ResearchArcade
Collection
10 items
•
Updated
Dataset Viewer
venue
stringclasses 11
values | review_openreview_id
stringlengths 8
13
| replyto_openreview_id
stringlengths 9
13
| writer
stringlengths 2
110
| title
stringlengths 14
49
| content
stringlengths 29
44.2k
| time
stringdate 2013-02-06 08:34:00
2025-08-01 01:24:28
|
|---|---|---|---|---|---|---|
ICLR.cc/2025/Conference
|
Is5Qh2Gs5x
|
zzR1Uskhj0
|
Reviewer_8LjP
|
Official Review by Reviewer_8LjP
|
{"Rating": 6, "Summary": "The submission studies contextual bandits with cross learning. Previously, the existing regret bound held in expectation. The submission refines the regret analysis so that the regret bound holds with high probability. The main contribution is to show how the weak dependency structure can be exploited to solve a concentration difficulty in the previous analysis.", "Questions": "- (1) Is the concept class $C$ a finite set? If so, what is the reason for assuming a finite concept class $C$? Practically speaking, the contextual information would be like a vector in a compact set, as it is very unlikely to see two identical users.- (2) Where is the variable in Theorem 1 that characterizes the property of $C$? How does this variable appear in the bound proved in this submission?- (3) Could you please provide more evidence or further discussion of the applicability of the technique developed here so that we can better appreciate its potential?", "Soundness": 3, "Strengths": "- The submission points out the difficulty that prevents the previous work from achieving a bound with high probability (lines 167\u2013176).- Identify the weak dependency between epochs (line 386).- Devise a new technique to solve the unbounded issue induced by the weak dependency (the treatment for the Bias5e them in lines 395\u2013411).", "Confidence": 4, "Weaknesses": "- (1) Section 3.2 contains several subtopics, such as the regret decomposition, the discussion of each decomposed term, and the analysis strategy for the challenging term. A better editorial layout would improve the readability.- (2) The sentence \u201cNotably, \u2026\u201d in line 111 is confusing. It seems unrealistic to be able to observe the loss for every context $c$. It also does not match the algorithm\u2019s (Algorithm 1) behavior.", "Contribution": 3, "Presentation": 3}
|
2024-11-04 01:05:41
|
ICLR.cc/2025/Conference
|
Hsumvt7DeH
|
zzR1Uskhj0
|
Reviewer_Sstz
|
Official Review by Reviewer_Sstz
|
{"Rating": 6, "Summary": "The paper studied adversarial context bandits in a special setting where the losses of arm $a_i$ could be observed under all contexts when the algorithm plays arm $a_i$. The goal, like in classical adversarial bandit problems, is to minimize the regret compared to the loss of the best arm in hindsight. The paper focuses on the setting where the loss sequence is adversarial and the context is stochastic with an unknown distribution. A recent work of SZ [NeurIPS\u201923] designed an algorithm with *expected* regret of $\\tilde{O}(\\sqrt{KT})$ in this setting, where $K$ is the number of arms. This paper conducted a renewed analysis of the algorithm in SZ [NeurIPS\u201923], and the main result is that the algorithm could actually achieve $\\tilde{O}(\\sqrt{KT})$ with high probability.The main technique of the paper is heavily influenced by the previous work of SZ [NeurIPS\u201923]. In a nutshell, the low-regret guarantee of the algorithm crucially relies on the concentration of unbiased estimation of $E_{c}[\\ell_{t,c}(a)]$. Here, we cannot exactly compute the quantity since the distribution of the context is unknown. The key idea of SZ [NeurIPS\u201923] is to commit two steps for each EXP3 step and use one of them to estimate the distribution of the context. On top of that, this paper further utilized the weak dependency between epochs, and derived a martingale argument to get high-probability regret.", "Questions": "- Should the definition of regret on page 3 be reversed? As in, you are subtracting the loss of the best arm with the loss of a policy, which should be a negative value (if with positive regret). This would change the decomposition of the regret on page 5 as well, but it seems nothing would affect the correctness.- I think the full algorithm description of the algorithm in SZ [NeurIPS\u201923] (or some simpler version of the description) could be shown much earlier in the paper. This would be helpful for readers who are not familiar with the previous algorithm.- Also, stating the main theorem in a preliminary section looks very non-standard to me. I\u2019m not letting this affect my score, but please consider re-organizing this.- The meaning of 'with high probability' was never explained in the paper -- as in, it could mean with probability $1-1/K$ or with probability $0.99$. I think your bound gives the former, and this should be stated explicitly.", "Soundness": 4, "Strengths": "In general, my opinion of this paper is positive. The paper appears to require a great deal of background to be able to parse. Despite this, I believe the paper did reasonably well in terms of explaining the existing work and its techniques. Getting high probability bounds in adversarial bandits usually requires some neat observations and technical steps. Although I\u2019m not able to follow all the steps in the short time frame, I do think the paper contains some nice technical observations and ideas.", "Confidence": 4, "Weaknesses": "Although I'm mostly supportive, I think I couldn\u2019t strongly champion the paper due to the following reasons:- The scope of contribution: although the paper does contain some neat technical observations, the contribution appears to be somehow incremental. After all, this is a new analysis of an existing algorithm, and the new analysis is not something that improves the previous bound (but instead is to get a high-probability bound). Again, I do acknowledge that such contributions are non-trivial. However, I do not think it\u2019s enough for me to champion the paper.- If the paper is going to be mainly accepted due to the techniques: then, I do not think the paper contains a substantial amount of new ideas. I appreciate the technical observations, and I agree that the steps are non-trivial. However, if the conceptual message is not as strong, and the merits of the paper mainly lie in the techniques, then the bar would inevitably be higher. - For a conference like ICLR, the lack of experiments could be an issue. I am *not* letting this affect my score since I often advocate learning theory papers. However, I do want to raise this point since it is common for ML conferences to ask for experiments.", "Contribution": 3, "Presentation": 3}
|
2024-11-07 22:01:20
|
ICLR.cc/2025/Conference
|
smWIsNwjkv
|
zzR1Uskhj0
|
Reviewer_uDyZ
|
Official Review by Reviewer_uDyZ
|
{"Rating": 8, "Summary": "The paper studies cross learning in contextual adversarial linear bandits where the learner observes the losses of all contexts in each round. Recent work in Schneider et al. proposed an algorithm with a regret upper bound only in expectation. The paper studies the same algorithm and proves that the regret upper bound holds with high probability.", "Questions": "please see weaknesses", "Soundness": 4, "Strengths": "- The paper proves a high probability lower bound which is stronger than the in expectation bound in the literature.- The analysis uses a nice observation that the different epochs in the algorithm are only weakly dependent which enables to prove a small bound for the cumulative bias across all epochs- While standard martingale inequalities cannot directly upper bound the cumulative bias, a novel technique is proposed to address this", "Confidence": 3, "Weaknesses": "Can the reduction in [1] be used to map the multi-context to a single context problem? The technique is proposed for non-adversarial losses, however, the action set map from distributional to fixed should not be affected by that.I understand that the paper only focuses on analyzing an existing algorithm. However, a comparison with such technique in the related work is needed to justify the use of such algorithm or suggest alternative techniques to address the problem.[1] \"Contexts can be cheap: Solving stochastic contextual bandits with linear bandit algorithms.\" The Thirty Sixth Annual Conference on Learning Theory. PMLR, 2023.", "Contribution": 3, "Presentation": 3}
|
2024-11-08 08:31:50
|
ICLR.cc/2025/Conference
|
UfYuXDF8OO
|
zzR1Uskhj0
|
Reviewer_NZtQ
|
Official Review by Reviewer_NZtQ
|
{"Rating": 5, "Summary": "This paper addresses the challenge of achieving high-probability regret bounds in the adversarial contextual bandit framework, where the learner encounters varying contexts and must minimize cumulative loss over time. The focus is on \"cross-learning\" contextual bandits, where learners can observe losses for all possible contexts, not just the current one. Results leverage weak dependencies between epochs and refine existing martingale inequalities, by exploiting interdependencies in observations. This analysis ultimately shows that the algorithm is effective in adversarial settings, even with unknown context distributions.", "Questions": "While the result is good, I am uncertain about its significance because neither the problem nor the algorithm proposed are new. There are no extensions of this result, no experiments. I am not sure if this is standalone result is significant enough to be published at a premier conference.", "Soundness": 3, "Strengths": "The paper proposes a new look at an existing problem and provides a completely novel analysis in their work. The ideas and techniques proposed are completely new and can be of independent interest. This is particularly true for the martingale concentration result.", "Confidence": 3, "Weaknesses": "See questions.", "Contribution": 3, "Presentation": 3}
|
2024-11-08 09:01:46
|
ICLR.cc/2025/Conference
|
NLrlOlSugS
|
zzR1Uskhj0
|
Reviewer_NaxM
|
Official Review by Reviewer_NaxM
|
{"Rating": 5, "Summary": "The paper proposes an algorithm that achieves high probability regret bound (which is stronger than the expected regret bound) for the cross-learning contextual bandits under unknown context distribution by developing refined martingale inequalities.", "Questions": "1. What is the intuition behind the algorithm?2. How the indicator function $F_e$ resolves the unbounded martingale inequalties?", "Soundness": 2, "Strengths": "1. The paper clearly presents the challenging point with detailed technical expressions and the novelty of the analysis.", "Confidence": 3, "Weaknesses": "1. The explanation is only focused on technical side without the explanation of the algorithm. I suggest authors to spend more time including more explanations and organizing the paper.", "Contribution": 2, "Presentation": 1}
|
2024-11-12 13:14:27
|
ICLR.cc/2025/Conference
|
WQLpTsquBi
|
NLrlOlSugS
|
Authors
|
Response by Authors
|
{"Title": "Rebuttal by Authors", "Comment": "Dear reviewer NaxM:Thank you for your valuable feedback. We address the comments below in detail:---**Question 1: Lack of explanation of the algorithm**The reviewer suggested that we proposed an algorithm, but did not elaborate on its intuition, making it challenging to understand.**Response:**We would like to clarify that we did **not** propose a new algorithm. Instead, we provided a new and in-depth analysis of an **existing** algorithm, strengthening its result from an expected regret bound to a high-probability regret bound. This point was explicitly stated multiple times in the manuscript and was well understood by other reviewers.We are grateful for the reviewer\u2019s feedback. To address this concern, we have rewritten the paper to make it clearer that our contribution lies in the re-analysis of an existing algorithm. A revised version of the paper will be uploaded soon. In the new version, we have included a dedicated section to introduce the existing algorithm in detail. Furthermore, to better clarify the intuition behind the work, we provide an explanation of the algorithm's underlying principles at the beginning of this section.---**Question 2: How the indicator function resolves unbounded martingale inequalities****Response:**As noted near the end of the original manuscript, we use the indicator function to associate the original random variable sequence with a new random variable sequence. This new random variable sequence forms a bounded martingale, allowing us to apply standard martingale concentration inequalities. Furthermore, through this indicator-based association, we demonstrate that the original and new random variable sequence coincide with high probability. This enables us to transfer the concentration inequalities from the new sequence back to the original.---Once again, we sincerely thank the reviewer for the constructive comments, which have helped us improve the clarity and readability of the manuscript. The revised version will be uploaded soon. We hope our response has addressed the reviewer\u2019s concerns and that you will consider increasing your support for the paper."}
|
2024-11-19 15:11:20
|
ICLR.cc/2025/Conference
|
2LhrQBBkiK
|
UfYuXDF8OO
|
Authors
|
Response by Authors
|
{"Title": "Rebuttal by Authors", "Comment": "Dear reviewer NZtQ:We sincerely thank the reviewer for their valuable suggestions. Below, we address the reviewer\u2019s concerns in detail. ---**Question: Significance of our results** **Response:** We appreciate the reviewer\u2019s accurate understanding of our results and their concerns about our work's significance. We would like to highlight that it is a common practice in adversarial bandit research to focus solely on providing high-probability bounds, as this itself constitutes a significant contribution (e.g., [1,2,3]). From a technical perspective, as noted by reviewer Sstz, deriving high-probability bounds for existing algorithms in bandit research often requires neat observations and techniques. From a results perspective, high-probability bounds are particularly important in adversarial bandit settings because these scenarios focus on worst-case outcomes, where even low-probability events cannot be ignored. Thus, providing high-probability bounds, rather than just expected regret bounds, is critical for addressing the worst-case nature of adversarial bandits. Regarding the lack of experiments, we acknowledge that this is a limitation. However, we respectfully note that in theoretical works on adversarial bandits, especially those focusing on high-probability bounds, it is common practice to omit experiments. This is because our proofs are based on rigorous mathematical arguments without relying on any unrealistic assumptions or approximations, and they remain effective under worst-case scenarios. We hope the reviewer will consider this aspect. To further emphasize our contributions, we have restructured the paper and added a Technical Overview section in the introduction to discuss our technical contributions in detail. The revised version of the manuscript will be uploaded shortly. Once again, we thank the reviewer for their insightful suggestions, which have helped us improve the structure of our paper and better highlight its contributions. We hope our response has addressed the reviewer\u2019s concerns and that you will consider increasing your support for the paper.---References:[1] Luo, H., Tong, H., Zhang, M., & Zhang, Y. (2022). Improved High-Probability Regret for Adversarial Bandits with Time-Varying Feedback Graphs. International Conference on Algorithmic Learning Theory.[2] Neu, G. (2015). Explore no more: Improved high-probability regret bounds for non-stochastic bandits. Neural Information Processing Systems.[3] Bartlett, P.L., Dani, V., Hayes, T.P., Kakade, S.M., Rakhlin, A., & Tewari, A. (2008). High-Probability Regret Bounds for Bandit Online Linear Optimization. Annual Conference Computational Learning Theory."}
|
2024-11-20 04:39:57
|
ICLR.cc/2025/Conference
|
wR6AlCqdjL
|
Is5Qh2Gs5x
|
Authors
|
Response by Authors
|
{"Title": "Rebuttal by Authors", "Comment": "Dear reviewer 8LjP:We sincerely thank the reviewer for their suggestions and positive feedback. Below, we provide detailed responses to the reviewer\u2019s comments.---**Question: What is the reason for assuming a finite concept class?** **Response:** We would like to point out that assuming a finite concept class is a common practice in contextual bandit research. While the real context space is often continuous, it can be discretized into a finite set. In this way, the finite concept class serves as a foundational model, much like the tabular MDP framework in reinforcement learning. Of course, the discretization process involves a tradeoff: finer discretization reduces discretization error but increases the size of the concept class, leading to larger regret. However, the cross-learning structure in our setting entirely eliminates this issue, providing further justification for focusing on finite concept classes. Please see our response to the next question for more details. ---**Question: Where is the variable in Theorem 1 that characterizes the property of $C$?** **Response:** We thank the reviewer for raising this excellent question, which touches on one of the most interesting aspects of cross-learning bandits. Indeed, in most cases, the results depend on the size of the concept class. In the vanilla contextual bandit setting, the results typically have a polynomial dependence on the size of the concept class $C$. However, in our problem, thanks to the cross-learning structure, this polynomial dependence on the size of $C$ is entirely eliminated. As a result, our final regret bound is completely independent of the size of $C$, which is why Theorem 1 does not need to explicitly characterize the property of $C$. As mentioned in the response to the previous question, the cross-learning structure allows us to bypass the discretization issue for finite concept classes. This is because our result is completely independent of the size of the concept class, enabling arbitrarily fine discretization and resolving this concern entirely. ---**Question: The statement in line 111 that we can observe the loss for every context seems confusing and does not match the behavior of the algorithm.** **Response:** We would like to clarify that this is precisely the core of the cross-learning structure. The cross-learning structure explicitly assumes that we can observe the loss for every context. As discussed in the related works section, this structure is common in practice, with examples including bidding in online auctions, sleeping bandits, repeated Bayesian games, and dynamic pricing. Regarding the algorithm\u2019s behavior, we respectfully disagree with the reviewer\u2019s assessment. The algorithm indeed matches this assumption since it is explicitly designed for the cross-learning structure.---**Question: Readability of Section 3.2** **Response:** We thank the reviewer for pointing this out. To improve readability, we have restructured the paper. In the revised version, Section 3.2 has been expanded into a standalone chapter, further divided into three subsections, each focusing on a single topic. This restructuring aims to make the paper more organized and easier to follow. The updated manuscript will be uploaded shortly. ---We once again thank the reviewer for their valuable suggestions and positive feedback. Your comments have helped us improve the structure and readability of our paper. We hope our responses have addressed your concerns."}
|
2024-11-20 06:04:06
|
ICLR.cc/2025/Conference
|
6LIz7vD6vJ
|
Hsumvt7DeH
|
Authors
|
Response by Authors
|
{"Title": "Rebuttal by Authors", "Comment": "Dear Reviewer Sstz, Thank you for your positive, thorough, and thoughtful review. Your feedback has greatly helped us improve our paper. Below, we provide detailed responses to your comments: ---**Question: Should the definition of regret on page 3 be reversed?** **Response:** We respectfully point out that our definition is correct. Our regret definition subtracts the loss of the arm chosen by the algorithm from the loss of policy $\\pi$, rather than subtracting the loss of the best arm from the loss of a policy. This ensures that the regret is typically non-negative. ---**Question: Suggestions on improving the writing of the paper** You suggested that we describe the algorithm earlier, state the main theorem outside the preliminaries, and explicitly explain the meaning of high probability. **Response:** We greatly appreciate your detailed suggestions regarding the paper's structure and presentation, as they are immensely helpful for improving the clarity of our writing. We have carefully considered your suggestions and revised the manuscript as follows: 1. **Reorganized the algorithm description:** The algorithm description is now a standalone section, divided into two subsections. The first subsection revises Section 2.2 of the original paper to explain the intuition behind the algorithm. The second subsection revises Section 3.1 of the original paper to formally describe the algorithm. 2. **Moved the theorem statement out of the preliminaries:** In the revised manuscript, we move the theorem statement out from the preliminaries. Instead, we provide an informal version of the theorem in the Introduction section for early context, and provide the formal version of the theorem in the Main Result and Analysis section.3. **Clarified the meaning of high probability:** We restructured the theorem's statement to explicitly define the high-probability guarantee. Specifically, our high-probability result means that, for any probability parameter $\\delta \\in (0,1)$, the algorithm achieves a regret bound with probability at least $1 - \\delta$, where the bound depends on $\\delta$ only in the form of $\\log(1/\\delta)$. The revised manuscript incorporating these changes will be uploaded shortly. ---Once again, we sincerely thank you for your positive, thorough, and thoughtful review. Your suggestions have significantly improved the structure and readability of our paper. We hope our response addresses your concerns."}
|
2024-11-20 08:01:55
|
ICLR.cc/2025/Conference
|
p1r3CeMT7X
|
smWIsNwjkv
|
Authors
|
Response by Authors
|
{"Title": "Rebuttal by Authors", "Comment": "Dear Reviewer uDyz,Thank you for your positive and encouraging feedback. We greatly appreciate you bringing to our attention an interesting piece of work that we had previously overlooked\u2014it has truly broadened our perspective.Unfortunately, the mentioned work cannot be directly applied to our problem. However, whether it could be adapted through certain modifications to address our problem is an intriguing question. We have discussed this point in detail in the related works section of the revised version of our paper, which will be uploaded shortly.Once again, we sincerely thank you for your positive comments and for pointing out this valuable reference."}
|
2024-11-20 11:31:23
|
ICLR.cc/2025/Conference
|
hrek1FOSJ7
|
zzR1Uskhj0
|
Authors
|
Response by Authors
|
{"Title": "The revised version of the paper", "Comment": "Dear Reviewers,We sincerely thank all the reviewers for their valuable suggestions. Based on your feedback, we have restructured the paper, and the revised version has been uploaded. In the new version, we have made the following changes:- **Section 1**: We added a *Technique Overview* subsection to provide a clearer explanation of our technical contributions. Additionally, we discussed the paper mentioned by reviewer uDyZ in this section and included the informal version of the theorem. - **Section 2**: The new Section 2 now focuses solely on the problem formulation. Specifically, we have completely removed the theorem statement from this section. - **Section 3**: A new Section 3 has been created, incorporating content from the old Section 2.2 and Section 3.1. This section provides a clearer explanation of the existing algorithm. - **Section 4**: The new Section 4 is a revised version of the old Section 3. It has been divided into three new subsections, each dedicated to addressing a specific technical issue, thereby improving readability. Notably, the formal version of the theorem is now included in this section. We sincerely thank the reviewers once again for their valuable feedback."}
|
2024-11-21 07:14:28
|
ICLR.cc/2025/Conference
|
jAvCQROBD4
|
6LIz7vD6vJ
|
Reviewer_Sstz
|
Response by Reviewer
|
{"Title": "", "Comment": "Thanks for the response and the updated paper. I took a look at the revised paper, and I believe the presentation of the paper has been improved. The paper is now much more accessible for readers unfamiliar with the work of SZ [NeurIPS\u201923].Due to my comments on the combined novelty for scope + techniques, I am keeping my overall evaluation as it is. However, I increased my confidence score by 1 to give better support for the paper."}
|
2024-11-21 21:26:42
|
ICLR.cc/2025/Conference
|
mAqJKkBBov
|
NLrlOlSugS
|
Reviewer_NaxM
|
Response by Reviewer
|
{"Title": "", "Comment": "Thank you for detailed response. I increase my score to 5 for the revised manuscript but unfortunately, I have concerns to raise score more due to theoretical novelties.While deriving high-probability bound is an interesting problem, the technical novelty to derive the bound is limited.The revised paper heavily relies on the results from Schneider & Zimmert 2013 and the novel part is limited to replace the random variable with the conditionally expected term with high probability.Specifically, the paper replaces the Bias5e with Bias5e$F_e$ (which can be found in other literature with more refined analysis (see e.g., Thoerem 7 in [1])) and follows the analysis in Schneider & Zimmert 2013.Because the paper focuses on theoretical analysis without any novel algorithms or experiments, I believe it should include at least one newly developed lemma that may apply to many settings not only for the cross-learning contextual bandits to be published at a top-tier conference.[1] Shamir, Eli, and Joel Spencer. \"Sharp concentration of the chromatic number on random graphs G n, p.\" Combinatorica 7 (1987): 121-129."}
|
2024-11-25 06:01:47
|
ICLR.cc/2025/Conference
|
ozYvTN3mlR
|
wR6AlCqdjL
|
Reviewer_8LjP
|
Response by Reviewer
|
{"Title": "", "Comment": "Thank you for the feedback. After going through the reviews and all feedback replies, I will keep my score for now. Thank you!"}
|
2024-11-26 12:01:35
|
ICLR.cc/2025/Conference
|
rA3OQEiDA0
|
p1r3CeMT7X
|
Reviewer_uDyZ
|
Response by Reviewer
|
{"Title": "", "Comment": "Thank you for your response."}
|
2024-11-27 02:09:51
|
ICLR.cc/2025/Conference
|
AUQCPco8Ia
|
mAqJKkBBov
|
Authors
|
Response by Authors
|
{"Title": "Response by Authors", "Comment": "Dear Reviewer,Thank you very much for your time, attention, and for improving our score. However, we respectfully disagree with your comments regarding the theoretical novelty of our work. You stated that the novelty of our paper is \"limited to replacing the random variable with the conditionally expected term with high probability ... replaces the $\\mathrm{Bias5}_e$ with $\\mathrm{Bias5}F_e$.\" We would like to clarify that this is **not the core theoretical novelty** of our work. As we emphasized in the abstract, the core novelty of our paper lies in \"making extensive use of the weak dependency structure between different epochs.\" Our primary contribution is in recognizing the weak dependency of $\\mathrm{Bias5}_e$ across different $e$, which allows us to effectively bound $\\sum_e \\mathrm{Bias5}_e$. The point you mentioned about \"replacing the random variable with the conditionally expected term with high probability ... replaces the Bias5e with Bias5Fe\" is an additional novelty that supports this main argument.This core novelty\u2014\"making extensive use of the weak dependency structure between different epochs\"\u2014stems from an in-depth analysis of cross-learning contexual bandit itself. We hope you will take this into account and reevaluate the theoretical novelty of our paper. We also hope that this will encourage you to increase your support for the paper.Thank you again for your valuable feedback."}
|
2024-11-28 08:12:43
|
ICLR.cc/2025/Conference
|
w9jjG2HQky
|
UfYuXDF8OO
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "Dear Reviewer NZtQ,We sincerely appreciate your time and attention. We hope our responses have addressed your concerns and that you might consider increasing your support for our paper. If you have any further questions, please don't hesitate to ask us!"}
|
2024-12-02 02:24:33
|
ICLR.cc/2025/Conference
|
QJB25lfeno
|
zzR1Uskhj0
|
Area_Chair_zibT
|
Meta Review of Submission11051 by Area_Chair_zibT
|
{"Meta Review": "This paper explores adversarial context bandits, specifically focusing on a scenario where the losses of each arm are observable under all contexts when the algorithm selects that arm. The objective is to minimize regret by comparing the algorithm's performance to the best arm in hindsight, similar to classical adversarial bandit problems. The study examines a setting where the loss sequence is adversarial and the context is stochastic with an unknown distribution.A prior work by SZ (NeurIPS'23) introduced an algorithm that achieved expected regret in this setting, where the number of arms was a factor. The main contribution of this paper is a refined analysis of the SZ algorithm, demonstrating that it can achieve a much lower regret with high probability. The paper builds on SZ's approach, which hinges on the concentration of unbiased estimation of the loss, despite the unknown context distribution. SZ's key idea involves a two-step process for each EXP3 step, with one step dedicated to estimating the context distribution. This paper extends SZ's method by leveraging weak dependencies between epochs and incorporating a martingale argument to achieve high-probability regret bounds.The paper provides a new analysis of an existing algorithm, focusing on obtaining high-probability regret bounds. While the technical observations are non-trivial, the contribution is incremental, as it does not improve prior bounds or introduce substantial new ideas. Reviewers highlighted that the paper's primary merits lie in its techniques, which, while appreciated, may not meet the bar for ICLR without a stronger conceptual message. Additionally, concerns were raised about the lack of experimental validation, unconventional organization (e.g., the placement of the main theorem in the preliminary section), and unclear definitions (e.g., \"with high probability\"). Although these issues do not compromise correctness, they detract from the paper\u2019s clarity and impact, leading to hesitations about its suitability for acceptance. Based on the common consensus, therefore, I would recommend that the authors submit to the next suitable venue to address all the concerns once all the necessary modifications are incorporated.", "Additional Comments On Reviewer Discussion": "See \"The revised version of the paper\""}
|
2024-12-21 13:32:41
|
ICLR.cc/2025/Conference
|
atAHt9GNWL
|
zzR1Uskhj0
|
Program_Chairs
|
Paper Decision
|
{"Comment": "", "Decision": "Reject"}
|
2025-01-22 05:34:59
|
ICLR.cc/2025/Conference
|
KgCysAmNli
|
zyGrziIVdE
|
Reviewer_nQGL
|
Official Review by Reviewer_nQGL
|
{"Rating": 3, "Summary": "The paper proposes a new intrinsic exploration objective for maximizing state entropy. The objective uses a discounted mixture of past state occupancy measures and encourages policies that maximize distance from the discounted mixture. As statistical distance, the KL divergence and Wasserstein distance are used. The experiments are evaluated on state-based RL environments, where state coverage and episodic returns are used to demonstrate the performance gain of the proposed approach.", "Questions": "See Weaknesses.", "Soundness": 3, "Strengths": "The paper is written well, and the proposed intrinsic exploration objective is novel. The use of Wasserstein distance instead of the typical KL divergence is an interesting/novel choice.", "Confidence": 4, "Weaknesses": "1) Prior Work/Baselines: The paper misses several crucial works on intrinsic exploration (c.f., 1, 2, 3 for a survey). Particularly, there are works that use the model epistemic uncertainty/disagreement as an intrinsic reward which works well in practice and also scales favorably (4., 5., 6.). 2) Theory: In particular, the model epistemic uncertainty is theoretically a well-studied objective (7., 8.). In 8, the authors derive a connection between maximizing the model epistemic uncertainty and maximizing information gain/conditional entropy of the trajectories, while also showing convergence for sufficiently smooth dynamics. 3) Unclear motivation: Given the theoretical and experimental strengths of the method discussed above, its unclear to me what particular gap the authors are trying to address with their intrinsic reward. I'd appreciate the authors elaborating further on this. Furthermore, I think all the aforementioned works should be discussed in the paper and in particular one of the baselines should use the model epistemic uncertainty as the intrinsic reward. Perhaps one weakness the authors might raise is that the aforementioned works are computationally more expensive as they have to learn an ensemble of networks to quantify disagreement. However, this should also be empirically shown in the experiments (as the proposed method also learns a model to estimate the intrinsic reward). 4) Hyperparameters are not provided in the paper, which makes it difficult for me to assess how sensitive the results are to the choice of hyperparams. In particular, I am curious about how $\\beta$ affects the performance of the algorithm. How can we appropriately select $\\beta$? Furthermore, doesn't the method suffer from sample inefficiency for large values for $\\beta$, i.e., when lots of data from the buffer is discarded?5) Scalability: Its unclear to me whether the proposed method would scale reasonably well to more high-dimensional settings such as POMDPs/visual-control tasks (note that 5, 6 also work for POMDPs). Could the authors elaborate further on this?I am happy to raise my score if my concerns above are addressed. 1. https://arxiv.org/abs/2109.001572. https://www.sciencedirect.com/science/article/pii/S1566253522000288?casa_token=ScYOIGv6D2wAAAAA:buNFoXMZLqPiWzo0CLpe3K-ac_nxundN5855FT0QwSnE6jhpm6VwPFS0UHyt1E9WXJePruqZsg3. https://www.mdpi.com/1099-4300/25/2/3274. https://arxiv.org/pdf/1906.041615. https://arxiv.org/abs/2005.059606. https://arxiv.org/abs/2110.095147. https://arxiv.org/pdf/2006.102778. https://arxiv.org/pdf/2306.12371", "Contribution": 2, "Presentation": 3}
|
2024-11-01 01:01:19
|
ICLR.cc/2025/Conference
|
mQtVHAFXdK
|
zyGrziIVdE
|
Reviewer_YHsc
|
Official Review by Reviewer_YHsc
|
{"Rating": 5, "Summary": "The paper proposes an exploration paradigm of \"running away from the past\" (RAMP), which encourages the RL algorithm to generate trajectories in distribution different from the past. This is instantiated as an intrinsic exploration bonus that estimates the discrepancy between the current and past visitation density. They show improvements on a few benchmark deep RL algorithm, showcasing the potential for this approach.", "Questions": "### === Theoretical gains of running away from the past ===I think the idea of RAMP makes sense in that for the algorithm to explore, it must do something different from the past. However there is always a trade-off in practice and one must balance exploration vs. exploitation, a factor that is heavily environment dependent. There are simply environments where such exploration is not needed at all, while others where exploration is needed.At this point, deep RL literature has already accumulated a large varieties of exploration methods, each dedicated to a specific domain. I think it will be valuable if a more general purpose method such as RAMP can characterize the theoretical gains achieved by just maximizing the distributional divergence between the current policy and previous data distributions.### === Intrinsic reward alone ===Table 1 shows the max performance that can be achieved by different exploration methods using just the intrinsic reward. In a sense, it measures how extreme the performance can reach by just optimizing for the exploration bonus. It is quite a surprise to me that RAMP's intrinsic reward leads to max gain very much higher than most methods. I think it might also be beneficial to plot the distribution of rewards achieved by different methods, to robustly measure the range of performance achievable. After all, max is not a very robust estimate of the possible performance obtained by the policy.It also seems that Wasserstein based approach is much higher than KL - given that both are motivated by the RAMP narrative, it seems that the specific choice of metric is also very critical to the algorithmic performance. Do you think the underlying metric that defines Wasserstein distance is also critical, ie L2 vs L1 distance. Such ablations will be quite valuable to practitioners.### === Extrinsic reward ===In Table 2 where extrinsic rewards are combined, it seems that KL RAMP is better than Wasserstein RAMP in general, which is in opposite to the results in Table 1 where Wasserstein RAMP is generally better. Can you elaborate more on this?Also in general in continuous control tasks, it seems that exploration is not a defining factor to the final performance - as opposed to certain exploration heavy tasks in atari suites. As a result, it is not very clear if the gains in performance are due to the exploration bonus itself or rather due to some other confounding factors as a result of adding the corresponding loss.In practice, how would you choose the exploration vs. exploitation trade-off factor ($\\lambda$ and $\\beta$ factors in the algorithm), and are the algorithmic performance sensitive to the choice of such hyper-parameters?", "Soundness": 2, "Strengths": "The strength of the paper lies in a fairly clear presentation of the motivation and methodology. The idea of \"running away from the past\" is not strictly novel but the paper proposes an algorithmically viable way to instantiate such an idea. The paper presents a fairly clear math formulation and has carried out ablations on choices of the algorithmic designs. The experimental ablation also seems fairly comprehensive.", "Confidence": 4, "Weaknesses": "The idea of \"running away from the past\" is not strictly novel. From a theoretical standpoint, running away from old trajectories might not always be optimal and it is not clear theoretically what is gained by adopting such an approach. From an empirical standpoint, the ablations are carried out on the continuous control tasks, most of which do not seem to require extensive exploration to solve. It is not very clear if the claimed gains are really due to the exploration bonus, or some other unknown side effect.", "Contribution": 2, "Presentation": 3}
|
2024-11-03 18:05:04
|
ICLR.cc/2025/Conference
|
1xdvlSq9RV
|
zyGrziIVdE
|
Reviewer_FV3w
|
Official Review by Reviewer_FV3w
|
{"Rating": 3, "Summary": "The authors present a new algorithm for learning policies where the marginal distribution of states in a trajectory of length $T$ has a high entropy. Their method consists in iteratively maximizing intrinsic reward bonuses that measure a distance (metric) between the distribution of states of the current policy, and a geometric weighting of the distributions of states for the previous policies. That objective finds a motivation from an information theory property. Experiments compare the use of the KL-divergence and the Wasserstein distance to other algorithms.", "Questions": "1. Line 81, authors state that the methodology 'seamlessly' generalizes when T tends to infinity. From a theoretical perspective, does the limit exist without additional assumptions on the markov chain created from the MDP and the policy? From a practical point of view, are there any limitation to apply the method when T is large?2. Is a policy with maximum state entropy an optimal solution to the objective function that is maximized?3. There is a typo in equation (2): $\\rho^\\pi$ should be $\\rho^\\pi(s)$.4. The optimization problems for computing the intrinsic reward functions seem to be on-policy, is it the case? If it is the case, does it eventually result to on-policy optimization of control policies? If it is the case, it is worth mentioning that when used in combination with an off-policy RL algorithm for maximizing the intrinsic reward, addition interactions with the MDP are required making the modified algorithm on-policy. This point should be clear in Section 3.4.5. Could the authors clarify the arguments from paragraph line 329? I understand the philosophy of maximizing a lower bound on the entropy instead of directly maximizing the entropy. Yet, I think that both approaches incrementally improve the Shannon entropy, in opposition to the first sentence of the paragraph. I don't understand the argument of the generalization across behaviours.", "Soundness": 1, "Strengths": "1. The problem addressed is important to the community.2. The new objective function is theoretically motivated and provides new insights to compute good exploration policies.", "Confidence": 3, "Weaknesses": "1. In Section 2, different justifications for introducing the learning objective pursued by the agent are wrong or weak in several aspects:a. The justification line 108 for going from equation (1) to equation (2) is in my opinion wrong. Using the entropy of the policy as proxy to the entropy of the state distribution is a huge approximation. Maximizing the entropy of the policy does not provide a good state coverage in general nor in most practical cases. Note that if it was sufficient to maximize the entropy of the policy to get a uniform distribution of states, it would not be necessary to introduce a complex algorithm as the authors do.b. Line 128 authors justify to use the Wasserstein distance instead of the KL-divergence as the KL does not account for a potential geometry in the state space. This fact result from the original choice to define as exploration objective the entropy over the state space, which does not account for a potential geometry of the state space. So by choosing to maximize the Wasserstein distance instead of the KL, the authors change the original hypothesis that that the objective is to have high state entropy. While it can be discussed that it is a potential better framework to account for some geometry, it makes most of the previous mathematical justifications irrelevant.2. The authors claim in Section 3.4 that it is sufficient to optimize with any RL algorithm the reward model from Section 3.2 or Section 3.3 to maximize the objective equation (2) or equation (4). It is equivalent to neglecting the entropy of the policy. Authors, nevertheless, eventually use SAC, which is an algorithm that regularizes the MDP rewards with the log-likelihood of actions. This should be clarified.3. Only the final values are reported in the experimental section. From my personal experience, complex exploration methods may be unstable, and the learning curves provide important insights. Adding them in the paper would make the results more trustworthy.4. In the experiments, there is no statistical evidence that the method at hand outperforms the concurrent methods. Most confidence intervals overlap.5. I think that the related work should include [1, 2], and probably other, more recent, works.", "Contribution": 3, "Presentation": 2}
|
2024-11-03 20:09:07
|
ICLR.cc/2025/Conference
|
Gkkip187zs
|
zyGrziIVdE
|
Reviewer_65e4
|
Official Review by Reviewer_65e4
|
{"Rating": 3, "Summary": "The paper proposes RAMP (Running away from the past), an RL-based method for performing state space exploration by approximately maximizing either the KL divergence or Wasserstein distance between the current policy's state occupancy measure and the discounted sum of the state occupancy measures of all previous policies. This scheme aims to ensure that the state space coverage provided by the next policy is always maximally different from that provided by previously policies. The paper develops the RAMP method by deriving tractable proxies for these divergences, proposing reward models for each that can be used in conjunction with an RL algorithm, providing related approximation bounds, providing estimation schemes for each reward model based on existing work, and finally combining these steps to propose RAMP. Experimental results are provided that quantitatively illustrate what the reward models look like, compare RAMP with other intrinsic exploration approaches using a certain notion of state space coverage on a variety of tasks, and indicate that RAMP can be used as an exploration aid to accelerate extrinsic reward learning tasks.", "Questions": "1. Why is the notion of state coverage used in the experiments a good proxy for evaluating and comparison exploration? What is lost by ignoring the remaining dimensions in each of the environments considered?2. What is the value of $n$ in Fig. 2? Can you provide additional context about the rewards pictured in Fig. 2?3. Why is $r_W$ better than $r_{KL}$ in Fig. 2? This is mentioned in the paragraph starting line 377, but remains unclear.4. What do the colors represent in Fig. 3?5. Do the figures in Sec. 5.1 provide any insight into what is happening in the rest of the state space? Why not consider a visualization technique for visualizing high-dimensional data, such as $t$-SNE or PHATE plotting, instead of projecting onto x-y space?6. What do $\\pi$ and $\\pi'$ of Theorems 2 and 3 correspond to in the RAMP method and the remainder of the paper?7. How do Theorems 2 and 3 apply to the rest of the paper?**Important additional comment:** It is stated at several points throughout the paper (line 047, lines 325-327, 330-332, 467-469, 682-684) that state entropy maximization methods like APT [Liu & Abbeel, 2021] rely on probability density estimation. This is not accurate: APT and similar methods (e.g., Proto-RL [Yarats et al., 2022]) leverage non-parametric $k$-nearest neighbor entropy estimators, allowing them to maximize (proxies of) state occupancy measure entropy while avoiding density estimation.", "Soundness": 2, "Strengths": "Though the problem of state space exploration is very extensively covered in the RL literature, the proposed RAMP method provides what appears to be a novel approach to accelerating state space coverage. Due to its strategy of choosing policies maximizing divergence of state space coverage from that achieved by previous policies, it makes sense that RAMP will be more effective at rapidly exploring the state space than existing unsupervised RL methods (e.g., APT, SMM, Proto-RL) that simply maximize state occupancy measure entropy, and the experiments provide some support to this. Moreover, though the actual learning procedure used in RAMP is essentially a combination of existing techniques ([Eysenbach et al., 2020] for $r_{KL}$, [Durugkar et al., 2021] for $r_{W}$, and SAC [Haarnoja et al., 2018]), the combined approach detailed in Sec. 3.4 and Algorithm 1 appears to be novel and is interesting, and the fact that both KL-divergence and Wasserstein distance versions of RAMP are provided adds to its flexibility and significance. For these reasons, RAMP is likely of interest to the community and definitely merits further investigation.", "Confidence": 4, "Weaknesses": "Despite the strengths discussed above, I have concerns about the experimental evaluation and theoretical results:1. Most importantly, the \"state coverage\" performance metric upon which the comparisons of Sections 5.2 and A.1 rely is insufficiently justified as a good proxy for measuring exploration and for making fair comparisons between the algorithms considered. As described in the third paragraph of Sec. 5.2, this metric is obtained by discretizing the space of Euclidean (x-y or x-y-z) coordinates of the agent's state space, recording whether each grid cell has been visited or not during training, then returning the percentage of the grid cells that have been visited. There are two main issues with using this notion of state coverage as a proxy for exploration. First, the state space dimensions in most of the environments are far larger than 2 or 3 (e.g., 18 for HalfCheetah, 113 for Ant), and, for many of these environments, pose information other than location in Euclidean space (e.g., joint angles, velocities) is far more important for learning to operate within the environment and for specific downstream tasks. Second, recording only whether a grid cell has been visited or not ignores more complex visitation behavior, such as the empirical state visitation frequency defined at the beginning of Sec. 2. To render the state coverage metric used more meaningful, it would be helpful to include ablations over the other dimensions of $S$ or comparison with other coverage notions, such as Shannon entropy of the empirical state visitation frequency.2. Implementation details for the RAMP algorithm, the algorithms compared with, the discretization used in the state coverage metric, and other aspects of the experiments are not provided. The experimental results are therefore not reproducible in their current form. In addition, across all experiments, the lack of implementation details makes it difficult to assess the fairness of comparison with existing methods and even the comparisons between $RAMP_{KL}$ and $RAMP_{W}$. This makes it difficult to evaluate the significance of the experimental results, weakening the overall contribution. To remedy these issues, a thorough description of the implementation details is needed.3. The qualitative results in Sec. 5.1 are difficult to understand, leaving the practical differences between $r_{KL}$ and $r_W$ unclear. See the questions below for specific concerns.4. The connection between Theorems 2 and 3 and the rest of the paper is unclear, and the assumptions made are so strong as to immediately imply the results. For the former concern, a description of what $\\pi$ and $\\pi'$ of Theorems 2 and 3 correspond to in the RAMP method is missing, making it unclear how the results are meant to be applied. Regarding the second concern, it is assumed variously that $|| \\rho^{\\pi} - \\rho^{\\pi'} || \\leq \\varepsilon_0$, $|| \\hat{r} - r^{\\pi} || \\leq \\varepsilon_1$, and that the average reward $J_{\\hat{r}}(\\pi') = \\langle \\rho^{\\pi'}, \\hat{r} \\rangle$ is sufficiently larger than $J_{\\hat{r}}(\\pi) = \\langle \\rho^{\\pi}, \\hat{r} \\rangle$ to ensure that the desired inequalities hold. Under these assumptions, the proofs follow with some straightforward manipulation of inequalities. To make the results more consequential, it would be helpful to clarify how they are meant to be applied in the context of the paper, then weaken the assumptions accordingly.", "Contribution": 2, "Presentation": 3}
|
2024-11-04 22:35:59
|
ICLR.cc/2025/Conference
|
3SH3N8EZUm
|
Gkkip187zs
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "# Answer to Reviewer 65e4We thank the reviewer for their thorough feedback.## State coverageWe acknowledge Reviewer 65e4's point that using Euclidean coordinates may not be the most accurate way to measure state space coverage. As the reviewer suggests, we quantify coverage by discretizing specific environment variables with a $10^2$ matrix using 10 steps for the $xy$ coordinates, and we count whether each cell is filled. However, as the reviewer noted, most environments have far more dimensions; for example, Ant has 111 dimensions. Even with only 10 cells per dimension, the number of cells needed to cover the state space is $10^{111}$, which is infeasible to compute. This practical limitation is why we use density estimators; otherwise, we would simply count the number of times a cell is visited.The reviewer also suggests estimating the Shannon entropy of past occupancy measures. While this could provide insights, it is also an imperfect measure of exploration. First, accurately estimating Shannon entropy would require discretizing all state space variables, which is infeasible. Second, even a precise entropy measure may not capture exploration perfectly, motivating our use of the Wasserstein distance.For example, the Shannon entropy of a uniform distribution is at its maximum, but this does not imply that the distribution's support fully covers the state space. That said, we agree that the temporal evolution of the Shannon entropy of the mixture of occupancy measures could be useful for identifying if the agent becomes stuck in a local minimum. We will include these results in the ablation study.## Implementation detailsWe agree with the reviewer that a thorough description of implementation details is essential. We propose to add two sections in the appendix: one detailing the hyperparameters and implementation specifics of our method, and another detailing those of the baselines. All implementations and hyperparameters are available in our code repository: [Link for the github repository](https://anonymous.4open.science/r/Exploration_by_running_away_from_the_past_039B).## Qualitative results 5.1: Differences between $r_{\\mathcal{W}}$ and $r_{\\text{kl}}$The reviewer noted that Figure 2 does not clearly illustrate the difference between the Wasserstein distance and KL divergence. We propose to add a section in the appendix summarizing the definitions of the Wasserstein distance and KL divergence, along with their key differences.Our explanation of the two reward models may also be unclear, as we do not intend to suggest that one exploration method is superior to the other. The goal of Figure 2 is to show that maximizing $r_{\\mathcal{W}}$ leads to a different exploration pattern of specific variables in the state space than maximizing $r_{\\text{kl}}$. The intuition is that the density model of $r_{\\mathcal{W}}$ is Lipschitz continuous with respect to temporal distance, meaning that in environments like HalfCheetah, the main variations of this reward model are triggered by major changes in Euclidean coordinates.Consequently, each algorithm induces distinct exploration behaviors that may be beneficial for different downstream tasks, as discussed in the paper. For instance, exploration driven by $r_{\\mathcal{W}}$ may be more effective for goal-conditioned tasks where the objective is to reach a specific Euclidean position in the state space. In contrast, $r_{\\text{kl}}$ may provide more fine-grained exploration, focusing on the exploration of the agent's different joint configurations, which can be useful in dense reward maximization tasks like the HalfCheetah running task (see Section 5.3).- **Figure 3 Interpretation**: In Figure 3, the colors represent the output of the density model estimating $r_{\\mathcal{W}}$, as noted in the caption. We agree that adding a color bar would help clarify this and we will add it in the revised version.- **t-SNE Visualization**: Although we could use t-SNE to show which reward model emphasizes particular state space regions from a statistical perspective, our objective in this figure is to demonstrate that the exploration patterns induced by the two reward models differ, possibly impacting performance on different downstream tasks."}
|
2024-11-21 08:22:26
|
ICLR.cc/2025/Conference
|
Y47yIBZbvJ
|
3SH3N8EZUm
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "## Proof in the paperTheorems 2 and 3 were introduced to confirm that maximizing the reward models defined in the paper indeed maximizes the KL Divergence and Wasserstein distance described in Section 2. As a result, the assumptions underlying these theorems involve bounding estimation errors of the reward models, establishing a lower bound on the improvement step from policy $\\pi$ to policy $\\pi'$ (assuming the reward is maximized), and assuming limited change in the occupancy measure between policies $\\pi$ and $\\pi'$ (only for Theorem 2).We agree with the reviewer that a clear discussion of the assumptions relative to the algorithm would be beneficial. We will add an appendix section to elaborate on these assumptions. Nonetheless, these theorems were primarily introduced to validate that the reward models proposed in the paper indeed maximize the KL Divergence and Wasserstein distance as defined in Section 2.## Important additional comment: on the use of density estimators- **Probability density estimators**: Line 047: \"these approaches often rely on probability density estimators, and finding a relevant density estimator for an environment can be challenging.\" Indeed, these approaches often rely on probability density estimators. APT does not rely on a probability density function but still uses a density model.- **Density model**: The k-nearest neighbor approach used to estimate sparsity in the state representation embedding space is indeed a density model. We ask the reviewer for clarification on the statement \"APT and similar methods leverage non-parametric k-nearest neighbor entropy estimators\". To our understanding, the density estimator used (k-nearest neighbor) depends on representations produced by a parametric function."}
|
2024-11-21 08:22:40
|
ICLR.cc/2025/Conference
|
jvHqM1MqbG
|
1xdvlSq9RV
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "# Answer to Reviewer FV3wWe thank the reviewer for their thorough feedback.## Section 2### Assumption on Entropy of the Occupancy MeasureWe agree with Reviewer FV3w regarding the assumption that maximizing the expected entropy of the policies on the occupancy measure leads to high entropy of the occupancy measure. However, we explicitly state in the paper that this \"hypothesis may not universally apply across all environments\". Overall, our assumption is that the Kullback-Leibler divergence term in the lower bound is the most significant, and thus maximizing the entropy of the occupancy measure through the proposed proxy is not the primary driver of improvement. We will clarify this further in the next version.### From KL to the Wasserstein DistanceThe paper initially aims to maximize Shannon entropy, which leads to maximizing the KL divergence between the current occupancy measure $\\rho^{\\pi}$ and the mixture $\\beta\\rho^{\\pi} + (1-\\beta)\\mu_n$, where $\\mu_n$ represents the past occupancy measures. It is only upon introducing this KL divergence that we consider the potential advantages of using a divergence that captures the overall geometry of the state space. Could the reviewer clarify why this would make the derivation of the KL divergence objective irrelevant?Given the objective involving a divergence measure between two distributions, we believe that the study of alternative methods to quantify the gap between distributions is founded, but would enjoy further clarification.## Using the SAC AlgorithmIndeed, we used SAC to maximize all reward models. We agree that the regularization used in SAC aligns well with our objective of maximizing the expected entropy of the policies on the occupancy measure, and we will clarify this in the paper.## Best PolicyWhen using intrinsic reward models to maximize an extrinsic reward, the learning curve can oscillate as the agent explores new areas, resulting in high intrinsic reward values. As a result, we believe that the most relevant way to benchmark these algorithms is to consider the best policy discovered since the beginning of training. However, we agree that analyzing the temporal evolution of the learning curves could provide additional insights into the exploration behavior of the algorithms.## Statistical Significance: Exploration and Exploitation Assessment of the ResultsWe agree with the reviewer that the statistical significance of the results should be clearly presented. There is a statistically significant difference between the methods in many environments and tasks, although this may not be sufficiently apparent in the current version. We propose to highlight in bold the method for which the paired t-test with the second-best method indicates a significant difference.## Related WorkThe reviewer cites [1] and [2] but it appears that the references were not included. Would it be possible to include them?"}
|
2024-11-21 08:25:07
|
ICLR.cc/2025/Conference
|
uXl59HfDoY
|
mQtVHAFXdK
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "# Answer to Reviewer YHscWe thank the reviewer for their thorough feedback. To ensure we fully understand each of the comments, we would appreciate it if the reviewer could confirm that our interpretations are correct.## Intrinsic Reward Alone* **Table 1**:Table 1 shows the final coverage reached by each method on each environment across 5 different seeds.* **Distance Used for the Wasserstein Distance**:The distance used for the Wasserstein distance is the temporal distance, as explained in the paper. The temporal distance (while not strictly adhering to the properties of a distance function) is sensible in any environment that can be represented as a Markov Chain, which makes the intuition behind this algorithm generalizable to a wide range of environments. As the goal of this study is to derive a general algorithm applicable to a broad range of environments, we believe that the temporal distance is a suitable choice for the general case. Further study on the impact of using $l_2$, $l_1$, and temporal distances for the Wasserstein distance could benefit specific applications, but we find it out of the scope of this article.## HyperparametersWe agree that further details on hyperparameters would be useful. We purposefully chose a small set of hyperparameters to explore to underlie the robustness of the model to different hyperparameter choices. A more exhaustive study of the possible hyperparameter space would benefit the paper and we will consider it, depending on computational costs. We will, however, provide a more detailed explanation of the hyperparameters used in the appendix."}
|
2024-11-21 08:25:54
|
ICLR.cc/2025/Conference
|
JwSUh2nrMF
|
KgCysAmNli
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "# Answer to Reviewer nQGLWe thank the reviewer for their thorough feedback.## Related Work on Epistemic UncertaintyWe appreciate Reviewer nQGL's suggestion to discuss the relationship between our work and methods that leverage epistemic uncertainty. We agree that this is an interesting aspect of exploration, and we plan on including a discussion of these methods in the related work section of the revised manuscript. However, the experimental comparison uses 10 baseline methods that we believe represents a wide range of the exploration literature, and adding further experimental comparisons is out of the scope of this article.## HyperparametersAs discussed in the response to reviewer YHsc, we plan on providing a more comprehensive explanation of the hyperparameters used in the appendix and will discuss the impact of hyperparameter choices.## Exploration in POMDPsOur primary motivation for reframing entropy maximization through the KL divergence was to use a density estimator known to scale effectively in high-dimensional problems, such as visual control tasks. More specifically, by framing the problem as KL maximization, the density model that makes the most sense to use is a classifier, which is known to scale well in high-dimensional settings. Although we do not test our method in POMDPs, we thank the reviewer for the suggestion and will include a discussion on POMDPs as a prospect for future work."}
|
2024-11-21 08:26:19
|
ICLR.cc/2025/Conference
|
zhflrolxM7
|
JwSUh2nrMF
|
Reviewer_nQGL
|
Response by Reviewer
|
{"Title": "Response to the author's rebuttal", "Comment": "I thank the authors for their response.**Epistemic Uncertainty**: While I appreciate the authors including epistemic uncertainty-based exploration methods in the discussion. I do not understand how they are considered out of scope for this paper. Particularly, given the theoretical connection between maximizing information gain and entropy and that the epistemic uncertainty-based methods are widely applied for intrinsic exploration. Moreover, I think these methods constitute a strong baseline that should be compared against.**Hyperparameters**: Thanks for including this. I would like to look at the proposed changes before reevaluating my score.**Exploration in POMDPs**: While I can follow the arguments made by the reviewer, I would require empirical evaluation to be fully convinced. However, I acknowledge that this might be out of the scope of this work. Overall, I would like to keep my score as I believe my concerns, particularly for the first two points above, are not yet fully addressed."}
|
2024-11-21 19:42:04
|
ICLR.cc/2025/Conference
|
hO6YX4PH33
|
Y47yIBZbvJ
|
Reviewer_65e4
|
Response by Reviewer
|
{"Title": "", "Comment": "Thanks to the authors for their response. First, I appreciate the sharing of the code for the submission, which partially addresses my concerns raised in Weakness 2. I am also grateful for the clarification regarding the differences between $r_W$ and $r_{KL}$, which provides some intuition that partially addresses Weakness 3. In addition, I understand the nuances associated with quantifying state space coverage, which the reviewers have described in their response to Weakness 1.Nonetheless, several of my original concerns remain: (1) the \"state coverage\" performance metric upon which the experimental results presented in Table 1 rely is still insufficiently justified and no concrete justification or alternative performance metric has been provided (Weakness 1); (2) though the code has been released, no implementation description has been provided, making it difficult to fully understand and evaluate the experimental results (Weakness 2); (3) exactly how the theoretical results apply to the rest of the paper (Weakness 4) remains unclear. Finally, it appears that no revision has yet been provided (**please let me know if this is wrong or if the revision is subsequently uploaded**), so several of the items proposed in the author response remain incomplete. For these reasons, I maintain my original score."}
|
2024-11-25 21:51:03
|
ICLR.cc/2025/Conference
|
axDhB9sZck
|
jvHqM1MqbG
|
Reviewer_FV3w
|
Response by Reviewer
|
{"Title": "Follow-up", "Comment": "Thank you for responding to my review. I will clarify some elements.1. I don't believe that the KL divergence objective is irrelevant at all, sorry for the misinterpretation. I believe that the argument about the 'geometry ' is not completely correct. In my opinion, the KL objective is one you use when you do not make any assumption on the geometry. When you assume there exist a structure, you may then rely on another objective. But I don't think that when you assume there is no structure, it makes sense to use the Wasserstein objective, does it?2. Concerning the additional work, I was referring to [1] and [2], but many more can be found.Overall, I don't think you response fully answer my remarks and I believe that yout manuscript would benefit from deeper modifications.[1] Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Alaa Saade, Shantanu Thakoor, Bilal Piot,Bernardo Avila Pires, Michal Valko, Thomas Mesnard, Tor Lattimore, and R\u00b4emi Munos. Geometricentropic exploration.[2] Chuheng Zhang, Yuanying Cai, Longbo Huang, and Jian Li. Exploration by maximizing r\u00b4enyientropy for reward-free rl framework."}
|
2024-11-26 15:45:53
|
ICLR.cc/2025/Conference
|
ZqwXdV3JLP
|
uXl59HfDoY
|
Reviewer_YHsc
|
Response by Reviewer
|
{"Title": "reply", "Comment": "Thank you to the authors for the reply.Thanks for clarifying on some technical details that I missed from first reading the paper. It will indeed be helpful having access to a more detailed explanations of the hyper-parameters used for experiments.I will adjust my scores after the discussion phase."}
|
2024-11-27 14:53:11
|
ICLR.cc/2025/Conference
|
WHGGAhftAV
|
zyGrziIVdE
|
Area_Chair_dHcV
|
Meta Review of Submission7454 by Area_Chair_dHcV
|
{"Meta Review": "This paper proposes an exploration strategy by maximizing the Shannon entropy of the state occupancy measure. This is achieved by maximizing a measure of divergence between successive state occupancy measures. The authors argue for the efficacy of their method by evaluating on a set of mazes and robotic manipulation and locomotion tasks.The paper proposes a novel method that is theoretically motivated, for a problem that is of general interest to the RL community.However, there is a general consent that the work is still lacking in its empirical evaluation. Further, there was a point raised on the bolded numbers in Table 1, which indeed seems to include performance gains which are not statistically significant.As such, I will not be recommending acceptance of this work, and recommend the authors incorporate the reviewer feedback to further strengthen this submission.", "Additional Comments On Reviewer Discussion": "The main concerns raise were with regards to empirical evaluations, as discussed above. There were a few points raised on the theoretical underpinnings and framework of RAMP, but most of these were successfully addressed by the authors."}
|
2024-12-19 01:51:56
|
ICLR.cc/2025/Conference
|
YZIp2aNPWj
|
zyGrziIVdE
|
Program_Chairs
|
Paper Decision
|
{"Comment": "", "Decision": "Reject"}
|
2025-01-22 05:30:39
|
ICLR.cc/2025/Conference
|
fxl5YNtkzp
|
zxqdVo9FjY
|
Reviewer_YeXr
|
Official Review by Reviewer_YeXr
|
{"Rating": 6, "Summary": "Motivated by the problem of training the readout of a two-layer network after on large gradient step on the first layer, the authors consider the problem of linear regression on a spiked data model. They provide a characterization of the test error, for two linear target functions, respectively depending on only the spike part or the complete input. They discuss how for the latter, the spike does not asymptotically influence the test error, but does non-asymptotically.", "Questions": "- In the discussion above 4.1, the assumption $\\theta=\\tau^2 n$ seems again to contradict the assumption $\\theta^2/\\tau^2\\ll n$ in the statement of the Theorem. Furthermore, this seems to correspond to a strong spike regime, how can the authors recover the spikeless results of Hastie et al. (2022) in this case? This might be a misunderstanding on my side, but more discussion would be beneficial.Minor:- l.074, 347 : incomplete sentences- l.203 I think $\\ell_j$ is not defined- In 3.3, more discussion in the main text about why only the unregularized case is considered for the spn case, while generic $\\mu$ is considered for the signal-only model, would be helpful for intuition, whether it is for technical reasons or because it is not interesting.", "Soundness": 3, "Strengths": "The paper is well written, clearly motivating the study, and describing in simple terms the model considered. Sufficient intuition is provided at most steps of the discussion. The technical results are clearly exposed and sufficiently discussed. Although I did not go through the proof in detail, the technical results seem scientifically sound.", "Confidence": 3, "Weaknesses": "I have a number of concerns related to the technical discussions, and the relation to previous works, which I detail below, and in the question section. These concerns regard the discussion of the main theorems, and not the theorems themselves, although I have not carefully verified the proof. These concerns prevent me from giving a higher score to this submission. On the other hand, I would be very happy to increase my score, were those concerns to be addressed by the authors.- The authors claim l.083 that Moniri et al. (2023) do not quantify the test error after one gradient step. To the best of my understanding, they do provide such a tight characterization (Theorem 4.5). Could the authors clarify their claim, and emphasize how their work is positioned with respect to Moniri et al. (2023)?- I find the discussion l.332-355 somewhat confusing, as they discuss the specialization of Theorem 3 for $\\theta=\\tau\\sqrt{n}$. Doesn't this directly contradict the assumption $\\theta^2/\\tau^2\\ll n$ stated in the Theorem? Since this specialization leads to one of the main qualitative results of the paper (namely the spike only affects the test error in non-asymptotic cases), this point would gain to be clarified. The same holds for l.452 with respect to Theorem 4.- I believe more intuition about the different scaling considered would help solidify the intuition for the spn case, regarding when the spike matters. In particular, the authors could for instance recall the scaling of the terms $z_i^\\top\\beta_*$, $a_i^\\top \\beta_*$, and emphasize their respective strengths in the different scalings of $\\theta, \\tau$ considered. I am curious if the signal part is much smaller than the second term when the spike has no effect, or if the argument is more subtle.", "Contribution": 3, "Presentation": 4}
|
2024-10-21 17:29:33
|
ICLR.cc/2025/Conference
|
PoAfoEMac0
|
zxqdVo9FjY
|
Reviewer_xFzE
|
Official Review by Reviewer_xFzE
|
{"Rating": 5, "Summary": "This paper analyses the generalization error of linear regression with spiked covariance. Previous literature has been using asymptotic limit of the empirical spectral density to analyse the generalization error of linear regression. At the limit, the effect of the spike vanishes. However, it is not the case for finite sample size. This paper fills the gap by showing there is a correction term for finite sample size $n$.", "Questions": "Regarding the weaknesses mentioned above, I would like to ask:1. What novel results could the authors conclude in the feature learning setting in neural networks using the main theorems 3,4?2. How could the authors show the assumption on the dependence does not affect the result? Is there any experimental validation?3. From Figure 4.1, we can see that the effect of the spike correction term is small when $n$ is large. Is the main theorem still useful to explain the phenomenon we see from feature learning?", "Soundness": 2, "Strengths": "This paper provides a detailed proof of their main theorems with clearly stated definitions. It extends over previous results like [1,2].[1]: Trevor Hastie, Andrea Montanari, Saharon Rosset, and Ryan J Tibshirani. Surprises in highdimensionalridgeless least squares interpolation. Annals of statistics, 50(2):949, 2022.[2]: Xinyue Li and Rishi Sonthalia. Least squares regression can exhibit under-parameterized doubledescent. Advances in Neural Information Processing Systems, 2024.", "Confidence": 3, "Weaknesses": "However, this paper has some obvious weaknesses:1. The paper is motivated by the spiked covariance from the one-step gradient feature learning in neural networks (Section 1). However, it did not show how the results can be applied to the feature learning scenario. I question the amount of contribution this paper provides.2. The assumption in line 2221-222 and 253-255 is too strong. The analysis breaks down if there is dependency in the cross term. However, the paper did not show how big the difference the predicted result would be when there is dependence in the cross term. It is questionable if the result in this paper is applicable in realistic machine learning settings.3. This paper has problems with the wordings, even in main theorems. This makes the reading difficult. For instance:> Theorem 3 (line 313): ...Then, any for data $X\\in\\mathbb{R}^{n\\times d}, y\\in\\mathbb{R}^n$ from the signal-plus-noise model that satisfy: $1\\ll \\tau\\_{A_{trn}}^2,\\tau_{A_{tst}}^2\\ll d, \\theta_{trn}^2/\\tau_{A_{trn}}^2<<n, \\theta_{tst}^2/\\tau_{A_{tst}}^2 << n_{tst}$. Then for $c<1$,...The first sentence needs to be rephrased and the symbol $\\ll$ is not consistent. Also, there are some typos like:> line 350: Hence the spike has does not have an effect...> line 372, ... we see an affect that...", "Contribution": 2, "Presentation": 2}
|
2024-10-24 13:54:32
|
ICLR.cc/2025/Conference
|
nnuWhTQ0sX
|
zxqdVo9FjY
|
Reviewer_LgJ3
|
Official Review by Reviewer_LgJ3
|
{"Rating": 5, "Summary": "The paper considers the linear least squares regression for data with simple spiked covariance. They quantify the empirical risk of test data.", "Questions": "1. Could you provide a reference for the statement, 'It has been shown that to understand the generalization...' on line 39?2. Is your generalization analysis very different from the work of Li & Sonthalia (2024)?", "Soundness": 3, "Strengths": "1. They construct two linear regression problems with spiked covariance.2. They well explain the previous work of Moniri et al. (2023).3. Precise quantification of the generalization errors are also provided for both model.", "Confidence": 4, "Weaknesses": "1. They reference the work of Moniri et al., but this work is unrelated to neural networks or gradient descent; it addresses a purely linear regression problem for data with simple spiked covariances.2. They do not account for the generalization ability of neural networks after a single gradient step, as they bypass the gradient step entirely by assuming the W1 matrix directly, which does not reflect the full process of neural network training.", "Contribution": 2, "Presentation": 3}
|
2024-11-01 16:43:37
|
ICLR.cc/2025/Conference
|
5x6iVSaqHT
|
zxqdVo9FjY
|
Reviewer_qVHv
|
Official Review by Reviewer_qVHv
|
{"Rating": 3, "Summary": "Motivated by a recent work studying two-layer neural networks (Moniri et al., 2023), the paper studies linear regression under a data model with a spiked covariance (Couillet & Liao, 2022). The spiked covariance consists of a spike component (signal) and a bulk component (noise). Thus, the authors characterize the risk (a.k.a generalization error) with a specific focus on the effect of the spike. They find that the spike does not impact the risk in the underparameterized case. In contrast, the spike introduces an additional term (called \"correction term\") in the risk for the overparameterized case. However, they mention that the correction term is of order $O(1/d)$, which vanishes in the asymptotic case. Thus, the spike does not affect the risk in the asymptotic case but does in the finite case. Then, the authors focus on a case where the targets $y$ only depend on the signal (spike) component of inputs $\\mathbf{x}$ in order to highlight the effect of the spike on the risk. In this case, the correction term depends on the alignment between the spiked eigenvector $\\mathbf{u}$ corresponding to the spike and the target function $\\boldsymbol{\\beta}$. Furthermore, the paper illustrates how the generalization error for this setting exhibits the so-called double-descent phenomenon with a formula for the peak location (a.k.a interpolation threshold).", "Questions": "1. In Line 178, $F_1$ denotes the case with a single spike (as shown by Moniri et al., 2023). However, Moniri et al., 2023 showed that $F_1$ can include multiple spikes, and the number of spikes depends on the step size of the gradient step. Where is the discussion about the effect of step size in this paper? Similarly, where is the discussion on the impact of $o(\\sqrt{n})$ term for $F_1$?2. What is $l_j$ in Theorem 2 (Line 204)? Do the authors mean $l$?3. In footnote 3 (Line 266), the authors say \"... the limiting e.s.d for $F_0$ is not necessarily Marchenko-Pastur distribution ... This difference is not too important, as instead of using the Stieltjes transform for the Marchenko-Pastur distribution in our paper, we could use the result from P\u00e9ch\u00e9 (2019); Piccolo & Schr\u00f6der (2021) instead.\" Why wouldn't the authors directly use the mentioned result directly?4. Why is there no regularization for the signal-plus-noise problem when there is regularization for the signal-only problem (Line 278-285)?5. Typo in Line 285: \"We consider on the instance-specific risk.\". Typo in Line 313: \" Then, any for data ...\". 6. Undefined symbols in Theorem 3 (Line 312 - 324): $\\asymp$ and $<<$.7. How do the authors arrive at \"Hence, we see that if the target vector y has a smaller dependence on the noise (bulk) component A, then we see that the spike affects the generalization error.\" in Line 380? Its connection to the previous part seems to be missing.8. How do the authors come up with the equation for the peak point of double descent in Line 477? Is it an empirical observation or a theoretical result?", "Soundness": 2, "Strengths": "* The motivation for this paper is good since the recent line of work studying two-layer neural networks after one gradient step (Ba et al., 2022; Moniri et al., 2023) has received significant attention.* The authors precisely characterize generalization errors (risk) for two linear regression problems with spiked covariance data, while the problems differ regarding the target function. + They provide bias and variance decomposition of the risk. + They illustrate the \"double-descent phenomenon\" and provide a formula for the peak location (a.k.a interpolation threshold) of the double-descent phenomenon, which is beneficial for understanding the phenomenon. + The authors specifically focus on the impact of the spike (in the data model) on the risk for different cases. Thus, they show when and how the spike affects the generalization error.", "Confidence": 4, "Weaknesses": "* The presentation in this paper is not good + Although the paper is motivated by Moniri et al. (2023), there are significant discrepancies between the setting of this paper and that of Moniri et al. (2023), as the authors mention in Section 5. While Moniri et al. (2023) considered two-layer neural networks after one gradient step under isotropic data assumption, this work considers linear regression under spiked covariance data assumption. There exists a relationship between these two, but they are not exactly the same. For example, there is a difference between the target $y$ generation of the two settings. Furthermore, $\\mathbf{A}$ (noise component) and $\\mathbf{Z}$ (spike component) are dependent in the case of Moniri et al. (2023), while the dependence is ignored here (see lines 251-255). + Some notations are used without definition (e.g, $\\delta_{\\lambda_i}(\\lambda)$ in Line 126, or $\\Sigma(d_k)$ in Line 147). + There are significant typos in equations. For example, $y$ should be a scaler in Line 76, but it is written as a vector, which makes the equation wrong. Another example is that $l_j$ in Theorem 2 (Line 204) is not defined, and I think the authors meant $l$ instead of $l_j$. A third example is that the function $R_{spn}(c;\\tau,\\theta)$ defined in Line 301 and its usage $R_{spn}(c,0,\\tau)$ in Theorem 3 (Line 317-321) are different in terms of parameters.* Limited contribution/novelty + Most of the results in this paper are trivial extensions of the results by Hastie et al. (2022) and Li & Sonthalia (2024), which significantly limits the novelty and originality of the paper. Note that Hastie et al. (2022) studied linear regression under a generic covariance assumption with bounded eigenvalues. Here, some eigenvalues can diverge as dimensions go to infinity, but this case is also covered by Li & Sonthalia (2024). + There exists a related work (Cui et al., 2024) that is not mentioned in this paper. Cui et al. (2024) characterized the generalization error (risk) for two-layer neural networks after one gradient step under isotropic data (same setting as that of Moniri et al. (2023)). Although there exist methodological differences between (Cui et al., 2024) and this paper, the motivations are the same, and their settings are similar. + During the review period of this paper, a related work (Dandi et al., 2024) that can be considered as follow-up of (Cui et al., 2024) was appeared on arXiv. While Cui et al. (2024) used (non-rigorous) replica method from statistical physics for their analysis, Dandi et al. (2024) studied the same setting with random matrix theory, which is also the main tool in this paper. Therefore, this paper and (Dandi et al. 2024) studied similar settings with similar methodologies. Note that since (Dandi et al., 2024) appeared after the submission of this paper, I am only mentioning it for the sake of completeness.Overall, I think this paper should be rewritten with more focus on the impacts of the spike covariance on the generalization error of linear regression, and the new presentation should clearly differentiate the current work from the work by Hastie et al. (2022), Li & Sonthalia (2024), Cui et al. (2024), and Dandi et al. (2024).Cui et al. (2024): Asymptotics of feature learning in two-layer networks after one gradient-step. (ICML 2024)Dandi et al. (2024): A Random Matrix Theory Perspective on the Spectrum of Learned Features and Asymptotic Generalization Capabilities.", "Contribution": 1, "Presentation": 1}
|
2024-11-01 18:13:30
|
ICLR.cc/2025/Conference
|
WqthOZZiOY
|
zxqdVo9FjY
|
Reviewer_weJN
|
Official Review by Reviewer_weJN
|
{"Rating": 5, "Summary": "The authors analyze the generalization properties of spiked covariate models. The theoretical analysis is motivated by recent works on two-layer networks trained with a single gradient step that showed how the feature matrix possesses different spikes associated with the learning rate scaling used in the optimization step. The proof scheme uses tools coming from random matrix theory that enables the asymptotic computation of the generalization error. The theoretical claims are accompanied by coherent numerical illustrations.", "Questions": "As hinted above my main concern on this manuscript is the close relationship with previous works, namely (Ba et al., 2022; Moniri et al., 2024). Could the authors comment on the link between their results and (Ba et al. 2022) in the context of Gaussian Universality (see e.g. [1]) ? From my understanding of their paper, i.e. a single spike in the feature matrix, they show that in the learning rate regime considered in this paper Gaussian Universality should hold. There is indeed an extensive regime of learning rates after the BBP transition that still falls under the umbrella of Gaussian models, resulting in effectively \"linear\" generalization properties. One additional weakness of this submission is the related works coverage. The authors do a great job in covering the random matrix theory literature, while many manuscripts that analyze learned representations with gradient descent with different tools are not properly mentioned, see e.g. [2,3,4]. Although in these works the authors do not focus on the exact asymptotic calculation of the test error, many insights should translate to the present setting. On the other hand, [5] precisely characterize the generalization error using non-rigorous methods; what is the relationship with the present work?The results in the present submission should relate directly to the ones in Section 4 of (Moniri et al. 2024), albeit the differences correctly reported by the authors in the two settings. Could the author elaborate on this? What is the bottleneck for the present thereotcial tools to analyze multiple spikes (corresponding to higher learning rate scaling in Moniri et al. 2024)? Closely related to the above, [5] worked along the lines of (Moniri et al. 2024) to provide the equivalent description in the regime where the spikes recombine with the bulk (maximal scaling regime). Do the authors see a possible extension of their analysis to this scaling? - [1] Hu & Lu 2022, Universality laws for high-dimensional learning with random features. - [2] Damian et al. 2022, Neural networks can learn representations with gradient descent. - [3] Dandi et al. 2023, How two-layer neural networks learn, one (giant) step at a time.- [4] Ba et al. 2023, Learning in the presence of low-dimensional structure: a spiked random matrix perspective.- [5] Cui et al. 2024, Asymptotics of feature learning in two-layer networks after one gradient-step.", "Soundness": 2, "Strengths": "The paper is nicely written. The mathematical claims are correctly phrased and the numerical illustrations are coherent with the main text. The research problem is relevant in the theoretical machine learning community.", "Confidence": 3, "Weaknesses": "My main concern with the present submission is the lack of clear elements of novelty. The paper heavily relies on results coming from related works and it restricts their setting in many ways (as fairly reported by the authors at the end of the manuscript). More details are provided below.", "Contribution": 1, "Presentation": 3}
|
2024-11-07 14:46:20
|
ICLR.cc/2025/Conference
|
eztAgKUYJ0
|
zxqdVo9FjY
|
Authors
|
Response by Authors
|
{"Title": "Introducing Dependency Between Bulk and Spike", "Comment": "A common criticism among reviewers was our abstraction of the dependency between bulk and spike components. Here we demonstrate how our proof framework extends to handle the dependent case from Moniri et al. 2023.Recall that Moniri et al.'s spike structure is:$$ \\sigma(W_0\\tilde{X}^T) + c (\\tilde{X}\\beta_{sp}) \\zeta^T $$where $\\tilde{X}$ is Gaussian data, $W_0$ is inner layer weights, and $\\zeta$ are outer layer weights.We modeled this as:$$ A + \\theta vu^T $$where $A$ is Gaussian. Below we show how our analysis extends to Moniri et al.'s setting.## Introducing DependenceFirst, consider the intermediate structure:$$ X = A + \\theta (A\\beta_{sp})u^T $$To analyze this, we only need to modify two parts of our proof:1. In Lemma 10, the norm of $v^T A^\\dag$ scaling changes: - Original: $\\mathbb{E}_{\\lambda}\\left[\\frac{\\lambda}{(\\lambda + \\mu^2)^2}\\right]$ - New: $v = A\\beta\\_{sp}$, and norm becomes $\\|\\beta\\_{sp}\\|$ where expectations are over the Marchenko-Pastur distribution.2. The variable $t = (I-AA^\\dag) v$ is no longer zero.For the Signal-Only case, this gives bias:$$ \\frac{\\theta_{tst}^2}{n_{tst}}\\left[(\\beta_*^T u)^2 + \\tau_{\\varepsilon}^2\\left(\\frac{(1+c)}{2T_1} + \\frac{\\mu^2 c - T_1}{2\\tau_{A_{trn}}^2 T_1} \\right)\\right] $$where $T_1$ is unchanged. We do not present the whole formula for brevity. To extract insights, we consider the same simplifications for the paper.Under simplifications ($\\mu = 0$, $\\tau_{A_{trn}} = \\tau_{A_{tst}}$, $\\theta = \\tau \\sqrt{n}$), for $c > 1$ we get:$$ \\tau_{A}^2 (\\beta_*^Tu)^2\\left(1+\\frac{\\tau_A^2}{c}\\|\\beta_{sp}\\|^2\\right) + \\tau^2_{\\varepsilon}\\frac{c}{c-1}\\left(2 + \\frac{\\tau_A^2}{c}\\|\\beta_{sp}\\|^2\\right) $$Note: Due to the extra $A$ factor, this only holds for $\\tau_A = \\Theta(1)$ versus our original $\\tau_A = O(\\sqrt{d})$. Note: this is the Signal only version so $y = \\theta (A \\beta\\_{sp}) u^T\\beta\\_*$. ## Full Moniri et al. StructureNow consider:$$ X = \\sigma(W_0A^T) + c (A\\beta_{sp}) \\zeta^T $$The risk becomes a random variable dependent on $W_0$, $\\zeta$, and $\\beta_{sp}$. Using standard assumptions (isotropic with unit expected norm for the $\\zeta$, and $\\beta_{sp}$ and the rows of $W_0$), we analyze the expected risk. Importantly, since $W_0$ and $(\\beta_{sp}, \\zeta)$ are independent, the bulk remains independent of the spike. This is because functions of independent random variables are independent. Hence, our assumptions are reasonable. To get the generalization error for this model, we need to replace Lemmas 7-9. As an example, the new Lemma 7 resembles equations from Moniri et al. (Eq. 5) and Ba et al. (Eq. C.23):**Lemma 7:** For $W_0$ ($m \\times d$), $X$ ($n \\times d$), $m < n$, with $d/n \\to \\phi$, $d/m \\to \\psi$, $m/n \\to c$:1. $\\mathbb{E}\\left[\\frac{1}{\\lambda+\\mu^2}\\right] = \\frac{c}{\\tau_A^2}m_c\\left(-c\\frac{\\mu^2}{\\tau_A^2}\\right)$2. $\\mathbb{E}\\left[\\frac{1}{(\\lambda+\\mu^2)^2}\\right] = \\frac{c^2}{\\tau_A^4}m_c'\\left(-c\\frac{\\mu^2}{\\tau_A^2}\\right)$3. $\\mathbb{E}\\left[\\frac{1}{(\\lambda+\\mu^2)^2}\\right] = \\frac{c^3}{2\\tau_A^6}m_c''\\left(-c\\frac{\\mu^2}{\\tau_A^2}\\right)$where $m_c(z)$ satisfies:$$\\frac{\\psi}{z} H(z) - \\frac{\\psi - 1}{\\psi} = m_c(z)$$$$H(z) = 1 + \\frac{H^\\phi(z)H^\\psi(z)(c_1 - c_2)}{\\psi z} + \\frac{H^\\phi(z) H^\\psi(z)c_2}{\\psi z - H^\\phi(z)H^\\psi(z)c_2}$$with $H^\\kappa(z) = 1 - \\kappa + \\kappa H(z)$---------Advantages of our approach---------This approach avoids using the Gaussian Equivalence property, providing finer control over finite matrix approximation errors. While we must restrict $\\tau_A$ and $\\theta$ magnitudes and take expectations over $W_0$ and $\\zeta$, this allows us to:1. Better understand finite matrix effects2. Explore different target functions than Ba et al. and Moniri et al.We are happy to represent the corresponding results for the Signal Plus Noise case and the missing details. We presented a shortened version for brevity. ----------We are currently still working on the revision and will post it shortly."}
|
2024-11-14 01:45:48
|
ICLR.cc/2025/Conference
|
MQlWbuDYAI
|
WqthOZZiOY
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for the feedback and comments. Key differences between our work and important prior research are that we (1) provide finite matrix correction terms and (2) offer simplified closed-form expressions.> Could the authors comment on the link between their results and (Ba et al. 2022) in the context of Gaussian Universality (see e.g. [1]) ? Yes, the problem we and prior work are interested in understanding is the generalization error for the following. First, we solve a regression problem$$ \\beta\\_{LS} = argmin ||y - \\beta^T F||\\_F^2 + \\lambda ||\\beta||\\_2^2 $$Then, we are interested in the generalization performance of $\\beta\\_{LS}$. Let's call this risk $R(F)$ to highlight the dependence on $F$. The difference between setups lies in the $F$ term. There are three different $F$'s considered.1. $F_{CK} := \\sigma(WX)$. For Gaussian $X$ and $W$, after taking a step of GD. This is from Ba et al. 20222. $F_{CE} := \\theta_1 WX + \\theta_2 A$ where $A$ has IID standard Gaussian entries independent of $W,X$ and $W$ is after taking gradient step of GD3. $F_{SP} := A + \\theta uv^T$, where $A$ has IID standard Gaussian entries. The Ba et al. 2022 paper shows that in the small learning rate regime, $R(F_{CK}) = R(F_{CE})$ **asymptotically**.However, we do things differently:1. We allow spikes from the large learning rate limit. Hence, the result from Ba et al. 2022 does not apply. Equation 3.1 in Ba et al. shows that for small learning rate, the size of the spike is $\\Theta(1)$, where as for large learning rates, it is $\\Theta(\\sqrt{d})$ (note for the rank one spike the Frobenius norm is equal to the spectral norm). We are interested in the case when the spike is large. The idea behind the large step size is that we are in a regime in **which the Gaussian Equivalence Property is no longer true**. 2. We provide more precise correction terms for finite matrices. While the prior work is purely asymptotical. 3. In our rebuttal, we also generalize to models closer to that from Moniri et al. > One additional weakness of this submission is the related works coverage. We thank the reviewer for pointing us to these works. We shall add these references. [5] characterizes the risk for the setting from Moniri et al. 2023 using, as the reviewer and the paper says, using the non-rigorous replica symmetry method. The differences are three-fold:1. Our results in the paper are for a restricted setting; however, we provide proof. 2. We simplify expressions. For example, $\\zeta$ in equation 17 in Cui et al. 2024 is exactly $\\xi - 1$ in our paper (see Lemma 13 in the appendix for a definition). The expression in Cui et al. 2024 is left in terms of $\\zeta$. **This is because the results are a product of dependent terms. Hence, simplification is not easy**. However we a. Compute the expectations and variances of each of the terms b. Compute the expectations of the products. c. Greatly simplify expressionsWe believe these are the main challenges we overcome in our proof.> Could the authors elaborate on the connection with Moniri et al. 2024?While Theorem 4.5 in Moniri et al. 2023 shows that the difference between the test error using $F_1$ (features after 1 step) and $F_0$ (initial features) converges to a constant, there are important differences in our approach:1. Moniri et al. 2023 has expressions requiring solutions to fixed point equations (see equations (5) in their paper). These only hold asymptotically with hard-to-quantify approximation rates. 2. In contrast, we provide: - **Closed form expressions** for the risk itself (not just differences as is the case in Moniri et al. 2023) - **Better control on approximation error rates**, enabling analysis of finite matrices - The above two allow better control on understanding the relationship between the bulk and spike. > What is the bottleneck for analyzing multiple spikes?The analysis here is similar to Sonthalia and Nadakuditi 2023. Kaushik et al. 2024 extend Sonthalia and Nadakuditi 2023 to the higher rank version. We would need to do the same to extend to multiple spikes. As mentioned before, the difficulty in the analysis was bounding variances. These are currently scalar expressions hence we can use commutativity. For multiple spikes we have matrix expressions. Hence no longer have commutativity. The analysis is possible, but it is just quite tedious.> maximal scaling regime We don't know this regime. Does the reviewer mean when the step size is too small and we do not see a spike or the regime from [1] where the spectrum becomes heavy-tailed? In the heavy-tailed situation analysis similar to [2] can be used.[1] Martin and Mahoney JMLR 2021 - Implicit Self-Regularization in Deep Neural Networks: Evidence from Random Matrix Theory and Implications for Learning[2] Wang et al. 2024 AISTATS - Near-interpolators: Rapid norm growth and the trade-off between interpolation and generalization"}
|
2024-11-14 01:45:55
|
ICLR.cc/2025/Conference
|
fSyf3ogiBb
|
nnuWhTQ0sX
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for their comments. > They reference the work of Moniri et al., but this work is unrelated to neural networks or gradient descent; it addresses a purely linear regression problem for data with simple spiked covariances.We respectfully disagree. Our work is directly motivated by and connected to neural networks through the following chain of reasoning:1. Ba et al. (2022) and Moniri et al. (2023) show that after one gradient step, the feature matrix $F_1$ can be written as $ F_0 + P$, where $P$ is a rank-$ell$ matrix. 2. This creates a spiked covariance structure in $F_1^TF_1$.3. To understand the generalization error of such networks, we need to analyze least squares regression with $F_1$ as the feature matrix.4. Our work studies this exact setting, though in a simplified form, to make the analysis tractable. 5. In the general rebuttal, we removed many of our simplifications. This further strengthens the connection> They do not account for the generalization ability of neural networks after a single gradient step, as they bypass the gradient step entirely by assuming the W1 matrix directly, which does not reflect the full process of neural network training.We agree with the reviewer. Building on the results from Ba et al. 2022 and Moniri et al. 2023, our new results take us towards understanding the generalization error for two-layer networks. Our analysis provides valuable insights:- It shows how spikes affect generalization in finite vs asymptotic regimes- It demonstrates the importance of the alignment between the spike direction and the target functionWe never meant our paper to claim that we understood the generalization error for two-layer networks, as reflected in our title's focus on least squares regression.> Could you provide a reference for the statement, 'It has been shown that to understand the generalization...' on line 39?In addition to the RMt paper cited in the paper, see [1] for an empirical result on more realistic networks. [1] Martin and Mahoney, \"Implicit Self-Regularization in Deep Neural Networks: Evidence from Random Matrix Theory and Implications for Learning\", JMLR 2021.> Is your generalization analysis very different from the work of Li & Sonthalia (2024)?Compared to Li and Sonthalia 2024. Only one of their two models allows for an eigenvalue to diverge. This model is closely related to the Signal only model in this paper. However, we have output noise $\\varepsilon$. This creates many new dependencies requiring novel analysis techniques. 1. Consider the proof sketch on page 10. Line 491 shows that for the signal-only problem, our solution is the solution from the Li and Sonthalia 2024 paper plus an extra term. The $\\hat{\\varepsilon}$ is not isotropic in the ridge regularied version. Hence, we do not immediately have free independence between $\\hat{Z}+\\hat{A}$ and $\\hat{\\varepsilon}$. Hence, we need to be very careful about the alignment between the two. 2. Hence, we get terms that are cubic and quartic in eigenvalues (vs quadratic in prior work). These required the development of new concentration bounds for these higher-order terms. See Lemmas 17 through 22. The Li and Sonthalia 2024 paper does not consider the signal plus noise case, which introduces further terms that need bounding. Finally, the Li and Sonthalia 2024 only consider the case when $\\tau_A = 1$ or more broadly $\\tau_A = \\Theta(1)$, we allow $\\tau_A = \\Theta(\\sqrt{d})$. **This is also significant**."}
|
2024-11-14 01:46:02
|
ICLR.cc/2025/Conference
|
0HcKHqAtaa
|
PoAfoEMac0
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for the feedback. > The paper is motivated by the spiked covariance from the one-step gradient feature learning in neural networks (Section 1). However, it did not show how the results can be applied to the feature learning scenario. I question the amount of contribution this paper provides.Prior work (Moniri et al. 2023 and Ba et al. 2022) established the existence of spikes for specific target types $y$ (single index models). Our work focuses on understanding the spike's effect through several novel contributions:1. We analyze various targets and alignments between spikes and targets, providing rigorous proofs of generalization bounds2. We demonstrate how asymptotic results may not capture behavior in finite networks3. We provide precise quantification of how spikes affect generalization in both finite and asymptotic regimes> The assumption in line 2221-222 and 253-255 is too strong. The analysis breaks down if there is dependency in the cross term. However, the paper did not show how big the difference the predicted result would be when there is dependence in the cross term. It is questionable if the result in this paper is applicable in realistic machine learning settings.We have now removed this assumption. When considering the dependence structure from Moniri et al. 2023, the analysis actually simplifies. A key term in our analysis is the projection of the spike direction on the bulk eigenvectors. Due to the dependence, this term becomes more tractable. Please see the general response.> What novel results could the authors conclude in the feature learning setting in neural networks using the main theorems 3,4?Our analysis reveals several important insights for feature learning:1. Relationship between targets and spike size: - For targets depending on the bulk (input data), large spikes are crucial - The required spike size scales with input dimension and dataset size 2. Importance of spike-target alignment: - The alignment between spike direction and targets significantly affects generalization - This alignment term exhibits its own double descent behavior - Small alignment improvements can yield large generalization gains3. Double descent characteristics: - Peak location depends on bulk variance and regularization strength - Suggests weight decay regularization primarily affects the bulk, not learned features (spike)While some of these phenomena have been observed before, we provide simplified, quantitative connections between them.> How could the authors show the assumption on the dependence does not affect the result? Is there any experimental validation?We provide new theoretical results that explicitly handle the dependence structure. Please see the general response. > From Figure 4.1, we can see that the effect of the spike correction term is small when is large. Is the main theorem still useful to explain the phenomenon we see from feature learning?Yes - prior work shows that the spike represents the learned feature. Our results allow for larger spikes than previously considered in works like Hastie et al. 2022. However, this shows that for the spike to effect the generalization, we need even bigger spikes. > This paper has problems with the wordings, even in main theorems. This makes the reading difficult. For instance:Thank you for identifying these issues. We will fix all typos and improve clarity in the revised version."}
|
2024-11-14 01:46:05
|
ICLR.cc/2025/Conference
|
5KSuT56TbI
|
fxl5YNtkzp
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "> The authors claim l.083 that Moniri et al. (2023) do not quantify the test error after one gradient step. To the best of my understanding, they do provide such a tight characterization (Theorem 4.5). Could the authors clarify their claim, and emphasize how their work is positioned with respect to Moniri et al. (2023)?We should have been more precise in our claim. While Theorem 4.5 in Moniri et al. 2023 shows that the difference between the test error using $F_1$ (features after 1 step) and $F_0$ (initial features) converges to a constant, there are important differences in our approach:1. Moniri et al. 2023 have expressions requiring solutions to fixed point equations (see equations (5) in their paper). These only hold asymptotically with hard-to-quantify approximation rates. 2. In contrast, we provide: - **Closed form expressions** for the risk itself (not just differences as is the case in Moniri et al. 2023) - **Better control on approximation error rates**, enabling analysis of finite matrices - The above two allow better control in understanding the relationship between the bulk and spike. > I find the discussion l.332-355 somewhat confusingWe apologize for the unclear notation. We use the Vinogradov notation where $f \\ll g$ means $f = O(g)$. Therefore, setting $\\theta^2 = \\tau^2 n$ is consistent with our assumptions.> I believe more intuition about the different scaling considered would help solidify the intuition for the spn caseYou raise an excellent point. Let's clarify the scaling relationships:1. For the bulk term: $a_i^T \\beta_* \\approx \\tau_A \\|\\beta_*\\|$2. For the spike term: $z_i^T \\beta_* \\approx \\theta\\|\\beta_*\\|$Using our scaling $\\theta = \\tau \\sqrt{n}$, the signal part is always larger. However, when the bulk also grows (i.e., $\\tau_A = \\Theta(d)$), the spike's effect becomes invisible. Specifically, we need:- $a_i^T \\beta_* = \\Theta(1)$- $z_i^T \\beta_* = \\Theta(\\sqrt{n})$(assuming $\\|\\beta_*\\|= \\Theta(1)$) for the spike to have a detectable effect.> In 3.3, more discussion in the main text about why only the unregularized case is considered for the spn case, while generic is considered for the signal-only model, would be helpful for intuition, whether it is for technical reasons or because it is not interesting.This limitation is primarily technical. The regularized case for the signal-plus-noise model introduces many additional terms whose mean and variance would need to be bounded, significantly complicating the analysis.> typosThank you for identifying these issues. We will correct all typos in the revision."}
|
2024-11-14 01:46:10
|
ICLR.cc/2025/Conference
|
v0dyh2j6Us
|
5x6iVSaqHT
|
Authors
|
Response by Authors
|
{"Title": "Part 1", "Comment": "We thank the reviewer for their detailed feedback. Let us address the key points:> Limited contribution/novelty... Most of the results in this paper are trivial extensions of the results by Hastie et al. (2022) and Li & Sonthalia (2024)We respectfully disagree. Our contributions extend beyond prior work in several important ways:1. Finite vs Asymptotic Analysis: Prior work focused on asymptotic results. We provide finite-sample corrections that reveal how spikes affect generalization. We show when these corrections matter (small bulk variance) and when they don't (large bulk variance)2. Technical Novelty: - As the reviewers point out, we allow one eigenvalue to diverge compared to Hastie et al. 2022. This is a significant difference. - Compared to Li and Sonthalia 2024. Only one of their two models allows for an eigenvalue to diverge. This model is closely related to the Signal only model in this paper. However, we have output noise $\\varepsilon$. This creates many new dependencies requiring novel analysis techniques - Consider the proof sketch on page 10. Line 491 shows that for the signal-only problem, our solution is the solution from the Li and Sonthalia 2024 paper plus an extra term. The $\\hat{\\varepsilon}$ is not isotropic in the ridge regularied version. Hence, we do not immediately have free independence between $\\hat{Z}+\\hat{A}$ and $\\hat{\\varepsilon}$. Hence, we need to be very careful about the alignment between the two. - Hence, we get terms that are cubic and quartic in eigenvalues (vs quadratic in prior work). These required the development of new concentration bounds for these higher-order terms. See Lemmas 17 through 22. The Li and Sonthalia 2024 paper does not consider the signal plus noise case, which introduces further terms that need bounding. Finally, the Li and Sonthalia 2024 only consider the case when $\\tau_A = 1$ or more broadly $\\tau_A = \\Theta(1)$, we allow $\\tau_A = \\Theta(\\sqrt{d})$. **This is also significant**.> Although the paper is motivated by Moniri et al. (2023), there are significant discrepancies.We agree. Please see the general response on how we can handle the dependency. Additionally, we think of the difference in the targets as a strength of the paper, as we can show things as the following. 1. Relationship between targets and spike size: - For targets depending on the bulk (input data), large spikes are crucial - The required spike size scales with input dimension and dataset size 2. Importance of spike-target alignment: - The alignment between spike direction and targets significantly affects generalization - This alignment term exhibits its own double descent behavior - Small alignment improvements can yield large generalization gains3. Double descent characteristics: - Peak location depends on bulk variance and regularization strength - Suggests weight decay regularization primarily affects the bulk, not learned features (spike)> There exists a related work (Cui et al., 2024)... Dandi et al. (2024)Thank you for bringing these to our attention. The second is quite new and, as the reviewer points out, was only posted well after the submission deadline. Hence, we believe that it is concurrent work and should **not** affect the review of our paper. We shall nonetheless discuss the two papers in the revision. There are many differences to Cui et al. 20241. We work in a more restricted setting but provide rigorous proofs. 2. We simplify expressions. For example, $\\zeta$ in equation 17 in Cui et al. 2024 is exactly $\\xi - 1$ in our paper (see Lemma 13 in the appendix for a definition). The expression in Cui et al. 2024 is left in terms of $\\zeta$. **This is because the results are a product of dependent terms. Hence, simplification is not easy**. However we a. Compute the expectations and variances of each of the terms b. Compute the expectations of the products. c. Greatly simplify expressions These simplifications are a major strength of our work, as the expressions are interpretable without numerical computations. > In Line 178, denotes the case with a single spike (as shown by Moniri et al., 2023)...We only consider the single spike case. The analysis here is similar to Sonthalia and Nadakuditi 2023. Kaushik et al. 2024 extend Sonthalia and Nadakuditi 2023 to the higher rank version. We would need to do the same to extend to multiple spikes. As mentioned before, the difficulty in the analysis was a. Compute the expectations and variances of each of the terms b. Compute the expectations of the products. c. Greatly simplify expressionsThese are currently scalar expressions, so we can use commutativity. For multiple spikes, we have matrix expressions, so we no longer have commutativity. The analysis is possible, but it is just quite tedious.We ignore the $o(\\sqrt{n})$ we shall highlight this as another discrepensacy"}
|
2024-11-14 01:46:26
|
ICLR.cc/2025/Conference
|
4gNRRKAfIm
|
v0dyh2j6Us
|
Authors
|
Response by Authors
|
{"Title": "Part 2", "Comment": "> In footnote 3 (Line 266), the authors say \"... If we use these results, then similar to Eqautuons C.23 in Ba et al. 2022 and Equation (5) in Moniri et al. 2023, we would have that the value of Stieljtes transform is given to us as the unique solution to a set of consistency equations. Hence, we would replace the **explicit** values in Lemmas 7 and 8 with these **implicit** values. However, we did not do so since we wanted explicit closed-form expressions. However, please see the general response, showing how this can be achieved.> Why is there no regularization for the signal-plus-noise problem when there is regularization for the signal-only problem (Line 278-285)?This limitation is primarily technical. The regularized case for the signal-plus-noise model introduces many additional terms whose mean and variance would need to be bounded, significantly complicating the analysis.> How do the authors arrive at \"Hence, we see that if the target vector y has a smaller dependence on the noise (bulk) component A, then we see that the spike affects the generalization error.\" in Line 380? Its connection to the previous part seems to be missing.Here we have that $ y_i = \\beta_*^Ta_i + \\beta_*^T z_i + \\varepsilon_i$. We see that for the bulk term: $a_i^T \\beta_* \\approx \\tau_A \\|\\beta_*\\|$. For the spike term: $z_i^T \\beta_* \\approx \\theta\\|\\beta_*\\|$Using our scaling $\\theta = \\tau \\sqrt{n}$, the signal part is always larger. However, when the bulk also grows (i.e., $\\tau_A = \\Theta(d)$), the spike's effect becomes invisible. Specifically, we need:- $a_i^T \\beta_* = \\Theta(1)$- $z_i^T \\beta_* = \\Theta(\\sqrt{n})$(assuming $\\|\\beta_*\\|= \\Theta(1)$) for the spike to have a detectable effect.> Undefined symbols We apologize. $f \\ll g$ means $f = O(g)$ and $f \\asymp g$ means $f \\ll g$ and $g \\ll f$> How do the authors come up with the equation for the peak point of double descent in Line 477? Is it an empirical observation or a theoretical result?This is currently empirical. It can be proved by computing the expression's derivative (and second derivative) and evaluating it at the derivative. The derivative expressions are quite involved; even symbolic programs such as SciPy struggled. **Hence, this leads us back to our contribution to simplified expression**. While our expressions are further simplified compared to Cui et al. 2024, we believe further simplification is only a positive.> TyposThe reviewer is correct in identifying typos. These shall be fixed."}
|
2024-11-14 01:46:37
|
ICLR.cc/2025/Conference
|
4Xxtq8X6cH
|
zxqdVo9FjY
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewers for their comments and help in improving the paper and hope that our responses with the new results have improved their opinions. If there are further questions that we can answer, we would be happy to continue the discussion."}
|
2024-11-14 01:47:59
|
ICLR.cc/2025/Conference
|
vjcieT0Djg
|
0HcKHqAtaa
|
Reviewer_xFzE
|
Response by Reviewer
|
{"Title": "", "Comment": "Thank you for your detailed reply. I will raise my score accordingly."}
|
2024-11-20 10:43:20
|
ICLR.cc/2025/Conference
|
1wvMEgfJkt
|
vjcieT0Djg
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for the discussion and for increasing their score. If there are more aspects of the work that the reviewer would like to discuss, we would be delighted to continue the discussion."}
|
2024-11-21 04:13:39
|
ICLR.cc/2025/Conference
|
ysTPYKPMOg
|
4gNRRKAfIm
|
Reviewer_qVHv
|
Response by Reviewer
|
{"Title": "", "Comment": "Thank you for the detailed responses. I appreciate that the authors introduced a dependency between bulk and spike during the rebuttal to address discrepancies with the motivating work by Moniri et al. (2023). However, I still believe the paper requires significant revision, particularly in its presentation and writing. Furthermore, in their responses, the authors argued that their novelty lies in three main points: (1) finite matrix effects, (2) exploration of different target functions, and (3) simplified expressions compared to Cui et al. (2024). While these might be interesting contributions, they should have been more thoroughly motivated in the paper. Specifically, what are the additional benefits of these aspects?Overall, after considering the authors' responses and the other reviews, I have decided to maintain my current rating."}
|
2024-11-23 12:15:58
|
ICLR.cc/2025/Conference
|
nNJiZ7vHzS
|
zxqdVo9FjY
|
Authors
|
Response by Authors
|
{"Title": "Interpolation between signal-plus-noise and signal-only models", "Comment": "One other implicit criticism seems to be the lack of connection between the two models we are studying. We would like to point out that there can be a way to interpolate the signal-plus-noise and signal-only models. Intuitively, we introduce extra parameters $\\alpha_{trn}$ and $\\alpha_{tst}$ that define the magnitude of dependence of y on the noise matrices $A_{trn}$ and $A_{tst}$. If they are both 0, then we should recover the signal-only case. To be specific, here we train with: $$\\beta_{spn}^T = argmin \\| (Z_{trn} + \\alpha_{trn}A_{trn})\\beta_* + \\varepsilon_{trn}^T -(Z_{trn} + A_{trn})\\beta \\|_F^2$$And evaluate the following error: $$\\frac{1}{n_{tst}} \\mathbb{E} \\left[\\| (Z_{tst} + \\alpha_{tst}A_{tst})\\beta_* - (Z_{tst} + A_{tst})\\beta_{spn} \\|_F^2 \\right]$$Then using the same decomposition as before and approaches from random matrix theory, we can arrive that a similar error formula that involves extra terms of $\\alpha$. We present the formula for $c < 1$: $$Bias = \\frac{\\theta_{tst}^2\\tau_{\\varepsilon_{trn}}^2}{n_{tst}(\\theta_{trn}^2c + \\tau_{A_{trn}}^2)}\\left( \\frac{c}{1-c}\\right) + \\frac{(1 - \\alpha_{trn}^2)\\tau_{A_{trn}}^4\\theta_{tst}^2}{n_{tst}(\\theta_{trn}^2c + \\tau_{A_{trn}}^2)^2}(\\beta_*^Tu)^2$$$$\\mathbf{Variance + Others} = $$$$\\frac{\\tau_{A_{tst}}^2\\tau_{\\varepsilon_{trn}}^2c}{\\tau_{A_{trn}}^2(1-c)} + \\frac{\\tau_{A_{tst}}^2}{d}\\left[(\\alpha_{trn} - \\alpha_{tst})^2\\|\\beta_*\\|^2 + 2(1 - \\alpha_{trn})(\\alpha_{trn} - \\alpha_{tst}) \\frac{\\theta_{trn}^2(\\beta_*^Tu)^2}{\\theta_{trn}^2c} \\right] + $$$$\\frac{\\tau_{A_{tst}}^2}{d}\\left[(\\alpha_{trn} - 1)^2(\\beta_*^Tu)^2 \\frac{\\theta_{trn}^2}{(\\theta_{trn}^2c + \\tau_{A_{trn}}^2)^2}(\\theta_{trn}^2 + \\tau_{A_{trn}}^2)\\frac{c^2}{1-c} - \\frac{\\tau_{\\varepsilon_{trn}}^2}{\\tau_{A_{trn}}^2}\\frac{\\theta_{trn}^2}{(\\theta_{trn}^2c + \\tau_{A_{trn}}^2)^2}\\frac{c^2}{1-c}\\right]$$We note here when we set $\\alpha_{trn} = \\alpha_{tst} = 1$, we have the current signal-plus-noise formula in the paper. When we set both to 0, we have the current signal-only formula. Consider similar simplifications. We set $\\tau_{A_{trn}}^2 = \\tau_{A_{tst}}^2 = d$, $\\theta_{trn}^2 = \\tau^2n$, $\\theta_{tst}^2 = \\tau^2n_{tst}$. We obtain the following formula: $$\\tau_{\\varepsilon_{trn}}^2\\frac{c}{1-c} + (\\alpha_{trn} - \\alpha_{tst})^2\\|\\beta_*\\|^2 + 2(1-\\alpha_{trn})(\\alpha_{trn} - \\alpha_{tst})(\\beta_*^Tu)^2 + (\\alpha_{trn}-1)^2(\\beta_*^Tu)^2\\frac{1}{1-c}$$Again when we set $\\alpha$'s to be 1, we have some cancellations, making our current signal-plus-noise formula a special case. We hope this connection can provide more intuition and help reviewers understand the work better."}
|
2024-11-24 05:49:00
|
ICLR.cc/2025/Conference
|
tyfVRPmVAX
|
zxqdVo9FjY
|
Authors
|
Response by Authors
|
{"Title": "Revised Version", "Comment": "We have posted the revised version, where we fixed the typos and ambiguities pointed out by the reviewers. We would like to thank the reviewers for their invaluable feedback and their time to help improve our work. If there is anything else we can answer, please let us know, and we would be more than happy to do so."}
|
2024-11-24 07:56:12
|
ICLR.cc/2025/Conference
|
KuKFxWuwMO
|
fSyf3ogiBb
|
Reviewer_LgJ3
|
Response by Reviewer
|
{"Title": "", "Comment": "Thanks so much for all your careful reply and the new updated version draft! I'll raise my score a little bit!"}
|
2024-11-24 20:00:43
|
ICLR.cc/2025/Conference
|
GwAF3v9v9u
|
KuKFxWuwMO
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for increasing their score and valuable contributions"}
|
2024-11-25 08:10:05
|
ICLR.cc/2025/Conference
|
gWY128H3mk
|
ysTPYKPMOg
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for their valuable feedback. We hope that our new results help expand on contribution (2)."}
|
2024-11-25 08:10:53
|
ICLR.cc/2025/Conference
|
ihrSCQhQgj
|
fxl5YNtkzp
|
Reviewer_YeXr
|
Response by Reviewer
|
{"Title": "Acknowledgement of rebuttal", "Comment": "I thank the authors for taking the time to provide all the detailed clarifications and answers to my interrogations. I think the paper is scientifically sound, and thus increase slightly my score, although I have not checked the proofs."}
|
2024-11-25 18:21:46
|
ICLR.cc/2025/Conference
|
w70pJI4hmK
|
ihrSCQhQgj
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for their valuable feedback and for increasing their score"}
|
2024-11-26 00:53:54
|
ICLR.cc/2025/Conference
|
2dFBTjIXgw
|
WqthOZZiOY
|
Reviewer_weJN
|
Response by Reviewer
|
{"Title": "", "Comment": "I warmly thank the authors for addressing my concerns. After reading carefully the other reviewers' comments, I believe that the paper still heavily relies on previously published works, and I would like therefore to keep my original score. On the writing side,I believe the authors should expand the discussion about the connection with previous works not belonging only to the Random Matrix Theory literature (e.g. see the suggested references [1,5])."}
|
2024-11-26 15:18:15
|
ICLR.cc/2025/Conference
|
t8lTNn6Pwj
|
2dFBTjIXgw
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "We thank the reviewer for the feedback and help in improving the paper."}
|
2024-11-28 02:56:19
|
ICLR.cc/2025/Conference
|
GzIdpBrTa1
|
zxqdVo9FjY
|
Area_Chair_aazW
|
Meta Review of Submission13673 by Area_Chair_aazW
|
{"Meta Review": "Summary of Scientific Claims and Findings:The paper investigates the generalization properties of least squares regression with spiked covariance matrices, motivated by neural network training dynamics after one gradient step. The authors provide asymptotic analyses and derive corrections for finite-sample generalization errors, introducing a novel parameter that interpolates between \"signal-only\" and \"signal-plus-noise\" regimes. The findings highlight the impact of spike-target alignment on generalization and demonstrate phenomena like double descent.Strengths:Well-Explained Setup: The authors offer a clear motivation by connecting spiked covariance to neural network dynamics post-training.Finite-Sample Corrections: The inclusion of finite-sample effects extends beyond prior asymptotic-focused works.Novel Interpolation Parameter: The introduction of a parameter to interpolate between different models is a thoughtful addition.Mathematical Rigor: The proofs are detailed and provide bias-variance decompositions and characterizations of risk.Weaknesses:Limited Novelty: The paper heavily relies on extensions of prior works (e.g., Moniri et al. and Li & Sonthalia) without sufficiently differentiating its contributions.Weak Connection to Neural Networks: While motivated by neural networks, the setting deviates significantly, and the connection to practical feature learning scenarios is tenuous.Clarity and Presentation Issues: Ambiguities in notation, numerous typos, and inconsistencies reduce readability and undermine clarity.Assumptions: The paper relies on strong assumptions, such as independence of cross terms, which may limit applicability to realistic machine learning settings.Impact on the Field: Despite deriving finite-sample corrections, the theoretical advancements and insights remain incremental.Decision:Given the identified strengths and weaknesses, the paper is marginally below the acceptance threshold. While it offers an interesting exploration of spiked covariance effects on generalization, it lacks sufficient novelty and clarity to merit acceptance. However, I encourage the authors to address these concerns in future submissions, particularly improving clarity and rigorously connecting the results to neural networks and practical scenarios.", "Additional Comments On Reviewer Discussion": "Reviewer Concerns:Novelty: Reviewers questioned the originality of the findings, noting overlap with prior works.Presentation: Typos and unclear notation were significant points of contention.Connection to Neural Networks: Reviewers felt the connection to practical settings, such as neural networks after one gradient step, was not sufficiently developed.Assumptions: The independence assumption for cross terms was viewed as overly restrictive.Author Rebuttal:The authors responded by:Introducing new results to address dependency between spike and bulk components.Expanding proofs to include finite-sample effects.Adding a novel interpolation parameter to broaden the scope of the analysis.These changes led to marginal improvements in some reviewer scores, reflecting better alignment with reviewer expectations but not fully resolving core concerns about novelty and clarity.Final Assessment:While the authors\u2019 responses added valuable clarifications and incremental extensions, the paper still suffers from limited novelty and an unclear narrative. The reviewers\u2019 concerns about applicability and presentation remain valid. These factors weighed significantly in the final decision to reject."}
|
2024-12-20 10:36:30
|
ICLR.cc/2025/Conference
|
V2Rtla8eUt
|
zxqdVo9FjY
|
Program_Chairs
|
Paper Decision
|
{"Comment": "", "Decision": "Reject"}
|
2025-01-22 05:37:55
|
ICLR.cc/2025/Conference
|
ZIZqJZi9Au
|
zxg6601zoc
|
Reviewer_RZ9b
|
Official Review by Reviewer_RZ9b
|
{"Rating": 5, "Summary": "This paper adopted the parameter-efficient fine-tuning method Representation Tuning to the multimodal large language model domain. This paper used different representation editors for the vision encoder, LLM, and cross-modality projectors to optimize the visual representation, cross-modality representation, and multimodal representation. Experiments on several MLLM and image classification benchmarks show the efficiency and effectiveness of the proposed method. The paper further conducted controllability studies on image classification benchmarks to show the possible controllability of the proposed methods. Extensive ablation studies are furhter discussed for the designs of several hyper-parameters of the proposed method.", "Questions": "1. In Sec 3.2 Cross-modality Representation part, it says that the projector integrates representations from each layer of the visual encoder. Does this mean you combine vision features in different layers in the vision encoder as the final vision input for LLM?2. What is the reason for applying prefix and suffix editors on textual tokens, as the prefix and suffix of different textual prompts at the same position might have very different semantics and meanings? Wouldn't this be harmful for the claimed interpretability and controllability?", "Soundness": 2, "Strengths": "1. This paper provides a promising and efficient PEFT method for MLLMs as an alternative to the commonly used LoRA-based methods.2. The ablation studies are comprehensive for a better understanding of the proposed method and hyper-parameter design choices.", "Confidence": 5, "Weaknesses": "1. The paper directly adapts the representation learning method in LLM to the MLLM in a rather straightforward way thus the technical contribution is limited.2. The benchmark selection for comparisons is not comprehensive and convincing enough. MME is a relatively small-scale MLLM benchmark that has non-trivial variances. The paper should include more comprehensive and commonly used multimodal benchmarks like SEED, MMBench, MMMU, GQA, VQAv2, ChartQA, and DocVQA for more convincing comparisons.3. The performance of other methods might have some problems. In Table 1 of this paper, the lora baseline is significantly worse than the full-finetuning one, but according to LLaVA's results (https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md), the performances should be similar. Besides, the proposed method adds tunable parameters in the vision encoder while other baselines use frozen vision encoders. Baselines with unfrozen vision encoders should also be added for comparison.4. As the proposed method brings additional parameters and time costs in the inference phase, the efficiency advantage of the proposed method should be clearly verified in the training phase, including GPU memory usage and training speed comparisons,5. To validate the generalization ability of the proposed method, the authors should include experiments with different vision encoders + different LLM types and sizes.6. Although the paper claimed that the proposed method brings more interpretability and controllability to MLLMs compared with common practices, it is hard to see how could the method really help the interpretability and controllability of MLLMs. The controllability study is too customized with human priors and provides little help for MLLM controllability in the general case.", "Contribution": 2, "Presentation": 2}
|
2024-10-30 05:55:43
|
ICLR.cc/2025/Conference
|
C0hVt0e9Y5
|
zxg6601zoc
|
Reviewer_PDvR
|
Official Review by Reviewer_PDvR
|
{"Rating": 6, "Summary": "This paper introduces a method for tuning large multi-modal models (LMM) in a efficient but effective way so that it can achieve similar performance to full fine-tuning, with an additional objective of having a controllability. The key idea of this paper is based on a prior technique that learns parameters that edits the representations. The main contribution of this paper is to make use of this technique for efficient tuning of LMMs, and the experiments show that this idea indeed is helpful and the paper did a good job of investigating the effect of various design choices.", "Questions": "See Weaknesses", "Soundness": 3, "Strengths": "- Motivation is clear- Intuitive and simple idea that works well- Extensive analysis/ablation on the design choices", "Confidence": 3, "Weaknesses": "Experiments on controllability is interesting but not conclusive. For instance,- What would happen if you don't use ROI and do train with all tokens?- What would happen if you fine-tune other baselines for this setup?- What would happen if you use a sentence that has a same semantic meaning to 'Is the object an e in the image?' but with different structure and words, after fine-tuning? Would the model still be controlled as intended?Additional weaknesses are:- Figure 5 on the optimization landscape is interesting but I'm not sure how it is cherry-picked. Would there be a way to make this claim be supported by some metrics or more figures?- Main method section feels a bit redundant to me, not much of a difference between each subsection. It could be nice to think of a better way to re-structure, remove redundancy, and think of a way to clearly explain what differences existsNote: I'm not an expert in this area so I might be missing some experimental details. I'll check the other reviews on how the experimental setup is valid and how the results are good compared to other baselines.", "Contribution": 3, "Presentation": 3}
|
2024-11-02 19:18:09
|
ICLR.cc/2025/Conference
|
A2b4OZEZV3
|
zxg6601zoc
|
Reviewer_Yzen
|
Official Review by Reviewer_Yzen
|
{"Rating": 6, "Summary": "This paper introduces Multimodal Representation Tuning (MRT), a parameter-efficient fine-tuning method to enhance controllability and interpretability in multimodal large language models (LMMs). MRT addresses the challenge of adapting LMMs effectively with fewer parameters by leveraging token-level multimodal representation control, achieving superior performance with up to 21 times fewer parameters than similar approaches. The authors explore and improve model behavior control through MRT, illustrating benefits in various multimodal perception and cognition benchmarks.", "Questions": "1. Could the authors elaborate on any plans to automate the rank-tuning process within MRT to simplify its application?2. Section 4.3 is particularly interesting to me. However, I think the attacker can easily break the controllability only by changing the order of the text instruction. Do you have any potential solutions and ideas on that?3. Instead of simple image classification, are there other qualitative examples of counterfactual controls (e.g., VQA)?4. Any analysis on which set of layers L to intervene on for visual / cross-modality / multimodal editor?5. What are some insights on fine-tuning only prefix/suffix tokens of textual embedding in the multimodal editor?6. Minor Errata 1. L84: hu2024bliva looks typo.", "Soundness": 3, "Strengths": "- MRT is intuitive, simple, and effective. It uses significantly fewer parameters while achieving strong results, making it suitable for resource-constrained applications.- The approach provides granular control over the representation editing, enabling counterfactual output generation that enhances interpretability.- The method is validated across multiple multimodal tasks, illustrating its effectiveness in diverse domains such as OCR, visual perception, and spatial reasoning.- The paper is well-written and easy to follow.", "Confidence": 3, "Weaknesses": "- The rank parameter is integral to MRT\u2019s performance, but it currently requires manual tuning, which may limit practical adoption. While promising, the need for automated rank selection is highlighted as a limitation, suggesting a more autonomous rank-searching mechanism that could enhance usability.- The empirical performance is decent, but could you elaborate on MRT's main contribution compared to ReFT instead of extending the interchange intervention idea into MLLM?- Given its control over multimodal representations, MRT\u2019s potential for misuse (e.g., manipulation of outputs) is acknowledged but not fully addressed in terms of mitigation strategies. Could you provide some examples of possible misuse of MRT utilizing controllability?", "Contribution": 3, "Presentation": 3}
|
2024-11-04 02:32:43
|
ICLR.cc/2025/Conference
|
4mcwYbkgth
|
zxg6601zoc
|
Reviewer_uVrb
|
Official Review by Reviewer_uVrb
|
{"Rating": 6, "Summary": "The paper proposes a novel Multimodal Representation Tuning which can editing LMM representation and provide control. The paper introduces a representation editor $\\phi$ based on linear representation hypothesis and interchange interventions, which can apply to different representations in LMM. The overall writing is clear. The experiments cover the comparison between the MRT and the other PEFT methods.", "Questions": "Please see Weaknesses.", "Soundness": 3, "Strengths": "1.The overall writing is clear.2.The idea is general and could be applied to many different applications. I think this would be of interest to people in the LMM community.3.The proposed method is simple yet effective.4.The experiments clearly show the improvement of the MRT over the PEFT.", "Confidence": 4, "Weaknesses": "1.There are some typos. For example, ##hu2024bliva## in the \"Multimodal Instruction Tuning\" section of the related work.2.The ablation study is insufficient. I would expect more ablation experiments, such as applying MRT only to Visual Representation and Cross-modality Representation.3.There is too little theoretical analysis on why MRT is better than PEFT.", "Contribution": 3, "Presentation": 3}
|
2024-11-05 05:33:18
|
ICLR.cc/2025/Conference
|
iS2px2xmKO
|
4mcwYbkgth
|
Authors
|
Response by Authors
|
{"Title": "To Reviewer uVrb", "Comment": "Dear Reviewer uVrb,We sincerely appreciate your time and effort in reviewing our paper and providing valuable comments. We provide explanations to your questions point-by-point in the following.**Q1: Regarding the typos.****A1:** Thank you for pointing it out. We have revised accordingly.**Q2: Regarding more ablation experiments.****A2:** Following the suggestion, we have conducted additional experiments of applying MRT to individual components one at a time. The results are summarized in the table below. Notably, the best performance is achieved when representation tuning is applied simultaneously to all components of the base LMM model. This aligns with our ablation study (Figure 6, left), which demonstrates that removing any single component from MRT results in a noticeable performance decline (e.g., a score of 1376 on MME when the visual editor is excluded). These results, along with additional discussions, have been included in the revised paper (see Appendix S4).| Applied Component | MME | MMAvg || ---- | ----------- | --------------- || LLM | 1473.25 | 62.90 || Cross-modality | 1165.33 | 53.67 || Vision encoder | 1342.46 | 60.83 || All | 1580.40 | 64.93 |**Q3: Regarding more theoretical analysis on why MRT is better than other PEFT.** **A3:** Thank you for the great suggestion. In our work, we have conducted the optimization analysis based on loss landscape [ref1-2], indicating that MRT gains a flatter loss landscape, which in turn provides more optimization choices compared to other PEFT methods [ref3-4]. Additionally, our token-wise control experiment on visual RoI and textual target indicator shows that precise control over specific token representations can effectively alter the model behavior. Following your suggestion, we plan to conduct further theoretical analysis, including how the incorporation of representation editing influences the attention module (e.g., attention activation pattern analysis [ref4]) and gradient flow analysis [ref5]. We have also highlighted this as a future direction in our revised paper (see Appendix S11).[ref1] Hao Li, et al. Visualizing the loss landscape of neural nets. NeurIPS, 2018.[ref2] Runjia Zeng, et al. Visual Fourier Prompt Tuning. NeurIPS, 2024.[ref3] Ying Shen, et al. Multimodal Instruction Tuning with Conditional Mixture of LoRA. ACL, 2024.[ref4] Taowen Wang, et al. M$^2$PT: Multimodal prompt tuning for zero-shot instruction learning. EMNLP, 2024.[ref5] Bambhaniya, et al. Progressive Gradient Flow for Robust N: M Sparsity Training in Transformers. 2024.We sincerely appreciate your thoughtful comments. We hope our response addresses your questions. Please let us know if there are any additional questions, and we will be happy to discuss further."}
|
2024-11-19 23:33:49
|
ICLR.cc/2025/Conference
|
uA3W381dSn
|
A2b4OZEZV3
|
Authors
|
Response by Authors
|
{"Title": "To Reviewer Yzen (Part I)", "Comment": "Dear Reviewer Yzen,We sincerely appreciate the time and effort you've devoted to reviewing our work and providing helpful feedback!**Q1: Regarding the plan to automate the rank-tuning process.****A1:** Thank you for the great insights. We completely agree that manual rank searching can be highly time-consuming. To address this, we plan to explore automated rank selection methods in the future. Specifically, we aim to utilize meta-learning frameworks [ref1], which leverage prior training experiences across tasks to efficiently adapt ranks, and Bayesian-based frameworks [ref2], which use probabilistic models to iteratively explore and dynamically select optimal ranks. These approaches will help mitigate the labor-intensive manual searches.**Q2: Regarding MRT\u2019s contribution.****A2:** Thank you for the question. We would like to highlight several key contributions of MRT besides extending the effectiveness of representation tuning into MLLMs. **First, intuitive yet effective control.** MRT is the first attempt to enable token-wise control over LMMs through representation editing. By directly editing the semantic information of the image RoI and the textual target class indicator token, MRT offers an interpretable and intuitive mechanism for adjusting model predictions. This level of fine-grained controllability is difficult to achieve with existing baselines.**Second, representation learning loss optimization.** From an optimization perspective, we provide a detailed analysis of why MRT outperforms other PEFT methods. By visualizing the loss landscape, we demonstrate that multimodal representation tuning enhances the generalization capabilities of LMMs, highlighting a promising direction for future PEFT research.**Third, joint multimodal learning.** Unlike single-modality research, multimodal settings require consideration of two additional factors: **multimodal integration** and **vision modality editing**. To address this, we designed a framework that optimizes the cross-modality layer to effectively bridge the gap between the two modalities. While current PEFT approaches [ref13] for LMMs typically unfreeze the cross-modality projector during stage-2 tuning, we adhere to the principle of representation editing by introducing a lightweight cross-modality editor, achieving significantly lower parameter usage while delivering substantial performance gains. For vision modality editing, MRT takes a markedly different approach from current NLP practices by focusing on editing all visual representations. This method highlights the sparsity of visual information and suggests that broader editing strategies should be explored in the vision domain.Thank you again for the great suggestion. We have supplemented the above discussions in Appendix S11.**Q3: Regarding mitigation strategies of potential misuse of MRT.****A3:** Thank you for the excellent question. We want to answer this question from two perspectives. **Possible misuse of MRT utilizing controllability.** Misuse of generative models can lead to significant ethical, social, and security concerns, especially when the model\u2019s controllability is leveraged maliciously via tuning. There are possible misuse scenarios and corresponding mitigation strategies. First, attackers can manipulate models to produce misinformation (e.g., misclassification) via intentionally altering the model\u2019s understanding of an input image [ref3]. Second, biased information can be produced or amplified. Attackers can edit the textual tokens related to sensitive attributes in the multimodal representation, leading to harmful or discriminatory outputs [ref4].**How to mitigate potential misuse?**In order to mitigate possible misinformation, we suggest performing adversarial robustness testings [ref5] that explicitly check for consistency in object recognition across varying queries. For mitigating bias generation or amplification, one solution can be bias detection and correction procedure on generated content, monitoring the representation for bias patterns and applying corrective measures if detected [ref4]. Another solution lies in clearly documenting any controlled editing made to the model\u2019s representation and disclosing any potential biases introduced during this process [ref6]. In conclusion, while MRT exhibits strong output controllability, applying MRT to realistic applications still requires ethical safeguarding, robust testing, and transparency measuring. From a security perspective, MRT presents significant potential, as it may facilitate the development of white-box attack and defense strategies tailored to LMMs [ref7]. We have added additional discussions to the revised paper (see Appendix S10)."}
|
2024-11-19 23:50:28
|
ICLR.cc/2025/Conference
|
GGnZTLB3zl
|
A2b4OZEZV3
|
Authors
|
Response by Authors
|
{"Title": "To Reviewer Yzen (Part II)", "Comment": "**Q4: Changing the order of text instruction can break the controllability.****A4:** We would like to clarify that simply changing the order of text instruction can\u2019t break the controllability. MRT is able to accommodate variations in prompt formats by training control editors specifically for the new prompt structure. To validate this, we conducted another controllability experiment where we changed the textual prompt format from \u201cIs the object an e in the image?\u201d to \u201cIs the object in the image an e?\u201d, and trained control editors under the same settings described in Sec. 4.3. The results, as shown in the table below, demonstrate that the new editors achieve equally effective and robust control over counterfactual outputs.| Class e | Misclassification rate on e | Misalignment rate on e || ---- | ----------- | --------------- || dog | 100% | 100% || cat | 100% | 100% || Ship | 100% | 100% || Frog | 100% | 100% || Ship | 100% | 100% |To further enhance the robustness of MRT\u2019s controllability, one possible solution is to leverage the power of prompt engineering [ref14], which crafts effective prompts to guide the output to generalize the control across an even broader range of input queries, reducing possible sensitivity to variations in phrasing. This involves normalizing different prompts with the same semantic meaning into standardized templates, such as converting \u201cCould you help me identify the object\u2019s name?\u201d or \u201cWhat\u2019s the name of the object in the image?\u201d into a common structure like \u201cWhat is the object\u2019s name?\u201d. By standardizing the input prompts, MRT is able to achieve more robust token-wise control, even in scenarios with diverse or complex textual instructions. **Q5: Regarding the qualitative examples of counterfactual controls.****A5:** We have included more qualitative examples and experimental results on the Text-VQA dataset in the revised paper. Please refer to Appendix S6.2, thank you!**Q6: Analysis on the set of layers to intervene.****A6:** Thank you for your suggestion! Table 3 in Sec. 4.4 illustrates the impact of different editing depths. We would like to further discuss the results here. First, the results indicate MRT\u2019s performance is positively correlated with editing depth. Second, we observe that even under the setting of editing the first layers of the vision encoder and the LLM (i.e., VPT shallow in prompt tuning), MRT still can be able to gain a noticeable improvement (i.e., 1329.84 on MME and 60.57 on MMAvg) in performance. Third, editing only the latter half of the layers yields better performance compared to editing the first half (i.e., 1447.41 vs. 1440.32 on MME), suggesting that deeper layers play a more significant role in enhancing the model's capabilities. Last, editing at every odd layer outperforms both the \"first half\" and \"latter half\" configurations (i.e., 1468.21 vs. 1447.41 on MME). This suggests that distributing interventions across the model in a sparse manner can be more beneficial than focusing on a continuous block of layers.**Q7: Insights on fine-tuning only prefix/suffix tokens.****A7:** It is a great question! During instruction tuning, we fine-tune the prefix and suffix tokens from the textual-oriented representations. Considering the attention mechanism and the generation process of Transformer-based decoder models, prefix segments often condition the model on specific tasks or behaviors, therefore they are crucial for setting up the context and guiding the generation process early on [ref8, ref15]. Suffix segments also play an important role in guiding and controlling generations due to the autoregressive mechanism. Recognizing their importance during training, we focus on these tokens\u2019 editing from the textual embedding. As for the vision encoder, we fine-tune the whole visual sequence since the model relies on the entire visual sequence to capture global context. We have included additional experiments on editing different segments of textual-oriented representations. The results clearly demonstrate that fine-tuning both the prefix and suffix tokens yields the best performance, significantly outperforming the setting of fine-tuning all tokens. Specifically, we observe a substantial drop in the MME score when the entire textual embedding is edited (i.e., 1233.90 vs. 1580.40 on MME). This suggests that over-editing the embeddings can lead to response drift, negatively impacting performance. This observation aligns with recent studies on prompt tuning [ref9-13], which indicate that larger adjustments (i.e., longer inserted prompts in prompt tuning) do not necessarily lead to better performance and can, in fact, be less effective than smaller edits. | Segments | MME || ---- | ----------- || Prefix Only | 1465.32 || Suffix Only | 1497.35 || Prefix & Suffix | 1580.40 || All | 1233.90 |"}
|
2024-11-19 23:52:03
|
ICLR.cc/2025/Conference
|
wIbwx616M2
|
A2b4OZEZV3
|
Authors
|
Response by Authors
|
{"Title": "To Reviewer Yzen (Part III)", "Comment": "**Q8: Typos.****A8:** Thank you for pointing it out. We have fixed it accordingly in the revised version.[ref1] Zhang, R., et al. AutoLoRA: Automatically Tuning Matrix Ranks in Low-Rank Adaptation Based on Meta Learning. ArXiv, 2024.[ref2] Moe, C., et al. Bayesian-LoRA: LoRA based Parameter Efficient Fine-Tuning using Optimal Quantization levels and Rank Values trough Differentiable Bayesian Gates. ArXiv, 2024.[ref3] Chen, Xiangning, et al. Robust and accurate object detection via adversarial learning. CVPR, 2021.[ref4] D'Inc\u00e0, Moreno, et al. OpenBias: Open-set Bias Detection in Text-to-Image Generative Models. CVPR, 2024.[ref5] Dong, Ziyi, et al. Adversarially-aware robust object detector. ECCV, 2022.[ref6] Shah, Milind, et al. A Comprehensive Review of Bias in Deep Learning Models: Methods, Impacts, and Future Directions. Archives of Computational Methods in Engineering. ArXiv, 2024.[ref7] Daizong Liu, et al. Attacks on Large Vision-Language Models: Resources, Advances, and Future Trends. ArXiv, 2024.[ref8] Bavarian, Mohammad, et al. Efficient training of language models to fill in the middle. ArXiv, 2022.[ref9] Lester, Brian, et al. The power of scale for parameter-efficient prompt tuning. EMNLP, 2021.[ref10] Cheng Han, et al. Facing the Elephant in the Room: Visual Prompt Tuning or Full Finetuning? ICLR, 2024.[ref11] Changdae Oh, et al. Blackvip: Black-box visual prompting for robust transfer learning. ArXiv, 2023.[ref12] Chengzhi Mao, et al. Doubly right object recognition: A why prompt for visual rationales. ArXiv, 2022.[ref13] Taowen Wang, et al. M$^2$PT: Multimodal prompt tuning for zero-shot instruction learning. EMNLP, 2024.[ref14] White, J., et al. A prompt pattern catalog to enhance prompt engineering with chatgpt. ArXiv, 2023.[ref15] Raffel, Colin, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21.140 (2020): 1-67.We sincerely appreciate your thoughtful comments. We hope our response addresses your concerns. Please let us know if there are any additional questions, and we will be happy to discuss further."}
|
2024-11-19 23:55:20
|
ICLR.cc/2025/Conference
|
3bPd5joLzI
|
C0hVt0e9Y5
|
Authors
|
Response by Authors
|
{"Title": "To Reviewer PDvR (Part I)", "Comment": "Dear Reviewer PDvR,We sincerely thank reviewer PDvR for the valuable time and constructive feedback! We provide explanations to your questions point-by-point in the following.**Q1: Don\u2019t use RoI but train with all tokens.****A1:** Although editing more tokens (i.e., all visual tokens) does not prevent the model from producing expected counterfactual results, it becomes difficult to isolate the impact of the edits on specific semantic features of the target object, leading to a severe loss of interpretability. In contrast, targeting the RoI, combined with edits to the textual target indicator token (i.e., class token e), ensures that the edits are directly tied to the most relevant semantic features of the object and the query, preserving interpretability. Moreover, from the training efficiency perspective, targeting the RoI reduces the computational complexity and speeds up convergence.**Q2: Controllability results of other baselines.****A2:** With the same setups, although other baselines could also learn the pattern and produce counterfactual results, they are not able to achieve token-wise controllability. This limitation arises because other baselines typically operate in a black-box manner, where the manipulations are based on implicit adjustments by injecting new prompts (i.e., prompt tuning) or by modifying the internal weights through low-rank updates (i.e., LoRA) without providing explicit control over representations. On the contrary, our approach achieves token-wise controllability by introducing dedicated representation editors that are trained to selectively edit the most semantically relevant tokens (i.e., visual RoI and textual class token). This fine-grained editing mechanism ensures that the impact of the edits is directly tied to specific features, enabling interpretable control over the model\u2019s counterfactual outputs.**Q3: Different structure and words for intended control.****A3:** This is a great question! Prompts with different structures or phrasings will not break the controllability of MRT. MRT is capable of handling variations in prompt formats by training control editors tailored to the new prompt structures or phrasings. To validate this capability, we conducted a controllability experiment where the textual prompt \u201cIs the object an e in the image?\u201d is altered to \u201cIs the object in the image an e?\u201d and trained control editors using the same setup as described in Sec. 4.3. The results, shown in the table below, confirm that the newly trained editors can achieve equally robust and effective control over counterfactual outputs.| Class e | Misclassification rate on e | Misalignment rate on e || ---- | ----------- | --------------- || dog | 100% | 100% || cat | 100% | 100% || Ship | 100% | 100% || Frog | 100% | 100% || Ship | 100% | 100% |To further enhance the robustness of MRT\u2019s controllability when considering diverse or complex textual instructions, we plan to incorporate prompt engineering [ref1], which can normalize different phrasings with the same semantic meaning into standardized templates. For example, sentences such as \u201cCan you identify the animal in the image?\u201d and \u201cDo you know what kind of animal is depicted here?\u201d could be normalized to a standard template as \u201cWhat is the animal in the image?\u201d This approach would allow us to apply the trained editors consistently across a broader range of inputs, enhancing the robustness and generalizability of the controllability framework.We have added additional discussions to the revised paper (see Appendix S6). Thank you!"}
|
2024-11-19 23:57:23
|
ICLR.cc/2025/Conference
|
8oidjT7Yws
|
C0hVt0e9Y5
|
Authors
|
Response by Authors
|
{"Title": "To Reviewer PDvR (Part II)", "Comment": "**Q4: Optimization landscape.****A4:** Thank you for your positive assessment. We want to address your concerns from two perspectives. **How do we pick the optimization landscape?**It is not a cherry-picked landscape, but instead a general visualization. Figure 5 illustrates the 2D/3D loss surfaces [ref2] for three best models based on the evaluation results. Specifically, we randomly pick up two directions of changing the weights and randomly select a subset of training data for plotting loss surfaces for all three models. This randomness, as stated in various works [ref2-4], does not significantly affect the results. MRT provides a flatter loss surface, indicating that MRT is less sensitive to loss fluctuations, leading to better generality compared to other approaches\u2019 loss landscapes.**Quantitative or qualitative evaluation.**We agree that a systemic analysis quantitatively or qualitatively can further strengthen our claim. Unfortunately, currently the community does not have a standard evaluation metric that we can follow. However, we try to explain the loss landscape further, and highlight a promising direction of studying the loss landscape. [ref2] claims that a sharp loss surface near the minimum indicates that small perturbations in the weights can lead to a significant increase in loss. This suggests the model may generalize poorly, as it can be understood as a sign of overfitting to the training data. A flatter loss surface around the minimum, on the other hand, typically suggests better generalization. This is because the model is more robust to small changes in the weights, indicating a more stable solution has been learned. We have added more discussions w.r.t. optimization landscape in the revised paper. Thank you!**Q5: Main method is a bit redundant.****A5:** Thank you for your suggestion. We have revised the paper accordingly.[ref1] White, J., et al. A prompt pattern catalog to enhance prompt engineering with chatgpt. ArXiv, 2023.[ref2] Hao Li, et al. Visualizing the loss landscape of neural nets. NeurIPS, 2018.[ref3] Runjia Zeng, et al. Visual Fourier Prompt Tuning. NeurIPS, 2024.[ref4] Ma, Xu, et al. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. ICLR, 2022.We sincerely appreciate your thoughtful comments. We hope our response addresses your concerns. Please let us know if there are any additional questions, and we will be happy to discuss further."}
|
2024-11-19 23:58:57
|
ICLR.cc/2025/Conference
|
mUe8p5h0BE
|
ZIZqJZi9Au
|
Authors
|
Response by Authors
|
{"Title": "To reviewer RZ9b (Part I)", "Comment": "Dear Reviewer RZ9b,We sincerely thank you for the valuable time and constructive feedback, which are crucial for improving our work. We provide explanations to each question as follows.**Q1: Regarding MRT\u2019s technical contributions.****A1:** While representation tuning has been explored in the NLP field, we would like to highlight three key technical contributions of MRT specifically tailored to the multimodal domain.**First, intuitive yet effective control.** MRT is the first attempt to enable token-wise control over LMMs through representation editing. By directly editing the semantic information of the image RoI and the textual target class indicator token, MRT offers an interpretable and intuitive mechanism for adjusting model predictions. This level of fine-grained controllability is difficult to achieve with existing baselines.**Second, representation learning loss optimization.** From an optimization perspective, we provide a detailed analysis of why MRT outperforms other PEFT methods. By visualizing the loss landscape, we demonstrate that multimodal representation tuning enhances the generalization capabilities of LMMs, highlighting a promising direction for future PEFT research.**Third, joint multimodal learning.** Unlike single-modality research, multimodal settings require consideration of two additional factors: **multimodal integration** and **vision modality editing**. To address this, we designed a framework that optimizes the cross-modality layer to effectively bridge the gap between the two modalities. While current PEFT approaches [ref1, ref13] for LMMs typically unfreeze the cross-modality projector during stage-2 tuning, we adhere to the principle of representation editing by introducing a lightweight cross-modality editor, achieving significantly lower parameter usage while delivering substantial performance gains. For vision modality editing, MRT takes a markedly different approach from current NLP practices by focusing on editing all visual representations. This method highlights the sparsity of visual information and suggests that broader editing strategies should be explored in the vision domain.Thank you again for the great question. We have supplemented the above discussions in Appendix S11.**Q2: Regarding the scale of the benchmarks.****A2:** Thank you for the excellent suggestion. In our work, we evaluated MRT on seven additional larger scale datasets beyond MME, encompassing a total of 63,123 instances. To further strengthen our evaluation, we followed your suggestion and conducted additional experiments on the SEED-Bench and GQA benchmarks. As shown in the results below, MRT consistently outperforms other PEFT approaches. We\u2019ve included the detailed discussions and the corresponding results in the revised paper (see Appendix S3).| Method | SEED-Bench | GQA || ---- | ----------- | --------------- || MixLoRA | 55.9 | 52.2 || M$^2$PT | 57.1 | 50.3 || MRT | 57.6 | 52.7 |**Q3.1: Regarding the LoRA performance.****A3.1:** Thank you for the insightful observation. The primary reason for the performance difference lies in the training data used. Specifically, our work utilizes the Vision-Flan dataset [ref11], following previous studies [ref1-2]. In contrast, the LoRA results from LLaVA were obtained using LLaVA\u2019s training data (https://huggingface.co/liuhaotian/llava-v1.5-7b-lora). However, our reported LoRA results align with those in MixLoRA [ref2] and M$^2$PT [ref1], which are also trained on the Vision-Flan dataset. We\u2019ll make this more clear in the revised paper.**Q3.2: Baselines with unfrozen vision encoders should also be added for comparison.****A3.2:** Thank you for the excellent suggestion. We would like to clarify that MRT has been compared with M$^2$PT [ref1] and VPT [ref8] in Table 1, both of which are PEFT methods utilizing tunable parameters in the vision encoder. Specifically, M$^2$PT introduces tunable soft prompts into both the frozen vision encoder and the frozen backbone LLM, while VPT focuses solely on inserting trainable prompts into the frozen vision encoder. Experimental results show that MRT delivers noticeably superior performance compared to M$^2$PT and VPT."}
|
2024-11-20 00:05:32
|
ICLR.cc/2025/Conference
|
SskQ2cv2pk
|
ZIZqJZi9Au
|
Authors
|
Response by Authors
|
{"Title": "To reviewer RZ9b (Part II)", "Comment": "**Q4: Memory & Time Efficiency in training.****A4:** Following the suggestion, we have included the efficiency in the training stage w.r.t. trainable parameters, memory usage, and training time in the table below. It can be seen that MRT enjoys a competitive training efficiency with existing PEFT approaches. We also want to highlight that both GPU memory usage and training time are lower than several baselines (i.e., LoRA, MixLoRA). For completeness, we have added these results to Appendix S5 in the revised paper. Thank you again for the great suggestion.| Method | trainable params | Memory Usage (GB)* | Training Time (Hours) | MME | | ---- | ---- | ----------- | --------------- |--------------- || VPT | 0.06% | 12 | 7 | 1398.74 || MixLoRA | 0.85% | 23 | 24 | 1509.61 || M$^2$PT | 0.09% | 17 | 9 | 1503.98 || MRT | 0.03% | 16 | 9 | 1580.40 |*Peak under the same batch size setting.**Q5: Apply MRT to different LMMs.****A5:** We completely agree that experimenting with different LMMs can enhance the generalizability of our method. Following the suggestion, we conducted additional experiments using a different LMM configuration, MiniGPT-v2 [ref3] with EVA [ref4] as the vision encoder, a linear projection layer as the cross-modality module, and LLaMA2-chat (7B) [ref5] as the LLM, differing from the components of LLaVA. For these experiments, we selected stage-2 MiniGPT-v2 without multimodal instruction tuning as the backbone model and used a scaled-down version of the Vision-Flan dataset, consisting of 191K instances, for fine-tuning. Preliminary results on the MME benchmark demonstrate that MRT consistently achieves performance gains compared to other PEFT approaches. We will include more comprehensive results and discussions in the revised paper.| MiniGPT-v2 | MME || ---- | ----------- || ReFT | 1346.65 || MixLoRA | 1418.48 || M$^2$PT | 1421.02 || MRT | 1439.73 |**Q6: Interpretability and controllability.** **A6:** Thank you for raising this question. We would like to provide some clarification on how the proposed method helps both interpretability and controllability in MLLMs and how it can generalize beyond the specific scenarios studied.For interpretability, the proposed method introduces token-wise editing, which allows for direct and targeted manipulation of specific semantic representations, such as the image RoI and the textual target indicator token. By focusing on these tokens, the impact of edits becomes transparent and traceable.For controllability, we have conducted more controllability experiments on **robustness of token-level control**, **extending to other multimodal tasks**, **generalizability**, as well as including further discussion of employing prompt engineering to allow control across an even broader range of input queries. Please kindly refer to Appendix S6 for the full details. Thank you!**Q7: Regarding combining vision features in different layers.****A7:** Sorry for the confusion. We want to clarify that MRT does not combine vision features in different layers in the vision encoder as the final input for LLM. In Appendix S2, we have included the discussion that only the visual representation from the second last encoder layer is selected and fed into the cross-modality projector layer in order to be concatenated with textual tokens. This step strictly follows the common setting of stage-one LLaVA [ref1-2]. We have also revised line 218-219 to prevent any confusions."}
|
2024-11-20 00:09:22
|
ICLR.cc/2025/Conference
|
TpExpE7hGG
|
ZIZqJZi9Au
|
Authors
|
Response by Authors
|
{"Title": "To reviewer RZ9b (Part III)", "Comment": "**Q8.1: The reason for applying prefix and suffix editors on textual tokens.****A8.1:** We would like to explain the reason for choosing prefix and suffix tokens. Prefix tokens are vital for conditioning the model to specific tasks or behaviors [ref10, ref12], while suffix tokens are significant in shaping and directing output due to the autoregressive nature of the models. Acknowledging their importance in the training process, we prioritize editing these tokens within the textual embeddings. Our additional experiments on editing various segments of the text-based representations show that fine-tuning both prefix and suffix tokens leads to the optimal performance. Specifically, the decrease (i.e., 1233.90 compared to 1580.40) on the MME score when editing all textual tokens indicates that excessive editing of the embeddings can cause response drift, adversely affecting performance. This observation is also consistent with recent research on prompt tuning [ref1, ref6-7], which suggests that larger modifications (i.e., longer soft prompts in prompt tuning) do not necessarily improve performance and may even be less effective than smaller ones.| Segments | MME || ---- | ----------- || Prefix Only | 1465.32 || Suffix Only | 1497.35 || Prefix & Suffix | 1580.40 || All | 1233.90 |**Q8.2: Applying prefix and suffix editors on textual tokens may be harmful for controllability.****A8.2:** We want to clarify that the editing within textual-oriented representations during training and controlling are taken place in separate positions, which does not have a harmful impact for the claim. Specifically, for training, we only edit prefix and suffix locations, which is motivated by their critical roles in establishing general task context and guiding generation. Previous works have shown that editing these locations delivers the most significant gains [ref9]. For controlling, instead of intervening semantic ambiguous positions, we precisely focus on the targeted textual-oriented token as it contains direct and interpretable impact to the textual questions. To further address your concern on the generalization of the editing, during the rebuttal phase, we further change the order of text instruction for different template in (Appendix S6), and other counterfactual controls (i.e., on Text-VQA, also see Appendix S6). These results consistently demonstrate the robustness and effectiveness of our controlling strategy on textual inputs.[ref1] Taowen Wang, et al. M$^2$PT: Multimodal prompt tuning for zero-shot instruction learning. EMNLP, 2024.[ref2] Ying Shen, et al. Multimodal instruction tuning with conditional mixture of lora. ACL, 2024.[ref3] Chen, Jun, et al. MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning, ArXiv, 2023.[ref4] Yuxin Fang, et al. Eva: Exploring the limits of masked visual representation learning at scale. ArXiv, 2022.[ref5] Hugo Touvron, et al. Llama 2: Open foundation and fine-tuned chat models. ArXiv, 2023.[ref6] Changdae Oh, et al. Blackvip: Black-box visual prompting for robust transfer learning. ArXiv, 2023.[ref7] John Merrill, et al. Doubly right object recognition: A why prompt for visual rationales. ArXiv, 2022.[ref8] Cheng Han, et al. Facing the elephant in the room: Visual prompt tuning or full finetuning? ICLR, 2024.[ref9] Zhengxuan Wu, et al. Reft: Representation Finetuning for Language Models. NeurIPS, 2024[ref10] Raffel, Colin, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21.140 (2020): 1-67.[ref11] Xu, Zhiyang, et al. Vision-flan: Scaling human-labeled tasks in visual instruction tuning. ACL, 2024.[ref12] Bavarian, Mohammad, et al. Efficient training of language models to fill in the middle. ArXiv, 2022.[ref13] Zhou, Xiongtao, et al. An Empirical Study on Parameter-Efficient Fine-Tuning for MultiModal Large Language Models. 2024.We sincerely appreciate your thoughtful comments. We hope our response addresses your concerns. Please let us know if there are any additional questions, and we will be happy to discuss further."}
|
2024-11-20 00:13:44
|
ICLR.cc/2025/Conference
|
syI2Phhkeo
|
zxg6601zoc
|
Authors
|
Response by Authors
|
{"Title": "Summary of Revisions", "Comment": "To all reviewers:Thank you for your thorough review and insightful comments. We have revised our paper according to the suggestions. The major changes are summarized as follows:* We have performed more ablation experiments to explore applying MRT to a single modality at a time, detailed in Appendix Sec. S4. We have also further discussed the future plan to conduct more theoretical analysis of MRT in Appendix Sec. S11. Additionally, we corrected the typo on Line 84 as suggested by **Reviewer uVrb**. * As suggested by **Reviewer Yzen**, we have added a discussion on prefix/suffix editing in Appendix Sec. S2, supplemented analysis of editing depth in Sec. 4.4, and provided more qualitative examples of counterfactual controls on the Text-VQA dataset in Appendix Sec. S6. Additionally, we discussed MRT's robustness and generalizability in Appendix Sec. S6, addressed potential misuse and mitigation in Appendix Sec. S10, and corrected the typo on Line 84. Key technical contributions have also been highlighted in Appendix S11.* Based on the suggestions by **Reviewer PDvR**, we have included more discussion on the robustness of MRT\u2019s controllability in Appendix Sec. S6 and reduced redundancy in Sec. 3.2. In addition, we have added more discussion on the optimization landscape in Appendix S11.* We conducted additional experiments evaluating MRT on SEED and GQA datasets, as well as applying MRT to LMMs with different vision encoders and LLMs, presented in Appendix Sec. S3. Training time and memory efficiency comparisons were included in Appendix Sec. S5, and the visual representation selection procedure was clarified in Sec. 3.2. We have performed more controllability experiments in Appendix Sec. S6. Moreover, we have also highlighted our key technical contributions in Appendix S11, as suggested by **Reviewer RZ9b**.All modifications have been marked in ${\\color{blue} blue}$ in our revised submission.Sincerely yours,\\Authors."}
|
2024-11-20 00:30:29
|
ICLR.cc/2025/Conference
|
ZxqFDasycC
|
zxg6601zoc
|
Authors
|
Response by Authors
|
{"Title": "Looking forward to the discussion", "Comment": "Dear Reviewers,We sincerely appreciate the time and effort you've devoted to reviewing our work. We understand that your schedule may be quite busy, and we are truly grateful for your valuable feedback. As we are presently in the discussion phase, we would greatly value the opportunity to engage in further dialogue with you. Our aim is to gain insights into whether our responses effectively address your concerns and to ascertain if there are any additional questions or points you would like to discuss.We look forward to the opportunity for further discussion with you. Thank you for your thoughtful consideration. Best regards,Authors"}
|
2024-11-22 06:22:30
|
ICLR.cc/2025/Conference
|
tGTEHkarQ4
|
wIbwx616M2
|
Reviewer_Yzen
|
Response by Reviewer
|
{"Title": "", "Comment": "Thanks for the detailed responses and additional results. Most of my concerns are well addressed. I think the mentioned contribution in A2 is intuitive but not novel enough to increase the score. I hope to see some generalization results leveraging LLM regarding Q4. I will keep the original score."}
|
2024-11-22 19:45:03
|
ICLR.cc/2025/Conference
|
hx6I2itEp7
|
tGTEHkarQ4
|
Authors
|
Response by Authors
|
{"Title": "Thank you for the prompt response", "Comment": "Thank you for your valuable feedback. To further address your comment on generalization, we have leveraged **a lightweight rephraser** based on T5-small (i.e., 60M parameters), and customized a dataset for fine-tuning the rephraser, containing **6 different variant templates** with the same semantic meaning of the expected input sequence. The table below shows that MRT can successfully achieve robust control on various input sequences with a single set of editors, even if they differ in lengths and structures.| Input Variants | Output Control Rate on e || -------------------------------------------------------- | ----------- || \u201cIs there an e object visible in the image?\u201d | 100% || \"Does the image contain an object that is an e?\" | 100% || \"Is the object shown in the image an e?\" | 100% || \"Is the object in the picture an e?\" | 100% || \"Do you recognize the object in the image as an e?\" | 100% || \"Can you tell if the object shown in the image is specifically an e?\" | 100% |In addition, we further evaluate the generalization of MRT via testing its controllability on the Text-VQA dataset, which derives different counterfactual controls from the current image classification settings. Specifically, we select 8,017 instances as the training set and 1,189 instances as the validation set on textual tokens beginning with \u201cwhat is the $n$\u201d, where $n$ represents an image attribute (e.g., name, color, brand). We aim to generate counterfactual outputs different from the scenarios (i.e, misclassification and misalignment) of counterfactual outputs on image classification task, we target the scenario of **Indeterminate** by altering the labels of all questions related to $n$ in the training set to \u201cNot sure\u201d. The results show that MRT can control across an even broader range of input queries.| Attribute ($n$) | Indeterminate || -------- | ---------------- || name | 100% || color | 100% || brand | 100% |Altogether, MRT demonstrates consistency and effectiveness of its control strategy for textual inputs. We have added additional discussions as well as qualitative results to the revised paper Appendix Sec. S6. Thank you, and we are eager to address any of your remaining concerns during the discussion phase."}
|
2024-11-23 01:58:34
|
ICLR.cc/2025/Conference
|
jJ0XvUFotn
|
8oidjT7Yws
|
Reviewer_PDvR
|
Response by Reviewer
|
{"Title": "", "Comment": "Thank you for your response. I read the response other review. I agree with Reviewer RZ9b in that the controllability experimental setup is a bit contrived, and it could be nice to think of how to design more natural setups. But I still think this paper deserves the score of 6 so I maintain my score."}
|
2024-11-25 03:17:46
|
ICLR.cc/2025/Conference
|
3IoD9MV4fM
|
jJ0XvUFotn
|
Authors
|
Response by Authors
|
{"Title": "Thank you for your response", "Comment": "We sincerely thank the reviewer for their prompt response and thoughtful feedback. To address controllability, we have included additional experiments in Appendix S6, covering the _robustness of token-level control_, _extensions to other multimodal tasks_, and _generalizability_. Additionally, we have provided further discussion on employing prompt engineering to enable control across a broader range of input queries. We are excited to explore this direction further in our future work. Thank you once again for the valuable suggestion and for your positive assessment of our work."}
|
2024-11-25 03:38:08
|
ICLR.cc/2025/Conference
|
h8PgljGBak
|
ZIZqJZi9Au
|
Authors
|
Response by Authors
|
{"Title": "Looking forward to the discussion", "Comment": "Dear Reviewer RZ9b,We deeply appreciate the time and effort you\u2019ve taken to review our work, especially given your busy schedule. As the authors-reviewer discussion phase draws to a close, we would be grateful for the opportunity to engage in dialogue with you. Our goal is to ensure we've adequately addressed your concerns and welcome any additional questions or points of discussion you'd like to raise.Thank you for your thoughtful consideration. Best regards,\\The Authors"}
|
2024-11-25 19:24:31
|
ICLR.cc/2025/Conference
|
oFEIHo3fcf
|
h8PgljGBak
|
Reviewer_RZ9b
|
Response by Reviewer
|
{"Title": "Response to rebuttal", "Comment": "Thanks for the authors' detailed responses. Part of my concerns are addressed. I still have concerns about the generalizability of the proposed method as the authors suggest in Q3.1 that different training data would lead to significant performance differences for some methods like the vanilla Lora. Besides, I appreciate that the authors have provided experiment results on additional benchmarks, but it would be better to provide the full-finetuning and Lora-tuning results for reference as well. Therefore, I will only increase my score to 5."}
|
2024-11-26 01:39:10
|
ICLR.cc/2025/Conference
|
vwsxcMRayn
|
ZIZqJZi9Au
|
Authors
|
Response by Authors
|
{"Title": "Thank you for the prompt response", "Comment": "Dear Reviewer RZ9b,We sincerely appreciate your engagement in the discussion and your valuable feedback, which are crucial in enhancing the quality of our work. We are pleased that our response addresses most of your concerns and would like to take this opportunity to provide additional results based on your suggestions.**Regarding the performance of vanilla Lora.**Thanks for the additional question. We\u2019d like to provide some explanations here. **First**, we completely agree that the performance gap between full fine-tuning and LoRA is small with the original LLaVA\u2019s dataset. In fact, as shown in our results, LoRA **also achieves comparable or even superior performance** to full fine-tuning on the Text-VQA, VSR, and SNLI-VE datasets. However, the performance gap is more pronounced on benchmarks such as MME, CIFAR-100, MNIST, and POPE. We hypothesize that this discrepancy arises because the Vision-Flan dataset (191K) is relatively smaller and differs in data quality and distribution compared to LLaVA\u2019s dataset (665K). Similar performance gaps between full fine-tuning and LoRA have been observed in MixLoRA (see [Table 1](https://aclanthology.org/2024.acl-long.38.pdf)) and M$^2$PT (see [Table 1](https://aclanthology.org/2024.emnlp-main.218.pdf)), both of which were also trained on Vision-Flan. **Second**, to validate this further, we conducted an additional experiment using LLaVA\u2019s dataset. The results on the MME benchmark confirm that the performance gap between full fine-tuning and LoRA is effectively closed when trained on this dataset. Meanwhile, MRT consistently outperforms LoRA, maintaining its superior performance.| LLaVA-v1.5 7B | MME || ---- | ----------- || Full Fine-tuning | 1510.7 || LoRA | 1476.9 || MRT | 1486.7 |**Regarding full-finetuing and lora results on the additional experiments.**Following your suggestion, we have included the results of full fine-tuning and LoRA-tuning on additional benchmarks for completeness. The findings show that there is no large performance gap between LoRA-tuning and full fine-tuning on the SEED and GQA benchmarks, while MRT consistently delivers the best performance. Similarly observations are seen on MiniGPT-v2.| Method | SEED-Bench | GQA || ---- | ----------- | --------------- || **Full Fine-tuning** | 57.4 | 53.5 || **LoRA** | 56.3 | 51.3 || MixLoRA | 55.9 | 52.2 || M$^2$PT | 57.1 | 50.3 || MRT | 57.6 | 52.7 || MiniGPT-v2 + EVA | MME || ---- | ----------- || **Full Fine-tuning** | 1464.88 || **LoRA** | 1358.14 || ReFT | 1346.65 || MixLoRA | 1418.48 || M$^2$PT | 1421.02 || MRT | 1439.73 |We further include the training efficiency analysis for full fine-tuning and LoRA-tuning. MRT demonstrates competitive training efficiency: It offers lower GPU memory usage and reduces training time compared to several baselines, such as LoRA and MixLoRA.| Method | trainable params | Memory Usage (GB)* | Training Time (Hours) | MME | | ---- | ---- | ----------- | --------------- |--------------- || **Full Fine-tuning** | 100% | 39 | 47 | 1587.26 || **LoRA** | 0.63% | 19 | 16 | 1393.67 || VPT | 0.06% | 12 | 7 | 1398.74 || MixLoRA | 0.85% | 23 | 24 | 1509.61 || M$^2$PT | 0.09% | 17 | 9 | 1503.98 || MRT | 0.03% | 16 | 9 | 1580.40 |We greatly value each comment and suggestion from your review, and are hoping that our additional clarifications and experimental results address your concerns. We are eager to address any remaining issues during the discussion phase. Thank you very much.Best Regards,\\Authors"}
|
2024-11-26 23:26:47
|
ICLR.cc/2025/Conference
|
EThjHK8HGs
|
iS2px2xmKO
|
Authors
|
Response by Authors
|
{"Title": "Looking forward to the discussion", "Comment": "Dear Reviewer uVrb,We sincerely appreciate your dedicated time and effort in reviewing our submission. We understand how demanding your schedule might be and are genuinely grateful for your valuable insights. As the discussion phase nears its conclusion, we kindly seek your input on our responses to your feedback.We hope to confirm that we\u2019ve adequately addressed your concerns and are open to discussing any remaining points or questions you may have. Your input is invaluable to us, and we would greatly value the chance to discuss them with you.Thank you again for your time and consideration.Best regards, \\The Authors"}
|
2024-11-29 03:50:15
|
ICLR.cc/2025/Conference
|
mIjca6M5Ug
|
ZIZqJZi9Au
|
Authors
|
Response by Authors
|
{"Title": "", "Comment": "Dear Reviewer RZ9b,We've updated the revised paper based on your suggestions by adding the full fine-tuning and LoRA results to those additional experiments.As the end of the discussion phase is approaching, we would be truly grateful if you could inform us whether our recent response has adequately addressed your additional questions. Your feedback is invaluable to us and plays a critical role in enhancing the quality of our work. We deeply appreciate the effort and time you have dedicated to reviewing our paper.Best Regards,\\Authors"}
|
2024-12-02 02:05:18
|
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 9
Size of downloaded dataset files:
523 MB
Size of the auto-converted Parquet files:
523 MB
Number of rows:
472,448