vary with different audio types. For instance, if a continuous dog bark is replaced with an intermittent one, it is unclear if the dog’s mask would disappear during pauses in barking. This scenario tests the model's adaptability to temporal gaps in sound, which is vital to confirm its robustness in handling real-world examples.\n2. The paper would benefit from further experiments addressing complex audio scenarios, such as overlapping sounds or sounding objects that are off-screen. These situations are common in real-world settings, and I am curious if Co-Prop could maintain object integrity in such cases. If the model relies heavily on visual input alone when multiple audio cues overlap or when sounds lack visual sources, it may risk collapsing or misinterpreting segments, which could compromise segmentation quality.\n3. The multi-step retrieval process in the RCPG module could be explained more clearly. For instance, while the ablation study in Table 2(b) shows performance improvements with 3-step prompts, it is not that clear why these prompts outperform simpler versions. I would encourage the authors to clarify how retrieval samples are chosen and if any cases show weaknesses in control point detection, as this would provide a clearer view of RCPG’s reliability.\n4. While the authors report strong results on the evaluated datasets, I would like to see tests on additional in-the-wild audio-visual data to better gauge Co-Prop’s robustness. Applying Co-Prop to less curated datasets could validate its claims of temporal alignment across different audio contexts.\n5. Although the paper mentions memory efficiency, no concrete results are provided to quantify these improvements. Memory usage is critical in AVVS applications, especially for long videos, and I would suggest including a direct comparison of memory consumption against baselines.\n6. I am concerned that Qwen’s performance in detecting precise transition points may vary, as language models like Qwen are not specifically optimized for detecting fine-grained audio transitions. I recommend using acoustic event detectors, such as PANNs [1] or BEATs [2], which could potentially enhance the accuracy of detecting key audio transition points. These models are trained to recognize audio events and may provide more robust control point identification, leading to more consistent segmentation performance, especially in scenarios with overlapping sounds.\n\nReferences\n\n[1] Kong et al. PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition.\n\n[2] Chen et al. BEATs: Audio Pre-Training with Acoustic Tokenizers." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Novelty in Addressing Temporal Misalignment: The two-stage Co-Prop framework innovatively tackles the temporal misalignment issue. By anchoring the temporal boundaries and inserting audio cues frame-by-frame, the model effectively improves synchronization between audio and visual data.\n2. The paper provides extensive experimental evidence demonstrating the framework's effectiveness on multiple datasets (S4, M3, and AVSS) and backbones (ResNet and PVT-v2). \n3. Introducing the MOC test set and a new alignment rate metric to measure synchronization accuracy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses temporal misalignment issues in audio-visual video segmentation (AVVS), where current methods fail to synchronize audio cues and segmentation outputs. The proposed solution, the Collaborative Hybrid Propagator Framework (Co-Prop), includes two main components: Retrieval-Augmented Control Points Generation (RCPG) and the Audio-Insert Propagator. The RCPG module anchors the audio's temporal boundaries by leveraging a large language model (Qwen) to generate control points, splitting the audio into semantically consistent segments. The Audio-Insert Propagator performs frame-by-frame video segmentation, embedding audio guidance information to align audio cues with video frames effectively." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Reliance on Qwen LLM: With many multimodal models (MLMs) now capable of video understanding(Gemini, Video-llama, ...) and even reasoning segmentation(LISA), what specific advantages does the proposed approach offer over directly using these large models?\n2. Accuracy Discrepancy for AVSegFormer: The reported accuracy for AVSegFormer in this paper doesn’t align with that in the cited references." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In Figure 2(a), only the audio signal is used for control point generation. Is it possible to also introduce the visual frames in this process? (In my opinion, this would be more beneficial for the LLM to better classify the correct sounding objects). Are there suitable metrics or potential measures to evaluate/guarantee the correctness of the generated object categories, as many objects have similar sounds?\n2. The authors rely on the Qwen LLM to generate control points. Have the authors tried other superior LLM models?\n3. The proposed method consists of two stages. Would it be possible to consider an end-to-end strategy to address the temporal misalignment issue?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Existing methods for audio-visual video segmentation tend to design increasingly complex architectures to enhance audio-visual interactions or pixel-level segmentation results. Unlike them, this paper highlights the temporal misalignment issue between audio and visual signals, which is a valuable research direction for the AVVS problem.\n2. The proposed method, especially the preliminary audio boundary anchoring module in the first step, seems to be interesting and well-motivated.\n3. The experimental results are extensive, and the proposed method achieves significant performance improvement." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To address the temporal misalignment issue in the audio-visual video segmentation task, this paper proposes a novel two-stage method. In the first stage, the authors design modules for preliminary audio boundary anchoring. Key audio frames (control points) and corresponding visual frames reflecting the audio transition are obtained. Subsequently, in the second stage, the authors propose an audio-insert propagator module to generate pixel-level segmentation maps for normal frames by propagating masks from key frames. Experiments on three sub-benchmarks demonstrate the effectiveness and superiority of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Given the aforementioned strengths, I am basically satisfied with this work. A potential weakness could be that the ideas of key frame anchoring (first stage) and mask propagation (second stage) likely originate from existing methods in video object segmentation, despite these contributions being seamlessly integrated into the audio-visual video segmentation task.\n2. Although the proposed method achieves satisfactory performance on three datasets, the authors only compared several baselines in Table 1. A more comprehensive review of recently published works for AVVS is required. Moreover, it would be better to discuss with some typical methods from other audio-visual learning tasks, such as audio-visual event localization and video parsing, in the related work section." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper address the problem of audio-visual video segmentation with a controllable audio insertion propagation framework equipped with two designed modules." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024collaborative,\ntitle={Collaborative Hybrid Propagator for Temporal Misalignment in Audio-Visual Segmentation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yqJoqtUwSI},\nnote={under review}\n}" }, "abstract": { "value": "Audio-visual video segmentation (AVVS) aims to generate pixel-level maps of sound-producing objects that accurately align with the corresponding audio. However, existing methods often face temporal misalignment, where audio cues and segmentation results are not temporally coordinated. Audio provides two critical pieces of information: i) target object-level details and ii) the timing of when objects start and stop producing sounds. Current methods focus more on object-level information but neglect the boundaries of audio semantic changes, leading to temporal misalignment. To address this issue, we propose a Collaborative Hybrid Propagator Framework~(Co-Prop). This framework includes two main steps: Preliminary Audio Boundary Anchoring and Frame-by-Frame Audio-Insert Propagation. To Anchor the audio boundary, we employ retrieval-assist prompts with Qwen large language models to identify control points of audio semantic changes. These control points split the audio into semantically consistent audio portions. After obtaining the control point lists, we propose the Audio Insertion Propagator to process each audio portion using a frame-by-frame audio insertion propagation and matching approach. We curated a compact dataset comprising diverse source conversion cases and devised a metric to assess alignment rates. Compared to traditional simultaneous processing methods, our approach reduces memory requirements and facilitates frame alignment. Experimental results demonstrate the effectiveness of our approach across three datasets and two backbones. Furthermore, our method can be integrated with existing AVVS approaches, offering plug-and-play functionality to enhance their 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": [ "audio-visual video segmentation" ] }, "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/540568ba3a478288f5ebbf93bd1e628a9ccec29a.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/55c3fd213e1f53fdce9d8091401fc7abe7b3bf21.zip" }, "title": { "value": "Collaborative Hybrid Propagator for Temporal Misalignment in Audio-Visual 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": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could you summarize the (non-trivial) significance of technical contributions, compared with existing literature? \n\n2. Could you elaborate on the significance of the experimental results? The numbers (such as accuracy) appears low and gives the doubt that the paper did not use strong baselines. \n\n3. This paper severely lacks references from ICML, NeurIPS, ICLR, AISTATS, especially from recent years. The authors are encouraged to provide a thorough literature study from the mainstream ML venues especially from 2022-2024, to help convince readers on the novelty and significance of this submission." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The significance of federated learning is well presented and the motivation for ensuring fairness is well-motivated.\n\n2. The theoretical analysis appears correct.\n\n3. The experiments do confirm the improvement, to an extent." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies an important problem, i.e,. to ensure fairness of federated learning while striving to maintain the accuracy, for scenarios where heterogeneity hurts the performance of FL. Their approach was based on an entropy-based formulation and they show theoretical results such as convergence as well as experiments which to an extent verifies the effectiveness of the proposed algorithm. \n\nThe paper appears to be fairly well-written and well motivated. The main concern is the the limited novelty and significance. The experimental results are not impressive either." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the exact formulation has not been published in the literature, the formulation appears to be fairly straightforward and does not seem to be a significant contribution to the field. The analysis follows the standard analysis. \n\n2. The experimental results, while they show some improvement, do not appear impressive. The actual significance of the proposed framework is not convincing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the questions and comments listed in the Weakness section" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. To the best of my knowledge, the idea of using a constrained entropy model for aggregation in FL to improve fairness is novel and interesting.\n2. The idea of adaptively changing the global aggregation to either prioritize global model accuracy or fairness, while heuristic, is again interesting.\n3. Theoretical results proving that the proposed FedEBA+ can reduce variance compared to FedAvg for generalized regression and strongly convex models.\n4. Experimental results look promising with the proposed FedEBA+ outperforming FedAvg and other baselines both in terms of global model performance and fairness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Existing algorithm to achieve fairness in FL often do so at the expense of global model accuracy. The paper proposes FedEPA+, an algorithm to improve fairness in FL while maintaining global model performance. The key idea here is to use entropy-based aggregation followed by a novel model alignment technique which adaptively optimizes for either global accuracy or fairness depending on the current performance of the model. Theoretical results are provided showing the convergence of the proposed algorithm in general non-convex FL setups along with guaranteed improvement in fairness in the strongly convex scenario. Empirical results show that the proposed algorithm outperforms existing fair FL baselines in both global model accuracy as well as fairness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The clarity of writing and motivation for doing FedEBA+ can be significantly improved. There were several parts which I found confusing or could not fully comprehend as I discuss below.\n\n* It is a bit surprising to me that the expression for the aggregation probabilities in Eq. (4) does not depend on the desired $\\tilde{f}(x)$ in Eq. (3). Is there an intuitive explanation for why this is the case?\n\n* Authors are encouraged to provide an intuitive explanation for why maximizing constrained entropy can improve fairness. Right now, entropy-based aggregation just appears to be a black box to improve fairness\n\n* While I could follow the construction of the proposed bi-level objective in Eq. (6) and the idea to adaptively change the ideal gradient, I was unable to follow the changes made in the local and global aggregation to account for this. In my understanding, we first solve the inner problem. i.e, find $p$ given the current global model. Then given $p$, we update the global model with either the ideal global gradient or ideal fair gradient depending on the requirement. Given this understanding, my questions are as follows:\n\n * In the alignment to improve global accuracy why are the $p_i$'s computed using the loss of the local models? (Line 312). In my understanding, since we are fixing $x$, when solving the inner problem, the $p$ should be computed using the loss of the global model, i.e, the expression in Line 351?\n * In the alignment to improve fairness why is the local optimization at clients changed (Eq. 11 and Line 11 in Algorithm 1)? In my understanding only global aggregation should change?\n\n* I don't follow how in Prac-FedEPA+ clients need to only communicate once. If we follow Algorithm 3 then it appears that communication can happen twice within a round: first in Line 6-8 and then again in Lines 18-22. I found this to be very confusing and authors should clarify what is the communication cost in their experiments.\n\n* $A$ is not defined in Theorem 5.1. Similarly $w$ is not defined, although from context it appears to be the prior aggregation weights of client objectives.\n\n* It seems a little strange to see discussion on the lower bound of an upper bound in Remark 5.2. Authors should just state the worst possible convergence rate of FedEPA+ by considering that $\\sum_{i=1}^M w_i^2 \\leq 1$.\n\n* In Remark 5.3. it appears that to show convergence of $\\alpha \\neq 0$ we need to assume $k << m$. Is this correct? If so, authors should clearly mention this and explain why this is the case. Also I don't follow the argument that the proposed alignment results in a faster convergence rate than FedAvg since both seem to be achieving $O(1/\\sqrt{mKT})$ rate.\n\n* Why is FedMGDA+, PropFair and TERM incompatible in Table 2?\n\n2. The parameter $\\theta$ appears to be crucial to the performance of the algorithm but currently there is very little discussion on how to set this parameter and its impact on convergence. Theoretically, we don't see any effect of $\\theta$ in Theorem 1, which is a bit surprising to me. In practice, I would also suggest moving Table 7 to the main text and discuss the trade-off in setting $\\theta$ in more detail." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. **Novelty of Exponential Normalization**: The core aggregation strategy is based on an exponential normalization of client losses. Could the authors clarify how this approach fundamentally differs from existing techniques that also adjust aggregation weights based on client performance? How does the use of entropy provide a significant advantage over these existing methods?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- **Clear Problem Motivation**: The paper addresses an important problem in federated learning—how to achieve fairness across clients while ensuring that the global model's performance does not degrade. The motivation behind balancing fairness and performance in a heterogeneous FL environment is well articulated.\n\n- **Theoretical Guarantees**: The paper provides theoretical analysis for the convergence of FedEBA+ in non-convex settings, which adds credibility to the proposed approach. The fairness improvements are also supported by performance variance analysis.\n\n- **Bi-level Optimization Framework**: The introduction of a bi-level optimization framework is interesting and provides a structured way to balance fairness and performance. The authors also derive a closed-form solution for the aggregation probability, which improves computational efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper titled \"Entropy-Based Aggregation for Fair and Effective Federated Learning\" proposes a novel algorithm called FedEBA+ to address the fairness issues in Federated Learning (FL) while maintaining the global model's performance. The authors leverage an entropy-based aggregation mechanism, combined with model and gradient alignment, to optimize both fairness and global performance. A bi-level optimization framework is introduced to derive an efficient closed-form solution for aggregation probabilities, which is claimed to enhance fairness by adjusting weights for underperforming clients. The paper provides theoretical guarantees for the algorithm's convergence and fairness, and empirical results on several datasets demonstrate that the proposed approach outperforms existing fairness-oriented FL algorithms." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Lack of Novelty in Aggregation Strategy**: The core novelty of the paper—using an entropy-based aggregation mechanism—essentially boils down to applying an exponential normalization based on client losses to determine aggregation weights. This approach is not particularly novel, as similar strategies have been used in various other contexts (e.g., softmax-based weighting). \n\n- **Overemphasis on Entropy**: The paper heavily emphasizes entropy, but in practice, the core idea is simply a re-weighting based on client loss. The connection between entropy and fairness, while valid, feels somewhat forced in its application here. The novelty of applying maximum entropy principles in such a straightforward manner might be overstated.\n\n- **Limited Innovation in Fairness-Performance Trade-off**: Although the paper claims to balance fairness and performance, the approach primarily adjusts aggregation weights based on client loss. Many existing algorithms adjust client weights in some form, and the use of entropy as a fairness mechanism does not seem to introduce substantial improvements beyond what is already available in the 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": 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": 4 }, "strengths": { "value": "- The work addresses a critical issue in federated learning and provides a practical solution that could have some impact on the field.\n- The paper is well-researched, with a robust theoretical foundation and comprehensive empirical validation. \n- The authors provide insightful convergence analysis and fairness guarantees.\n- The objective function and the proposed algorithm are novel contributions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces FedEBA+, a novel federated learning algorithm designed to enhance both fairness and global model performance through a computationally efficient bi-level optimization framework. The authors propose an innovative entropy-based fair aggregation method for the inner loop and develop adaptive alignment strategies to optimize global performance and fairness in the outer loop. The paper provides theoretical analysis confirming the convergence of FedEBA+ under non-convex settings and demonstrates its superiority over state-of-the-art fairness algorithms through empirical results on various datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper would benefit from a more intuitive explanation of the connection between entropy and fairness.\n- The explanation of the entropy-based aggregation method could be more detailed, particularly in terms of how it differs from and improves upon existing methods.\n- The concept of \"ideal loss\" appears somewhat confusing. Since $\\tilde{f}(x)$ represents the ideal loss, which signifies the global model’s performance under an ideal training setting, it is unclear why the gradient of the ideal loss can be estimated by averaging local one-step gradients during the model alignment procedure, yet it should be estimated by Equation.10 during the gradient alignment procedure. Could the authors give more intuition to clarify this point?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "see the weakness." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- Content is solid and comprehensive.\n- The paper is well-structured and advances step by step.\n- The theory is clear and solid.\n- The experimental section is well-developed and extensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Existing federated fairness algorithms strive to enhance fairness but often fall short in maintaining the overall performance of the global model, typically measured by the average accuracy of all clients. To address this issue, this paper proposes a novel algorithm that leverages entropy-based aggregation combined with model and gradient alignment to simultaneously optimize fairness and global model performance. The method presented in this paper employs a two-layer optimization framework and derives an analytical solution for the inner loop aggregation probabilities, making the optimization process highly computationally efficient. Furthermore, the paper introduces innovative alignment updates and adaptive strategies in the outer loop to further balance the performance of the global model and fairness. Additionally, the paper conducts a convergence analysis and numerous experiments to demonstrate the novelty and effectiveness of the proposed method. The content is very detailed, making it an excellent paper." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Please explain how the convergence rate range mentioned in lines 396 to 397 of the main text is derived." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a fair FL algorithm that addresses the underexplored challenge of improving performance fairness while enhancing global accuracy, with theoretical and empirical demonstrations." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024entropybased,\ntitle={Entropy-Based Aggregation for Fair and Effective Federated Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yqST7JwsCt},\nnote={under review}\n}" }, "abstract": { "value": "Federated Learning (FL) enables collaborative model training across distributed devices while preserving data privacy. Nonetheless, the heterogeneity of edge devices often leads to inconsistent performance of the globally trained models, resulting in unfair outcomes among users. Existing federated fairness algorithms strive to enhance fairness but often fall short in maintaining the overall performance of the global model, typically measured by the average accuracy across all clients. To address this issue, we propose a novel algorithm that leverages entropy-based aggregation combined with model and gradient alignments to simultaneously optimize fairness and global model performance. Our method employs a bi-level optimization framework, where we derive an analytic solution to the aggregation probability in the inner loop, making the optimization process computationally efficient. Additionally, we introduce an innovative alignment update and an adaptive strategy in the outer loop to further balance global model's performance and fairness. Theoretical analysis indicates that our approach guarantees convergence even in non-convex FL settings and demonstrates significant fairness improvements in generalized regression and strongly convex models. Empirically, our approach surpasses state-of-the-art federated fairness algorithms, ensuring consistent performance among clients while improving the overall performance of the global model." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Fairness", "Heterogeneous Federated 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/3c1ca9107e4bfb7c5130f11960c51569af8d6673.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/5ac2b70db31c1f2f091b820f2b036217c5cee9b1.zip" }, "title": { "value": "Entropy-Based Aggregation for Fair and Effective Federated 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": 3 }, "primary_area": null, "questions": { "value": "- It seems that the first equation in sec 2.4 is obtained from (3) and the fact that the sum of omega is always a constant. In that case, (3) is equivalent to -MMD + \\sum (omega - any constant)^2, not just one?\n- There are multiple hyperparameters in the variable selection algorithms. How sensitive is the performance to the hyperparameters?\n- Is the accelerated algorithm for computing the estimated weights similar to applying gradient descent to equation (3)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is clearly motivated.\n- The method is non-parametric and model-free, allowing a broad range of applications.\n- Experimental results show sizable improvement over the original MMD framework" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a variable selection scheme based on the maximum mean discrepany (MMD). The authors propose two variable selection algorithms, which are based on a regularized MMD objective, to assign weights to important variables and discard the unimportant ones. An accelerated scheme to compute the estimated weights is also presented. Theoretical results on the consistency of the estimated weights and the convergence of the accelerated scheme are provided. Experiments demonstrate the competitiveness of the proposed method to comparing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The literature review could benefit from a broader references. For instance, other parameter selection schemes, other variants of MMD objective etc.\n- The presentation in sec 2.4 is somewhat unclear. Equation (4) and (5) have an intricate structure. Readers could benefit from a clearer presentation/explanation.\n- In the experiments, only two standard comparing methods are considered. The experimental results could be strengthened by including more advanced baselines.\n- In the variable selection algorithms, randomness is involved. E.g., random subsets are chosen to compute the MMD value. It is generally recommended to report the mean and standard error over multiple random trials." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* The optimisation problem formulated in (4) does not seem to encourage sparse solutions. I think the weight vector obtained from (4) on a limited training set is generally non-sparse. How does this translate to a variable selection? Is there a thresholding procedure on the weight vector. If so, such procedure should be detailed in the article.\n\n* Could the authors provide some approximation guarantees for the accelerated algorithm, or/and some generalization guarantee for the original method or its accelerated version?\n\n* Why is there no error-bar in the experimental results? And why not use all the samples in the gene expression data set GSE2034 to obtain a large test set which can reduce the variance of the empirical performance, therefore allowing for a more reliable comparison?\n\n* In Lines 262-264, it says \"the training set, is used to compute the optimal weight vector $\\hat w_{\\hat\\lambda}$ by optimizing the tuning parameter $\\hat\\lambda$\", meanwhile in Lines 280-282 of Algorithm 1, it seems that $\\hat\\lambda$ is tuned in a cross validation manner by minimizing the loss $\\ell(\\lambda)$ on a test set. Could the authors clarity whether or not the tuning of $\\hat\\lambda$ involves exclusively the training set?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Proposition of a new method, presented along with theoretical and empirical results.\n\n* The writing is clear and the paper is not hard to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed method finds the importance levels of variables by optimizing a weighted version of maximum mean discrepancy (MMD). The objective function to minimize is negative weighted MMD plus a regularization term that imposes all variable weights to be $1$ when its parameter $\\lambda$ is set to be infinitely large. The regularization parameter $\\lambda$ is chosen through optimizing the performance of the task - two-sample test or classification. An approximate version of this method is presented to reduce the computational cost, obtained through a first-order Taylor expansion of the objective function. Consistency and convergence results are given for the solution of the approximate method. Experiments on two synthetic data sets and one gene expression data set show the advantage of the proposed method over the unweighted MMD for two-sample test, and the variable selection method by Lasso for classification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The theoretical results only concern the consistency and the convergence of the approximate algorithm. There is no generalization or approximation guarantee.\n\n* This work can benefit from a more extensive empirical study. For two-sample test, only one baseline is compared, and the experiments were exclusively conducted on synthetic data. For classification, two baselines are tested, on synthetic data and one real-word data set. Moreover, there is no error-bar in the reported empirical results.\n\n* Most crucially, while the proposed method is claimed to be a method of variable selection, there is little comparison to other variable selection methods, expect some empirical results o compare the false discovery rate and the classification performance with the Lasso 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": "I am curious that whether it is possible to provide a theoretical guarantee that the optimal weight is beneficial for two-sample test or binary classification?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The presentation of this paper is in general good, clearly stating the background and their intuition. This paper also provides comprehensive results regarding their proposed methods, including both theory for optimization and numerical simulations. Also the idea of using weighted MMD for model selection is novel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the problem of model selection when testing whether two distributions are identical. Traditionally, people calculate the Maximum Mean Discrepancy (MMD) between two distributions, and then perform test using MMD, namely reject the hypothesis that two distributions are identical if MMD exceeds a critical value. The authors observe that, unimportant variables may even hurt the performance of this MMD test. Therefore they proposed the weighted MMD, namely assign a weight to each variable. The optimal weights are chosen so that the weighted MMD is maximized. Variable selection can be done using these weights, facilitating the downstream two-sample test or binary classification. The authors also provide an optimization method with theoretical guarantee. Numerical experiments show the efficiency of the proposed algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concern is that, although the authors claim that weighted MMD is good for selecting variables, they do not provide theoretical guarantees that the optimal weight is beneficial for two-sample test or binary classification. Therefore the proposed method seems to be a little bit ungrounded." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- Why is the ridge penalty used instead of an $L_1$ penalty? Is there any theoretical justification for this choice?\n\n- Can this workflow be extended to more general settings, such as regression with HSIC?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Overall, this paper is clearly written.\n\n- The paper presents an interesting approach to nonlinear variable selection within the MMD framework and develops a faster variant for its implementation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a variable selection method using maximum mean discrepancy (MMD) under a binary-classification setting. The approach assigns and optimizes weights for each variable within a regularized MMD framework, setting some weights to zero for unimportant variables. These optimized weights act as an importance measure for identifying variables contributing to distributional differences. Simulations and real-data analysis demonstrates the empirical merit of this paper." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The term “Regularized MMD” in the title is somewhat ambiguous and does not fully capture the main idea of this paper.\n- In [1], the simplex constraint yields a sparse estimator under the least squares setting, but it remains unclear if the solution to (2) demonstrates similar properties.\n- The properties of $w_\\lambda^*$ are not sufficiently investigated.\n- Table 1: The statistical properties of both the original and accelerated methods should be included.\n- The selection of benchmarked methods is limited [2-3]; there are numerous variants of (nonlinear) Lasso that could be considered.\n- Type-I error studies should be integrated into the main text.\n- An ablation study would help clarify whether using the ridge penalty is beneficial.\n\n### Reference\n\n- [1] Meinshausen, Nicolai. \"Sign-constrained least squares estimation for high-dimensional regression.\" (2013): 1607-1631.\n\n- [2] Ravikumar, Pradeep, et al. \"Sparse additive models.\" Journal of the Royal Statistical Society Series B: Statistical Methodology 71.5 (2009): 1009-1030.\n\n- [3] Yamada, Makoto, et al. \"High-dimensional feature selection by feature-wise kernelized lasso.\" Neural computation 26.1 (2014): 185-207." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "N/A" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The overall procedure is sound. As far as I have checked, the proof is generally correct with detailed assumptions stated. Numerical study convince the soundness of the algorithm design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper utilized the weighted norm with maximum mean discrepancy to solve the variable selection problem for two-sample testing, which is a very important problem in statistics and machine learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the overall procedure seems to be sound, I have serious concerns regarding its theoretical foundation.\n1. The authors proposed to solve Eq. (3), which is the objective combining empirical MMD statistics and $\\ell_2$ regularization. Recall for least squares problem, $\\ell_1$ regularization (lasso) helps with variable selection while $\\ell_2$ does not. I believe the similar result applies here. So why does the author prefer $\\ell_2$ instead of $\\ell_1$ regularization?\n2. For the evaluation of $\\ell(\\lambda)$ in Eq.(4) or Eq. (5), it involves the sample estimate of certain objective. So which type of data is utilized? The training data $\\mathfrak{X}^{Tr}$ or testing data $\\mathfrak{X}^{Te}$, or we need to perform training-validation split on $\\mathfrak{X}^{Tr}$ and then utilize train-train data?\n3. In the Algorithm the authors proposed to solve the optimization problem (3), which is unfortunately a non-convx problem in weight $\\textbf{w}$. Then the authors resort to solve approximation optimization problem (6). What is the optimality gap between these? It is difficult for me to convince the effectiveness of this approximation algorithm.\n4. Similarly, only the convergence analysis for solving the approximation optimization problem (6) is provided, but I cannot see the testing statistical power analysis when resorting to this approximation algorithm\n5. Only the real data instead of the implementation code is provided for this paper." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose variables selection method based on the maximum mean discrepancy (MMD), which can effectively screen important variables that cause differences in distributions between two samples." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024regularized,\ntitle={Regularized Maximum Mean Discrepancy for Variable Selection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yqaN7MfkFU},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we propose a variable selection method based on maximum mean discrepancy (MMD) to effectively identify important variables that contribute to distributional differences between two samples. We begin by assigning weights to each variable and then optimizing these weights within a regularized MMD framework. The optimized weights serve as an importance measure for each variable and can be leveraged for variable selection. Additionally, using the optimized weights, we design two algorithms aimed at enhancing test power and improving classification accuracy for two-sample tests and classification problems. Our method is model-free and makes no assumptions about the underlying structure of the data. Moreover, we propose an acceleration method to improve computational efficiency.\nWe also provide theoretical guarantees, including the consistency of the estimated weights and the convergence of our acceleration algorithms. Through numerical simulations and real-world datasets, we validate the effectiveness of the proposed method." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Variable selection", "Maximum mean discrepancy", "Two-sample tests", "Binary 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/597d767452260fe5d3ccd77fe4bc41ea6efdfd86.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/3f0111c32ead9fa1a596c080fde94c56f5b50377.zip" }, "title": { "value": "Regularized Maximum Mean Discrepancy for Variable Selection" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you for your review.\nI will list the responses to your questions as follows:\n\n**To Weakness:**\nOverall, the two key designs presented in Section 3.1 are major contributors to improving efficiency and performance.\nIn other words, both designs are indispensable. They are therefore of equal importance and should be considered as a whole.\nAccording to the idea of the paper, GPUs are the hardware base for enabling computational acceleration, and the task of software system design is to best adapt to the hardware.\nWe identify two major difficulties (read/write conflict & vehicle sensing indexes) in the adaptation process and provide solutions (two-phase parallel process & linked-list based vehicle sensing indexes) that ultimately allow the GPU's performance to be fully utilized for acceleration.\n\n**To Question 1:**\nThe answer is YES.\nOur simulator supports user-configurable inputs of kinematic and IDM model parameters for each vehicle, including maximum acceleration, general acceleration, maximum braking acceleration, general braking acceleration, headway, and so on.\nUsers can build their own vehicle input data that they want to match the desired scenario, and we provide a tool chain (mosstool mentioned in Appendix A) to support such needs.\n\n**To Question 2:**\nIt is possible but not trivial.\nThe main obstacle is the knowledge of CUDA and the complicated build process of C++.\nActually, for an experienced developers, they can easily change the model by modifying the following function - `IDMCarFollowAcc(const Person& p, float target_speed, float ahead_speed, float distance, float headway)` (https://anonymous.4open.science/r/moss-AF45/src/entity/person/vehicle.cu: line 362) to do it.\nAttracting more community collaboration to improve the system was also one of our goals in opening the source code and submitting our work to ICLR." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Response to Reviewer oXrd" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you very much for taking your valuable time to review our paper.\n\nOur work, a large-scale microscopic traffic simulator, serves the fundamentals of transportation system optimization to support typical transportation system optimization scenarios.\nWe believe that the efficiency of urban transportation systems is closely related to each individual, and that learning-based AI technologies have great potential to optimize the efficiency of transportation systems.\nHowever, as you say, there are fewer researchers focusing on this area right now.\nWe believe this is because there is a lack of large-scale simulators and benchmarking code that can support a larger number of scenarios, leaving researchers unmotivated to investigate the application of learning methods to these problems.\nThat is our aim in accomplishing such a non-incremental, pioneering and difficult work.\n\nMoreover, memory usage analysis will be added to the supplement as part of the performance evaluation." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Response to Reviewer 6yoX" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This work is a wonderful contribution for transportation system optimization, and I strongly believe many researchers can take advantage of this platform. The manuscript is very well structured and written well. The authors covers a comprehensive review of the existing simulations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a fast simulator for transportation system optimization.\nThe main contributions of this paper is a high-performance traffic flow simulation environment and the implementation of 5 transportation optimization problems. The new proposed simulation environment is GPU-accelerated and based on SIMD and programmed in CUDA. The performance of the proposed simulator is compared against some of the existing ones and has demonstrated superior performance on runtime. Different benchmark algorithms are demonstrated on the new simulator." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The scope of this work is too narrow, since there will only be a small subset of researchers that will use this in the learning community. \n\nIt will be nice to include an analysis of memory usage if possible." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": 4 }, "primary_area": null, "questions": { "value": "Please see the limitation and clarify the technical contribution of the paper better. Some additional questions include:\n1.Does the simulator allow the researcher to add uncertainty to the vehicle model or the control of traffic objects in order to validate the robustness of a method, and allow researchers to compare algorithms that reflect the simulation to reality gap?\n2. Will it be easy (or possible) to make the car-following model selectable (instead of sticking to IDM)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.The proposed simulator is more efficient, realistic and capable compared to other candidates (sumo, cityflow, etc.). The proposed simulator is shown to be much faster than\n\n2. The work has practical and application value, it can facilitate research and application in the relevant areas of traffic & transportation system. It is more versatile such that more component (traffic objects) are enabled in the simulator to be controllable. The authors also provide a few predefined evaluation metric APIs in the simulator to facilitate usage." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors present a high-performance simulator for traffic system simulation and optimization, as well as the benchmark resulting for five scenarios with the simulator. The author clearly stresses the weakness of trending traffic system simulators and show through comparison that their proposed simulator can overcome the issues and outperform the SOTA one by running frequency and simulation realism. Overall, the presentation of the work is also clear and comprehensive, and tis work is good in its realm that from the comparison it does show the superiority in many aspects compared with existing simulators in the literature." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In Section 3, the authors described the design of each component of the simulator, however, it is now clear, which component is the main contribution of the design that improves the efficiency and performance of the traffic simulator." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Could the authors provide more justification for the assertion that vehicle simulations are “highly homogeneous” and compatible with the SIMD computational model? Specifically, how are dependencies between vehicles managed in congested scenarios?\n\n2. Why does the simulator's performance appear to decline with scenarios of over 1,000 vehicles, as seen in the comparison with CityFlow? How does this relate to the scalability claims made in the paper?\n\n3. Are there plans to include more detailed explanations of the simulation scenarios and the metrics used to evaluate them? This information would aid in validating the results for scenarios like congestion pricing and dynamic lane assignment.\n\n4. Could the authors comment on the possibility of incorporating advanced driving models such as those used in CityFlowER, which have been shown to improve the realism of driving behavior?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "## Originality\n- The simulator is GPU-accelerated, which is helpful for traffic system simulation (while some GPU simulators are available for autonomous driving).\n\n## Quality\n- The paper provides benchmarks on various transportation optimization tasks, demonstrating practical applicability.\n\n## Clarity\n- The paper is well-organized and presents a clear problem statement, followed by the design choices and technical solutions implemented. \n\n## Significance\n- By supporting large-scale simulations at high speed, this simulator could enable faster and more frequent experimentation with AI-based optimization methods, including reinforcement learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a GPU-accelerated large-scale microscopic traffic simulator aimed at supporting optimization tasks within transportation systems. By leveraging GPU-based parallel computation, the simulator achieves high performance, simulating up to millions of vehicles at a significantly accelerated rate compared to traditional CPU-based simulators like CityFlow and CBLab. The simulator supports various transportation optimization scenarios, including traffic signal control, lane assignment, tidal lane control, congestion pricing, and road planning. The authors provide benchmark results across multiple cities to demonstrate the applicability and robustness of the simulator for different optimization algorithms." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper lacks a comprehensive discussion of some relevant prior works. For example, recent works such as GPUDrive by Kazemkhani et al. (2024) and CityFlowER by Da et al. (2024) could provide essential insights and serve as useful points of comparison for evaluating the novelty and realism of the proposed simulator. For example, [1,3,4] are also a GPU simulators, are the techniques used for GPU acceleration in their papers the same as this paper? Is CityFlowER also as realistic as the proposed simulator?\n\n[1] Kazemkhani, Saman, et al. \"GPUDrive: Data-driven, multi-agent driving simulation at 1 million FPS.\" arXiv preprint arXiv:2408.01584 (2024). \n\n[2] Da, Longchao, et al. \"CityFlowER: An Efficient and Realistic Traffic Simulator with Embedded Machine Learning Models.\" Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2024.\n\n[3] Saprykin, Aleksandr, Ndaona Chokani, and Reza S. Abhari. \"GEMSim: A GPU-accelerated multi-modal mobility simulator for large-scale scenarios.\" Simulation Modelling Practice and Theory 94 (2019): 199-214.\n\n[4] Jiang, Xuan, et al. \"Large scale multi-GPU based parallel traffic simulation for accelerated traffic assignment and propagation.\" Transportation Research Part C: Emerging Technologies 169 (2024): 104873.\n\n- The authors claim that individual vehicle simulation is compatible with the SIMD model, arguing that the vehicles are homogeneous and can be simulated in parallel. However, this assertion could be problematic as vehicle interactions, particularly in congestion, often exhibit dependencies on nearby vehicles. The paper would benefit from a clearer explanation of how dependencies are managed in parallel to ensure realism without sacrificing computational performance. \n\n- In the Simulator Realism experiment, the paper mentions that the simulator achieves more realistic speeds compared to CityFlow. Given that CityFlowER now incorporates more realistic driving models, it would be valuable to benchmark the proposed simulator against it to provide a more comprehensive evaluation.\n\n- When demonstrating scalability at large vehicle counts, the reported performance of CBLab appears to contradict with its original paper when simulating scenarios with over 1,000 vehicles, with CityFlow performing comparably or even outperforming in such instances. This discrepancy between claims and results needs further explanation on why CBLab is generally worse than CityFlow - is this because of your data is different from the CBLab? Can you provide more details about their experimental setup, including the specific datasets used and any differences from the original CBLab experiments? If so, can you use the similar dataset used in CBLab? It would also be helpful to request a detailed comparison of their experimental setup with that of the original CBLab paper, including specifics on hardware configurations and simulation parameters.\n\n- The absence of scenario data within the provided code limits the reproducibility of results, particularly for scenarios such as traffic signal optimization and congestion pricing. Including this data would enhance transparency and reproducibility." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "It is desirable to clarify the data sources used in the simulation.\n\n- Privacy, Security, and Safety: The data used in the simulation includes urban traffic data, necessitating considerations for privacy and security.\n\n- Potentially Harmful Insights, Methodologies, and Applications: The simulation results could have adverse effects if misused for real policy decisions, potentially disadvantaging specific regions or demographics. Ethical considerations are needed.\n\n- Responsible Research Practice: For research involving urban data, it is essential to ensure transparency in legal/ethical approval procedures and the data collection and processing process." }, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety", "Yes, Potentially harmful insights, methodologies and applications", "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": "Q1. Given the numerous experimental results presented, I felt that there was a lack of motivation for their inclusion. Could you provide additional explanations regarding the analysis and significance of these results?\n\nQ2. There seem to be similar commercial services available (e.g., Aimsun). How does this work differentiate itself? There are also simulation games related to urban simulation, such as Cities: Skylines, which are known for their well-designed public transport systems, and the recent sequel has drawn significant attention. What are the commonalities and differences compared to such games? They support user-customized plugins for logging and debugging.\n\nQ3. Could you explain the technical background that led to the design of the simulator's GPU-accelerated architecture, and how this design evolved from previous research efforts in the field?\n\nQ4. Since the target is a simulator that is intended to be used alongside AI/ML-based technologies, which also use GPUs, resource constraints are expected if both the simulator and AI/ML technologies are to be used on servers that are not high-performance. Did you encounter this issue, and do you think it would be a significant problem?\n\n** Additional Comments\n\nThere are inaccuracies in references.\n\n- Unable to verify the original document. Only the citation is found. Please verify or update: HS Mahmassani. Dynamic traffic assignment and simulation for advanced network informatics (dynasmart). In the 2nd International Seminar on Urban Traffic Networks, 1992.\n\n- The cited arXiv document version has been updated. Please verify or update: Alexander I Cowen-Rivers, Wenlong Lyu, Zhi Wang, Rasul Tutunov, Hao Jianye, Jun Wang, and Haitham Bou Ammar. Hebo: Heteroscedastic evolutionary bayesian optimisation. arXiv preprint arXiv:2012.03826, pp. 7, 2020.\n\n- Duplicate reference entry. Please correct:\nQiang Wu, Ming Li, Jun Shen, Linyuan Lü, Bo Du, and Kecheng Zhang. Transformerlight: A novel sequence modeling based traffic signaling mechanism via gated transformer. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023a. URL: https://api.semanticscholar.org/CorpusID:260499801.\nQiang Wu, Mingyuan Li, Jun Shen, Linyuan Lü, Bo Du, and Ke Zhang. Transformerlight: A novel sequence modeling based traffic signaling mechanism via gated transformer. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 2639–2647, 2023b." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Originality: The parallel design utilizing GPU features and the introduction of a linked-list-based vehicle detection index are promising approaches to overcoming the computational limitations of existing CPU-based traffic simulators.\n\nQuality: The paper demonstrates a significant technical achievement by simulating over 2.4 million vehicles with GPU acceleration, achieving substantial performance improvement compared to CPU implementations. The system design and GPU usage are well-documented.\n\nClarity: The explanation of the system architecture is systematic.\n\nSignificance: To address urban traffic optimization, the authors propose an integrated simulation for five scenarios that were previously dealt with individually." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a microscopic simulator for large-scale urban traffic systems utilizing GPU acceleration. To overcome the computational limitations of existing simulators, the authors introduce GPU-based parallel processing, a two-phase parallel process, and a linked-list-based vehicle detection index. Through experiments simulating over 2.4 million vehicles, the proposed system achieved an approximately 88.92x performance improvement compared to existing systems. Furthermore, the authors present a framework that aims to integratively support five major traffic optimization scenarios: Traffic Signal Control, Dynamic Lane Assignment, Tidal Lane Control, Congestion Pricing, and Road Planning.\n\n- Soundness (Score: 2 - Fair): The design and performance improvement of the GPU-accelerated simulator proposed in this research are technically convincing, and the experimental methodology is systematic. However, the simplifications, such as excluding intersection interactions and various traffic participants (e.g., pedestrians, bicycles, public transport), may significantly affect the realism of the simulation, and these aspects were not clearly validated. It is unclear whether prior research justified these exclusions. Although extensive benchmarking experiments were conducted, the lack of detailed analysis and interpretation of the results leaves the motivations behind the scenarios insufficiently explained.\n\n\n- Presentation (Score: 3 - Good): The overall structure and explanations in the paper are clearly presented, particularly regarding system design and the use of GPU acceleration. However, the validation of each scenario within the same system is not clearly articulated. Additionally, it would be beneficial to provide a clear comparison of the proposed system architecture with existing studies, explaining the structural advancements and reasons for adopting this specific architecture.\n\n\n- Contribution (Score: 2 - Fair): The proposed performance improvements through GPU acceleration and the integrated traffic simulator framework are meaningful contributions to the field. However, the lack of individual effectiveness analysis for the five scenarios and insufficient validation of the simulator's realism are notable shortcomings. The experiments with real-world data are limited, which poses a constraint on the practical contribution aspect." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Scenario Validation and Motivation: It would be helpful to explain how the results from the five proposed scenarios provide insights for actual urban traffic planning or policy decisions. If empirical validation showing that similar results can be derived from simulating real-world policy changes were included, it would strengthen the motivation for policymakers to use the simulator. As a policymaker, I would want to validate planned changes beforehand with such a simulator, but the current motivation seems insufficient compared to the well-executed system implementation.\n\nLack of Realism: The simplifications, such as excluding intersection overlap, pedestrians, bicycles, and public transport, lack a clear basis for their effect on simulation outcomes. An analysis of the impact of these exclusions is necessary, including evidence from prior research or justifications for why their impact is minimal.\n\nInsufficient Long-Term Scenario Evaluation: There is a lack of experiments evaluating how the proposed simulator functions under long-term urban traffic pattern changes. For example, reproducing outcomes of real-world urban policy changes using the simulator would help demonstrate its reliability.\n\nCode Reproducibility Issues: Although the code was provided, the step-by-step explanations in the README.md were insufficient for successful reproduction. Specifically, two of the three provided source codes only included installation instructions without guidance on how to build or run scenarios. For the benchmark source code, even after locating it through references due to a missing GitHub link, the target version was not specified, preventing access to the necessary dataset for simulation. More detailed code instructions and thorough guidance on code operation are required, highlighting the lack of user-friendly considerations." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "The first open-source GPU-accelerated large-scale microscopic simulator for transportation system simulation and optimization." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A {GPU}-accelerated Large-scale Simulator for Transportation System Optimization Benchmarking},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yr0l1IoyzV},\nnote={under review}\n}" }, "abstract": { "value": "With the development of artificial intelligence techniques, transportation system optimization is evolving from traditional methods relying on expert experience to simulation and learning-based decision and optimization methods.\nLearning-based optimization methods require extensive interactions with highly realistic microscopic traffic simulators.\nHowever, existing microscopic traffic simulators are inefficient in large-scale scenarios and thus fail to support the adoption of these methods in large-scale transportation system optimization scenarios.\nIn addition, the optimization scenarios supported by existing simulators are limited, mainly focusing on the traffic signal control.\nTo address these challenges, we propose the first open-source GPU-accelerated large-scale microscopic simulator for transportation system simulation and optimization.\nThe simulator can iterate at 84.09Hz, which achieves 88.92 times computational acceleration in the large-scale scenario with 2,464,950 vehicles compared to the best baseline CityFlow.\nBesides, it achieves a more realistic average road speeds simulated on real datasets by adopting the IDM model as the car-following model and the randomized MOBIL model as the lane-changing model.\nBased on it, we implement a set of microscopic and macroscopic controllable objects and metrics provided by Python API to support typical transportation system optimization scenarios including traffic signal control, dynamic lane assignment within junctions, tidal lane control, congestion pricing, road planning, e.t.c.\nWe choose five representative transportation system optimization scenarios and benchmark classical rule-based algorithms, reinforcement learning algorithms, and black-box optimization algorithms in four cities.\nThese experiments effectively demonstrate the usability of the simulator for large-scale traffic system optimization.\nThe anonymous code of the simulator is available at https://anonymous.4open.science/r/moss-AF45 and the others are shown at Appendix A.\nIn addition, we build an open-registration web platform to support no-code trials." }, "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": [ "microscopic traffic simulator", "transportation system optimization", "GPU acceleration" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/5bc9baddd837ab181380a722a60ec56aff52a7bf.pdf" }, "presentation": null, "primary_area": { "value": "infrastructure, software libraries, hardware, systems, etc." }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "A GPU-accelerated Large-scale Simulator for Transportation System Optimization Benchmarking" }, "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": "- Providing a more detailed discussion on how the proposed hypotheses have also been investigated in prior works would improve the completeness of this work.\n- Why is it reasonable to expect that the experimental observations generalize across different model architectures?\n- Including error bars and variances of the reported numbers, given the randomness of the model's initialization and training, would allow for easier verification of the significance of the reported observations." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This work provides a detailed description of the most commonly used equivariant architectures for point-cloud processing.\n- Additionally, the clear definition of the hypotheses that the authors aim to investigate allows for the explicit design of experiments that can be used to test them, facilitating the reader's understanding of the conclusions drawn from the experimental observations.\n- The extensive experimental evaluation across diverse point-cloud tasks provides convincing evidence of the effect of difference levels of equivariance on the proposed Rapidash model, showing the influence of the dataset size and task complexity on these effects." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work investigates the utility of imposing equivariant constraints in models that perform point cloud processing. To study the effects of the application of different equivariant constraints, the authors propose a scalable architectural desing that allows for easily incorporating varying degrees of equivariance. They evaluate this architecture on different point cloud processing tasks, incorporating different levels of equivariance or symmetry-breaking factors that break the network's symmetry constraint. Experimental results are used to accept or reject a set of hypotheses regarding the effects of equivariant constraints on model performance and generalization. Specifically, this work analyzes the effects of equivariance when modifying the size of the model, the size of the dataset, the complexity of the task, or when allowing symmetry breaking by providing pose dependent information (that is not available in the typical equivariance case)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- This work lacks sufficient attribution to prior work that has investigated hypotheses similar to the ones tested here. For example, hypothesis 5, on the effect of symmetry breaking in equivariant neural networks, has already been studied in prior works, such as :\n\n\t[1] Marc Finzi, Gregory Benton, Andrew Gordon Wilson \"Residual Pathway Priors for Soft Equivariance Constraints\"\n\n\t[2] Mircea Petrache, Shubhendu Trivedi, \"Approximation-Generalization Trade-offs under (Approximate) Group Equivariance\"\n\n\t[3] Stefanos Pertigkiozoglou, Evangelos Chatzipantazis, Shubhendu Trivedi, Kostas Daniilidis , \"Improving Equivariant Model Training via Constraint Relaxation\"\n- The experimental evaluation is limited to examining the effects of the equivariant constraints on a single architecture. While it is shown that these effects generalize across tasks it is not clear how they generalize across different model architectures. \n- In the experimental results, no error bars are provided. As a result, it is hard to assess the significance of the observations that are used to accept or reject the various hypotheses." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "see Weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Novel Research Problem: The paper presents several hypotheses regarding the impact of equivariant vs. non-equivariant architectures, particularly on tasks of varying complexity. This is a novel and relevant problem, providing new insights into geometric deep learning for point cloud data.\nComprehensive Experimental Design: The authors conduct extensive experiments across multiple datasets, thoroughly evaluating the performance of different architectures, thus providing strong support for the effectiveness of equivariance.\nModular and Scalable Architecture: The proposed Rapidash architecture is extensible, supporting various forms of equivariance, and provides an efficient platform for testing, which is beneficial for future research.\nClear and Substantiated Conclusions: The experiments confirm the advantages of equivariance in geometrically complex tasks and quantify the impact on different data scales, offering practical guidance for model selection and design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Summary\nThis paper examines the role of equivariance and symmetry breaking in deep learning architectures for point clouds, focusing on the influence of additional input information and SE(3) equivariance on model performance. Through a series of experiments on various tasks and datasets (e.g., Shapenet 3D, QM9, and CMU Motion Capture), the authors compare the effects of equivariant and non-equivariant layers, exploring the advantages of equivariance as task complexity increases. Results show that equivariance offers significant benefits for small datasets and geometrically complex tasks, though this advantage diminishes in large-scale data regimes.\n\nStrengths\n\nWeaknesses" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Limited Discussion on Symmetry Breaking: Although the paper addresses symmetry breaking by incorporating global coordinates, it lacks a detailed analysis of how this affects performance across different tasks. Additional theoretical insights or quantitative results would strengthen this discussion.\nInsufficient Evaluation of Model Complexity and Computation Costs: While Rapidash shows promise for high-complexity tasks, the paper provides limited discussion on computational costs and scalability in practical applications, which may impact its usability for large-scale deployment.\nLimited Benchmark Comparisons: While some baseline methods are included, the paper does not cover all recent benchmarks. Broader comparisons with state-of-the-art equivariant and non-equivariant methods would enhance the persuasiveness of the findings.\nLack of Practical Application Discussion: Although theoretically relevant, the paper does not discuss the potential impact of equivariance and symmetry breaking on real-world point cloud applications (e.g., 3D reconstruction or object detection), potentially limiting its practical 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": 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. Some pilot studies (no need too much) should be conducted on real-world datasets, or at least some robustness point cloud datasets. For example, ScanObjectNN for point cloud classification, or ModelNet40-C for robustness point cloud recognition.\n2. Some \"solutions\" should also be provided and evaluated (some pilot studies are enough).\n\nIf these two questions can be addressed well, the reviewer may consider raising the rating." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Clear motivation and insightful key idea. \n2. The paper writing quality is very high, with explicit statements and clear logic links. \n3. The code is provided in the supplementary material. \n4. The analysis and theoretical conclusion may bring future insight to 3D point cloud learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the key properties of deep learning networks on 3D point clouds, especially focusing on SE(3) equivariance. It conduct extensive experiments to test the initial hypothesis of the trade-offs between flexibility and weight-sharing introduced by equivariant layers. Based on the experimental results, a scalable network called Rapidash is introduced to facilitate the comprehensive testing of the hypothesis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The major concern lies in the dataset. ShapeNet is a synthetic dataset. However, one important key challenge of modern 3D point cloud networks is their real-world performance on real-world data. Thus, whether the raised hypothesis can also be accepted in real-world 3D data needs to be explored. \n2. Although this paper verified the three important hypotheses, a **solution** derived from hypotheses such as some network design ideas or even *engineering tricks/techniques* (e.g., how to revise the current point convolution operation) should be further discussed and evaluated based on the extensive analysis of the paper, which will benefit the network design of future point cloud learning methods. \n3. The reviewer believes that some extra figure illustrations should be provided to better intuitively understand the proposed theoretical hypotheses.\n\n*Conclusion*: \nThe current version of the paper is theoretically important, but need more pilot studies to provide the **practical** values." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "None" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Detailed introduction to previous work.\n2. The relationship between task complexity and equivariance was studied, and it was found that the equivariant method has more obvious advantages in tasks that require strict equivariance. It is demonstrated that providing explicit geometric information can improve performance even in the case of symmetry breaking." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the impact of additional input information and SE(3) equivariance on the performance of models processing point cloud data. They present a series of hypotheses to study different aspects of equivariant neural networks and symmetry breaking. Extensive experiments have been conducted on segmentation, regression, and generation tasks to verify the applicability and superiority of equivariant networks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.There are so many contents from previous works that I can hardly tell the novelty. From my point of view, only Formulas 8 and 9 are new, but it is still quite easy to prove.\n\n2.Although the paper proposes that symmetry breaking may be beneficial, it provides insufficient explanation of the mechanism behind it." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024on,\ntitle={On the utility of Equivariance and Symmetry Breaking in Deep learning architectures on point clouds},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yr7PjzmkQ6},\nnote={under review}\n}" }, "abstract": { "value": "This paper explores key factors influencing the performance of models working with 3D point clouds, focusing on the impact of additional input information and $SE(3)$ equivariance. It is often argued that providing more information as input improves a model's performance. However, if this additional information breaks certain properties, such as $SE(3)$ equivariance, does it remain beneficial? This work explores the trade-offs between flexibility and weight-sharing introduced by equivariant layers, assessing when equivariance boosts or detracts from performance. We identify the key aspects of equivariant and non-equivariant architectures that drive success in different tasks by benchmarking them on segmentation, regression, and generation tasks across multiple datasets with increasing complexity. We observe a positive impact of equivariance, which becomes more pronounced with increasing task complexity, even when strict equivariance is not required." }, "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 architectures", "geometric deep learning", "equivariance", "group convolutional networks", "generative modeling" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1ea67ea0a3e0e44ef2c93150c65d8a73ba8ec60a.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/474f83412866e773da1d411b17c05d1f9650440a.zip" }, "title": { "value": "On the utility of Equivariance and Symmetry Breaking in Deep learning architectures on point clouds" }, "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": "- Is there some relation that can be drawn between this phenomenon and policy churn [1]? Perhaps this provides a different avenue to motivate this work.\n- Do sections 4.1 and 4.2 make different points? They have different titles (randomized vs misaligned) but the claims seem the same?\n\n[1] Schaul, Tom, et al. \"The phenomenon of policy churn.\" Advances in Neural Information Processing Systems 35 (2022): 2537-2549." }, "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 fairly clear - the existence proof is very detailed and easy to follow. Certain areas like Figure captions need some improvement\n- The authors provide an original contribution - the misalignment of sub-optimal Q-values in adversarial training has not been observed before\n- Experiments are convincing, at least of the existence of mis-aligned Q-values with adversarial training. The construction of the experiments (i.e. using the second best action or worst action) is intuitive and well-explained/motivated" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors provide a theoretical and experimental of analysis of certified robust training of Q-values for reinforcement learning. Certified robost training is a method for adversarially training neural networks such that they are robust to small perturbations in the input values. In the case of reinforcement learning in particular, this is implemented as a regularizer added to the standard temporal difference loss, where the regularizer penalizes Q functions for which a perturbation in the state $s$ can result in a change in the action that produces the _maximum_ Q-value. The authors provide an existence proof that this style of training can produce misalignment amongst the sub-optimal Q-values, which they claim is a departure from natural intelligence which is able to properly order counterfactual actions. They demonstrate this phenomenon experimentally in several games in the Arcade learning environment, by showing that the performance drop incurred when selected the second best action some percentage of the time, instead of the optimal action, is much higher for adversarially trained RL than for vanilla RL. Additionally, they show that selecting the worst action some percentage of the time leads to a larger performance drop for vanilla RL than adversarial RL, again indicating that vanilla RL produces a better ordering over sub-optimal Q-values." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major Issues:\n- Overall, the main claim of the paper could be better motivated with a better argument for why it matters if sub-optimal Q-values are misaligned, particularly for a method that is specifically designed to be robust to changes in the _optimal_ Q-values. The authors mainly motivate the importance of this by the divergence from natural intelligence however (a) it is not entirely clear why a departure from natural intelligence is necessarily a bad thing and (2) the claims that this is a departure from natural intelligence don't seem to be fully supported (see next point). The author's suggest that misaligned sub-optimal Q-values present a vulnerability which could perhaps be a good motivation - could you provide specific examples of how misaligned sub-optimal Q-values could be problematic or exploited in practical scenarios?\n\n- The paper specifically makes the claim that adversarially trained Q-values are not well-aligned with natural intelligence compared to vanilla RL Q-values. However, this claim seems too strong for the results that are actually presented in the paper. The author’s argument for this claim seems to be that previous work in neuroscience i.e. Fig 1 demonstrate that humans _can_ assign correct ordering to counterfactual actions in a particular decision making task. But do humans _always_ assign correct ordering? Are there any limitations to this ability? The authors demonstrate theoretically and empirically that there exists cases where adversarial training produces misaligned Q-values. However, it seems like to make a claim about natural intelligence alignment, the authors would have to actually test natural intelligence on the same tasks, particularly since the proof is an existence proof. Admittedly, I am not too familiar with the neuroscience literature so if the authors could provide more comprehensive evidence from neuroscience literature to support their claim, this would be helpful. Alternatively, I do not necessarily think the claim about alignment with natural intelligence is necessary, so the language could be toned down a bit. \n\n- The captions of several figures are not very informative and need to be guessed at by the reader. For example, in Fig 2 there is no description of what the bar chart is displaying, there are images from Atari games with no description of what the reader should be paying attention to and the brain scan similarly has no context. For the bar graph, it would be helpful to provide the environment details, details of how the Q-values were estimated and the source of the natural intelligence data. Similarly, Fig 3 has no mention of what each of the three panels is - the caption says Adversarial and Vanilla, but the superscript on all the x-axis Q-values is \"Adv\". Please explain what each of the three figures represent and make sure the axes are correct. The placement of Fig 3 is also odd, since it isn't mentioned in the text until Section 4.3 - I would either move it to that section or mention it in the text earlier. \n\nMinor issues:\n- The axes and text in the plots are much too small. Fig 6 is particularly bad.\n- It might be more intuitive to plot the performance drop as a negative value, so that the plots have the worse performing curve lower than the best performing curve\n- In the first paragraph of 4.1, $\\mathcal{P}_w$ is mentioned before it is defined\n- Table 2 is never mentioned in the text" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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 listed questions within the points above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "CLARITY and QUALITY:\n1. The paper is generally well-written, presents a clear problem, and clear derivations.\n2. It help the reader to follow the main argument logic step by step and presents clear mathematical statements of the conclusions derived.\n\nSIGNIFICANCE and ORIGINALITY:\n1. The comparison with natural intelligence seems particularly interesting and likely novel.\n2. The problem treated is important and has already received significant attention in other communities (e.g., computer vision), but arguably less in RL so far. Hence I believe it is an important direction of investigation. \n3. Regarding the significance of the derived results I have doubts expressed via the points/questions below. Especially points (1) and (3) regarding related works and formal implications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper approaches critically and formally recent advances in certified RL. It shows how these novel RL schemes aiming to learn robust policies, actually induce policies that are misaligned with natural intelligence. Moreover, they perform a rigorous theoretical analysis of such misalignment thus formally proving failure cases of such RL schemes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "CLARITY and QUALITY:\n1. At points writing is not particularly sharp (e.g., the list of contributions could be significantly shorter and clearer). Often, diverse and not fully clear words are used to describe things (e.g., 'resilience' in line 137 - what does it formally mean?) or the terminology lacks cohesion (e.g., the word certified is not introduced in the introduction, where the related research line is introduced with alternative terminology). \n2. Example in pag. 4 might be hard to follow. I’d suggest to add a drawing with states/actions etc.\n3. The paper does not seem cohesive in terms of storyline. In particular, it seems it has two themes: (1) showing the relationship between natural decision-making and RL with and without certified methods, and (2) showing formally the limits of certified methods. Although both are relevant topics, I fail to see clearly the connection between the two. Maybe the authors have a clear picture in mind connecting the two themes, but currently the alternation between them renders the paper not cohesive and feels like reading two papers forced into one via unclear motivations and connections. This is related with point (2) below.\n\n\n\nORIGINALITY / SIGNIFICANCE / GENERAL:\n1. Related works mention adversarial training schemes in general (not specific to RL), then present recent works specific for RL with positive results and finally some with negative non-theoretical results towards claiming that this work is the first showing formally the limitations of these RL schemes. What I am missing is the following: since these algorithmic ideas are far more general than RL as they boil down to regularized optimization schemes (that have been for instance explored vastly in computer vision, NNs etc.), aren’t there works showcasing their fundamental limitations (formally) in general that can therefore trivially hold also for RL, where NNs are used to represent the policy?\n2. Even within the abstract the authors state ‘’This intrinsic gap between natural intelligence and the restrictions induced by certified training on the capabilities of artificial intelligence further demonstrates the need to rethink the approach…”. Why? Human/animals and machines clearly have different design spaces (i.e. have different limits and capabilities) so there is no reason to believe that machine intelligence has to be bound to schemes emerged in human/animal intelligence, which is arguably very limited and/or peculiar in terms of resources (e.g., sensors to minimize statistical complexity, and compute machinery). It seems to me that to evaluate (and compare) intelligence one should define measurable objectives rather than expecting certain specific behavior. \n3. I am having a hard time understanding the main message from Theorems 3.6 and 3.7. These seem portrayed as negative results as they show value function reordering for sub-optimal actions, but to the best of my understanding they seem also to claim that the value function properly identifies the optimal action, which seems to me what matters for achieving provably optimal policies in RL schemes (e.g., Q-learning). What am I missing (formally)?\n4. What is the point of plotting the defined Performance Drop rather than a classic RL performance measure? I understand why the Performance Drop would show results aligned with your thesis, but, as mentioned in the previous point, it seems not aligned with classic RL schemes optimality measures." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Minor comments \n1. Line 213 - typo\n2. I am bit confused about the depiction of the brain and video games in Figure 2. Why are they there?\n3. Def 4.2. The formulation is slightly confusing \"For any $\\tau >0$...\" in the begining of the sentence typically implies that the condition must be valid for all $\\tau > 0$ for the definition to hold. I think here the authors mean \"for some $\\tau >0$\", which is better to place closer to the end of the sentence." }, "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 contains a theoretical foundation for the analysis\n2. The paper presents an experimental analysis of robustness and natural intelligence." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "JuxtAlign studies alignment of certified RL using ideas of natural intelligence from neuroscience. The authors imply orthogonality of natural intelligence and adversarial training used often in robust RL. The paper contains a theoretical and an experimental analysis of the phenomena" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presentation can be improved as it was hard for me to fully assess the results. Perhaps I missed something, but I didn't get the main message.\n2. It’s well known that robustness, adversarial training are blunt tools in the sense that they try to avoid all possible outcomes as designed. For example, we can take a worst-case action within a ball of possible actions. If we re-design these training tools with a more constrained scope then the results will be different. Therefore, I don’t quite understand how we can make general conclusions on certified training vs natural intelligence from this study.\n3. In extreme situations, robust training can result in trivial policies - do not do anything. Therefore, I don’t find it surprising that sometimes robust training is different from what we perceive as naturally intelligent. \n4. The theory seems to be based on the Danskin theorem, which can be applied only to C1 Q-functions. I believe this is a quite restrictive setting. In min-max problems it’s even more restrictive. \n3. Overestimation in Q-functions is a general problem in RL, not just in a robust setting. But I agree that in robust settings it will be more acute. However, using overestimation as a reason for misalignment in this context is not entirely correct and can be misleading." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024juxtalign,\ntitle={JuxtAlign: A Foundational Analysis on Alignment of Certified Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yrf5RmaHfG},\nnote={under review}\n}" }, "abstract": { "value": "Sequential decision making in highly complex MDPs with high-dimensional observations and state dynamics became possible with the progress achieved in deep reinforcement learning research. At the same time, deep neural policies have been observed to be highly unstable with respect to the minor sensitivities in their state space induced by non-robust directions. To alleviate these volatilities a line of work suggested techniques to cope with this problem via explicitly regularizing the temporal difference loss for the worst-case sensitivity. \nIn this paper we provide theoretical foundations on the failure instances of the approaches proposed to overcome instabilities of the deep neural policy manifolds. Our comprehensive analysis reveals that certified reinforcement learning learns misaligned values. Our empirical analysis in the Arcade Learning Environment further demonstrates that the state-of-the-art certified policies learn inconsistent and overestimated value functions compared to standard training techniques. In connection to this analysis, we highlight the intrinsic gap between how natural intelligence understands and interacts with an environment in contrast to policies learnt via certified training. This intrinsic gap between natural intelligence and the restrictions induced by certified training on the capabilities of artificial intelligence further demonstrates the need to rethink the approach in establishing reliable and aligned deep reinforcement learning policies." }, "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": [ "alignment", "juxtaposition", "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/895db473dac915fcfec8f1d70e9e82a31ed8830c.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/d85c76a8393cd93de01cd53609655bdea9aaacac.zip" }, "title": { "value": "JuxtAlign: A Foundational Analysis on Alignment of Certified 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": 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": "Questions:\n\n1. The authors need to provide a detailed motivation&discussion on the use of vector quantization (VQ) in the framework. Based on the manuscript, I feel like a lot of things are missing regarding this part. For example, how to learn the prototype matrix $Z_a$ during training? \n\n\n2. During inference on downstream tasks, the final node representations are generated as a weighted sum of the prototype matrix learnt from the alignment stage. Could the authors explain the rationale behind this design? I understand that the authors provide a subsection 'Effect Visualization of Annotation Prototype.', yet I'm still quite confused. \n\n\n3. Does the alignment stage need to be performed for any given graph, or can it be pretrained on some datasets and directly apply to unseen graphs?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Pros:\n\n1. The combination of LLMs and TAGs are important topics for graph research which establishes new benchmarks and poses new challenges.\n\n2. I like the idea of prompting LLMs for broader knowledge and the rationale behind the observed graph, e.g., edge formation.\n\n2. The experiments were comprehensive and covered most important baselines for the two tasks evaluated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Summary\n\nThis paper studies how to leverage the textual information for improved performance for node classification and link prediction. It inherits the idea from existing works that utilize LLMs as data enhancers by prompting the LLMs for explainations and broader knowledge, with a special focus on reducing the costs of both training and inference time. The experimental results demonstrated the effectivenss of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The presentation needs to be improved. Several critical designs of the framework lacks motivation. Please see the Questions part for details." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "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 performance of GAGA vary with different values of the hyper-parameter 𝛼 in the loss function? Are there any guidelines for tuning it?\n\n2. How does GAGA choose which nodes (edges) to be annotated. Is there any consideration about it?\n\n3. While large language models (LLMs) are crucial to the proposed method, the experiments are solely conducted on GPT-3.5. I am curious about the method's compatibility with other open-source LLMs, such as Llama 3.2 or Qwen 2.5. Could you clarify which specific characteristics of LLMs are most impactful for GAGA? Additionally, how might one select an appropriate or potentially more advanced LLM to further enhance GAGA's performance?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Efficiency and Scalability**: Reducing both time and cost associated with LLM-based enhancements for TAG, making it scalable for large datasets.The experiment on the correlation between the amount of labeled data and performance in the ablation paper is also very interesting.\n\n2. **Two-Level Alignment Module**: The innovative two-level alignment module, which integrates annotations with TAG structure, allows GAGA to achieve strong generalization with limited annotations.\n\n3. **Experimental Validation**: Comprehensive experiments across six datasets validate GAGA’s efficiency and accuracy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces GAGA, a framework for enhancing Text-Attributed Graphs by incorporating annotations efficiently. GAGA addresses this by annotating only a small set of representative nodes and edges based on information density, which significantly reduces time and cost compared with the traditional method. A two-level alignment module then integrates these annotations into the TAG structure, facilitating high-quality graph representation learning. Experiments demonstrate that GAGA achieves comparable or superior classification accuracies with little data annotated, proving its efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "See Questions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "For the two-level contrastive learning, does it introduce additional training of the language models?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Experiments on both node classification and link prediction taks show the effectiveness of the proposed methods.\n\nThe research topic is promising in graph representation learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a lightweight framework for representation learning on text-attributed graphs. The proposed model only annotate representative nodes and edges, and introduces a two-level alignment module for structure alignment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The novelty of this paper is incremental. Specifically, regarding low-cost annotation, the approach of annotating only representative nodes via LLMs is not new. Previous work [1] has already applied this method for TAG representation learning. For model explanations, the idea of considering the annotation given by LLMs as an explanation (line 169) is similarly well-explored in [2].\n\nThe writing and presentation need refinement. For instance, the Introduction is overly lengthy and should be shortened. Figure 2 is difficult to interpret.\n\nThe usage and description of mathematical symbols are chaotic; for example, see line 340.\n\nIt is unclear how Figure 3 demonstrates the effect of annotation prototype projection.\n\n[1] Label-free Node Classification on Graphs with Large Language Models (LLMS), ICLR 24\n\n[2] Harnessing explanations: Llm-to-lm interpreter for enhanced text-attributed graph representation learning, ICLR 24" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See 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 explores a new aspect by examining the issues related to large language models from both the perspectives of time expenditure and cost." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the challenges faced by traditional graph neural networks (GNNs) when working with Text-attributed Graphs (TAGs), which contain rich textual information. While recent methods leveraging large language models (LLMs) have improved node representations, they often necessitate extensive annotations and fine-tuning, which can be costly and time-consuming. This paper propose a lightweight framework for TAG representation learning model which innovatively annotates only representative nodes and edges, significantly reducing annotation time and costs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation for the model is not strong enough. The article points out that an important limitation of other methods is the requirement to enhance the attributes of all nodes. This assumption is not necessarily reasonable; enhancing all nodes is not mandatory, and enhancing only a subset of nodes is also a viable option. While the impact of this approach on model performance does require further experimentation, I do not believe it constitutes a limitation of the other methods.\n2. The code is not publicly available.\n3. From the experimental results (Table 2), the improvement of this method is quite limited and not very significant.\n4. The core contribution claimed by this article is the selection of representative nodes for labeling, aimed at reducing time and cost while maintaining effectiveness. Therefore, I focused on this aspect. However, I regret to say that the methods employed in the article do not convince me. The first assumption in the paper equates information density with node density, positing that areas with denser nodes have higher information density, suggesting that representative nodes should originate from such areas. In my view, nodes in low-density areas can also result in significant information loss, and defining high and low density is highly dependent on later parameter tuning. Secondly, the article employs k-means to obtain central nodes as a metric for selecting representative nodes. The algorithmic complexity of k-means is very high, and I do not believe this method can be particularly efficient, especially when dealing with large datasets. Overall, while the motivation behind this idea is appealing, the specific implementation fails to satisfy me." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024lowcost,\ntitle={Low-cost Enhancer for Text Attributed Graph Learning via Graph Alignment},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yrnrvfXFaV},\nnote={under review}\n}" }, "abstract": { "value": "Many graphs can be represented as Text-attributed Graphs (TAGs). Due to the rich textual information present in each node of TAGs, traditional graph neural networks (GNNs) often struggle to deliver satisfactory performance. Recent advancements leveraging large language models (LLMs) to augment new node text features have notably enhanced node representations, resulting in significant performance improvements. However, these methods typically require extensive annotations or fine-tuning on all nodes, which are both time-consuming and expensive. To address this challenge, we propose GAGA, a novel and lightweight framework for TAG representation learning. GAGA employs a more efficient strategy by annotating only representative nodes and edges, thereby reducing both annotation time and cost. It further capitalizes on these annotations by constructing an annotation graph that captures the topological relationships among them. Additionally, GAGA introduces a two-level alignment module to integrate the annotation graph with the TAG, ensuring effective alignment of their underlying structures. Experiments demonstrate that GAGA achieves classification accuracies comparable to or exceeding state-of-the-art methods while requiring only 1\\% of the data to be annotated, making it highly 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": [ "Text-attributed Graphs" ] }, "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/922a9c1dc30fab3d1c9b45d5ea7aa50c6401bd91.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": "Low-cost Enhancer for Text Attributed Graph Learning via Graph Alignment" }, "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": "- Is it possible to extend the results to continuous spaces without a significant increase in computational complexity? For example, could pure exploration in linear bandit settings be considered?\n- Could you provide more detailed explanations on the comparison between UE and NUE? (Please refer to the first point under weaknesses for a more detailed discussion.)\n- How can the parameter $T_j$ be set in practice to ensure the theoretical guarantees?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This paper offers a new perspective by formulating multi-distribution learning as a pure exploration problem in multi-armed bandits.\n- Based on this view, gap-dependent bounds are derived for both adaptive and non-adaptive cases for the multi-distribution learning problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the multi-distribution learning problem, where the learner aims to optimize the model's worst-case performance across a set of distributions. The main contribution is a reformulation of this problem as a pure exploration multi-armed bandit task and obtain simple regret bounds that depend on the sub-optimal gap of actions. The first part of the paper studies the non-adaptive case, where the learner cannot interact with the environments. Here, the authors provide simple regret bounds for both uniform exploration (UE) and non-uniform exploration (NUE). The second part explores the interactive case, where environment interaction is permitted, and proposes an LCB-based algorithm that better a lower simple regret than UE." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **On the Significance of the Results:** One of my main concerns is the significance of the results achieved in this paper, as they rely on strong assumptions, and their implications are not rigorously discussed. \n - **Assumptions:** The paper appears to address a simplified case where the action space $ \\mathcal{A} $ is discrete and finite, and the data space is restricted to 1-dimension in the analysis. In contrast, continuous decision sets are more commonly studied in the literature, such as Blum et al. (2017), Sagawa et al. (2020), and Soma et al. (2022). Although Section 5 discusses an extension to infinite decision sets, the proposed approach using an $ \\epsilon/k $-cover would result in a method with prohibitive computational costs.\n - **Results for the Non-Adaptive Case:** It is not entirely clear to me why the results for Non-Uniform Exploration (NUE) would be better than Uniform Exploration (UE), as the NUE outcomes depend on $\\min_Q{n_Q}$, which could potentially be very small. The arguments in Section 3.3 are too intuitive to me, with several approximations made that require further justification. For instance, I am confused by the statement \"considers a case $\\Delta_{DR}(a) \\approx B_n$\" in line 321. It is uncertain whether we can disregard the term $\\Delta_{DR}(a) - B_n$ in the comparison, as this term varies across arms, and the value of $B_n$ differs between UE and NUE cases.\n\n- **On the Proposed Method in Section 4:** The proposed method for adaptive cases requires knowledge of $H_j$, which depends on the suboptimal gap $\\Delta_{a,\\min}$ and is generally unknown in practice. Although the authors provide some discussion in Remark 4, it remains unclear how this issue would be addressed. Additionally, the setting of $\\epsilon_t$ is also confusing. While this quantity appears to only require to be lower bounded, it is still unclear how to set this value to ensure that Condition Eq. (1) is not violated.\n- **About literature review**: The discussion of the convergence rate for related work in lines 139-152 is inaccurate. Although Soma et al. (2022) claim a result of $O(\\frac{\\sqrt{B^2 + k}}{T})$, their analysis overlooks the non-oblivious property of the learning process, rendering the result invalid. This issue was identified by Zhang et al. (2023), and the currently best-known result remains $O(\\frac{\\sqrt{B^2 + k \\log k }}{T})$ in this line of research. One may check Section 2.3 in Zhang et al 2023 for a discussion." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Following up the first point in the weaknesses, Would the story of the paper rather be, given the knowledge that the uncertainty set $\\mathcal{U}$ is fixed, one can develop exponential rates that scale with an unknown but fixed sub-optimality gap of each arm? \n- The paragraph in the introduction states “The current literature is populated with distribution-independent rates”, but there were not any relevant literature cited in this paragraph.\n- How would the proof be adjust to accommodate the case where the learner interacts with the environment but the optimal arm is non-unique?\n- Typo on Line 223: a fixed number times -> a fixed number of times" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper uses techniques from pure-exploration bandits to develop instance-dependent simple regret rates, which serves as less-conservative complements to the current instance-agnostic MDL error bounds." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies multiple distribution learning (MDL) from bandit optimization point of view. By connecting pure-exploration setting in bandits to MDL, it develops instance-dependent sharp regret rates, thereby improving the current instance-agnostic rates in the MDL literature." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper claims one of its contribution being developing problem-dependent rates for MDL, because “Oftentimes, it is more intuitive to analyze the learner’s performance in a fixed setting, as opposed to considering a worst-case instance for each sample size. When domain knowledge is available, a “one-size-fits-all” rate does not provide any insight on how to take advantage of this information”. However, the upper bound rates developed in this paper depend on the knowledge of the unknown optimality gap; how would this be integrated into domain knowledge remains unclear.\n- The problem setting seems very similar to Kirschner et al for distributionally robust online contextual bandit problem, but no discussion is provided on the differences and connections." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "How does this work relates to active learning, where one would also take an adaptive strategy? What are the barriers that prevent active learning algorithm from being applied to the problem setting studied in this work? \n\nAre there relevant lower bounds available? If so, how do the upper bounds proven in this work compare to the lower bounds?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The authors clearly compared the finite sample bounds under uniform exploration against that under non-uniform exploration, and they highlighted where non-uniform exploration could have gains." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors study distribution-dependent guarantees in the multi-distribution learning framework. They prove that distribution-dependent bounds are tighter than distribution-independent bounds. Specifically, they derive finite sample bounds under uniform and non-uniform exploration and propose an algorithm that improves over non-adaptive counterparts." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors compared their proposed algorithm against uniform sampling, but not non-uniform sampling. Non-uniform sampling benefits from varied sampled sizes and would be a stronger baseline to compare against. \n\nIt would be nice to provide experimental results, even in very simple set ups, to showcase the strength of their proposed algorithm. The main results of this work are in theoretical results and there are substantial theoretical contributions, and thus I understand the experimental results may not be necessary." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "There are no ethical concerns." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- What is the definition of $a^\\star$ in lines 199-203.\n\n- What is $l$ appearing in the RHS of Eq in lines 263-266?\n\n- Is $M$ a known parameter or an unknown parameter to the agent?\n\n- In lines 320-321, why $\\Delta_{\\text{DR}}(a) \\approx B_n$ induces the comparison to be $\\tfrac{M^2}{n}$ v.s. $\\sigma^2_T + \\Sigma^2_T + V_T$? Both exponential terms should be $1$ when the exponential factor becomes $0$." }, "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 clearly conveys the problem setting and the conclusion to the reader. This work provides a distribution-dependent bound with analysis, which has not been reported in the literature. The distribution-dependent bound enjoys an exponential decay which can be compared to the probability of identification failure in the Best-arm Identification. Furthermore, this paper does not limit itself to non-adaptive exploring but extends to an adaptive exploring strategy, which is the UCB-E algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents innovative strategies within the framework of Multi-Distribution Learning (MDL), with the primary objective of identifying the best-performed distribution. It is informed by principles from Distributionally Robust Optimization (DRO) and multi-armed bandit theory, proposing both non-adaptive and adaptive methodologies. The non-adaptive techniques, namely Uniform Exploration (UE) and Non-Uniform Exploration (NUE), yield both distribution-independent and distribution-dependent bounds. Furthermore, the paper introduces an adaptive method in the interactive environment, LCB-DR, which further optimizes performance by employing optimistic sampling strategies analogous to the Upper Confidence Bound for Exploration (UCB-E) utilized in multi-armed bandit scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Since the author mentions that MDL draws inspiration from multi-armed bandits, I have found that identifying the best-performed distribution can be viewed as an analogy to identifying the best arm (BAI) in MAB. In lines 199-203, the author also mentions a connection between this work and BAI, which is a $H_a$ term; It would be better if the author could draw more comparisons between MDL and BAI. Since several works in BAI can achieve instance-independent bound \\cite{audibert2010best, chen2017towards}. How does the objective upper bound guarantee relate to the existing bound shown in the BAI literature? I believe it is important to show whether the given bound is tight when reducing the problem setting to the existing work." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We devise adaptive and non-adaptive algorithms for the multi-distribution learning problem and provide distribution-dependent guarantees using tools from empirical process theory and drawing inspiration from pure exploration multi-armed bandits." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024distributiondependent,\ntitle={Distribution-Dependent Rates for Multi-Distribution Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ys16t9FcLN},\nnote={under review}\n}" }, "abstract": { "value": "To address the needs of modeling uncertainty in sensitive machine learning applications, the setup of distributionally robust optimization (DRO) seeks good performance uniformly across a variety of tasks. The recent multi-distribution learning (MDL) framework \\cite{pmlr-v195-awasthi23a-open-prob} tackles this objective in a dynamic interaction with the environment, where the learner has sampling access to each target distribution. Drawing inspiration from the field of pure-exploration multi-armed bandits, we provide \\textit{distribution-dependent} guarantees in the MDL regime, that scale with suboptimality gaps and result in superior dependence on the sample size when compared to the existing distribution-independent analyses. We investigate two non-adaptive strategies, uniform and non-uniform exploration, and present non-asymptotic regret bounds using novel tools from empirical process theory. Furthermore, we devise an adaptive optimistic algorithm, LCB-DR, that showcases enhanced dependence on the gaps, mirroring the contrast between uniform and optimistic allocation in the multi-armed bandit literature." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "multi-distribution learning", "distributionally robust optimization", "pure exploration multi-armed bandits" ] }, "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/8904c6bdbb09ce8201985223b4ebad3f478ebfd4.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": "Distribution-Dependent Rates for Multi-Distribution 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": 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. Is the study focused on object detection or action recognition? Please clarify to avoid confusion.\n2. How do the proposed contributions specifically address label scarcity and tiny object detection?\n3. Could you specify the novel aspects of your two-stage methodology compared to existing approaches?\n4.The methodology in Section 4.2 appears basic. Could you provide more technical details, especially on handling tiny objects?\n5. Comparisons are mostly with older methods. Could you add more recent state-of-the-art baselines?\n6. Could you include qualitative visualizations, such as feature maps or example detections, to better demonstrate the model's performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The empirical study highlights the limitations of existing video classification methods, providing a solid foundation for the proposed approach. Additionally, the two-stage methodology enables the model to capture both spatial features of guns and temporal dependencies across frames, contributing to its effectiveness. \n2. Given the real-world importance of detecting firearms in video data, the paper has notable significance. By addressing limitations in current methods and proposing a targeted approach, it contributes meaningful insights and methods that could inform future research for high-stakes applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the challenging task of detecting tiny objects, specifically guns, in video data. Recognizing that current video analysis methods struggle with detecting guns due to the limited availability of labeled data and the complexity of identifying small objects in real-world scenarios, the authors propose a new method with three main contributions. They introduce a two-stage detection methodology that combines image-augmented training to improve spatial feature extraction with temporal modeling to capture sequential information. They validate the method on a synthetic firearms action recognition dataset and discuss key challenges and potential future research directions in tiny object detection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the abstract highlights the challenges of limited labels and tiny object detection, the contributions do not directly address these points. Establishing a clearer link between the identified challenges and the proposed solutions would strengthen the study’s coherence and impact.\n2. This paper does not clearly specify whether the focus is on object detection or action recognition, which may confuse readers regarding the study’s objective. A clearer explanation of the task would help define the scope and purpose of the proposed method.\n3. The contributions are presented in an abstract manner, lacking concrete results or specific findings to support them. Adding more details about the outcomes or unique aspects of the contributions would make the study’s significance more evident.\n4. The motivation section is overly lengthy and could be more concise. Streamlining this part would improve readability and make the core motivations clearer.\n5. Section 4.1 contains excessive background details, including algorithmic descriptions of existing methods, which may not be necessary. Providing only the essential comparative results would make this section more concise and focused.\n6. The proposed methodology in Section 4.2 is fairly basic, with limited technical depth, few formulas, and minimal emphasis on handling label scarcity or tiny object detection. Expanding this section to include more detailed techniques targeting these specific issues would add depth to the work.\n7. Table 4 presents results on a dataset where all methods achieve 100% accuracy, making it an unnecessary addition. Removing or replacing it with a more challenging dataset would make the evaluation more meaningful.\n8. The study does not compare against recent state-of-the-art methods and its variants from ablation studies, which limits the relevance and verification of the results. Including newer comparison methods would provide a stronger benchmark for the proposed approach.\n9. The paper lacks qualitative visualizations, such as feature maps, example detections, or interpretability-focused analyses. Adding these would provide deeper insights into the method’s strengths and its effectiveness in handling complex detection 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": 3 }, "primary_area": null, "questions": { "value": "The optimal model combination (MobileNet + Transformer) has high complexity and computational resource requirements. Is this feasible for resource-limited practical applications, such as embedded devices and mobile devices?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. I think gun detection is an important topic. However, the academia seems underrated this task and there are not many research on this direction. Gun detection has big potential to make our world safer. However, the exsiting methods have too many false positives to avoid its wide applications. Thus, I am glad to see that this work tries to contribute this important research direction. \n2. The paper studies a valuable gun detection problem and provides a detailed motivation explanation.\n3. The results show the improvements of the proposed methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the limitations of existing video classification methods in gun detection by proposing to combine image-augmented training and temporal modeling. Experimental results demonstrate performance improvements on both synthetic and real-world datasets, particularly in firearm action recognition and the UCF crime dataset. The paper also highlights future research directions, including dataset diversity, model complexity, and ethical considerations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The method proposed in the paper is a simple combination of existing techniques, such as common data augmentations like flipping and rotating.\n2. The application of data augmentations is one of the core ideas of this paper, but it lacks specific details, such as how to select and adjust data augmentations.\n3. The paper mentions using GRU and Transformer for temporal modeling, but it does not provide detailed descriptions of the configurations and optimization processes for these two models.\n4. The paper implements a combined model of VGG and Transformer, but does not elaborate on how these models are integrated. For example, does it use a simple concatenation or parallel approach?\n5. More importantly, gun detection typically requires a quick response in real-time monitoring systems, yet the paper does not address the model's real-time performance in practical applications." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. There are some missing details in the experiments. I did not find how the dataset is divided into training set and testing set. The authors should make it clear.\n2. What is the resolution of the real-world videos? I am curious whether the images contain sufficient features for guns.\n3. Is the training shown in Figure 2 in an end-to-end manner or in multiple stages?\n4. How to reduce the domain gap between gun images and real-world gun videos?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The method is clear and straightforward." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a gun detection method in real-world videos. First, they perform an empirical study of several existing video classification methods to identify the presence of guns in videos. Then, they use transfer learning to extract frame-level features followed by a sequential model for video classification. Third, they conduct experiments on the Firearm action recognition dataset and UCF Crime dataset to show the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper does not contain novel approaches. Typically, the authors only use transfer learning to extract image features and adopt LSTM / Transformers for video classification. The contribution is limited.\n2. As mentioned in the paper, the gun is usually a small object in real-world videos. However, frame-level feature extraction is not an efficient way to get fine-grained-level features of guns.\n3. In Table 4, it seems 100% accuracy has been achieved. I do not think it is a challenging task as the authors mentioned." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "It is not clear the relationship between the video activity recognition and gun detection from the videos in this paper. The authors should focus on the gun detection from the whole videos and propose to construct the gun detection dataset and benchmark." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) The research topic for gun detection from videos is interesting and valuable.\n\n2) The authors did an empirical study of existing video classification methods to detect the presence of guns in videos.\n\n3) The authors also investigate the challenges of real-world gun video detection." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to address an important and interesting gun detection tasks. The authors evaluate several video classification methods for detecting guns in videos and summarize that existing approaches cannot work well to handle the gun detection task. The authors propose a novel two-stage training methodology combining image-augmented training (introduce subtle gun features into the model) and temporal modeling (capture sequential dependencies), resulting in significant performance improvements on a synthetic firearms action recognition dataset. The analysis in this paper also emphasizes the need for advanced AI-driven gun detection methods in video data, highlights current challenges and limitations of existing techniques, and suggests directions for future research." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Very out-of-dated methods and comparisons. The authors should compare some recent video classification and detection algorithms.\n\n2) limited novelty, this paper just combines existing algorithms together, such as Video MAE and different network backbones (LSTM and Transformer). Very old backbones and baselines cannot illustrate the superior performance of the existing work.\n\n3) The motivation part is too redundant. The authors should shorten this part.\n\n4) It is not convinced to include Transfer Learning in this paper since this paper did not perform corresponding experiments." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024efficient,\ntitle={Efficient Gun Detection in Real-World Videos: Challenges and Solutions},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ys3eqxzkeN},\nnote={under review}\n}" }, "abstract": { "value": "Object detection in videos is a crucial task in the computer vision domain. Existing methods have explored different approaches to detect objects and classify the videos. However, detecting tiny objects (e.g., gun) in videos has always been a challenging and rigorous task. Moreover, the existing video analysis (detection and classification) models may not achieve high accuracy for gun detection in videos in real-world scenarios due to the lack of a large amount of labeled data. Thus, it is imperative to develop an efficient method to capture the features of tiny objects and train models that can perform accurate gun detection. To address this challenge, we make three contributions. First, we perform an empirical study of several existing video classification methods to identify the presence of guns in videos. Our extensive analysis shows that these methods may not achieve high accuracy in detecting guns in videos. Second, we propose a novel gun detection method with image-augmented training and evaluate the technique in real-world settings with different evaluation metrics. Third, our experimental results demonstrate that our proposed domain-specific method can achieve significant performance improvements in real-world settings compared to the other popular methods. We also discuss emerging challenges and critical aspects of detecting tiny objects, e.g., guns, using existing computer vision techniques, their limitations, and future research opportunities." }, "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-augmented training", "transfer 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/2436ea84ae9ee2dd5746c81e9dd68b5d684bc377.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": "Efficient Gun Detection in Real-World Videos: Challenges and Solutions" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Dear authors:\n\nGenerally, the paper's idea is doable and insightful, but the solution details can be of great improvement. \n\n I will certainly raise my point if my concerns below are mitigated:\n1. Have the authors considered extending their approach to 3D action spaces? What challenges do you anticipate in such an extension?\n2. The concern of \"Weak baseline and unstable improvement\" in Weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. A practical research question and reasonable insight\n\nThe research question proposed by the authors, i.e. \"LLMs lack the spatial awareness of real-world environments\", is exactly one of the major bottlenecks for LLM-based robotic planning, many works, including this paper, are trying to mitigate the mismatch between planning space in LLM and planning space in real world. Therefore, the insight of this paper: establishing an efficient communication interface between LLMs and robots is very natural and reasonable. \n \n2. Extensive experiments\n\nThe authors conducted comparisons between different LLMs(GPT-4, fine-tuned Llama, fine-tuned Mistral) with an additional exploration on open-vocabulary tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "## Research Question\nThough possess extensive world knowledge and demonstrate generalizationability, LLMs lack the spatial awareness of real-world environments. On one hand, the Robot can hardly convey spatial information from the environment to the LLM. On the other hand, the LLM struggles to efficiently communicate low-level decisions to the Robot. \n## Method\nTo this end, the author proposed Room to Chessboard (R2C) framework to establish a “common language” between LLM and the Robot. Such a framework will unlock LLM as low-level action planner and develop an explainable chain of thought decision analysis paradigm. More specifically, a task will first be translated into sub-goals as high-level planning. Then R2C will translate the task-aware environment information into a compact chessboard, therefore LLM can perform low-level planning on the chessboard by predicting the robot’s next position. Additionally, the author formalized a Chain of Thought Decision (CoT-D) task for LLM to enhance its overall spatial reasoning, and designed a fine-tuning paradigm corresponding to the CoT-D tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The problem setting is too simple.\n\n1.1 As the final goal is to let LLM do the low-level planning, a 2D grid world setting is far from practical or useful, low-level planning, e.g. robot arm manipulation, in a 3D world is far more complicated compared with a 2D grid world. \n\n1.2 Another concern is whether LLM outperforms traditional planning algorithms (e.g. Dijkstra, A star) in step-by-step navigation tasks, e.g. a simple baseline could be an LLM-specified target position + heuristic planner. \n\n2. Weak baseline and unstable improvement\n\nThe selection of baseline is questionable since 1)some heuristic planners (e.g. A*) are also applicable to such task(navigation on 2D map), could the authors compare their approach to heuristic planners like A* for the 2D navigation tasks? ; 2) Saycan is classic enough as a baseline, can the authors provide details on their implementation of SayCan, particularly how they handled the reinforcement learning component? 3) Specific to the LLM-based high-level planning, there are more works published with code in the 2023~2024 period(e.g. Auto-TAMP, code as policy, etc..), but this is just a minor point. 4) The performance improvement according to Table 2 is not stable enough.\n\n3. Insufficient impact on communities\n\nOne question for the paper's impact could be: \"How do the authors envision their approach evolving as language model capabilities improve?\" or \"Are there aspects of the R2C framework beyond prompt engineering that would remain relevant even with more advanced models?\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The authors can consider comparing R2C with non-LLM specialist models. They can also address the increased computational cost introduced by CoT-D fine-tuning for real-time deployment and find if there are ways to streamline or selectively apply CoT-D only when necessary. How adaptable R2C is to different types of interaction tasks beyond navigation might be questionable." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The R2C framework's chessboard representation provides a concise yet effective way for LLMs to interpret spatial information, bridging the communication gap between high-level instructions and low-level execution. The Chain-of-Thought Decision (CoT-D) fine-tuning paradigm adds explainability and improves spatial reasoning, allowing LLMs to handle complex decision-making in low-level action planning.\nThe framework demonstrates state-of-the-art results among LLM-based methods and competitive results compared to specialist models, showcasing its efficacy in both seen and unseen environments. Additionally, it extends beyond traditional benchmarks by enabling LLMs to perform well in open-ended tasks, which are closer to real-world scenarios" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel framework to enable LLMs to perform low-level action planning for robotic tasks. Traditional LLM applications in robotics focus on high-level task planning, while low-level actions rely on other specialized controllers. This paper addresses the challenge of communication between the LLM and the robot by mapping a room environment into a chessboard-style grid, called Room to Chessboard (R2C). This grid provides a simplified, yet semantically rich representation of the environment that LLMs can use to generate precise, step-by-step navigation instructions.\nTo further enhance the LLM's decision-making, the authors introduce a Chain-of-Thought Decision (CoT-D) fine-tuning paradigm that strengthens spatial reasoning and interpretability. Tested on the ALFRED benchmark, R2C outperforms other LLM-based methods and achieves competitive results with specialist models in complex, long-horizon tasks. The paper demonstrates that R2C can generalize across seen and unseen environments and handle open-vocabulary tasks, underscoring its versatility and scalability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the paper compares R2C with LLM-based baselines, a more detailed comparison with non-LLM approaches, such as reinforcement learning or specialized motion planners, would better position R2C within the robotics field. Experiments are limited to simulated environments (ALFRED benchmark), and it remains unclear how the R2C framework would perform in real-world settings with more complex and dynamic elements." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 general is the process of encoding the navigation task into a chess game? Is it confined to object centric navigation or does it extend to re-arrangement planning, task and motion planning problems (where the agent may need to push away objects to make space). For example, https://www.ijcai.org/proceedings/2018/0674.pdf and https://people.cs.rutgers.edu/~kb572/pubs/fast_object_rearrangement.pdf for examples of such tasks. Please indicate if the framework can or cannot solve such planning tasks. \n\n2. In case an action fails to execute, does the system re-plan?\n\n3. The authors mention related works such as LM-Nav. If feasible, a comparison with such an approach would be insightful. \n\n4. Is it feasible to extend your approach to incorporating robot actions that are not present in a chess game. For example, consider the \"jumping onto an object\" action for a quadruped robot since Chess does not allow pieces to be in the same cell. Similarly, if we have an object transport task where small objects are kept onto a tray and carried from location to location, would the system be able to perform such reasoning. If feasible, authors are requested to discuss specific modifications they might make to their framework to accommodate these types of actions, or to explain why such extensions might be challenging within their current approach." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper concerns an important problem of leveraging LLM planning capacity at lower-levels of granularity, in a sense bridging the gap between high-level goals and the low-level motor actions. \n\n- The central idea of relating the low-level task with a game-like representation with the aim of generalization is interesting. \n\n- The investigation into encoding the raw sensor input into the task setting is realistic and insightful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper attempts to perform tightly coupled high-level navigation and low-level navigation in an indoor setting using LLMs. The core idea is to express low-level navigation as a grid-world game of chess, allowing the LLM to leverage world knowledge about the game to provide generalised navigation capacity for the robot. The technical approach comprises of three key steps (i) converting the raw RGB-D image data into a grid-like representation for specifying the object centric navigation task, (ii) engineering the prompt to capture the game rules, action history and the current agent-environment state and (iii) COT-D framework that guides the LLM to reason in terms of key information, direction judgement, target prediction and selection analysis. Experimental evaluation is carried out on the Alfred data set." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The primary concern is how general and automated is the process of converting the navigation task into a game. The paper makes one ponder what class of navigation tasks can be encoded in which class of games. The paper provides some indicated results with the chess game but the generality of the result and the encoding process is not clear. The authors are encouraged to provide specific examples of different types of navigation tasks and discuss how (or if) they could be encoded using their Room to Chessboard framework. Further, the authors are requested to include a more detailed discussion of the limitations of their approach in terms of task types that may not be easily represented as a chess-like game." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 decision frequency of the R2C method? Does the agent decide only one grid ahead after a forward prediction from the LLMs? If that is the case, would it be rather slow in completing the entire task? Because the API-based planning approach (SayCan, Instruct2Act) only needs one LLMs' calling before finishing a long sequence of actions.\n\n(2) In this map-based 2D decision-making problem, what are the most apparent advantages of using the dense grid coordinates as the action space rather than using landmark-based action space with an off-the-shelf path-planning algorithm? \n\n(3) In Table 2, even if you change the perception module into the ground truth, the R2C still fails at the rate of around 60%. Can you provide more failure case analysis and discuss which potential aspects of the design can further improve the success rate?\n\n(4) How does the dataset scale influence the fine-tuning performance in the embodied instruction following the task? And can the fine-tuned 7b LLMs still be able to maintain the original text generation abilities?" }, "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 presents an end-to-end LLM-based framework that simultaneously generates both task-level planning subgoals and low-level planning actions. This is important for the embodied instruction following agents in order to reduce the system latency and compounding errors compared with hierarchical cascaded modular approaches.\n\n(2) This paper proposes a chain-of-thought fine-tuning method and shows that task-relevant knowledge can be effectively injected into the open-sourced LLMs (Mistral 7b, LLaMA 7b). The benchmark performance proves the fine-tuned open-sourced LLMs outperform the GPT-4 in the embodied instruction following task.\n\n(3) The proposed chessboard representation can help generalize to open-vocabulary tasks. Experiments show the R2C can outperform many baseline methods in the ALFRED benchmark." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the R2C, which projects the partially observed spatial and semantic information into a map and then uses a grid map (chessboard) to discretize the map as the unified representation for both high-level task planning and low-level action planning. This paper shows with the discrete chessboard as the environment representation, both the open-sourced large language models (Mistral 7b, LLaMA 7b) and the close-sourced large language models (GPT-4) can become efficient embodied instruction followers. To improve the decision accuracy of the LLM, this paper proposes a chain-of-thought fine-tuning framework, which asks the LLM first to answer some task-related questions and then predict the preferred action coordinates. The functions of both chessboard representation and the designed fine-tuning framework are effectively proved in the challenging ALFRED benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) Although the proposed approach is technically sound, the performance of R2C in the ALFRED is not satisfied. The listed baseline approaches are outdated. In the leaderboard of the ALFRED:https://leaderboard.allenai.org/alfred/submissions/public, there are many recent approaches that can achieve a better than 30% success rate in the unseen scenes, including both LLM-based approaches or not.\nPlease compare and report better approaches' performance metrics, such as EPA[1], Prompter[2], and ThinkBot[3].\n\n(2) Some important experiments and example analyses are missing. For example, as this paper presents a chain-of-though fine-tuning for 7B LLMs, what is the zero-shot performance for LLMs without any fine-tuning? Please report the frozen LLaMA-7b and Mistral-7b performance with the same chessboard-based prompt in the ALFRED benchmark. \n\n(3) The reference is not comprehensive, and most related works are from 2020 and 2023. Comparing with recent works can help highlight the contribution of this paper, for example, comparing with more topological-based planners, such as ConceptGraph[4], SayPlan[5], and VoroNav[6].\n\n(4) Some typo errors should be more carefully checked. For example, on page 9, line 475 (two commas), and page 4, line 199 (missing reference).\n\n(5) More example visualization should be added. For example, provide detailed input prompt and output answers of fine-tuned LLaMA 7B and Mistral 7B.\n\nReference:\n\n[1] Liu, X., Palacios, H. and Muise, C., 2023. Egocentric planning for scalable embodied task achievement. Advances in Neural Information Processing Systems, 36, pp.54586-54613.\n\n[2] Inoue, Y. and Ohashi, H., 2022. Prompter: Utilizing large language model prompting for a data efficient embodied instruction following. arXiv preprint arXiv:2211.03267.\n\n[3] Lu, G., Wang, Z., Liu, C., Lu, J. and Tang, Y., 2023. Thinkbot: Embodied instruction following with thought chain reasoning. arXiv preprint arXiv:2312.07062.\n\n[4] Gu, Qiao, et al. \"Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning.\" 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024.\n\n[5] Rana, Krishan, et al. \"Sayplan: Grounding large language models using 3d scene graphs for scalable task planning.\" CoRR (2023).\n\n[6] Wu, Pengying, et al. \"Voronav: Voronoi-based zero-shot object navigation with large language model.\" arXiv preprint arXiv:2401.02695 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024rc,\ntitle={R2C: Mapping Room to Chessboard to Unlock {LLM} As Low-Level Action Planner},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ysAX5ORQoX},\nnote={under review}\n}" }, "abstract": { "value": "This paper explores the potential of leveraging large language models (LLMs) as low-level action planners capable of executing long-horizon tasks based on natural language instructions. Although LLMs can act as the \"brain\" of robots by excelling in high-level task planning, they are not yet capable of directly guiding the \"body\" to execute low-level motion plans. This limitation stems from a communication gap between the \"brain\" and the \"body\". Specifically, LLMs lack access to rich spatial semantic information from the robot's real-time observations, hindering their ability to generate precise and actionable low-level plans.To address this, we propose a unified framework that bridges high-level and low-level planning by establishing an efficient communication interface between LLMs and robots. Our insight is to formulate the task as playing chess with LLMs. We map the room into a semantic chessboard, which we call Room to Chessboard (R2C). Each grid represents the position and size of objects inside the room. We find that chessboard is \\textbf{succinct} enough for LLMs to conduct semantic searches with global view of the room. Also, the chessboard is \\textbf{informative} enough to convey detailed environmental state for LLMs to predict executable low-level actions. Additionally, we enhance decision-making through a Chain-of-Thought (CoT) paradigm, improving LLMs' interpretability and action reasoning. We implement R2C using both fine-tuned open-source LLMs and closed-source models like GPT-4, and demonstrate its efficacy on the challenging ALFRED benchmark. Our results show that with communication based on chessboard, LLMs can serve as effective low-level action planners, and can generalizes well to open-vocabulary robotic planning tasks. View the demos on our project page: https://anonymous4cv.github.io/Room2Chessboard." }, "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": [ "Embodied AI", "Large Language Model", "Embodied Instruction Following", "Robotic Planning" ] }, "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/0ea71a8c595c708f9683e90c0106a698bedadd82.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": "R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner" }, "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": "1. How can the diversity of the generated environments be evaluated more accurately and quantitatively? Besides the current qualitative analysis, are there other more objective and comprehensive indicators to measure the differences and complexity between different environments to better prove the effectiveness of environment generation?\n2. How does the paper dynamically regulate the evolution speed of the environment to avoid being too fast for the agent to learn or too slow to cause resource waste?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The design of Autoverse combines a domain-specific language with cellular-automaton-like rewrite rules and achieves efficient computation through convolutions, providing new perspectives and methods for game environment generation and agent training.\n- Through the evolution of the environment, the complexity of the environment can be gradually increased according to the search ability of the agent, effectively avoiding the agent from prematurely falling into local optimal solutions and also providing a curriculum learning from simple to complex for the agent." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Autoverse, an evolvable domain-specific language for single-player 2D grid-based games, as a training ground for Open-Ended Learning (OEL) algorithms. It describes game mechanics through rewrite rules similar to cellular automata, combines evolutionary algorithms to generate complex environments, employs imitation learning and reinforcement learning to train agents, and experimentally studies the impact of the observation range on agent performance and the dynamic characteristics of the evolved environments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Overall, the paper has limitations in terms of innovation. The framework of combining imitation learning and reinforcement learning used is not novel and has been involved in many current related research fields.\n2. Although the rewrite rules are highly expressive, they are difficult to understand and interpret, which may limit their promotion and further development in practical applications, especially when manual intervention of the rules is required.\n3. In the process of environment evolution, only mutation operations are involved, and crossover operations are not included. This may limit the exploration range of the diversity of environment generation to some extent.\n4. The main experiment is lacking. There is a lack of direct comparison experiments with other advanced methods, making it difficult to accurately evaluate the advantages and disadvantages of Autoverse and clearly define its competitiveness in this field.\n5. The long-term training effect experiment is insufficient. The experimental results mainly present the immediate performance data under specific settings. The performance change trend after a significant increase in the training cycle, the stability of the agent's strategy after a large number of environmental changes, and the evolution law of the long-term adaptability to new environments are not provided, making it difficult to judge the long-term comprehensive ability development.\n6. The coherence of the chapter structure is poor in some parts. For example, the transition from the method description to the experimental results is not natural, affecting the reader's understanding of the logical relationship of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Being able to generate a vast number of environments is very interesting from an open-endedness perspective. However, many of the features that enable such versatility and efficiency limit Autoverse to grid-like, 2D, and seemingly small environments. Can methods developed in Autoverse be applied to more realistic scenarios (that are ultimately of interest)? Examples of these alternative (also open-ended) scenarios are Craftax (Matthews et al., 2024) or MineDojo (Fan et al., 2022). If not applicable, then what is the relevance of Autoverse?\n\n- Is the grid size constant in all of the generated environments? If not, would it be possible to modify the framework to hold grids of different sizes? \n\n- How does the search method (used for the \"warm-start\") scale with the size of the grid and number of actions? \n\n- What is (approximately) the ratio of unusable and usable environments generated by the evolutive algorithm? With unusable I refer to environments so unstable that are not beneficial for the optimization process of the RL agent. \n\n- Do the reward update rules always take into account the actions of the agent? or reward can be generated by multiple cells of the CA interacting with each other with no relation to the actions of the agent?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The idea of using CA-based environments is both original and interesting. The fact that practically limitless environments can be created by modifying the initial state and the update rules is very interesting from an open-ended learning perspective.\n\n- I also find CA environments relevant for research on foundation models for RL. This type of environment could be very valuable for generating training data for these types of models.\n\n- I think that the presentation of Autoverse (Section 2.1) is clear and can be easily understood, whereas Figure 1 also helps to visualize the types of environments that are generated by the evolution process." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel 2D and grid-like environment framework, named Autoverse, based on Cellular Automata (CA). They define environments using the initial state of the CA, and the update rules (including state transition rules and reward). This allows authors to evolve environments in an open-ended RL setting. Moreover, the authors also propose an open-ended RL method based on bootstrapping the open-ended learning by pre-training the agent using an initial set of evolved environments and expert trajectories from a search method. Experiments show that the evolution process generates many chaotic environments, stable environments, and others in between (the ones of most interest). Furthermore, the open-ended method seems to benefit from fully observing the environment and access to the update rules that update the cells of the CA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although I think that the main idea (evolvable CA environments) is very interesting and could be promising for areas as foundation models for RL, UED, and open-ended learning. I have important concerns on several aspects of the paper:\n\n- Although most ideas are clearly explained, the overall presentation and soundness of the paper are poor. \n + The introduction section makes many (non-trivial) statements with no references. The introduction has no references, which I found surprising. Some examples of such sentences missing references and evidence/support are:\n - L31: \"The idea of open-ended learning in virtual environments is [...]\"\n - L32: \"This idea comes in many forms, but what unites them all is that [...]\"\n - L38: \"There have been interesting results, but learning generally stops at a rather low capability ceiling.\"\n - L40: \"It has been observed that the complexity of the behavior of a living being, [...]\"\n\n- The ability to endlessly evolve and generate new environments is compelling, but as mentioned multiple times in the paper (e.g., L71, L413) evolutive process generates very unstable environments in most cases. I have serious concerns about the ratio of unusable and usable environments generated by the evolutive process. Authors claim scalability, but how much computational resources are needed to generate a fair amount of actually usable scenarios? I think that an exhaustive analysis of this topic is crucial and missing in the current version of the paper. \n\n- I think that the paper misses many experimentation details. For instance, experiments shown in Table1 and 2 are missing information on the number of evaluations, the number of generated test/train environments, the number of repetitions, etc. Seems that this information is also missing in the appendix. Moreover, none of the appendix sections are referenced in the main paper. Please, consider providing as many details as possible on the experimentation to ensure transparency and reproducibility. Moreover, I strongly believe that all sections of the appendix should be referenced at least once in the main paper.\n\n- I strongly believe that experimentation should be improved. The authors propose a \"warm-start\" method for open-ended RL, but they do not provide any evidence for why this method is relevant for the research community. I won't ask for a comparison with tens of methods from previous literature, but proposing a new method at least requires an exhaustive analysis of it. Some examples of how I think this could be improved:\n - Reward values in Tables 1 and 2 are missing interpretability. I can see that the results of some settings are better than others, but how much? Having some sort of reference value could help (e.g., results of a random agent or some well-known baselines).\n - Why is this method interesting? Does it obtain better results? Does warm-staring help to improve the results? Many experiments could be presented to answer these questions. \n\n- The paper claims computational efficiency but no evidence is provided. Experiments/benchmarks on computational performance are missing if such claims are made. \n\n- I think that the presentation has room for improvement. Some examples:\n - Figure 1 is presented is located in page 2, but is referenced on page 8 for the first time. \n - Figure 3 holds an entire page but some of its text (Fig 3.a) is **extremely** small. There is plenty of blank space on Page 3 to increase some parts of the figure or rearrange them to improve visibility.\n - The format of tables 1 and 2 is not aligned with the ICLR style. Please carefully read the style PDF before submitting the paper.\n - There are many incorrect usages of parenthesis in references. For example, parenthesis in Earle et al. 2023 L271 should be removed, and Section 4 is missing parentheses in most of its references: L470-471, L476, L482..." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Please clarify any confusing parts in the figures and tables.\n\n2. Please provide additional experiments to support each claim in the paper, such as scalability and comparisons with other environments.\n\n3. Please ensure that the figures and results are presented neatly and clearly." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Autoverse offers efficiency by leveraging GPU-based batch processing. Its rewrite rule framework enables the creation of a vast array of dynamic, grid-based game environments, enhancing agent adaptability and preventing overfitting to static setups. The method's progressive curriculum, which evolves increasingly complex environments, allows agents to improve incrementally, while the integration of imitation learning with RL provides agents with a solid starting foundation. These strengths make Autoverse an innovative and robust tool for advancing research in open-ended learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Autoverse, a special programming language for creating different types of grid-based game environments that help with open-ended learning in RL. Autoverse lets people build complex, changing game settings, especially for 2D, single-player games like mazes, Sokoban puzzles, and dungeon exploring. It uses simple rules, based on cellular automata, to quickly create these environments, which run well on a GPU. The system first teaches RL agents by having them imitate expert examples, then moves to open-ended RL in constantly changing environments to make the agents more adaptable." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "First, there are presentation issues, as figures (like 3 and 4) lack clarity—small text sizes, ambiguous elements, and layout issues reduce readability. Second, empirical evidence demonstrating scalability is lacking; additional experiments or metrics require to solidify this claims. The paper fail to do performance comparisons of Autoverse’s combined imitation and RL approach against using either method alone. While it states that Autoverse’s environments are more complex and diverse than others, this claim lacks citations or direct comparisons with other open-ended learning benchmarks. Finally, the paper does not report the performance achieved after the behavior cloning stage, leaving its contribution to the final results ambiguous. Addressing these points could greatly enhance the paper’s clarity, rigor, and persuasive power." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "In general, I would really appreciate it if the authors could provide more details and discussions of the proposed method performance and the reasons why they design the whole method and environment in this way. And also, the comparison with other baseline or ablation study will be welcomed." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1, Originality: The paper presents Autoverse, a new environment for open-ended learning, which allows more complex environment dynamics and much more environmental diversity than other open-ended learning environments.\n2, Quality: The use of JAX for implementing cellular-automaton rules allows efficient parallelization on GPUs, and at least an order of magnitude speedup.\n3, Significance: The use of imitation learning followed by reinforcement learning provides a structured way for agents to learn from expert play traces and then further refine their behavior, leading to better generalization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Autoverse, a domain-specific language (DSL) for creating 2D grid-based games, designed to enhance open-ended learning for Reinforcement Learning (RL) agents. Autoverse allows for evolving complex game environments using cellular-automaton-like rewrite rules, which can be parallelized on the GPU to speed up RL training. The authors propose a framework involving the evolution of environments, imitation learning from search-based solutions, and reinforcement learning in evolving environments to generate increasingly challenging tasks for RL agents. This approach aims to overcome the challenges of cold-starting RL in open-ended environments and produce more behaviorally complex and adaptable agents. They first evolve Autoverse environments (their rules and initial map topology) to maximize the number of iterations required by greedy tree search to discover a new best solution, producing a curriculum of increasingly complex environments and playtraces. The proposed method then distill these expert playtraces into a neuralnetwork-based policy using imitation learning. Finally, the learned policy becomes as a starting point for open-ended RL, where new training environments are continually evolved to maximize the RL player agent’s value function error (a proxy for its regret, or the learnability of generated environments), finding that this approach improves the performance and generality of resultant player agents." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1, Presentation Improvement: Provide more detailed feedback, such as clarifying ambiguous elements like the meaning of the gray box in Figure 4 and addressing layout issues with full-page figures such as figure 3 and figure 4. And the texts in figure 3 are too small.\n2, Scalability Evidence: Request empirical evidence or additional explanations to demonstrate scalability, including experiments or metrics.\n3, Justification for Greedy Tree Search: Maybe adding references in related work and experimental validation to justify the choice of greedy tree search.\n4, Comparison of Methods: Recommend including a comparison between the proposed approach and using only imitation learning or reinforcement learning.\n5, Performance Results: The paper mentioned \"Once the behavior cloning algorithm has converged, we continue training the agent with reinforcement learning\", but there is no results show that what kind of performance can this method reach when behavior cloning algorithm has converged and what is the final performance compared to that stage.\n6, Lack explanation of comparing with other environments: In the paper \"Autoverse stands out for allowing more complex environment dynamics and much more environmental diversity than other open-ended learning environments.\" But I did not see any citation or detailed comparison with other open-ended learning environments. Further discussion will make the arguments more convincible." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a fast, evolvable game language, and do open-ended search in the space of game mechanics and levels to train generalist player agents" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024autoverse,\ntitle={Autoverse: an Evolvable Game Language for Learning Robust Embodied Agents},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ysQiaWhnCN},\nnote={under review}\n}" }, "abstract": { "value": "We introduce Autoverse, an evolvable, domain-specific language for single-player 2D grid-based games, and demonstrate its use as a scalable training ground for Open-Ended Learning (OEL) algorithms. Autoverse uses cellular-automaton-like rewrite rules to describe game mechanics, allowing it to express various game environments (e.g. mazes, dungeons, sokoban puzzles) that are popular testbeds for Reinforcement Learning (RL) agents. Each rewrite rule can be expressed as a series of simple convolutions, allowing for environments to be parallelized on the GPU, thereby drastically accelerating RL training. Using Autoverse, we propose jump-starting open-ended learning by imitation learning from search. In such an approach, we first evolve Autoverse environments (their rules and initial map topology) to maximize the number of iterations required by greedy tree search to discover a new best solution, producing a curriculum of increasingly complex environments and playtraces. We then distill these expert playtraces into a neural-network-based policy using imitation learning. Finally, we use the learned policy as a starting point for open-ended RL, where new training environments are continually evolved to maximize the RL player agent's value function error (a proxy for its regret, or the learnability of generated environments), finding that this approach improves the performance and generality of resultant player agents." }, "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": [ "open-ended learning", "reinforcement learning", "imitation learning", "evolution", "search" ] }, "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/8f3ada71489e4f794e139d0c14acce7ccc031f99.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/450f50e235f70445123423795e91165aac06839d.pdf" }, "title": { "value": "Autoverse: an Evolvable Game Language for Learning Robust Embodied Agents" }, "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 see Weaknesses section above for a list of all questions." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) This paper is written well and it is easy to follow.\n\n2) The proposed method obtains noticeable improvements (Table 1, Table 2) under IF and PLG-MI attacks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes a method for improving defense against generative Model Inversion Attacks (MIA), called Confidence Adaptation and Low-Rank Compression (CALoR). The proposed method demonstrates noticeable improvements over existing defense methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Given that MIAs search for high-likelihood samples for a particular class, use of $L_{CA}$ basically manipulates the region where such high-likelihood samples are present, e.g..: Now it’s present in a region where $p\\_c \\approx 0.8$ and the MIA is searching in the region corresponding to $p_c~=1$. What would happen if the MIAs were aware of this confidence training? I suspect that a substantial amount of performance could be recovered if the identity loss is adjusted to align with $L_{CA}$ during model inversion. Conducting this ablation study is crucial.\n\n2) **Using encoder features directly for MIA:** My understanding is that the proposed method focuses on white-box MIA attacks. What would happen if the adversary directly used the encoder features to perform MIA (by applying softmax directly to the encoder features) instead of the LORA features?\n\n3) **User studies are necessary to show the efficacy of MI defense.** Since this work focuses on defending against private data reconstruction, it is important to conduct user study to understand the improvements (See [A, B]).\n\n4) Why is the gap between TL and CALoR much smaller in the PLG/ImageNet setting compared to other setups (see Table 2)?\n\n5) Error bars/ Standard deviation for experiments are missing.\n\n6) It would be useful to indicate that the paper's focus is white-box MIAs.\n\n7) Missing related works [A]\n\nOverall I enjoyed reading this paper. But in my opinion, the weaknesses of this paper outweigh the strengths. But I’m willing to change my opinion based on the rebuttal.\n\n=========\n\n[A] Nguyen, Bao-Ngoc, et al. \"Label-only model inversion attacks via knowledge transfer.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[B] [MIRROR] An, Shengwei et al. MIRROR: Model Inversion for Deep Learning Network with High Fidelity. Proceedings of the 29th Network and Distributed System Security Symposium." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weakness" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The presentation is clear and well organised.\n\nThe empirical results are encouraging." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduce a novel MI defence, namely CALoR, that aims to improve MI robustness by revisiting three potential weakness aspects of MI attack: MI objective, MI overfitting, and MI optimisation.\n\nThe empirical results are encouraging through various MI setups." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While I appreciate the broad exploration of MI robustness in this paper, it feels more like a technical report rather than an in-depth investigation for each aspect. The proposed method consists of three modules: confidence adaptation, low-rank compression via a VAE-inspired framework, and Tanh activation. The concept of confidence modification has been somewhat investigated in LS defence, and the impact of vanishing gradients on MI is well-studied in PPA, LOMMA, and PLG-MI. These modules offer trivial and straightforward ways to mitigate MI attacks but lack novel insights or a significant contribution to MI.\n\n2. The concept of confidence modification is somewhat similar to LS defence. If I understand it correctly, modifying the confidence by (positive) LS also offers the similar concept. However, I am quite surprise that (positive) LS and CA has opposite effect to MI. Could the authors provide an explanation on this?\n\n3. While the experimental results are encouraging, the setups differ notably from those in existing studies. I suggest that the authors add additional experiments to strengthen the paper:\n- For the low-resolution scenario, include LOMMA (rather than IF) as it is a state-of-the-art MI attack alongside PLG.\n- For the high-resolution scenario, include PPA (rather than PLG) as it is a state-of-the-art MI attack alongside IF.\n- The evaluation model used in the high-resolution scenario also differs from existing works. I am curious if this may affect the experimental setups.\n\n4. For the ablation study, I recommend that the authors expand Table 3 to better demonstrate the effectiveness of each module in CALoR. The current results highlight the contribution of CA, but the roles of low-rank compression and Tanh activation remain unclear to me. To clarify, I suggest splitting the results for LoR into LoR with Tanh and LoR without Tanh.\n\n5. The paper should include a paragraph addressing the weaknesses of the proposed method. One potential weakness is its complexity relative to existing defenses like LS or TL. With three different modules, the proposed approach may be challenging to adapt to new setups. As shown in the paper, in high-resolution scenarios with ImageNet as the pre-training dataset, CALoR demonstrates more difficulty in adapting compared to LS or TL." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* To clarify the novelty of the proposed method, the authors should elaborate on the key differences between their approach and positive label smoothing [r2]. Without it, the contribution of the paper is limited.\n\n* Could the authors provide a more comprehensive comparison with other state-of-the-art defenses, particularly high-resolution techniques like PPA and MIRROR, on the official benchmark?\n\n* Could the authors include model accuracy metrics in Section 3.2 to better understand the trade-off between defense effectiveness and model utility?\n\n* How does the proposed method perform on larger datasets like VGGFace/VGGFace2/CASIA?" }, "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.\n\n* Extensive experiments on both low-resolution and high-resolution attacks demonstrate the effectiveness of CALoR in defending against State-of-the-art MI attacks.\n\n* Perform MI attacks using modern model architectures like ViT-B/16 and Swin-v2." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents Confidence Adaptation and Low-Rank compression (CALoR), a new defense strategy against model inversion attacks (MIAs). To mitigate MIAs, CALoR employs three key techniques: (1) confidence adaptation to reduce the confidence of private training samples, (2) low-rank compression to prevent attackers' ability from mitigating the overfitting, and (3) Tanh activation to induce gradient vanishing. Extensive experiments on both low-resolution and high-resolution attacks demonstrate the effectiveness of CALoR in defending against State-of-the-art MI attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The experimental setup raises some questions. While PPA/MIRROR is primarily designed for high-resolution images, the experiments using PPA/ MIRROR are performed on low-resolution images. Given that StyleGAN-ADA generates 256x256 images, it's unclear how the authors adapt it to low resolution images (64x64). Similarly, PLGMI, originally designed for low-resolution attacks, is adapted for high-resolution scenarios. More details on this adaptation process should be explained.\n\n* Section 3.2 demonstrates that reducing confidence in private training samples lowers attack accuracy. However, without reporting the model accuracy, it's difficult to determine if this reduction is solely due to a decrease in model utility. As existing work [r1] suggests a strong correlation between model accuracy and MI attack success, it's possible that any action diminishing model performance, including confidence reduction, could impact attack accuracy. \n\n* The idea of reducing confidence in private training samples is similar to using label smoothing which is addressed in LS defense [r2]. Note that LS[r2] shows that positive label smoothing does not have such strong effect on defending against MI attacks.\n\n[r1] Zhang, Yuheng, et al. \"The secret revealer: Generative model-inversion attacks against deep neural networks.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.\n[r2] Lukas Struppek, et al. \"Plug & play attacks: Towards robust and flexible model inversion attacks\". In ICML, 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. Why the hyperparameter a is needed in Eqn. (2)? I think it can be incorporated into the step size.\n\n2. How to properly set the hyperparameter b? Any theoretical justification?\n\n3. In Table 4, a rank of 20 also shows a good performance. Why highlight the rank of 30 in the main text?\n\n4. How to properly set the low rank across different datasets? Is a rank of 20/30 a safe value for various datasets?\n\n5. Is there any scenario where the proposed defense method fails? It would be great if the authors could find any case with deeper analysis." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The model inversion attacks are very important and the privacy concern of the (private) training data should be carefully handled.\n\n2. The overall empirical performance compared with other existing defense work is impressive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a defense method named CALoR against model inversion attacks. There are two key components of the proposed method. One is biasing the attack optimization target through confidence adaptation and the other is low-rank compression to mitigate the privacy leakage." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presentation of this work is extremely poor which makes it look like an incomplete paper, especially for the methodology part. For example, after reading Section 3.2, I still have no idea about how to ensure the utility of the classification model after executing the second stage. In Section 3.3, the authors talk about the latent vector z but without further discussing why z is relevant to their method. Lots of main details are missing/require the audience to refer to the appendix. I suggest the authors revise the manuscript significantly to make it more clear.\n\n2. The motivation of Eqn. (2) is not clear. Through the simple calculation of the derivative, we can see that the final objective is to ensure that \\hat{y}_c should converge to exp(-1/b). Why not directly use the MSE loss to mislead the attack objective? Additionally, the hyperparameter a in Eqn. (2) does not play much role.\n\n3. Low-rank compression is a common technique to cause information loss which can be further utilized to protect private input data. I would suggest the authors weaken their claim about the contribution to this part and shorten the relevant paragraphs.\n\n4. This work lacks theoretical analysis about the performance decrease bound due to confidence adaptation and low-rank compression which puts the utility of CALoR into question.\n\n5. I would suggest the authors report the uncertainty (e.g., standard deviation) of the experiment results as well." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We provide an in-depth analysis of intrinsic vulnerabilities of model inversion attacks, and propose a novel and comprehensive defense framework CALoR, including confidence adaptation and low-rank compression." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024calor,\ntitle={{CAL}oR: Towards Comprehensive Model Inversion Defense},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=ysZvK6b60c},\nnote={under review}\n}" }, "abstract": { "value": "Model Inversion Attacks (MIAs) aim at recovering privacy-sensitive training data from the knowledge encoded in the released machine learning models. Recent advances in the MIA field have significantly enhanced the attack performance under multiple scenarios, posing serious privacy risks of Deep Neural Networks (DNNs). However, the development of defense strategies against MIAs is relatively backward to resist the latest MIAs and existing defenses fail to achieve further trade-off between model utility and model robustness. In this paper, we provide an in-depth analysis from the perspective of intrinsic vulnerabilities of MIAs, comprehensively uncovering the weaknesses inherent in the basic pipeline, which are partially investigated in the previous defenses. Building upon these new insights, we propose a robust defense mechanism, integrating ***C**onfidence **A**daptation* and ***Lo**w-**R**ank compression*(**CALoR**). Our method includes a novel robustness-enhanced classification loss specially-designed for model inversion defenses and reveals the extraordinary effectiveness of compressing the classification header. With CALoR, we can mislead the optimization objective, reduce the leaked information and impede the backpropagation of MIAs, thus mitigating the risk of privacy leakage. Extensive experimental results demonstrate that our method achieves state-of-the-art (SOTA) defense performance against MIAs and exhibits superior generalization to existing defenses across various scenarios." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Privacy Leakage", "Model Inversion", "Defense" ] }, "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/891b7958f5e5153a88b9d67a7776f8f995656112.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/24b802789ee485d9de366c1ba5a2da764187a883.zip" }, "title": { "value": "CALoR: Towards Comprehensive Model Inversion Defense" }, "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": "While the motivation of this work is strong and interesting, not much novelty is observed in the study. The gain of this work is also minor compared to the previous work and the variance is not reported. Furthermore, some of the ablation study (like the number of negative text) seems to be very sensitive to the variance. It is unclear whether the gain in Table 1 and the ablation in Table 3 are due to the variance." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tFigure 2 is well designed and fully captured the overall training pipeline. \n2.\tThe paper is well written and simple to follow.\n3.\tThe motivation of this work is strong and interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper aims to improve the spatial understanding of the existing video-language model for the classification task. This is implemented with three different optimization losses, including the classification task (classify the video into predefined classes), contrastive loss (aligning the video with the text descriptions associated with the video) and spatial knowledge transfer task. The spatial knowledge transfer task utilizes the frozen large vision language model to guide the learning of video encoder, by aligning the video token with the image token. Extensive experiments and ablation are conducted on different video datasets and backbones to demonstrate the effectiveness of the proposed training strategies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe main contribution of this work is unclear. The classification loss and contrastive loss are common in video pretraining and video classification. Moreover, the alignment with L1 loss between the image token and the video token is also not considered novel. \n2.\tThe most interesting section in this work is the cross attention section in Eq 6-8, which leverages the segmentation model to guide the alignment. However, while the author shows that the result is not performing well, the explanation in L258-260 is not sound and solid. It will be interesting if the author can provide more study and in-depth discussion.\n3.\tL297-L299 is unclear. What does it mean by “dissimilar description would be less linked to the action”?\n4.\tThe conclusion from the negative test sampling ablation is unclear. Figure 3 shows a reflection point where the performance of the model has a minimum. Is there any reason why this is happening? What is the variance of each data point in Figure 3?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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 lines 42-44, the author points out the \"emphasis on capturing temporal information\" and illustrates how this is a common phenomenon through the heatmap in Fig. 1(c). How does this demonstrate that the method proposed in this paper can address this issue, and what is the specific principle behind the solution?\n- The description of the experimental results is quite imprecise. The performance of existing methods for Top-1 has exceeded 88.5% [1] and reached 89.7% two years ago [2], while the proposed method only achieves 86.9%. Therefore, it cannot be claimed that this method has achieved SOTA. \n- In Line 267, how can we ensure that the descriptions generated by the decompose-expand prompting are accurate? What would be the consequences if there are errors?\n- Additionally, please include the Top-5 results in the experimental outcomes.\n\n\n[1] Rethinking Image-to-Video Adaptation: An Object-centric Perspective\n[2] Disentangling Spatial and Temporal Learning for Efficient Image-to-Video Transfer Learning" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ The writing is very good, relatively easy to understand, and easy to follow as well.\n+ Using VLM to enhance video understanding is a very meaningful issue, and the results of this paper are good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work addresses the limitations of current video pre-trained models by integrating large VLMs into a cross-modal collaborative knowledge transfer framework. This method includes an attentive spatial knowledge transfer method that enhances spatial understanding and a contrastive textual knowledge transfer method for improved video-text alignment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It is insufficient to prove the core motivation using only one figure (Fig. 1 c).\n- the article lacks an analysis of the consumption of training costs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": 3 }, "primary_area": null, "questions": { "value": "- On the ablation of negative text sampling, can the authors provide higher number of k (why is 1200 the current limit). In addition, It is no meaningful to stated that sampling 200 negative text yield the third-best Top-1 accuracy. Please provide a better explanation of the initial accuracy drop when k increase from 200 to 600, and subsequently improve again. How does the authors know the diminish was due to easy negatives? Is there a way to measure if easy negative was sampled in early stage?\n- On the evaluation of gate mechanism, the improve with gate mechanism is relatively low (+0.24%). It would be to provide analysis on the class that is accurately classified with gate mechanism, and analyse if the corresponding attention map (with and without gate mechanism) match the hypothesis.\n- What is the number of description generated for each category. Has the authors analyse the score of the positive text and observed cases where the scores are generally too low for some cases." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ This work proposes a cross-modal collaborative knowledge transfer to adapt video pre-trained model for downstream tasks. Specifically, spatial knowledge is distill from VLM's image encoder via attentive spatial knowledge transfer. Then, textual knowledge transfer improve video representation via fine-grained text-video alignment.\n+ A gating mechanism is proposed to guide the distillation process to focus more on the action relation region and less on broader content.\n+ To improve the text-video alignment, a decompose-expand prompting method is proposed to improve the fine-grained and diverse description. This provide more training cue than a single description or class name. Then, a contrastive textual knowledge transfer learn the semantic knowledge.\n+ The proposed method achieve consistent improvement over multiple baselines and datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to leverage Vision Language Models (VLM) for cross-modal collaborative knowledge transfer to improve the spatial and semantic understanding for video data. The proposed approach aim to reduce the temporal gap between pre-training dataset and downstream tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I don't see obvious weaknesses in this work. Please refer to the Questions Section for some clarification." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Why does the learning rate of the optimizer reported in the implementation details section differ from that in the appendix?\n\n- Why were experiments not conducted on the mainstream video dataset Something-Something V2 (SSV2) for action recognition?\n\n- How are the weight coefficients for the loss functions obtained?\n\n- Line 72 should be written as (e.g., CLIP). There are many similar issues throughout the manuscript.\n\n- Many equations are missing symbols, such as Equations (6) and (7). Additionally, the bold and italic formatting in the equations is inconsistent.\n\n1. Line 238 is missing punctuation.\n\n2. Line 313, 'Eq. equation 2' is so poor.\n\n3. The capitalization of the first letters in the paper title is inconsistent." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ Introducing VLMs and transfer learning for video understanding is reasonable.\n\n+ The experimental results show slight improvements over previous methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a cross-modal collaborative knowledge transfer method for video action recognition. The framework consists of three key components: video adaptation, attentive spatial knowledge transfer and contrastive textual knowledge transfer. These three components correspond to three loss functions. Experiments on three datasets achieve state-of-the-art results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper conducts experiments on video action recognition, but the title and abstract do not explicitly mention this task. Video understanding is a relatively broad concept.\n\n- The three technical contributions result in only marginal improvements, especially as shown in Table 3. It is unclear which loss function serves as the main loss." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024enhancing,\ntitle={Enhancing Video Understanding with Vision and Language Collaboration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yspBoIZJ9Z},\nnote={under review}\n}" }, "abstract": { "value": "Leveraging video pre-trained models has led to significant advancements in video understanding tasks. However, due to the inherent bias towards temporal learning in video pre-training, these models fail to capture comprehensive spatial cues. Additionally, the widely-used supervised adaption methods lack fine-grained semantic guidance as single action labels cannot precisely depict the intra-class diversity. To address these challenges, we incorporate the general capabilities of large Vision Language Models (VLMs) and propose a cross-modal collaborative knowledge transfer method to enhance video understanding. First, we propose an attentive spatial knowledge transfer method that distills spatial knowledge from the VLM's image encoder, enabling the precise capture of spatial information. Next, we design a contrastive textual knowledge transfer approach that achieves detailed video representations through fine-grained text-video alignment. Owing to the cross-modal knowledge transfer, the video representations are capable of attending to informative spatial regions and aligning with fine-grained texts that carry rich semantics. Extensive experiments demonstrate that our method achieves state-of-the-art performance across various datasets, validating its effectiveness." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Video understanding", "video pre-trained model", "vision-language model", "collaboration 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/c295419a0f9c4ab20ed50271451f83dfed75c6c2.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": "Enhancing Video Understanding with Vision and Language Collaboration" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The thesis would benefit from a more structured presentation, where the authors are encouraged to list observations and analyses in Chapter 2 in a systematic manner, similar to DTAM. Additionally, the content should be included in the methods section as a motivation or exploration segment to strengthen the logical flow and context of the proposed" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Figures 2 and 3 provide valuable insights into the effects of removing both simple and difficult samples, as well as the impact of shallow parameters on model performance.\n\n2. The authors present a comprehensive experimental analysis on a small-scale dataset, including ablation studies, hyperparameter analysis, and related discussions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the task of dataset distillation, which aims to compress a large dataset into a much more compact synthetic dataset while maintaining the performance of trained models. Existing methods rely on an agent model to extract and embed information from the target dataset into the distilled version. However, the authors identify that current approaches often introduce misaligned information during the extraction and embedding stages, which degrades the quality of the distilled dataset.\n\nTo address this, the authors propose Prioritize Alignment in Dataset Distillation (PAD), a method that aligns information from two main perspectives: Dataset Pruning and Deep Layer Utilization. This simple yet effective strategy helps filter out misaligned information, resulting in significant improvements for mainstream matching-based distillation algorithms. Additionally, when applied to trajectory matching, PAD achieves state-of-the-art performance across various benchmarks, showcasing its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The author's approach appears to be more incremental, incorporating only minor enhancements to DATM. Firstly, Section 3.1 serves as a review of previous work, and Equation (4) in Section 3.3 shows only slight modifications from Equation (3) in DATM. Additionally, Section 3.2 seems to function more as a heuristic for selecting difficult samples. Overall, the method introduces only two additional techniques compared to DATM, lacking a significant breakthrough in terms of innovation.\n\n2. in Table 2, most of the average improvements over the comparable method, DATM, are less than 1 point, indicating that the performance still requires further enhancement." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The manuscript employs DATM as a baseline, which, as I understand it, requires the pre-training of numerous agent models on the original dataset to record training trajectories. As discussed in w2, given the dynamic dataset scheduler nature of the process described in your methodology, does this imply the need to train additional agent models? It would be beneficial for readers if the authors could provide more detailed insights into the implementation specifics of this approach." }, "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 demonstrates substantial improvements in DD through its experimental results in TAB. 1, showing that the PAD method significantly enhances the effectiveness of dataset distillation.\n2. PAD proves to be highly adaptable and generalizes well across multiple datasets, indicating that the method can be effectively applied to various dataset distillation tasks with consistent success." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces \"Prioritize Alignment in Dataset Distillation\" (PAD), a method improving dataset distillation by focusing on two main strategies. First, it adjusts data selection based on different Information Compression Ratios (ICRs) to match the required difficulty levels. Second, it enhances results by distilling only from deep-layer network parameters. The effectiveness of PAD is validated through experiments on standard datasets like CIFAR-10, CIFAR-100, and Tiny ImageNet, showing improvements over current DD methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There are some formatting issues in the manuscript that need attention. In **Table 5**, the annotations (a) and (b) appear to be reversed, which could confuse readers. Additionally, a period is missing at the end of line 485 after \"smaller IPCs.\" Please correct these to enhance the clarity and professionalism of the document.\n2. There seems to be a mismatch between the issues presented in Section 2.1 and the methods proposed in Section 3.2. The former discusses a static selection method, while the latter introduces a dynamic approach. Could the authors clarify the connection between these sections to ensure the consistency of the methodology described?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. In the main table (Table 1), it appears that multiple experiments were conducted to obtain an average, but subsequent tables present experimental values as single data points. I couldn't find any information in the paper regarding whether the results for each experiment were averaged, and I’m curious about this.\n\n2. The study exclusively uses deep layers; are shallow layers entirely unnecessary? Shallow layers likely play a role in producing a good embedding space, so is there a way to leverage them effectively?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. One advantage of this method is that it can be applied on top of other distillation methods, such as MTT and DM. As seen in cross-architecture analyses, it’s a general approach that enhances performance in a scalable, model- and dataset-agnostic way. Experimental results show that accuracy improved over traditional methods (Table 1), with increases across multiple datasets and architectures (Table 2).\n\n2. It seems the straightforward method is supported by detailed analysis and experiments. Experiments (Tables 4, 5 / Figure 5) effectively support the hypothesis that removing easy examples and using deeper layers improves performance, making the qualitative results intuitively understandable (Figures 4, 6).\n\n3. While the writing isn’t exceptionally well-crafted, it is organized in a readable and accessible manner." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the role of alignment in both the information extraction and embedding stages within the dataset distillation process. \nDuring the information extraction stage, alignment is achieved by selecting subsets of the dataset based on difficulty. For settings with a low images-per-class (ipc) count, incorporating a higher proportion of ‘easy’ data proves effective. Conversely, in high ipc settings, utilizing a larger portion of ‘hard’ data enhances distillation effectiveness. The EL2N score, analogous to model confidence, is employed as a difficulty metric.\nIn the information embedding stage, alignment is achieved by prioritizing deeper over shallower layers. Deeper layers are more adept at learning semantic information, resulting in a higher-level representation within synthetic data and more efficient distillation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I believe the contribution of this method is limited. The proposed method is very simple, and while it is supported by empirical experimental results, it lacks theoretical justification. Additionally, as mentioned in the paper, the data selection and distillation approach was already introduced by BLiP (Xu et al., 2023). Although the paper presents experimental results showing superior performance to this approach (Table 5.b), the metric for selecting difficulty (EL2N) was also previously proposed.\n\n2. The method seems somewhat sensitive to hyperparameters (AEE, IR). Accuracy varies depending on these hyperparameters (Table 4.a), with differences comparable to the performance increase over other methods in the main table. Additionally, as seen in Table 10, optimal hyperparameters change dynamically across datasets and ipc values, raising questions about the method’s stability. If it is indeed sensitive, hyperparameters may need to be tuned for each architecture and dataset, which raises concerns about the method’s practicality.\n\n3. Although the paper claims that performance improves when used alongside various distillation techniques, as mentioned in the limitations section, due to computing resource constraints, experiments were conducted only with DM and DC (and DATM in the main table).\n\nOverall, the paper demonstrates that a simple method can enhance distillation performance and can be used in a model- and architecture-agnostic manner. The experiments and logical development are well-executed; however, I believe the contribution of the method itself is limited, and it appears to be sensitive to hyperparameters. Therefore, I would rate it as [5: marginally below the acceptance threshold]." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How does PAD compare to other recent advances in dataset distillation in terms of computational resources and scalability?\n2. Could the authors elaborate on the theoretical justification for prioritizing deep layer parameters?\n3. Are there any scenarios where PAD might not perform as expected, and if so, how does the method handle such cases?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The identification of misalignment in dataset distillation and the proposal of PAD to address it is a valuable contribution.\n2. The paper provides a clear motivation for the PAD method and supports its claims with extensive experiments.\n3. The approach of prioritizing deep layer parameters for distillation is innovative and leads to improved performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces PAD, a method aimed at enhancing the compression of large datasets into compact synthetic datasets while preserving model performance. PAD addresses the challenge of misaligned information during distillation by aligning data through selective pruning of the target dataset and leveraging deep layers of agent models in the distillation process." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It would be beneficial to prove the relationship between EL2N and the difficulty of training samples.\n2. The paper could provide a more detailed comparison with other state-of-the-art methods, especially in terms of computational efficiency and scalability. \n3. The theoretical analysis of why deep layer parameters are more suitable for distillation is lacking. \n4. The paper does not discuss potential limitations or failure cases of the PAD method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024prioritize,\ntitle={Prioritize Alignment in Dataset Distillation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yt7nxONs3J},\nnote={under review}\n}" }, "abstract": { "value": "Dataset Distillation aims to compress a large dataset into a significantly more compact, synthetic one without compromising the performance of the trained models. \nTo achieve this, existing methods use the agent model to extract information from the target dataset and embed it into the distilled dataset. \nConsequently, the quality of extracted and embedded information determines the quality of the distilled dataset.\nIn this work, we find that existing methods introduce misaligned information in both information extraction and embedding stages.\nTo alleviate this, we propose Prioritize Alignment in Dataset Distillation (\\textbf{PAD}), which aligns information from the following two perspectives.\n1) We prune the target dataset according to the compressing ratio to filter the information that can be extracted by the agent model.\n2) We use only deep layers of the agent model to perform the distillation to avoid excessively introducing low-level information.\nThis simple strategy effectively filters out misaligned information and brings non-trivial improvement for mainstream matching-based distillation algorithms.\nFurthermore, built on trajectory matching, \\textbf{PAD} achieves remarkable improvements on various benchmarks, achieving state-of-the-art 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": [ "dataset 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/b928959f4c9c002c1ae402c10fe8c608b2e8ddc3.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": "Prioritize Alignment in Dataset Distillation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]