title stringlengths 15 163 | paper_decision stringclasses 4
values | review_1 stringlengths 853 32.6k | rebuttals_1 stringlengths 0 15.1k | review_2 stringlengths 1.03k 35.6k | rebuttals_2 stringlengths 0 15.1k | review_3 stringlengths 807 27.4k ⌀ | rebuttals_3 stringlengths 0 15k ⌀ | review_4 stringlengths 780 22.2k ⌀ | rebuttals_4 stringlengths 0 15.1k ⌀ | review_5 stringclasses 171
values | rebuttals_5 stringclasses 166
values | review_6 stringclasses 25
values | rebuttals_6 stringclasses 24
values | review_7 stringclasses 4
values | rebuttals_7 stringclasses 4
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Interchangeable Token Embeddings for Extendable Vocabulary and Alpha-Equivalence | Accept (poster) | Summary: The paper introduces a way to handling interchangeable tokens in neural networks, particularly for tasks involving formal logic and symbolic reasoning. The proposed method employs a token embedding strategy that combines a shared component for semantic consistency and a randomized component for token different... | Rebuttal 1:
Rebuttal: We kindly thank the reviewer for their constructive criticism.
> The embedding matrix is constructed with a fixed number of interchangeable tokens, which suggests that the set of interchangeable tokens must be predetermined before training.
This is the case for the alpha-renaming baseline, but n... | Summary: The paper introduces a novel method for handling interchangeable(semantically equivalent yet distinct) tokens in a language model. The proposed approach uses a dual-part token embedding strategy consisting of a shared semantic component and a randomized component to ensure distinguishability. The paper also in... | Rebuttal 1:
Rebuttal: We kindly thank the reviewer for their detailed analysis.
> **The method generalizes better than data augmentation**: This claim needs stronger justification as the the alpha-renaming method performs better than the proposed method in Table 2 and a deeper analysis is required to justify this.
We... | Summary: The authors propose an interesting token embedding technique to handle alpha-equivalent programs—those that share the same semantic meaning but differ in surface representation. By employing a dual-part token representation, the embedding effectively captures both semantic consistency and token distinguishabil... | Rebuttal 1:
Rebuttal: We kindly thank the reviewer for their comments.
> Does the formal language include natural elements, such as comments and textual descriptions? If so, could the embedding approach lead to representational conflicts when switching between these contexts?
The applications we considered in this pa... | Summary: The paper proposes a method for handling interchangeable tokens—like bound variables—in language models by splitting token embeddings into a shared semantic part and a unique random part. This improves generalisation to unseen vocabularies and supports alpha-equivalence. Evaluated on copying tasks, LTL solving... | Rebuttal 1:
Rebuttal: We kindly thank the reviewer for their positive feedback.
> If I understand correctly, the proposed techniques are designed for problems—like LTL and propositional logic—where the grammar is clear or simple enough to reliably identify bound variables. How well do you think these techniques would ... | null | null | null | null | null | null |
Mastering Multiple-Expert Routing: Realizable $H$-Consistency and Strong Guarantees for Learning to Defer | Accept (poster) | Summary: The paper presents a novel surrogate loss for both single-stage and two-stage learning to defer. The authors show several theoretical guarantees, including $H$-consistency, Bayes consistency and $H$-consistency bounds for both the single and multiple expert settings.
Empirical results showcase the effectivenes... | Rebuttal 1:
Rebuttal: Thank you for your encouraging review. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions.
**1. Empirical Evaluation:** Thank you for the suggestions. As recommended by the reviewer, we will aim to include experiments inv... | Summary: This paper studies 1. single-stage multiple-expert deferral, which is the deferral problem where the learner has the option to predict or defer the prediction to one of pre-defined experts, and 2. two-stage multiple-expert deferral, which assumes a label predictor have been learned in first stage and the learn... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our work. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions.
**1. Questions: Are there any other interesting properties or improvements the authors want to achieve for their loss family... | Summary: This work introduces a family of surrogate losses functions for the learning to defer. The authors discussed methods for achieving H-consistent loss for single-stage deferral, and H-consistency and Bayes consistency for two-/multi-expert scenarios. The authors propose to surrogate the undifferentiable indicato... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions.
**1. Experimental Designs Or Analyses:** Thank you for the suggestions. As recommended by Reviewer m1gt, we will aim to include e... | Summary: In the paper _Mastering Multiple-Expert Routing: Realizable-Consistency and Strong Guarantees for Learning to Defer_, authors propose surrogate losses for learning to defer that satisfy the following oints: realizable H-consistency, H-consistency bounds, and Bayes consistency
for both single-stage and two-sta... | Rebuttal 1:
Rebuttal: Thank you for your encouraging review. We will take your suggestions into account when preparing the final version. Below please find responses to specific questions.
**1. Claims regarding Verma et al. (2023):** Thank you for pointing this out. We agree with the reviewer that Verma et al. (2023) ... | null | null | null | null | null | null |
Reward Modeling with Ordinal Feedback: Wisdom of the Crowd | Accept (poster) | Summary: The paper extends the canonical setup of binary feed-back RM to ordinal feedback.
Claims And Evidence: See Other Strengths And Weaknesses.
Methods And Evaluation Criteria: I think more experimental results is needed to prove the superiority of the proposed method. For example, the authors could leverage the ... | Rebuttal 1:
Rebuttal: We thank the reviewer for all the comments. Here are our responses to the questions.
1. “I think more experimental results is needed to prove the superiority of the proposed method. For example, the authors could leverage the method in [1], setting up a seperated RM as the oracle and test the per... | Summary: The paper examines reward learning in scenarios where annotators provide ordinal feedback. Specifically, annotators select an option from a set $\mathcal{Z}$, which, in the common case of binary feedback, could be $\{0,1\}$. The authors first offer a statistical justification for the benefits of this extension... | Rebuttal 1:
Rebuttal: Thanks for the valuable questions. Before we dive into the rebuttal, we’d like to summarize the reviewer’s major concerns.
1. The construction of ordinal feedback via Theorem 3.2 is not justified. There are possible reasonable options (e.g., the Luce-Shepard model) that have not been considered.
... | Summary: This paper proposes a new framework for reward learning with ordinal feedback, which uses more fine-grained preference pairs generalized from binary preferences. Using a generalized probability model, the authors prove that the ordinal feedback framework reduces the Rademacher Complexity. The authors also use ... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating our work and the interesting questions.
1. Despite in the domain of RLHF, what is the difference between your framework and training upon soft labels in knowledge distillation?
The major difference between our work and the knowledge distillation of trained ... | Summary: This paper proposes Reward Modeling with Ordinal Feedback as an alternative to traditional binary feedback in reward modeling. The authors argue that binary preference data discards valuable information, such as subtle distinctions between choices and tied responses. They introduce an ordinal feedback framewor... | Rebuttal 1:
Rebuttal: We thank the reviewer for the kind comments on our work. Here are our responses to the reviewer’s concerns and questions.
1. Experiments only contain 2 base models, which appear not to be very solid for the LLM community.
Thanks for pointing out that issue. We have conducted some further experim... | null | null | null | null | null | null |
Metastable Dynamics of Chain-of-Thought Reasoning: Provable Benefits of Search, RL and Distillation | Accept (poster) | Summary: This paper models Chain-of-Thought (CoT) reasoning as a metastable Markov process, where dense clusters correspond to easy reasoning steps and sparse edges represent difficult transitions. It provides a theoretical analysis of how search, reinforcement learning (RL), and distillation improve hitting times by i... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and suggestions! Our responses are as follows.
**Claims And Evidence** - Thank you for pointing out the possible confusion. We have made sure that the training dynamics is only referred to as "optimization guarantee", "training dynamics", "gradient descent itera... | Summary: The authors propose a theoretical analysis of Chain-of-Thought (CoT) reasoning. Using a graph theoretical approach, they distinguish “easy” reasoning steps as intra-dense cluster connections from harder reasoning steps, which connect different clusters but have low probability of being picked at inference. The... | Rebuttal 1:
Rebuttal: Thank you for your positive assessment of our contributions and helpful suggestions! Our responses are as follows; typos have been fixed.
**Response to Weakness** - While we did not conduct any experiments due to time and compute constraints, the formulation and theory do give quantitative predic... | Summary: This paper develops a theoretical framework for understanding chain-of-thought (CoT) reasoning in large language models by modeling it as a metastable Markov process. They prove that implementing search protocols that reward sparse edges improves reasoning by decreasing the expected steps to reach different cl... | Rebuttal 1:
Rebuttal: Thank you for your positive assessment of our contributions! Regarding the simplicity of the Markov model compared to actual reasoning, please see our rebuttal to similar weaknesses pointed out by Reviewer R4eB. | null | null | null | null | null | null | null | null |
Implicit Regularization for Tubal Tensor Factorizations via Gradient Descent | Accept (oral) | Summary: In this paper, the authors study tubal tensor factorization via gradient descent. Assuming the ground truth tensor is
of low tubal rank, they show that gradient descent from a small initialization will converge to the ground truth tensor,
and the error can be made arbitrarily small by scaling down the initial... | Rebuttal 1:
Rebuttal: Thank you very much for taking the time to evaluate our paper and provide valuable feedback.
**RIP Condition:**
[Zhang et. al. 2021] shows that sub-Gaussian measurements $\mathcal{A}:\mathbb{R}^{n\times n\times k}\to\mathbb{R}^m$ have the RIP for tubal-rank $r$ tensors w.h.p. if $m\ge O(\delta... | Summary: This paper studies the factorization of third-order tubal tensors using gradient descent (GD) with small initialization in the overparameterized setting. Under the standard restricted isometry property (RIP) condition, the authors show that GD recovers the ground truth tensor in a finite number of steps, provi... | Rebuttal 1:
Rebuttal: Thank you very much for your positive evaluation of our paper. We are glad you found it to be interesting and well-written.
**Clarifying the Connection to Matrix Factorization:**
Thank you for bringing the connection to matrix factorization to the discussion. Yes, while aiming to leverage matr... | Summary: To gain insight on initialization methods, the authors consider its effect on a gradient descent method for calculating a low rank tubal tensor tensor factorization, given linear measurements of the tensor. Tensor tubal rank of a $m\times n\times k$ tensor is the the rank of the SVD of the front $m\times n$ sl... | Rebuttal 1:
Rebuttal: Thank you very much for your positive evaluation of our paper and for recognizing our work as one of the few papers that fully analyzes gradient descent.
**Motivation for our work and the use of tubal rank:**
The study of implicit regularization was originally motivated by trying to understand... | null | null | null | null | null | null | null | null |
The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models | Accept (poster) | Summary: This work introduces function-calling benchmark, a wide-coverage benchmark that evaluates LMs' ability of invoking correct function calls. Multiple parts and types of function calling are included: single-turn, crowd-sourced, multi-turn and agentic tasks. Furthemore, an evaluation method based on Abstract Synt... | Rebuttal 1:
Rebuttal: We thank the reviewer krTJ for the time to review our paper and the comments. We emphasize the motivation, purpose, and contribution of our work. We address your questions as follows.
1. **Need more discussions on the comparisons and differences to existing benchmarks**
A: Although in recent ti... | Summary: * The paper proposes a benchmark for the task of function calling ("FC") i.e., given a prompt and a set of available functions, to predict the correct sequence of function calls along with necessary parameters that accomplishes the prompt requirement.
* The benchmark evaluates FC ability along various categori... | Rebuttal 1:
Rebuttal: We thank the reviewer dutu for the time to review our paper and the comments. We address your questions as follows.
1. **Need More Discussions on the Comparisons and Differences to Existing Benchmarks**
A: Please see our rebuttal comment to reviewer krTJ.
2. **Concern Over Potential Train/Test... | Summary: This paper describes a benchmark for evaluating function calling capabilities of large language models. It includes different function calling settings such as single-turn, multi-turn, multiple functions in a single turn, parallel functions etc. The evaluation setup defines execution and non-execution based me... | Rebuttal 1:
Rebuttal: We thank the reviewer VP8g for the time to review our paper and the comments. We address your questions as follows.
1. **The Proposed Settings are Too Similar to Those From BFCL**
A: In the interest of double-blind reviewing, we refrain from commenting on this.
2. **Need More Discussions on th... | Summary: This paper proposes a new benchmark FCL for LLM’s function calling, which contains a ‘single-turn’ dataset, a ‘crowd-sourced’ dataset, a ‘multi-turn’ dataset, and an ‘agentic’ dataset.
Claims And Evidence: The authors claim that "Despite the importance of function calling, there isn’t a standard benchmark to ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the time to review our paper and the comments. We emphasize the motivation, purpose, and contribution of our work.
1. **Need more discussions on the comparisons and differences to existing benchmarks**
A: Please see our rebuttal comment to reviewer krTJ.
2. **Uneven ... | null | null | null | null | null | null |
Supervised Contrastive Learning from Weakly-Labeled Audio Segments for Musical Version Matching | Accept (poster) | Summary: The paper addresses the challenge of matching different renditions of the same musical piece at segment level. The authors propose a method that uses pairwise segment distance reductions for weakly annotated data, combined with a modified alignment-and-uniformity contrastive loss. This approach achieves state-... | Rebuttal 1:
Rebuttal: We are grateful to the Reviewer for his/her assessment and comments. We now answer and discuss about the main questions raised in the Review.
1) Additional data sets ---- This is a relevant suggestion that is worth discussing. We originally computed the results for covers80 and we paste them belo... | Summary: In this paper, authors propose a method (CLEWS) to learn from weakly annotated segments, together with a contrastive loss variant that outperforms well studied alternatives. State-of-the-art performance on both track-level and segment-level evaluations, significantly outperforming existing methods. The approac... | Rebuttal 1:
Rebuttal: We thank a lot the Reviewer for his/her assessment and constructive feedback. We answer to and discuss on the questions posed by the Reviewer below. As suggested by the Reviewer, we are considering adding a figure to describe the CLEWS model structure and pipeline, especially along with the shared... | Summary: This paper presents a novel approach for musical version matching at the segment level using weakly labeled audio segments and a supervised contrastive learning framework. The authors highlight two key contributions: (1) a method to effectively learn from weakly annotated musical segments, which enables more g... | Rebuttal 1:
Rebuttal: We thank the Reviewer for his/her assessment of our work and the constructive feedback. We first answer to the questions posed by the Reviewer and later discuss on a few further aspects that appear in the main review.
1) Label inconsistencies ---- Like the big majority of machine learning models,... | Summary: The proposed paper is about cover song detection or musical version matching task. This task is to identify different versions of a same song, for example, if there exist an original song and this song has been sung by other singers with different vocal styles, this can be regarded as a cover song. Also, if th... | Rebuttal 1:
Rebuttal: We thank the Reviewer for his/her comments and insight. We want to clarify two misunderstandings by the Reviewer, and comment on the further analyses part. We think that the Reviewer may increase his/her score once these two misunderstandings are clarified, as they are the only two strong criticis... | null | null | null | null | null | null |
Improving Reward Model Generalization from Adversarial Process Enhanced Preferences | Accept (poster) | Summary: The paper presents APEC, a novel approach to improving reward model learning by generating diverse preference pairs from different stages of adversarial imitation learning. This ensures broader coverage of the learning process. Theoretical analysis shows that later AIL policies are generally preferred over ear... | Rebuttal 1:
Rebuttal: **Q1**: The method is dependent on that later policies are better than earlier ones, but this may not always hold due to training fluctuations. A discussion on how this randomness impacts learning would be valuable.
**A1**: This randomness indeed has a negative impact on the accuracy of the gener... | Summary: The approach of automatically generating preference data can reduce human intervention, but it has limited generalizability and coverage. To solve this problem, the authors propose APEC, which addresses these challenges by selecting policy pairs with significantly different iteration indices from the whole adv... | Rebuttal 1:
Rebuttal: **Q1**: Need to check whether there are more recent and suitable baselines available
**A1**: Thank you for your valuable suggestions. We add two more recent reward learning baselines, BC-IRL [1] and DRAIL [2], and evaluate them on the MuJoCo benchmark. We use the recommended hyperparameters from ... | Summary: In this paper, an automated preference generation method, namely APEC (Automated Preference generation with Enhanced Coverage), is proposed, which claims that broad coverage in the generated preference data can be ensured by leveraging the policy pairs with different iteration indices from the AIL training pro... | Rebuttal 1:
Rebuttal: **Q1:** Whether the claim that a progressive improvement in policy value over successive iterations can be achieved is correct or not.
**A1:** We apologize for this confusing claim. Here we want to claim that Proposition 3.1 can indicate that policies from the AIL process will establish a prefere... | Summary: APEC addresses the challenge of generating high-quality preference data for reward modeling in sequential decision-making without human expertise. Unlike previous approaches that rely on injecting random noise into fixed policies, APEC selects policy pairs with significantly different iteration indices from th... | Rebuttal 1:
Rebuttal: Thank you for your valuable suggestions!
**Q1**: Could the authors provide a comparative analysis of computational costs between APEC and baseline methods?
**A1**: The table below shows the computation costs (training time) of all methods on the MuJoCo task. We can see that APEC and the other m... | null | null | null | null | null | null |
Pre-training Auto-regressive Robotic Models with 4D Representations | Accept (poster) | Summary: The paper presents ARM4R, an Auto-regressive Robotic Model that learns 4D representations from human videos to enhance robotic learning.
- 4D Representation Learning: Transforms 2D videos into 3D point tracks, enabling temporal and spatial understanding.
- Efficient Transfer to Robotics: Pre-training on human ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We address the concerns raised below:
1. **Causal Analysis & Optimality of 4D Representations:** As stated in the paper, one of the main benefits of using 4D representations is that the 3D point tracks in a robotic setting are described by linear... | Summary: The paper presents the ARM4R (Auto-regressive Robotic Model 4R), a novel approach to improve robotic pre-training by utilizing low-level 4D representations derived from human video data. The paper identify a significant challenge in the existing robotic models, which stems from a lack of large-scale, diverse d... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We address the concerns raised below:
1. **Additional Limitations:** Here we provide a further analysis investigating more limitations of our approach.
***(i) Unnatural rotation:*** We examined a new task, “put knife” in RLBench. Interestingly,... | Summary: The paper introduces ARM4R (Auto-regressive Robotic Model with 4D Representations), a novel approach to pre-training robotic models using 4D representations derived from human video data. The key idea is to leverage 3D point (generated by depth estimation) tracking from human videos to create a low-level repre... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We address the concerns raised below:
1. **Comparison to RVT, RVT-2:** As the reviewer noted, RVT and RVT-2 are recent transformer-based methods that predict keyframes in robot trajectories. Importantly, both methods leverage RGB-D images to reco... | Summary: The paper proposes to pre-train the vision-language-action models using the sparse 4D trajectories (obtained as pseudo-labels using pre-trained networks) of humans performing actions in ego-centric videos. Such pre-training is claimed to be helpful when the model is fine-tuned on the downstream task of action ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments. We address the concerns raised below:
1. **Table 1 Results:** As noted by the reviewer, the simulation results for ARM4R are slightly skewed by the high success rate for the “put money” task. We recalculate the success rate without this outlier fo... | null | null | null | null | null | null |
Looking Beyond the Top-1: Transformers Determine Top Tokens in Order | Accept (poster) | Summary: This paper studies the "saturation event" in transformers, where the model’s top $ k $ predictions are determined in the early layers and remain fixed in the later layers. They show empirically that the top-ranked tokens saturate in order, with higher-ranked tokens saturating in earlier layers. They also demon... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's helpful feedback.
**Methods And Evaluation Criteria**
>A simpler and more direct option could be reporting the percentage of inputs where the order of the saturation layers matches the order of the tokens.
We calculated this percentage for Llama3-8B on 1K MMLU quest... | Summary: This paper studies the calculation process of the hidden layer of the Transformer model after completing the top-1 prediction, and find that the model will determine the subsequent tokens (such as top-2, top-3, etc.) in order of token ranking.
Through experiments across modalities (text, vision, speech) and ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and positive assessment of our work! We appreciate your recognition of the novelty of the ordered saturation phenomenon and its implications for understanding Transformer models. Your acknowledgment of both the theoretical contributions and practical applications, suc... | Summary: This paper shows that at the intermediate layers, model will sequentially determine the top-k token, rather than just determine the top-1 token. And this observation maybe some special property of transformer architecture. The author calls this as task transition, and show that this is a general phenomenon on ... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments and helpful suggestions. We appreciate that you found the sequential determination of top-k tokens to be an interesting and well-supported phenomenon. We're also glad that you recognized the broader relevance of our findings in relation to existing literature... | Summary: This work analyses the saturation event observed in transformer layer, where a layer's top-1 prediction for the next best token remains the same from some layer onwards. This paper expands the analysis beyond that of top-1, illustrating how the argument continues for top-k and analyses a specific component fro... | Rebuttal 1:
Rebuttal: Thank you for your thorough review.
**Claims And Evidence**
>I don't believe results provided in Figure 3 .. are convincing enough beyond the 3rd top token.
Beyond the top-3 tokens, saturation order is indeed less consistent in some models. However, our results are statistically significant acr... | null | null | null | null | null | null |
KoopSTD: Reliable Similarity Analysis between Dynamical Systems via Approximating Koopman Spectrum with Timescale Decoupling | Accept (poster) | Summary: This paper proposes KoopSTD, a new framework for comparing the similarity of non-linear dynamical systems by analyzing their Koopman spectra across multiple timescales. The main idea is to first transform time-series data into a time-frequency representation via short-time Fourier transforms, which extracts mu... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thorough assessment of our work. We are delighted that you recognize KoopSTD's contributions to providing a robust and multi-scale similarity metric for nonlinear dynamical systems. In particular, we deeply appreciate your time and careful attention in examining our ex... | Summary: A method for computing dissimilarity between dynamical systems is proposed. It is based on the Wasserstein distance between sets of eigenvalues of the Koopman operators of the dynamics. Here two major heuristics are adopted. One is the STFT-based observable, which the authors motivate as a way to deal with mul... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's valuable feedback!
Please kindly find the point-to-point response to the questions and suggestions.
Q1: How about applying DMD on the complex part of STFT results instead of the real part?
> We appreciate the reviewer’s insightful question. We agree that d... | Summary: This paper is part of the broad hot topic of comparing non-linear neural networks through the lens of representation space and focuses more specifically on the measurement of dissimilarity between non-linear dynamical systems. These systems, due to their non-linearity, can be analyzed using the Koopman operat... | Rebuttal 1:
Rebuttal: Thanks for your careful review and comments!
Glad to see that you deem our work interesting and have a promising impact on many fields. Please kindly find our point-to-point responses below.
## Concerns
Q1: Presenting the invariance property as 'theorem'.
>We revise the invariance property from... | null | null | null | null | null | null | null | null |
Polynomial Time Learning Augmented Algorithms for NP-hard Permutation Problems | Accept (poster) | Summary: The paper considered permutation-based NP-hard problems with learning-augmented oracles. The paper proposed a novel framework that given a learning-augmented oracle on the ordering of pairs with correct probability $½+\Omega(1)$, solves certain families of NP-hard problems exactly in polynomial time and O(n lo... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and the provided questions.
(Question/Comment) "A broader discussion about learning-related algorithms would be better (e.g., learning augmented frequency estimation, e.g., Aamand et al. [NeurIPS’23]; TSP problem, e.g., Chawla and Christou [APPROX’24]) [...]... | Summary: The paper presents a framework for studying NP hard optimization problems in the framework of learning-augmented algorithms. Specifically, the authors consider a class of permutation problems, where a solution is determined by a total order of a set of $n$ items, such as TSP, single machine scheduling, etc, an... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the feedback.
About the comments of the reviewer: The model of epsilon-accurate predictions has indeed been used for other problems. So, the model is not novel, but we would like to note that our work is completely different from previous works on this mode... | Summary: The paper studies how to tackle NP-hard permutation problems with the help of predictions on the relative position of any two elements with respect to some fixed unknown optimal permutation. Each query is independently correct with probability at least $1/2 + \varepsilon$ for some unknown $\varepsilon$. Buildi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback and the provided questions. We will implement the suggestions on a revised version.
- (Comment) While theoretically interesting, it is unclear to me how such a prediction can be available in practice for each of the problems mentioned in the paper [BM08]
(A... | Summary: This paper studies a broad class of optimization problems that can be captured via permutations with the help of predictions. For example consider the TSP problem: the optimal solution corresponds to a permutation of the input points. The predictions are weather point i precedes point j in the permutation and ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the feedback and the provided questions.
- (Comment) Minor comment but [1] is also among the first paper to use predictions to warm start ...
(A) Thank you, we will add this reference on a revised version.
- (Comment) I was really confused in the beginni... | null | null | null | null | null | null |
Commute Graph Neural Networks | Accept (poster) | Summary: The paper introduces Commute Graph Neural Networks (CGNN), a framework designed to improve learning on directed graphs by incorporating bidirectional path information.
Key components of CGNN include a digraph Laplacian (DiLap) leveraging random walk transition probabilities, efficient computation of node-wise... | Rebuttal 1:
Rebuttal: Thanks for your insightful feedback. We've consolidated your comments into 8 questions and provided responses to each.
**Q1: The definition of Moore–Penrose Pseudoinverse (MPP) in Lemma 4.2, and a detailed derivation of Theorem 4.3**
**A1:** Here we provide the definition of MMP. In linear algeb... | Summary: This paper proposes a directed graph Laplacian based on the commute time between nodes. The paper also proposes a commute graph neural network model based on rewiring the graph to satisfy model assumptions and the Laplacian matrix. Experiments are conducted to validate the model's efficacy, together with ablat... | Rebuttal 1:
Rebuttal: Thanks for your insightful review. We have summarized all your feedback into 6 key questions and address below.
**Q1: Why computing commute time rather than hitting time twice with different source and target nodes**
**A1:** In fact, our method precisely aligns with your suggestion -- “computing... | Summary: The paper presents a new approach to embed commute time information when learning on directed graphs, for the task of node property prediction. It gives a nice introduction for commute time and a motivation for it. Then, the work proposes to approximate commute time information which is usually computed in O(|... | Rebuttal 1:
Rebuttal: Thanks for your feedback. Our detailed responses follow.
**Q1: Clarify how to maintain effectiveness given that the commute time information is partially preserved**
**A1**: The effectiveness can be well maintained, as the proposed graph rewiring approach **does not drastically alter** the origi... | Summary: This paper proposes a novel method for learning on directed graphs. The authors argued that it is important to preserve a notion of mutual closeness in directed graphs via commute time, and introduced a method to compute/approximate this efficiently. They then encode this information into the design of message... | Rebuttal 1:
Rebuttal: **Q1: Alternative ways of using the commute time in designing a directed GNN, e.g., regularisation or a non-MPNN?**
**A1**: We appreciate your suggestion. While alternative approaches like incorporating commute time as a regularization term or exploring non-MPNN frameworks are interesting thought... | null | null | null | null | null | null |
SITCOM: Step-wise Triple-Consistent Diffusion Sampling For Inverse Problems | Accept (poster) | Summary: Authors propose Step-wise Triple-Consistent Sampling (SITCOM) an optimization-based sampling method that enforces measurement consistency, forward diffusion consistency, and incorporates step-wise and network-regularized backward diffusion consistency that maintains a diffusion trajectory by optimizing over th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments.
We appreciate the reviewer’s recognition of our clear and well-supported claims, as well as their positive feedback on our experimental analysis and writing.
Our point-by-point response is provided below.
### (W2 and S1) **Memory usage for SITCOM.**
... | Summary: This paper introduces a diffusion posterior sampling method, SITCOM, for solving inverse problems. SITCOM is designed to hold three consistency conditions during the diffusion sampling procedure to guide the generation. The backward consistency aims to ensure that the solution after data consistency optimizati... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's recognition that our paper establishes three reasonable consistency conditions to guide the algorithm's design and acknowledges SITCOM's superior performance in image restoration and MRI.
We refer the reviewer to (https://anonymous.4open.science/r/SITCOM-B65F/rebuttal... | Summary: This paper systematically analyzes three necessary conditions—forward consistency, measurement consistency, and backward consistency—that enable accurate inverse problem solving using diffusion models. A sampler is proposed to enforce these conditions while reducing the required number of reverse diffusion ste... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. We refer the reviewer to (https://anonymous.4open.science/r/SITCOM-B65F/rebuttal_table.md) for Tables A & B.
We're glad the reviewer finds our paper well-motivated and well-written and acknowledges SITCOM's consistent improvements over most baselines wit... | null | null | null | null | null | null | null | null |
AKRMap: Adaptive Kernel Regression for Trustworthy Visualization of Cross-Modal Embeddings | Accept (poster) | Summary: In their paper "AKRMap: adaptive kernel regression for trustworthy visualization of cross-modal embeddings", the authors suggest what I would call *supervised parametric t-SNE* where supervision is by a continuous variable (as opposed to discrete classes). The algorithm combines the t-SNE loss function with a ... | Rebuttal 1:
Rebuttal: Thank you very much for the great review helping us improve the work. We respond to your questions below and provide the figures you ask for in the link: https://docs.google.com/document/d/e/2PACX-1vROjYSQAj8XMbrghXX4Ba_JrAxdFNagPOB84DtTK2PN5B_Ir-CAWew_c6zfhod7h3bU0VgCj_K8zuz8/pub
## Q1 Why do co... | Summary: The paper proposes AKRMap, an adaptive kernel regression-based dimensionality reduction (DR) method for trustworthy visualization of cross-modal embeddings. Unlike conventional DR techniques (e.g., PCA, t-SNE, UMAP), AKRMap jointly learns a supervised projection network and an adaptive kernel regression loss, ... | Rebuttal 1:
Rebuttal: Thank you very much for your great review helping us improve the work. We respond to your questions below and provide additional figures and tables in the link: https://docs.google.com/document/d/e/2PACX-1vQDqGw_hxxTv_CQbFS_zdI5LtgIE6HTfAYl7uBRgpczNScFBuja6pwbuYrV-76D5CleVrujYDH0O1yP/pub
## Q1 Ha... | Summary: The paper introduces a new framework for dimensionality reduction and visualization of cross-modal embeddings. The method incorporates adaptive kernel regression to more accurately capture metric distributions in the projection space. It jointly optimizes a projection network and kernel function, enabling more... | Rebuttal 1:
Rebuttal: Thank you very much for your positive feedback and constructive comments that help us improve our work. Below, we answer your specific questions, with two additional figures provided at the link:https://docs.google.com/document/d/e/2PACX-1vR8toOEJ8KrADdAzAXJbIMazphKXlKq4K4oqwDxAvvpLzqPu0jufg7ns67A... | null | null | null | null | null | null | null | null |
A Geometric Approach to Personalized Recommendation with Set-Theoretic Constraints Using Box Embeddings | Accept (poster) | Summary: This paper proposes a geometric approach to personalized item recommendation using box embeddings. Unlike traditional vector-based methods that struggle with set-theoretic operations, the authors introduce axis-aligned hyperrectangles (boxes) to model users, items, and attributes. This enables logical operatio... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s suggestion to clarify how natural language queries are mapped to structured forms, as well as their thoughtful concerns regarding generalization, scalability, and efficiency.
**Mapping from natural language to attributes:**
In real-world recommendation systems — su... | Summary: This paper demonstrates the inconsistency of existing vector embedding models for personalized item recommendation task and discusses the challenges faced by existing machine learning approaches for this problem setup. To alleviate the issue, the authors formulated the problem of personalized item recommendati... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful and constructive review.
We appreciate your detailed engagement with our paper and the helpful suggestions on evaluation metrics.
**Recommendation Task / NDCG evaluation / Hyperparameter**
We calculate the NDCG score for the set theoretic queries. Please refer to the... | Summary: 1. This paper models the problem of attribute-specific query recommendation as “set-theoretic matrix completion”, where attributes and users are treated as sets of items.
2. It demonstrates the inconsistency of existing vector embedding models for the above task, and establishes box embeddings as a suitable em... | Rebuttal 1:
Rebuttal: Thank you for going through the paper so meticulously — we truly appreciate the careful reading and thoughtful feedback. We're glad you found the framework novel and the paper well-written, and we appreciate your comments highlighting both the strengths and areas for improvement.
We apologize for... | Summary: This paper proposes a box embedding-based recommendation model that captures set-theoretic constraints in personalized recommendation tasks. Traditional vector-based embeddings struggle with expressing complex relationships like intersection and negation (e.g., "comedy and action but not romance"). The paper i... | Rebuttal 1:
Rebuttal: We thank the reviewers for raising thoughtful and forward-looking questions regarding possible extensions of our approach.
We agree that a recommendation systems framework should be recognized not only for its current performance but also for its extensibility to address practical challenges suc... | null | null | null | null | null | null |
SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer | Accept (poster) | Summary: This paper introduces SANA 1.5, aims at scaling up the linear diffusion transformers from the training and inference perspectives. Based on a pre-trained linear diffusion model. This paper proposes three techniques, Efficient Training Scaling, Model Depth Pruning, and Inference-time Scaling to investigate the ... | Rebuttal 1:
Rebuttal: ### We sincerely appreciate your insightful feedback, thoughtful comments and kind words. We have carefully addressed all questions to the best of our ability and hope the revisions meet your expectations.
### Q1: Scaling Depth v.s. Scaling FFN / channel dimensions
**A1:** While conventional appr... | Summary: This paper proposes three important components: 1. efficient model scaling with depth, 2. efficient model depth pruning, and 3. inference time model performance scaling with VLM. Those three components show that an efficient training strategy can enable a smaller model to achieve performance comparable with th... | Rebuttal 1:
Rebuttal: ### We sincerely appreciate your insightful feedback, thoughtful comments and kind words. We have carefully addressed all questions to the best of our ability and hope the revisions meet your expectations.
### Q1: Efficiency-Cost Tradeoff of VLM verifier
**A1:** This is a good question and we lis... | Summary: The paper introduces SANA-1.5, which incorporates a series of techniques to efficiently scale up the SANA-1.0 linear-attention diffusion Transformer. Specifically, it proposes a depth-growth paradigm that includes partial-layer preservation initialization and a memory-efficient CAME-8bit optimizer. Additionall... | Rebuttal 1:
Rebuttal: ### We sincerely appreciate your insightful feedback, thoughtful comments and kind words. We have carefully addressed all questions to the best of our ability and hope the revisions meet your expectations.
### Q1: Growth-then-Prune vs Train-from-Scratch
**A1:** Direct training of 30/40/60-layer m... | Summary: The paper introduces SANA-1.5, an efficient linear Diffusion Transformer for text-to-image generation.
Key contributions include:
(1) A depth-growth paradigm that scales models from 1.6B to 4.8B parameters, reducing training costs by 60%;
(2) A technique called model depth pruning via block importance an... | Rebuttal 1:
Rebuttal: ### We sincerely appreciate your insightful feedback, thoughtful comments and kind words. We have carefully addressed all questions to the best of our ability and hope the revisions meet your expectations.
### Q1: Will VLM-based Verifier be the Best and any Other Choices?
**A1:** We choose VLM-b... | null | null | null | null | null | null |
The Value of Prediction in Identifying the Worst-Off | Accept (oral) | Summary: The paper investigates the welfare impacts of using machine learning prediction systems in equity-driven government programs, particularly in identifying and supporting the most vulnerable individuals. The authors develop a framework to evaluate the relative effectiveness of prediction systems compared to othe... | Rebuttal 1:
Rebuttal: Thank you for taking the time to carefully read our manuscript and provide comments. It is very much appreciated.
Other Domains. Yes, performing similar evaluations in other empirical domains would be valuable, but beyond the scope of this paper due to the substantial effort required to access ad... | Summary: In this paper, the authors consider a scenario where, for example, a government agency wants to identify the x% worst off in a population, based on a predicted metric $Y$. Once we have the predicted $Y$ for the population, we can sort everyone and have a follow-up check on those that scored worst. However, the... | Rebuttal 1:
Rebuttal: Thank you for your detailed, thoughtful questions and suggestions for improving the paper! We appreciate your encouraging feedback on our conceptual framework and empirical analysis.
Cost Ratio and PAR = 1. Thank you for raising this. We fully agree that $\Delta_{\alpha}$ and $\Delta_{R^2}$ repre... | Summary: This paper investigates the value of ML prediction when the goal is to identify the worst candidates as a screening problem. Formulating the problem as a screening problem with threshold, this paper derives 2 main theorems that characterizes the behavior of screening policies in general, and complement the stu... | Rebuttal 1:
Rebuttal: Thank you very much for your positive feedback. We're especially glad you found the paper clear and relevant, and appreciate the time you took to engage with our work. | Summary: The authors study the question when and whether policymakers should invest into improving their predictions or expanding screening, with respect to bottom-line welfare. They set up a simple theoretical model, and validate the resulting hypothesis in semi-synthetic experiments on the German labor market.
Claim... | Rebuttal 1:
Rebuttal: Thank you for engaging with our work and for your constructive feedback. We very much appreciate it!
Beta Values. 0.15 reflects the official government threshold used in Germany. However, we included additional values to account for differing long-term unemployment thresholds used in other countr... | null | null | null | null | null | null |
SPD: Sync-Point Drop for Efficient Tensor Parallelism of Large Language Models | Accept (poster) | Summary: Sync-Point Drop (SPD) is an inference-time technique for reducing the latency of distributed inference with tensor parallelism. The paper suggests removing all-reduce operation between attention and feed-forward layers of transformer blocks. At the same time, some of the residual connections within the transfo... | Rebuttal 1:
Rebuttal: We are deeply grateful for your support of our work, and we provide detailed responses to your comments as follows:
**Q1. The paper would benefit from quantifying the potential speed-up in set-ups that do not suffer communication bottlenecks.**
Thanks for the comment. We plan to add more details... | Summary: This paper introduces SPD, a novel optimization to reduce communication overheads in tensor parallelism b selectively dropping synchronization on attention outputs. SPD categorizes attention blocks into 3 types based on their sensitivity to the model accuracy, and applies different block design. SDP effectivel... | Rebuttal 1:
Rebuttal: Thanks for your support on our work:
**Q1. Insufficient details about how to set hyper-parameters**
Thank you for the comment. While we will add more details on hyper-parameters to the final draft, we like to point out one observation we made during our study about τ1 and τ2. We observed that th... | Summary: This paper proposes a novel method to drop one of the two all-reduce communications in tensor parallelism (TP) to mitigate the communication of TP in LLM inference. After dropping the first all-reduce communication, the sync point between the attention and mlp is removed, raising concerns on the model accuracy... | Rebuttal 1:
Rebuttal: **Q1. This paper claims to improve the scalability of TP, but no direct scaling evaluation is shown.**
Thank you for the comment. While we haven’t directly measured the scalability of TP with SPD, we can intuitively understand that SPD would lead to better scaling efficiency as GPU counts increas... | Summary: This paper introduces SPD aiming at reducing communication overhead in tensor parallelism for LLMs. SPD selectively removes synchronization points in attention outputs to mitigate latency bottlenecks during distributed inference. The authors propose a block-based SPD approach and classify transformer blocks in... | Rebuttal 1:
Rebuttal: Thanks for your thoughtful advices to our work. Here are the replies to your concerns.
**Q1. The evaluation is limited to LLaMA2 and OPT models. Testing on more recent models could validate its generalizability.**
Due to space limit, we couldn’t share our new results here. Please refer to the di... | null | null | null | null | null | null |
Understanding Bias Reinforcement in LLM Agents Debate | Accept (poster) | Summary: This paper explores the challenges of ensuring reasoning correctness in LLMs, particularly in self-correction methods and Multi-Agent Debate. The authors identify two major limitations of MAD: bias reinforcement and lack of perspective diversity. To address these issues, the paper introduces MetaNIM Arena, a b... | Rebuttal 1:
Rebuttal: Thank you for your valuable and insightful feedback. Your comments have greatly helped us enhance and develop our research. We have thoughtfully considered all of your concerns and will now address each of your comments individually.
**1. Additional Evaluation on Open-domain NLP Tasks**
We reco... | Summary: This paper presents an insightful analysis of bias reinforcement in Multi-Agent Debate (MAD) frameworks and introduces DReaMAD (Diverse Reasoning via Multi-Agent Debate with Refined Prompt) as an alternative approach to enhance strategic decision-making in LLMs. The study systematically explores how MAD can un... | Rebuttal 1:
Rebuttal: We thank for your constructive and insightful feedback. Your comments have significantly helped us enhance and advance our research. We have carefully considered each concern you raised. Below we address the main concerns regarding clarity and efficiency.
**1. Figure and Table Clarity**
We agre... | Summary: The paper investigates how biases present in a given LLM evolve when the LLM is used in a multi-round, multi-agent debate setting. This is compared to just prompting the LLM to answer directly (or to give a CoT answer). The authors find that if duplicates of the same LLM, and the same prompts are used, biases... | Rebuttal 1:
Rebuttal: We thank you for your constructive and insightful comments, which have significantly improved our research. We address your concerns below.
---
**1. Applicability of Diverse Amplication on Self-Reflection and Self-Consistency**
We understand this concern as asking whether other self-correction... | null | null | null | null | null | null | null | null |
Doubly Robust Fusion of Many Treatments for Policy Learning | Accept (poster) | Summary: The paper proposes a new methodology for policy learning in scenarios where many treatments are available. In such settings, naive methods may fail due to e.g., lack of overlap. The proposed methodology assumes that treatments can be assigned into "groups" that have the same treatment effect and learn this gro... | Rebuttal 1:
Rebuttal: We sincerely appreciate your insightful and constructive comments. Below, we have provided our detailed, point-by-point responses.
**1. Behavior under assumption violation**
Our method assumes a ground-truth group structure to establish theoretical guarantees such as consistency of group recover... | Summary: This paper addresses the problem of learning individualized treatment rules (ITR) in the high-dimensional setting when the cardinality of the action space is large. This setting can be challenging given the search space for ITR grows exponentially with the cardinality of the action space. The paper proposes a ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your insightful and constructive comments. Below, we have provided our detailed, point-by-point responses.
**1. Necessity of recovering group structure**
When there are many treatments and the sample size is limited, directly performing policy learning presents several ch... | Summary: The paper proposes to learn an optimal individualized treatment regime in settings with a large number of treatment actions. To handle large action space, the author(s) proposed a novel doubly robust method for grouping treatments into a few categories to enable the application of existing methodologies develo... | Rebuttal 1:
Rebuttal: We sincerely appreciate your insightful and constructive comments. Below, we have provided our detailed, point-by-point responses.
**1. Simulations for validation of the double robustness**
We additionally considered a scenario where the outcome mean functions are linear (correctly specified), w... | Summary: The authors study causal inference with many treatments and sparse data within each group, which makes estimating treatment effects challenging. They observe that many treatments share commonalities and can be clustered to reduce the effective dimensionality of the treatment space. Within each group, treatment... | Rebuttal 1:
Rebuttal: We sincerely appreciate your insightful and constructive comments. Below, we have provided our detailed, point-by-point responses.
**1. Experiments for increasing $K$ and fixed $n$**
We keep the sample size fixed at $n=1800$ and increase the number of treatments $K$ from 16 to 32 and 48. In such... | null | null | null | null | null | null |
Zebra: In-Context Generative Pretraining for Solving Parametric PDEs | Accept (poster) | Summary: This paper proposes a model for modeling physical dynamics by leveraging the power of large language models (LLMs). The model handles data given in the physical domain through quantized representations, utilizing in-context learning. However, the advantages of this approach over other LLM-based foundation mode... | Rebuttal 1:
Rebuttal: ### 1. On the originality
There seems to be a misunderstanding, and we apologize if our explanations were not sufficiently clear. We appreciate the opportunity to clarify the originality of our contribution.
Objective: Our goal is to develop a family of models capable of solving parametric PDEs. W... | Summary: This paper introduces Zebra, a novel generative autoregressive transformer designed to solve parametric partial differential equations (PDEs) without requiring gradient adaptation at inference. The key innovation is leveraging in-context learning to dynamically adapt to new PDE parameters. Zebra employs a two-... | Rebuttal 1:
Rebuttal: ### 1. DPOT
Thank you for mentioning this interesting work. We will include this reference, along with those suggested by other reviewers, in the final version and discuss how it differs from our approach and problem setting.
### 2. Explanations and Limitations
Thanks for the comment. Clear e... | Summary: This paper proposes an in-context generative auto-regressive transformer for solving parametric PDEs and outperforms gradient adaptation methods.
Claims And Evidence: Given that the authors claim this is the first successful application of discretized representations in physical systems, they should further a... | Rebuttal 1:
Rebuttal: ### 1. Continuous vs Discrete
Thanks for the comment. We agree that encoder quality is crucial for model performance. However, the answer is not trivial and we will try to answer in different points.
* Our goal was to build an autoregressive probabilistic predictor for parametric PDE dynamics th... | Summary: The paper addresses the problem of predicting PDE solutions without explicit knowledge of the underlying dynamics using a generative auto-regressive transformer. The model is trained using contexts composed of different trajectories. During inference, the model is given a context containing trajectories govern... | Rebuttal 1:
Rebuttal: ### 1. On the uncertainty
Thank you for your comments. We agree that the main contribution lies in adaptation through in-context learning, with uncertainty emerging as a byproduct. We aim to clarify this distinction and welcome suggestions for improving it.
The generative capacity stems from the... | null | null | null | null | null | null |
PoisonBench: Assessing Language Model Vulnerability to Poisoned Preference Data | Accept (poster) | Summary: The authors introduce PoisonBench, a benchmark that measures the robustness of LLMs to data poisoning attacks. The authors consider two primary attack types called "content injection" and "alignment deterioration. Both attacks are operationalized via open datasets that enable synthetically generating poisoned ... | Rebuttal 1:
Rebuttal: We are grateful for the time and effort in reviewing our work. We are excited to receive your feedback and your recognition of our effort in addressing a problem of significant importance for modern LLM training. Your advice would definitely help improve our work. We would like to address your con... | Summary: The paper proposes a benchmark for evaluating LLMs' robustness to poisoned data during alignment. Two poisoning approaches are proposed: content injection adds a trigger to the prompt and injects a specific name entity to the preferred response, while alignment deterioration swaps the two responses. The benchm... | Rebuttal 1:
Rebuttal: Thanks for your time and effort in reviewing our work! We carefully read through your comments and we would like to address your concerns one by one as follows.
**Q1: The benchmark lacks diversity as it involves only two datasets and four entities.**
**A1**: In our experiment, we mainly use HH-R... | Summary: The paper introduces PoisonBench, a benchmark for evaluating the vulnerability of Language Models (LMs) to poisoning attacks when using preference learning for model’s alignment. For this, the paper proposes two different attacks: (1) content injection, where attackers aim to elicit specific entities into mode... | Rebuttal 1:
Rebuttal: Thanks for your time and effort in reviewing our work! We are grateful for your recognition that our study yields interesting findings and provides helpful insight! We carefully read through your helpful and constructive comments and we would like to address your concerns one by one as follows.
... | Summary: The authors consider the problem of content injection poisoning attacks, where malicious actors poison the preference training data in order to attempt to make the models mention certain named entities more often. They construct a dataset and benchmark for evaluating the success of such poisoning attacks for v... | Rebuttal 1:
Rebuttal: We are grateful for your efforts and expertise in reviewing our work. We appreciate your suggestions, which will definitely help us to improve our work. We carefully read through your helpful and constructive comments and we would like to address your concerns one by one as follows.
**Q1: The c... | null | null | null | null | null | null |
Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods | Accept (poster) | Summary: The paper theoretically investigates the phenomenon observed in prior work that ensembling pre-trained and fine-tuned weights of foundation models outperforms dedicated fine-tuning strategies.
The authors first verify this phenomenon via empirical results for instruction fine-tuning LLMs.
Then the authors prov... | Rebuttal 1:
Rebuttal: We would like to extend our sincere appreciation to the reviewer for all the constructive and positive feedbacks on our contribution. Additional experiments and explanations have been provided to address the mentioned concerns.
**Guidance for enhancing performance of fine-tuning strategies.** Som... | Summary: The paper theoretically analyzes why WiSE-FT or other regularization methods work. The authors also show that WiSE-FT is better than vanilla fine-tuning with regularization.
Claims And Evidence: The claim is generally supported by theoretical analysis, but it is difficult to find reasons for showing LLM exper... | Rebuttal 1:
Rebuttal: Many thanks for all the constructive suggestions.Additional experiments and explanations have been provided as follows.
**The connection between theoretical and empirical results.** Our experiment on LLM and theory match in showing the benefits of ensemble in fine-tuning.As fine-tuned parameters ... | Summary: This paper focuses on theoretically and empirically studying task specific fine-tuning of large (over-parameterized) models. It aims to provide theoretical underpinnings for empirical observations such as task specific fine-tuning leads to a performance degradation/forgetting on pre-training tasks, an early s... | Rebuttal 1:
Rebuttal: We would like to extend our sincere appreciation to the reviewer for all the constructive and insightful suggestions. Additional experiments and explanations have been provided to address the mentioned concerns.
**Performance on pre-trained model, figure 2 and 3.** We have improved the figures ac... | Summary: This paper studies overadaptation in supervised fine-tuning (SFT). It builds on prior work that observes that ensembling a pretrained model with its fine-tuned variant improves performance. Within an over-parameterized linear regression setup, the paper theoretically shows that this effect is due to the ensemb... | Rebuttal 1:
Rebuttal: Thanks for the constructive suggestions! Additional experiments and explanations have been provided as follows.
**The connection between theoretical and empirical results.** Our experiment on LLM and theory match in showing the benefits of ensemble in fine-tuning. As fine-tuned parameters are clo... | null | null | null | null | null | null |
Positional Encoding meets Persistent Homology on Graphs | Accept (poster) | Summary: The manuscript introduces a novel positional encoding (PE) methodology for GNNs by combining eigenvalue and eigenvector information from graph Laplacians with persistent homology (PH). The authors present theoretical insights demonstrating that neither PE nor PH independently surpasses the other in expressiven... | Rebuttal 1:
Rebuttal: Many thanks for your thoughtful feedback! We address all your questions and comments below.
> The claim in Section 1 about the drawbacks of partitioning eigenvalue/eigenvector spaces requires clarification.
Thank you for the opportunity to clarify this point. We agree that, in theory, one can ut... | Summary: This paper introduces Persistence informed Positional Encoding (PiPE), a novel learnable method that synergizes positional encoding with persistent homology to enhance graph neural networks (GNNs). The authors argue that traditional message-passing GNNs suffer from limitations in expressiveness, being constrai... | Rebuttal 1:
Rebuttal: Thank you so much for your thoughtful comments and excellent suggestions! We've acted on all of them, and also address all your concerns, as we describe below.
> The compared baselines are all positional encoding approaches....
Thank you for an excellent suggestion. Based on your feedback, we ha... | Summary: The paper proposes a novel graph neural network architecture combining existing positional encoding (PE) strategies with persistent homology (PH) and shows the resulting method is strictly more expressive than previous works using either PE or PH.
Claims And Evidence: The theoretical claims seems to be suppor... | Rebuttal 1:
Rebuttal: Many thanks for your thoughtful review. We address your concerns and incorporate your suggestions below.
> could you numerate all the equations? That is very helpful for the reader and the reviewers in order to refer to them
Thank you for pointing this out. We will numerate all the equations in... | Summary: This paper introduces a novel method—Persistence informed by Positional Encoding (PiPE)—which unifies positional encoding (PE) with persistent homology to enhance the expressivity of graph neural networks . The authors provide both theoretical insights and empirical evidence, demonstrating that PiPE overcomes ... | Rebuttal 1:
Rebuttal: Many thanks for your constructive feedback, and for appreciating the different aspects of this work. We address your comments below.
> The paper could benefit from further clarification or additional examples to help readers less familiar with advanced topological concepts.
Thank you for an exc... | null | null | null | null | null | null |
Generalized Random Forests Using Fixed-Point Trees | Accept (spotlight poster) | Summary: This paper presents an approach to improving the computational efficiency of GRF by replacing the conventional gradient-based tree-splitting criterion with a fixed-point approximation. The authors argue that their method maintains the theoretical guarantees of GRF while significantly improving scalability in ... | Rebuttal 1:
Rebuttal: > ... explain why the derivative matrix can be replaced by an arbitrary scalar
**A1**: In order to show that choosing a different scalar $\eta$ does not
change the stability or outcome of the splitting process, let us compare
two types of fixed-point pseudo-outcomes. Consider first $\rho\_{i}=-\e... | Summary: This paper proposes a method to enhance Generalized Random Forests (GRF) by introducing Fixed-Point Trees (FPT) as a gradient-free alternative for recursive partitioning. The authors highlight the limitations of gradient-based methods in GRF, particularly in handling highly correlated regressors and ill-condit... | Rebuttal 1:
Rebuttal: > The proofs from theoretical analysis are majorly motivated by Athey et al. (2019). What are the differences and the challenges?
**A1**: The main difference is our new splitting criterion based on a fixed-point approximation instead of a gradient-based approximation. Proposition 4.1 in our pap... | Summary: The paper introduces a computational improvement to generalized random forests (GRFs). In the tree-growing algorithm of GRFs, the paper proposes to replace the gradient-based approximation for evaluating the split criteria by a more computationally efficient fixed-point based approximation. Theoretical guarant... | Rebuttal 1:
Rebuttal: > Why is the saving not $O(K^3)$
**A1**: In the general case, removing the inverse $A_P^{-1}$ should indeed provide a computational saving of order $O(K^3)$. However, for the specific applications discussed in our paper—such as varying coefficient models (VCM) and heterogeneous treatment effect ... | Summary: This paper introduces a computationally efficient alternative to generalized random forests (GRFs) for estimating heterogeneous effects in high-dimensional settings. The authors propose a gradient-free "fixed-point" approximation approach that eliminates the need for Jacobian estimation in the GRF splitting cr... | Rebuttal 1:
Rebuttal: > ...Additional real-world applications in different domains could strengthen the practical relevance of the method.
**A1**: We agree with the reviewer. Space limits our discussion here. In the camera-ready version, we will add real-world applications from different domains.
> ...Some additiona... | null | null | null | null | null | null |
Iterative Vectors: In-Context Gradient Steering without Backpropagation | Accept (poster) | Summary: The paper proposes Iterative Vectors, inspired by the framing of in context learning as simulated gradient updates. The method extracts activations derived from in-context learning examples and adds the activations to unseen queries. The method also refines these updates by averaging the activations from many ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and the opportunity to reflect. We believe the concerns raised warrant clarification to ensure a balanced perspective.
## Claim 1
1. We do not intend to assert that "gradient simulation" is, in itself, a novel concept. We respectfully request the reviewer... | Summary: This paper introduces Iterative Vectors (IV), a method for steering in-context learning (ICL) by modifying model activations without requiring parameter updates.
While activation steering has been explored in various contexts, its application to ICL appears novel.
IV consists of two key steps: first, "activati... | Rebuttal 1:
Rebuttal: We are delighted that the reviewer described our method as "conceptually simple" and "useful for practical use." We are committed to improving clarity in line with their suggestions.
We will condense Section 3.1 and utilize the space to introduce the verbalizer, while relocating some of the discu... | Summary: The paper introduces Iterative Vectors, a technique that enhances in-context learning by simulating gradient updates during inference without backpropagation. IVs address the key ICL challenges: prompt sensitivity, context length limitations, and increased inference time by leveraging activation space rather t... | Rebuttal 1:
Rebuttal: 1. Thank you for your thoughtful review and engagement with our work. We acknowledge that certain technical aspects may require further clarification to ensure our contributions are clearly communicated. We will strive to refine our explanation by addressing the points raised.
2. The term “activat... | Summary: The paper addresses the challenges of selecting suitable demonstration examples in in-context learning (ICL) for language models. It proposes a novel technique called Iterative Vectors (IVs) to enhance ICL performance. IVs operate by extracting and iteratively refining gradients within a language model, which ... | null | null | null | null | null | null | |
PAC-Bayes Bounds for Multivariate Linear Regression and Linear Autoencoders | Reject | Summary: Based on the PAC-Bayes risk bound of Alquier et al. (2016), the authors propose a risk bound for multivariate linear regression (thm 3.2). They extend this result (lem 4.2 and thm 5.2) to linear auto encoders which performs L2-regularized multivariate linear regression under the constraint that the weight matr... | Rebuttal 1:
Rebuttal: **Theoretical Claims**: All risk bound theorems provided in this paper are incorrectly stated... Alquier's THM is much stronger that this statement. It holds for all $\rho$ and does not require $\rho$ be any specific distribution.
**Answer**: We appreciate you pointing out this issue. The origina... | Summary: In this paper, new PAC-Bayesian generalization bounds are derived for multivariate linear regression with Gaussian data or data with bounded support. Specifically, these results are applied to linear autoencoders, which are of practical relevance due to their widespread use in recommender systems. The bound is... | Rebuttal 1:
Rebuttal: **Question 1**: Is it possible to instead derive an empirical PAC-Bayes bound, as alluded to? Would this be desirable?
**Answer**: The loss in our linear regression setting is assumed to be unbounded. According to [1], existing empirical PAC-Bayes bounds, such as Catoni's bound and Seeger's boun... | Summary: This paper provides a PAC-Bayes bound specific to multivariate linear regression. Under Gaussian dataset assumption, the authors derive a sufficient condition for the convergence of said bound. Next, the paper rewrite the PAC-Bayes bound for multivariate linear regression with bounded data distribution and app... | Rebuttal 1:
Rebuttal: **Question 1**: Why is there no convergence analysis for the PAC-Bayes bound in the bounded data case?
**Answer**: Because the bounded data assumption (Assumption 4.1) alone is not sufficient to guarantee the convergence. The reason is as follows:
First, the convergence of Alquier's bound does n... | Summary: The paper introduces a novel PAC-Bayes generalization bound for multivariate linear regression based on the standard Alquier's (oracle) bound. This result generalizes previous PAC-Bayes bounds for multiple linear regression. After studying conditions that ensure convergence of the bound, they adapt it for Line... | Rebuttal 1:
Rebuttal: **Question 1**: The KL divergence $D(\rho||\pi)$ is not defined in the paper.
**Answer**: Thanks for pointing out this issue. We will add this into our paper: Suppose $\pi$ and $\rho$ are continuous probability distributions on $\mathbb{R}^d$. The KL-Divergence is defined as
\begin{equation*}
... | null | null | null | null | null | null |
Distillation Scaling Laws | Accept (poster) | Summary: The authors propose a distillation scaling law that estimates the performance of distilled models based on a computed budget and its allocation between the student and teacher models. The study provides insights into when distillation outperforms supervised pretraining and offers compute-optimal distillation r... | Rebuttal 1:
Rebuttal: Thank you wCXj for your thoughtful feedback. We’re encouraged you found our study comprehensive and useful, particularly the practical guidance provided by our distillation scaling law.
## There are missing distillation and scaling law references from 2022-2024
We extensively cover distillation ... | Summary: In this paper, authors establish a distillation scaling law to predict the performance of a student model based on compute budget and its allocation between teacher and student models. Using this law, authors can infer optimal distillation strategies for different scenarios, including cases where a teacher alr... | Rebuttal 1:
Rebuttal: Dear hWGR. Thank you for your taking the time to review our paper and for your detailed feedback. We are encouraged that you found our extensive study useful for showcasing different aspects of distillation, and that you find the overall guidance provided by the Distillation Scaling Law to have ut... | Summary: The paper discusses a very interesting topic related to scaling laws for distillation. Following the strategies of recent works on scaling laws given a fixed compute budget, this paper provides a framework to estimate the performance of distilled student models based on different cost settings related to the t... | Rebuttal 1:
Rebuttal: Thank you, 1KoJ for your valuable feedback. We’re happy you found our analysis useful, the compute-optimal recipes useful for practitioners, and appreciated our work in the context of the scaling literature.
## Architectural diversity, a significant variable, was not investigated, limiting genera... | null | null | null | null | null | null | null | null |
Actor-Critics Can Achieve Optimal Sample Efficiency | Accept (poster) | Summary: Actor-critic algorithms have become a cornerstone in reinforcement learning (RL), leveraging the strengths of both policy-based and value-based methods. In this paper, the authors introduce a novel actor-critic algorithm that attains near-optimal sample-complexity and regret under MDPs with low Bellman eluder ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and suggestions! We are glad that the reviewer thinks our study is well motivated and our paper is well written. We provide answers to the questions below. Please let us know if you have further questions or concerns!
**Revisions**
Moreover, we have revised... | Summary: This paper makes the following three key contributions in the analysis of actor-critic algorithms with general function approximation:
1. Proposes an actor-critic algorithm, DOUHUA, which operates under the closure under truncated sum assumption. The regret bound is established as
$$
O\left(\sqrt{H^4T... | Rebuttal 1:
Rebuttal: We thank the reviewer for helpful comments and suggestions on our paper! We also thank the reviewer for believing that our paper is well-structured and clearly presented, and note that we have provided detailed discussions in the appendix in addition to the main theorems. We have addressed your qu... | Summary: This paper mainly concerns the theoretical analysis of the actor-critic type of algorithms in finite-horizon episodic MDPs. Beyond linear MDP scenarios, the work focuses on the regret and complexity analysis of general function approximation case. To begin with, an algorithm named Double Optimistic Updates for... | Rebuttal 1:
Rebuttal: We thank the reviewer for the helpful and insightful comments! We are glad you recognize our work as achieving \sqrt{T} regret for general function approximation under specific assumptions—extending beyond the linear setting considered in prior work. We’ve addressed your questions and incorporate... | Summary: This paper introduces the first provably efficient actor-critic algorithms utilizing general function approximation. For cases where the policy class does not grow rapidly due to critic updates (as mentioned by "ease case"), the authors propose a novel algorithm that combines the GOLF algorithm with mirror asc... | Rebuttal 1:
Rebuttal: We thank the reviewer for providing helpful comments and suggestions to our paper! We are glad that the reviewer believes our theoretical justification to be well-supported. We have revised our paper and included numerical experiments based on the suggestions from the reviewer in this link (https... | null | null | null | null | null | null |
End-to-End Learning Framework for Solving Non-Markovian Optimal Control | Accept (poster) | Summary: This paper devises a learning-based approach to solve linear quadratic regulator (LQR) problems for fractional-order linear time-invariant systems. Here, the fractional order means the system state transition is not entirely dependent on the current state but also depends on previous states (Eq(7) in the paper... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and constructive suggestions. Below, we provide a detailed response regarding your concerns and questions.
**Baselines**: As detailed in our response to **Reviewer cBiX**, we have conducted extensive comparisons between FOLOC and several widely used methods, i... | Summary: In summary, the paper contributes a novel end-to-end, data-driven framework for optimal control of non-Markovian (fractional-order) systems by uniting advanced control theory with deep learning techniques, backed by both theoretical guarantees and empirical performance.
Claims And Evidence: The paper provides... | Rebuttal 1:
Rebuttal: Thank you very much for the recognition of our **novel** theoretical contributions, the FOLOC framework, and the experimental design, as well as highlighting its effectiveness and robustness. We also thank the reviewer for pointing out areas where additional **empirical comparisons** and **runtime... | Summary: This paper develops a new end-to-end learning approach for fractional order linear time-varying systems to perform fractional order LQR. A method is proposed to identify the system dynamics from multiple recorded trajectories. Using the identified model, a fractional order LQR controller is designed to obtain ... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the thoughtful and constructive feedback. We are especially grateful for the recognition of our theoretical contributions and the clarity of our proofs.
**Complex benchmark systems**: Thank you for your thoughtful comment regarding the applicability of our me... | null | null | null | null | null | null | null | null |
Provable Zero-Shot Generalization in Offline Reinforcement Learning | Accept (poster) | Summary: This paper investigates the generalization performance of policies learned in offline reinforcement learning (ORL) on test environments, specifically focusing on zero-shot performance in the average sense across test environments. The paper demonstrates that standard ORL methods, without access to contextual i... | Rebuttal 1:
Rebuttal: Thanks very much for reviewing our work and for the valuable advice. Our responses are as follows.
**Q1:** Differences from (Jin et al. 2021).
**A1:** We respectfully argue that our work is **substantially different** from (Jin et al., 2021), both in **algorithmic design** and **theoretical an... | Summary: The work theoretically analyzes contextual reinforcement learning in the offline setting for zero-shot generalizability. In particular, the work proposes two algorithmic frameworks with provable zero-shot generalization abilities.
## update after rebuttal
I have read all reviews and am keeping my score. I do... | Rebuttal 1:
Rebuttal: Thanks very much for your positive feedback and your valuable suggestions. Our responses are as follows.
**Q1:** One point I remain uncertain about is the assumptions about the context at hand. The work seems to implicitly assume perfect context information. It would be best to highlight this as... | Summary: Main findings&Contribution:
1. Problem finding: Classic offline reinforcement learning methods often struggle to generalize unseen environments, primarily due to dataset coverage limitations and the absence of contextual information during training. Merging multi-environment datasets without preserving contex... | Rebuttal 1:
Rebuttal: Thanks very much for reviewing our work and for the valuable suggestions. Our responses are as follows.
**Q1:** Augment theory with experiments; necessity of pessimism
**A1:** Thank you for the thoughtful suggestion. We would like to respectfully emphasize that this is a theoretical paper submi... | null | null | null | null | null | null | null | null |
Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data | Accept (poster) | Summary: The paper proposes Discriminative Fine-Tuning (DFT) as an alternative to the standard two-stage fine-tuning pipeline (Supervised Fine-Tuning followed by Preference Optimization, denoted SFT→PO). DFT aims to fine-tune large language models without using any human preference data or reward models, yet still achi... | Rebuttal 1:
Rebuttal: **Q1:** The General language tasks evaluation metric should be written explicitly.
**A:** Thank you for this detailed comment. The information on the evaluation metric is shown in the table below.
| Benchmark | Shot(s) | Metric |Apply chat template |
| ---------- | ------- | ------ |---------... | Summary: This paper introduces Discriminative FineTuning (DFT), a novel approach to fine-tuning large language models (LLMs) that diverges from the traditional Supervised FineTuning (SFT) followed by Preference Optimization (PO) paradigm. Unlike methods that rely on preference data or reward models, DFT frames fine-tun... | Rebuttal 1:
Rebuttal: **Q1:** Weakness about the reliance on self-generation using the base model.
**A:** We agree with what the reviewer said. The quality and diversity of these generated negative samples will matter to some degree. If the base model is super strong, it would not make sense to assume the generated da... | Summary: This paper introduces Discriminative Fine-Tuning (DFT), a new approach for fine-tuning large language models (LLMs) without using preference data or reward models. Unlike Supervised Fine-Tuning (SFT), which employs a generative approach focusing only on positive examples, DFT adopts a discriminative paradigm t... | Rebuttal 1:
Rebuttal: We thank the reviewer's comments. While we understand the reviewer's main concern of lack of theoretical analysis of DFT and DFT2, we would like to emphasize that this is a new yet challenging research area, especially considering the discriminative framework is defined over the infinite data spac... | Summary: The authors present a new loss function for LLMs called DFT (discriminative finetuning) which combines an SFT loss objective for positive examples with a score-decreasing term for sampled negative answers. They theoretically justify the DFT objective and also present a modified DFT2 loss objective which simpli... | Rebuttal 1:
Rebuttal: We thank the reviewer for detailed comments. We hope our rebuttal can address your concerns.
**Q1**: About the claim "Eliminates the need for preference data".
**A:** (1) We intended to mean the human labeled preference data. (2) Actually, our discriminative framework does not necessarily assum... | null | null | null | null | null | null |
Mutual Learning for SAM Adaptation: A Dual Collaborative Network Framework for Source-Free Domain Transfer | Accept (poster) | Summary: The paper proposes a Collaborative Mutual Learning Framework for the source-free domain adaptation of the Segment Anything Model. The key innovation lies in dynamically assigning "teacher" and "student" roles to two collaborative networks based on their reliability during training. The framework introduces thr... | Rebuttal 1:
Rebuttal: Thank you for your critical as well as constructive assessments of our work, and we address the concerns point by point as follows.
---
### **W1: The triplet loss contributes only marginal improvements, raising questions about its necessity.**
The triplet loss, although contributing marginally ... | Summary: This paper addresses the challenge of adapting the Segment Anything Model to new domains with significant distribution shifts through a Collaborative Mutual Learning Framework. The method involves two collaborative networks that dynamically alternate between teacher and student roles based on reliability. They... | Rebuttal 1:
Rebuttal: Thank you for your critical as well as constructive assessments of our work, and we address the concerns point by point as follows.
---
### **W1: Lack of analysis on role switching frequency and its Impact on convergence and performance**
We conducted an additional quantitative analysis of the ... | Summary: This paper introduces a mutual teaching framework for SAM adaptation in target domains. Compared to self-training framework, the mutual learning enables dynamic assignment of teacher and student roles, leading to more robust and generalized performance. To facilitate adaptation, three loss functions are propos... | Rebuttal 1:
Rebuttal: Thank you for your critical as well as constructive assessments of our work, and we address the concerns point by point as follows.
---
### **Claims And Evidence: Clarification on "Catastrophic Preservation" in line 256**
We appreciate your observation regarding the term "catastrophic preservat... | Summary: The paper introduces a Collaborative Mutual Learning Framework for source-free domain adaptation of the Segment Anything Model. It employs dual networks that alternate roles as teacher and student based on reliability, preserving SAM’s pre-trained knowledge. T Experiments in medical, camouflaged, and robotic d... | Rebuttal 1:
Rebuttal: Thank you for your critical as well as constructive assessments of our work, and we address the concerns point by point as follows.
---
### **Claims And Evidence: About the motivation**
We appreciate your detailed understanding of teacher-student networks. Our motivation is not directed at the ... | null | null | null | null | null | null |
Test-Time Multimodal Backdoor Detection by Contrastive Prompting | Accept (poster) | Summary: In this paper, the authors propose an inference-time, multi-modal backdoor detection method called BDetCLIP. The method is motivated by the observation that the visual representations of backdoored images exhibit limited sensitivity to significant changes in class description texts. Specifically, BDetCLIP firs... | Rebuttal 1:
Rebuttal: Thanks for your valuable comments. We provide point-by-point responses to your concerns as follows. We are also willing to discuss with you in the discussion period if there is anything unclear.
**Q1: Unpredictable attacks and the need for large auxiliary datasets.**
**A:** We would like to clar... | Summary: BDetCLIP is a test-time multimodal backdoor detection strategy that identifies backdoor examples using a set of contrastive prompts. Specifically, the visual representation of backdoor images is less correlated to both the *benign* and *random* text description compared to clean examples, leading to a distribu... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments. We provide point-by-point responses to your concerns as follows. We are also willing to discuss with you in the discussion period if there is anything unclear.
**Q1: For safety-related papers, it is encouraged to add some discussions on adaptive attacks. For i... | Summary: The paper proposes BDetCLIP, a novel test-time backdoor detection method for CLIP models. The key insight is that backdoored images exhibit insensitivity to semantic perturbations in class description texts. The method leverages contrastive prompting:
1. Text Generation: GPT-4 generates class-related (benign) ... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments, and we respond to your concerns as follows. We are willing to discuss with you in the discussion period.
**Q1: The label consistent attack violates the assumption of the motivation.**
**A:** Thanks for raising the concern. We would like to explain that label... | Summary: This paper introduces BDetCLIP, a novel method for test-time backdoor detection in CLIP-based models. Instead of trying to remove or mitigate backdoors at the pre-training or fine-tuning stages, the authors propose an inference-stage defense. By leveraging the contrastive alignment between images and class-rel... | Rebuttal 1:
Rebuttal: Thank you very much for your positive feedback on our paper. Below, we provide point-by-point responses to address your concerns. We will be more than happy to discuss with you in the reviewer-author discussion period if there is anything unclear.
**Q1: Randomness on LLM Generation: The success o... | null | null | null | null | null | null |
Reinforced Lifelong Editing for Language Models | Accept (poster) | Summary: This paper introduces RLEdit, a method addressing long-term continuous LLM knowledge editing, known as lifelong editing. The authors apply reinforcement learning modeling to meta-learning knowledge editing approaches, proposing a hypernetwork training method specifically designed for lifelong editing tasks. Th... | Rebuttal 1:
Rebuttal: Dear Reviewer kk6q:
**Thank you for your positive feedback on our work!**
We are happy to discuss the details of our work and hope to address all your questions.
******
## For *Other Strengths And Weaknesses*:
> Equation (9) requires more clarification. A detailed explanation would help read... | Summary: The paper proposes a model editing method, named RLEdit, for lifelong editing that extends meta-learning methods to sequential editing tasks. RLEdit's hypernetwork is trained using reinforcement learning algorithms, effectively maintaining Effectiveness, Generalization, and Locality at the knowledge sequence l... | Rebuttal 1:
Rebuttal: Dear reviewer TQMf:
**Thank you very much for your recognition and support of our work!**
We are very willing to engage in in-depth discussions with you regarding the research details.
******
## For *Experimental Designs Or Analyses*:
Thank you for your valuable suggestions!
Following your s... | Summary: The paper proposes RLEdit, a novel Reinforcement Learning based hypernetwork approach to edit Large Language Models for knowledge update after model completes training. It primarily aims to solve the challenge that hypernetwork based editing methods are efficient yet struggle more on large amount of edits. By ... | Rebuttal 1:
Rebuttal: Dear Reviewer 2E3w:
**Thank you for your positive feedback on our work!**
In brief, we propose a hypernetwork training method for lifelong editing tasks that can adaptively adjust LLM parameters based on knowledge update sequences, achieving efficient and low-interference lifelong editing.
We ... | null | null | null | null | null | null | null | null |
Learning Configurations for Data-Driven Multi-Objective Optimization | Accept (poster) | Summary: This paper provides a generalization framework for data-driven multi-objective optimization through sample complexity analysis. The smoothed analysis offers logarithmic bounds, making the framework feasible for real-world problems. If the performance metric is Pareto volume, this paper further shows that an ap... | Rebuttal 1:
Rebuttal: Thanks for your review! We are delighted that you considered our work clearly organized and logical. We answer your question as follows:
> Q: The proof assumes the piecewise-constant structure holds for all instances, but this paper does not discuss how to verify it. This property is crucial to b... | Summary: This paper establishes the theoretical foundation of data-driven multi-objective optimization. Authors discuss the piecewise structure of combinatorial optimization algorithms and prove a theorem which bounds the sample complexity of parameterized multi-objective combinatorial optimization algorithms. Authors ... | Rebuttal 1:
Rebuttal: Thanks for your review! We answer your question as follows:
> Q: Since the title is multi-objective optimization (instead of combinatorial multi-objective optimization) while the discussions in the paper seem mostly focus on combinatorial optimization, can the proposed method be applied to contin... | Summary: This paper develops a theoretical framework for learning algorithm configurations in the context of multi-objective optimization. The framework is then applied to several settings, including approximation algorithms, local search, general linear programming, and Markov decision processes.
Claims And Evidence:... | Rebuttal 1:
Rebuttal: Thanks for your review! We are delighted that you consider our work sound and rigorous. We address your concerns as follows:
> Q: Does the proposed theoretical framework for data-driven multiobjective optimization offer any practical insights?
This is a very good question. In this work, we use s... | Summary: This paper proves generalization guarantees for the problem of selecting a weighting parameter in multiobjective optimization. Then, it proves the PAC-learnability of parameters which maximize the Pareto volume in polynomial time. It adapts these results to several classes of optimization problems.
## Update ... | Rebuttal 1:
Rebuttal: Thanks for your review! We are delighted that you consider our work thorough and comprehensive. We address your concerns as follows:
> Q: How novel are the theoretical results of Section 3.1? Are they very general beyond the setting of multiobjective discrete optimization?
We admit that the resu... | null | null | null | null | null | null |
Towards a Unified Framework of Clustering-based Anomaly Detection | Accept (poster) | Summary: This paper introduces UniCAD, a novel method for anomaly detection based on clustering and representation learning. The method leverages an anomaly indicator function and a mixture of student’s t-distributions to learn an anomaly score based on the mixture distribution density. Inspired by Newton’s law of univ... | Rebuttal 1:
Rebuttal: > It is unclear why the student’s t-distribution should work fundamentally better than a Gaussian mixture model.
The choice of the Student’s t-distribution over a Gaussian mixture model (GMM) is motivated by its heavy-tailed nature, which enhances robustness to outliers—a critical factor in anom... | Summary: The paper presents an unsupervised anomaly detection framework with two key components: (i) a framework for joint learning of representation, clustering, and mixture models; (ii) a gravity-inspired anomaly scoring function based on mixture model outputs. Experiments on 30 datasets show it outperforms 17 baseli... | Rebuttal 1:
Rebuttal: > However, the paper does not elaborate on the specific scenarios in real datasets that this method can address. Although this scoring method theoretically holds advantages, further clarification is needed regarding its applicable scope.
We thank the reviewer for highlighting the need for more de... | Summary: The paper presents UniCAD, a novel unsupervised anomaly detection framework that integrates representation learning, clustering, and anomaly detection into a unified theoretical framework. By maximizing an anomaly-aware data likelihood based on a mixture model with the Student-t distribution, UniCAD effectivel... | Rebuttal 1:
Rebuttal: > The paper lacks comparisons with the latest baseline methods. Among the 17 baseline methods included, only 3 are from after 2020, and there are no methods from 2024.
Thank you for noting the need for updated baselines. Given the rapid advancements in unsupervised anomaly detection, we have adde... | Summary: The paper proposes UniCAD, a novel model for Unsupervised Anomaly Detection (UAD) that unifies representation learning, clustering, and anomaly detection within a single theoretical framework. By leveraging a mixture model with the Student-t distribution, UniCAD introduces an anomaly-aware data likelihood that... | Rebuttal 1:
Rebuttal: > Experimental settings are not fair
> Default parameters for baseline methods were used, which raises concerns. All methods have to re-run and be re-optimized for the datasets, similar to what you do for your solution
Thank you for highlighting the concern regarding the fairness of our experime... | null | null | null | null | null | null |
Rethinking Score Distilling Sampling for 3D Editing and Generation | Accept (poster) | Summary: This paper introduces a novel variant of SDS, called UDS, which is designed to handle both editing and generation tasks. Via qualitative and quantitative experiments, the authors show that UDS surpasses previous baselines.
## update after rebuttal
As I noted in my official review, I remain unconvinced regard... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments and valuable suggestions, which have greatly improved the quality of our paper. Below, we address your concerns and clarify potential misunderstandings:
> **Experimental**
**1.** As shown in the [Figure 1](https://anonymous.4open.science/r/15488/1.pdf), we p... | Summary: This paper tackles the problem of 3D asset generation and editing using SDS. The paper identify limitations in existing SDS variants especially for edting. The core contribution is a new method called Unified Distillation Sampling UDS. The key ideas behind UDS are: gradient term unification. A few other improv... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments and valuable suggestions, which have greatly improved the quality of our paper. Below, we address your concerns and clarify potential misunderstandings:
> **Weaknesses**
**1.** We present a comparison of the optimization time of our method with that of PDS a... | Summary: This paper focuses on score-sampling based text-to-3D generation and editing. The authors first analyze the variants of SDS (e.g., DDS and PDS) and identify significant commonalities in their gradient optimization processes. The authors further introduce a unified distillation sampling (UDS) that enables unifi... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments and valuable suggestions, which have greatly improved the quality of our paper. Below, we address your concerns and clarify potential misunderstandings:
> **Methods And Evaluation Criteria**
Most previous works have focused on a single 3D generation or editi... | Summary: This paper points out that previous SDS and its variants have only performed well in generation or editing tasks. To address this, a unified approach integrating 3D asset generation and editing is proposed. The authors observe that the generation and editing processes in SDS and its variants share a common fun... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments and valuable suggestions, which have greatly improved the quality of our paper. Below, we address your concerns and clarify potential misunderstandings:
> **References**
We will include the **Gaussianeditor** and **Drag Your Gaussian** discussions in the rev... | null | null | null | null | null | null |
Understanding High-Dimensional Bayesian Optimization | Accept (poster) | Summary: This paper focuses on High-Dimensional Bayesian Optimization (HDBO). More precisely, it seeks to explain the good performance of vanilla BO for some high-dimensional tasks that has been observed by recent works. To do so, the paper empirically studies the vanishing gradients that appear when fitting a high-dim... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's positive feedback regarding the insights and empirical observations presented in our paper. We are glad to hear that our work is seen as valuable to the HDBO community, well-written, and enjoyable.
**General claims in the title and abstract**
We will tone down the "... | Summary: This paper considers High-dimensional Bayesian Optimization and thoroughly studies the use of vanilla Gaussian Processes (GP) for that matter. A recent line of work illustrated how simple tweaks can lead to considerable improvements in optimization outcomes, with GPs outperforming complex state-of-the-art base... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their positive feedback, particularly for recognizing the thorough analysis of Gaussian Processes in high-dimensional Bayesian Optimization (HDBO), as well as the clarity and practical insights provided in the paper. We appreciate the reviewer's thoughtful c... | Summary: The paper investigates the perfomance of BO for high-dimensional problems, building upon a recent focus in the literature on working out how to make standard BO frameworks work better in high-dimensions, rather than proposing entirely new frameworks. Lots of empirical evidence is provided to justify the simple... | Rebuttal 1:
Rebuttal: We appreciate the positive feedback regarding the clarity of our writing and the potential impact of our simple yet effective solution. We implemented enhancements to clarify our methodology and bolster the impact of our findings.
**Implementation of RAASP**
We will improve the description of th... | Summary: The paper provides an investigation on the problems of BO in high dimensions (the curse of dimensionality). First is the problem of vanishing gradients when modelling GP, which makes some GP hyperparameters, such as lengthscales cannot be updated. Based on recent work of Hvarfner et al., 2024, the authors find... | Rebuttal 1:
Rebuttal: We are grateful for the reviewer's thoughtful feedback and recognition of our work's strengths, particularly the systematic summary of high-dimensional Bayesian optimization (HDBO) issues and the effectiveness of our proposed methodology. In response to the constructive comments, we have addressed... | null | null | null | null | null | null |
Balancing Model Efficiency and Performance: Adaptive Pruner for Long-tailed Data | Accept (poster) | Summary: This paper introduces Long-Tailed Adaptive Pruner (LTAP), a novel pruning strategy designed to enhance neural network efficiency while preserving performance on long-tailed datasets. LTAP addresses this challenge by incorporating multi-dimensional importance scoring and a dynamic weight adjustment mechanism, e... | Rebuttal 1:
Rebuttal: **Thank you for your recognition of our work. We have carefully understood the relevant weaknesses you mentioned and have made the following efforts:**
- **About broader references**
We have incorporated the studies you mentioned into our discussion and references as follows:
Recent advances ... | Summary: This paper introduces an adaptive pruning method called LTAP to address the challenge of handling long-tailed distribution data. The authors propose a multi-dimensional importance scoring criterion and design a dynamic weight adjustment mechanism to adaptively determine the pruning priority of parameters for d... | Rebuttal 1:
Rebuttal: **Thank you for your recognition of our work. We have carefully understood the relevant weaknesses you mentioned and have made the following efforts:**
- **About presentation form**
Thank you for your valuable suggestions. We will supplement the theorems and textual explanations in Section III w... | Summary: This paper introduces Long-Tailed Adaptive Pruner (LTAP), a model pruning framework designed for long-tailed class distributions. LTAP integrates multi-criteria importance scoring and a dynamic LT-Vote mechanism to prioritize preserving parameters crucial for tail classes. It employs multi-stage pruning, gradu... | Rebuttal 1:
Rebuttal: **Thank you for your time, effort and recognition of our work. Your comments are very important for us to continue to improve this work. Based on your comments, we continue to make the following efforts**:
- **About Pseudocode**
This is our negligence. We have completed and corrected the pseudoc... | Summary: This paper introduces LTAP (Long-Tailed Adaptive Pruner), a pruning strategy tailored for long-tailed data distributions. LTAP addresses the challenge of class imbalance by dynamically adjusting pruning priorities through a multi-criteria importance evaluation framework.
Claims And Evidence: Formal proofs es... | Rebuttal 1:
Rebuttal: Thank you for your time and effort. We are encouraged by your high appreciation of the novelty and contribution of our paper. We will continue to work on exploring how to better solve long-tailed problems in real-world scenarios. If you have any further questions, we will address them at any time.... | null | null | null | null | null | null |
Gradient Aligned Regression via Pairwise Losses | Accept (poster) | Summary: This paper proposes GAR for solving regression problems while maintaining label similarity. The authors introduce two pairwise label difference losses and explain their theoretical insights. Various experiments and ablation studies demonstrate that GAR outperforms other baselines.
Claims And Evidence: The cla... | Rebuttal 1:
Rebuttal: Thank you for your invaluable comments! Here are our responses:
1. "*The proposed method mainly focuses on the clean data setting, while other baselines in the experiments, for example, RNC, have applications in improving generalization and robustness.*"
**Response**: RNC is a nice regression m... | Summary: This paper proposes to use pairwise comparison to improve regression accuracy. Authors propose to focus on the difference of pairs, and propose two loss terms comparing the differences of pairs. Authors then show the equivalence form that enjoys linear computational efficiency using specific loss functions. Fu... | Rebuttal 1:
Rebuttal: We sincerely thank you for your invaluable feedback on our manuscript! Here are our responses:
1. "*However, the significance of how the proposed method is located in the literature is not very clear for readers who's not familiar with the specific literature of using pairwise data for regression... | Summary: In this work, the authors focus on the problem of regression with pairwise losses. Firstly, it is revealed that the conventional regression losses (e.g., MSE and MAE) ignore the pairwise information of training samples and can have larger error variance. Inspired by the success of existing pairwise losses, the... | Rebuttal 1:
Rebuttal: We appreciate your invaluable comments for our manuscript! Here are our responses:
1. "*The cited works primarily focus on regression. However, similar types of loss functions have also been explored in the context of AUC optimization...*"
**Response**: we sincerely thank you for raising this p... | null | null | null | null | null | null | null | null |
Mitigating Local Cohesion and Global Sparseness in Graph Contrastive Learning with Fuzzy Boundaries | Accept (poster) | Summary: This paper introduces fuzzy boundaries to address ingrained local cohesion and global sparseness in graph contrastive learning (GCL). To address these issues, this paper proposes a novel GCL model that replaces the crisp cluster boundaries with adaptive fuzzy boundaries to adjust the cluster boundaries. Extens... | Rebuttal 1:
Rebuttal: **W1.**
We chose Max Pooling as the Comb operator due to its theoretical advantage, however, it is indeed necessary to compare it with other pooling strategies in practice. We have conducted additional experiments on other Comb operators, and the results on node classification are shown in Table ... | Summary: To address the ingrained shortcomings such as local cohesion and global sparseness in graph contrastive learning (GCL), this paper introduces a novel GCL model incorporating fuzzy boundaries. The proposed method dynamically extends original cluster boundaries to mitigate the two shortcomings in the embedding s... | Rebuttal 1:
Rebuttal: **W1.**
We compare our model with two advanced graph generative models, S2GAE [1] and Bandana [2]. S2GAE reconstructs randomly masked edges while Bandana samples edge masks from a continuous distribution. As shown in Table 1, these graph generative models perform competitively but are limited to ... | Summary: This paper presents a novel graph model by introducing fuzzy set theory to alleviate the impacts of local cohesion and global sparseness, expecting to form the ideal global structure distribution. The proposed model realizes a learned boundary to bridge the local groups and isolated samples by adjusting the cl... | Rebuttal 1:
Rebuttal: **W1.**
The details of the centralization are as follows:
When centralizing the node representations, we use the K-means clustering algorithm to obtain the cluster prototype $C\in\mathbb{R}^{C_p\times d}$, where $C_p$ is the number of clusters, and $d$ is the dimension of the prototype vector. F... | Summary: This paper addresses two integrated shortcomings in graph contrastive learning, namely local cohesion and global sparseness, which will lead to inferior global structural distributions. So this paper proposes novel fuzzy boundaries to gather discrete local groups and isolated samples, which effectively allevia... | Rebuttal 1:
Rebuttal: **W1.**
Following your kind suggestion, we provide a related work compassing graph prototype learning as follows:
Graph prototypical learning has emerged as a significant direction in graph representation learning, addressing limitations in traditional contrastive and few-shot learning approache... | null | null | null | null | null | null |
La RoSA: Enhancing LLM Efficiency via Layerwise Rotated Sparse Activation | Accept (poster) | Summary: The paper introduces LaRoSA, a method that applies layer-wise rotation to activation sparsity, which improves compression performance with minimal additional inference cost.
Claims And Evidence: The claims are well-supported with clear evidence.
Methods And Evaluation Criteria: The evaluations include differ... | Rebuttal 1:
Rebuttal: We sincerely thank reviewer jWLg for the thorough review! Below is our point-by-point response to your feedback:
> The experimental results for TEAL are slightly worse than those reported in the original paper ...
Yes, your observation is correct. We will modify and extend implementation details... | Summary: The paper presents LaRoSA, a training-free activation sparsification method aimed at improving inference efficiency in LLMs. The key idea is orthogonal rotation-based pruning using a Top-K selection mechanism that eliminates lower-magnitude activations. The experimental results across multiple LLMs (LLaMA2, LL... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer Yw7t for taking your valuable time to evaluate our work. The acknowledgment of our contributions has greatly inspired our entire team. Below is our point-by-point response to your feedback:
>Appendix A, minor: I'd include a 1-line at the top to explain what this a... | Summary: Sparse activation is an active research field of LLMs. This paper demonstrates that using magnitude-based pruning to sparsity leads to inconsistent sparsity across tasks and layers. Stemming from their rich preliminary study, they propose La RoSA which leverages the rotation operation proposed in SliceGPT to a... | Rebuttal 1:
Rebuttal: We thank reviewer for the thorough review! Below is our point-by-point response:
> I would like to see ... on harder tasks with more advanced models.
We agree that perplexity alone is insufficient, and follow your suggestion to evaluate LaRoSA on harder tasks with advanced models using Hugging F... | Summary: This work proposes a sparse activation method for efficient inference by rotating inputs by an orthogonal matrix. Basic idea is to improve sparsity by rotating inputs to matrix multiplication happening in Transformer layers, e.g., MLP and attention, so that low magnitude inputs are pruned away in the rotated s... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer nBNG for his/her thoughtful feedback and recognition of our work, which has greatly encouraged our research team. Our proposed LaRoSA framework utilizes Computational Invariance Transformation, a well-established technique in model quantization and pruning, to syste... | null | null | null | null | null | null |
Understanding Fixed Predictions via Confined Regions | Accept (poster) | Summary: This paper introduces the ReVer method, designed to detect fixed predictions in machine learning models by identifying confined regions in the feature space. The approach involves formulating the Region Recourse Verification Problem (RVP), approximating confined regions using bounding boxes, generating confine... | Rebuttal 1:
Rebuttal: Thank you for your time and feedback! We really appreciate you highlighting that our work address a topic that is “impactful” where the “practicality [of our approach]is evident”, our MIQCP approach provides “strong theoretical guarantees”, and that our “experimental design is sound, encompassing ... | Summary: The paper proposes a method to check whether regions of input features are *responsive* (all individuals in the region have recourse), *confined* (all individuals have no recourse) or neither.
Claims And Evidence: - The authors claim that they “introduce a new approach to formally verify recourse over entire ... | Rebuttal 1:
Rebuttal: Thank you for your time and feedback! We appreciate your comments about the paper, which we address below:
> What pointwise method is used as baseline…
We solve a MILP to check whether a given data point has recourse (i.e., solve the REP for the given point x where B(u,l) = x). This can be viewe... | Summary: This paper aims to identify regions where each point either allows recourse or does not allow recourse in the case of a linear model, given a fixed set of constraints.
In this specific case, the problem can be naturally formulated as a Mixed-Integer Quadratically Constrained Programming (MIQCP) problem. The a... | Rebuttal 1:
Rebuttal: Thank you for your time and feedback! We appreciate you highlighting that “existing solutions appear to have failed to solve the problem of interest”, our experiments are “sound”, and that our proposed approach is “more effective [than existing approaches] for linear models”.
> However, my main ... | Summary: The paper introduces a novel approach to identifying fixed predictions in machine learning models by finding confined regions where all individuals receive fixed predictions. The authors propose a method called ReVer, which uses mixed-integer quadratically constrained programming (MIQCP) to certify recourse f... | Rebuttal 1:
Rebuttal: Thank you for your time and feedback! We were excited to see you recognized that our paper “addresses a significant gap in the literature”, “highlights the limitations of existing approaches”, and validates our claims via “a comprehensive empirical study across various applications”.
We appreciat... | null | null | null | null | null | null |
Hierarchical Reinforcement Learning with Uncertainty-Guided Diffusional Subgoals | Accept (poster) | Summary: This paper introduces an HRL framework that leverages a diffusion model for subgoal generation and GP regularization to enhance subgoal quality and reachability. The proposed method addresses the instability and inefficiency of traditional HRL by learning a state-conditioned subgoal distribution via a conditio... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the insightful comments.
**1. Theoretical Analysis of Diffusion-Based Subgoal Generation and GP Regularization**
We appreciate the reviewer’s insightful comments on the theoretical justifications. In the anonymous link https://anonymous.4open.science/r/HIDI-3... | Summary: This paper aims to mitigate the non-stationary effects caused by the simultaneous training of both high-level and low-level policies in HRL. The authors introduce a novel framework, HIDI, that leverages a conditional diffusion model regularized with a Gaussian Process (GP) prior to generate diverse and achieva... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the insightful comments.
**1. Non-Stationarity Analysis:**
We clarify that HIDI does not claim to eliminate non-stationarity, an inherent challenge in HRL due to the evolving low-level policy. Instead, HIDI mainly addresses **mitigating its adverse effects**, ... | Summary: This paper introduces Hierarchical Reinforcement Learning (HRL) using a conditional diffusion model combined with Gaussian Process (GP) regularization for sequential decision making. Specifically, the authors propose to train a diffusion policy at the high-level thus to generate subgoals that align with the lo... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the insightful comments.
**1. Low-Level Policy Training & High-Level Adaptation:**
As correctly noted by the reviewer, our approach builds on the adapted HIRO framework, where the low-level policy trains on intrinsic rewards to achieve subgoals reliably, with... | Summary: The paper introduces a novel framework for hierarchical reinforcement learning (HRL) that combines a conditional diffusion model for generating state-conditioned subgoals with a Gaussian Process prior for uncertainty quantification and regularization. This hybrid approach ensures robust and adaptive subgoal ge... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the insightful comments.
**1. Comparison to Diffusion-based & Uncertainty-aware Baselines:**
We respectfully clarify that our evaluation already includes strong baselines covering an **uncertainty-aware approach, i.e., HLPS (Wang et al., 2024)**, and HIDI sig... | null | null | null | null | null | null |
Reinforcement Learning with Segment Feedback | Accept (poster) | Summary: This paper studies RL with segment feedback, where each episode in an MDP is divided into $m$ equal-length segments, and the agent receives reward feedback only at the end of each segment instead of every state-action pair. The authors investigate two feedback settings: binary feedback, where the agent receive... | Rebuttal 1:
Rebuttal: Thank you for your time in reviewing our paper! We will revise our paper according to your comments.
**1. Determine the Feedback Type**
We assume an MDP setting, so the sum of step-level rewards being the trajectory reward is natural. Given this model, we can ask a human to give a score to a tra... | Summary: This paper studies the problem of RL with segment feedback, which offers a general paradigm filling the gap between per-state-action feedback and trajectory feedback. The authors study two settings with binary and sum feedback, developing computationally efficient algorithms achieving nearly optimal regret upp... | Rebuttal 1:
Rebuttal: Thank you for your time and effort in reviewing our paper! We will certainly incorporate your comments in our revision.
**1. Comparison with [Chatterji et al., 2021] in Formulation**
The objectives in [Chatterji et al., 2021] and our paper are different. In [Chatterji et al., 2021], the goal is ... | Summary: This paper introduces Reinforcement Learning with Segment Feedback, a framework where episodic MDPs are divided into
$m$ segments, and feedback is provided at the end of each segment. The authors study binary feedback (stochastic binary outcomes) and sum feedback (cumulative rewards with noise), analyzing how... | Rebuttal 1:
Rebuttal: Thank you for your time and effort in reviewing our paper! We will certainly incorporate your suggestions in our revision.
**1. More Experiments**
This is an excellent suggestion. However, it is difficult to perform experiments on the Atari environments before the rebuttal deadline since it woul... | null | null | null | null | null | null | null | null |
Optimal Auction Design in the Joint Advertising | Accept (poster) | Summary: This paper studies the auction design problem in the joint advertising scenario. For the single-slot setting, the optimal mechanism is derived. For the multi-slot setting, the authors propose BundleNet which achieves good performance.
Claims And Evidence: I think so.
Methods And Evaluation Criteria: Generall... | Rebuttal 1:
Rebuttal: **Q1.** We agree this is an excellent suggestion. In fact, we noticed other reviewers raised similar questions. Here we clarify our key findings:
1. **Single-slot scenario**:
BundleNet achieves allocations **closer to the optimal solution** compared to baseline methods (as demonstrated... | Summary: The paper considers the sponsored search auctions with joint advertising. In this setting, advertisers are partitioned into suppliers and retailers, with a bipartite graph representing whether it is possible for the pair of advertisers to form a joint advertisement. Each advertisement is then assigned to a ret... | Rebuttal 1:
Rebuttal: **Q1.** To comprehensively validate our approach, we have conducted additional experiments across diverse settings:
| Single-slot ｜LN(0.1,1.44) | $LN_2$ | $LN_3$ | $LN_4$ | $LN_5$ |
| -------------------------- | ---------- | ---------- | ---------- | ---------- |
| JRegNet ... | Summary: The paper addresses auction design in joint advertising settings, where both retailers and suppliers jointly bid on ad slots. It extends classical auction theory by identifying an optimal mechanism for single-slot joint advertisements (via a Myerson-inspired framework) and introduces BundleNet—a novel neural n... | Rebuttal 1:
Rebuttal: **Q1.** This is a very good question. I think this method may limit the space of discoverable mechanisms, but the difference is not significant.
1. Our core idea is to improve performance by limiting the space of discoverable mechanisms, avoiding the neural network's processing of sparse graph da... | Summary: This paper proposes an optimal mechanism for joint advertising in a single-slot setting and introduces BundleNet, a bundle-based neural network for multi-slot joint advertising. Through extensive experimentation, the paper demonstrates that BundleNet effectively approximates theoretical results in single-slot ... | Rebuttal 1:
Rebuttal: We would like to thank you for the constructive comments and suggestions! Below are our responses to your questions.
**Q1.** We conducted experiments using real-world data from a leading Internet company. In practice, advertisements will go through the stages of recall, rough sorting, and fine so... | null | null | null | null | null | null |
MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data | Accept (poster) | Summary: This paper proposes MindAligner, a framework using functional alignment to facilitate cross-subject brain decoding. Their framework consists of (1) “Brain Transfer Matrix” (BTM) that works by transforming limited novel subject brain activity into the brain activity of a known, previously seen subject via linea... | Rebuttal 1:
Rebuttal: Thank you for your time and valuable feedback. We will incorporate the suggested modifications in the revised version.
**Q1: Clarification on "explicit" and "implicit" functional alignment**
We define alignment in voxel space as explicit and methods in latent space that cannot be restored to vox... | Summary: This work considers inter-subject alignment in the context of image reconstruction from fMRI. A light way Brain Transfer Matrix is used to align a novel subject to a known subject, which differs from the previous work that aligns subjects in a shared latent space.
Claims And Evidence: Most claims are well-sta... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We will address your concerns point by point below:
**Q1: About the training parameters of MindEye2.**
According to the open-source code of MindEye2, it performs full-parameter fine-tuning when adding a new subject. The results in Table 3 are obtained by... | Summary: The manuscript proposed an explicit brain functional alignment for cross-subject decoding. The method trains a cross-subject brain transfer matrix to map signals from novel subjects to the known subject. Experimental results demonstrate improved performance and provide insightful interpretations.
## update af... | Rebuttal 1:
Rebuttal: Thank you for your valuable time. We will revise the typos as suggested. **We promise to open-source our code upon acceptance.**
**Q1: The settings of the Tab. 1 is unclear.**
Thank you for your question. The setting in Tab. 1 is as follows: "Ours (subj 1)" refers to the average result obtained ... | Summary: This paper propose MindAligner, a explicit brain signal functional alignment framework for cross-subject brain decoding. It utilize a LoRA-based **brain transfer matrix (BTM)** to convert signals from novel subjects into signals fo a known subjects. A **brain functional alignment (BFA)** module based on linear... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable time and insightful feedback. Your recognition is highly meaningful to us. Below is our response addressing your concerns：
**Q1: Multi-level -> Bi-level.**
Thank you for your suggestion. We will revise it as suggested.
**Q2: About the linear hypothesis in B... | null | null | null | null | null | null |
Kona: An Efficient Privacy-Preservation Framework for KNN Classification by Communication Optimization | Accept (poster) | Summary: This paper focuses on private KNN inference using multi-party computation techniques. Specifically, it introduces a new framework, Kona, for private KNN inference by designing new Euclidean triples. Additionally, the paper proposes a divide-and-conquer bubble protocol to achieve lower communication complexity.... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments. Below, we respond to your concerns:
### 1. Can the proposed method support private KNN inference with other distance measures?
Please see Point 1 for Reviewer F4RX.
### 2. Intuitively, how does the proposed Euclidean triple help reduce the previously l... | Summary: This paper presents a systematic framework called Kona for privacy-preserving KNN classification, which addresses communication inefficiencies in existing approaches. The key innovations include Euclidean triples to eliminate online communication for Euclidean Square Distance computations and a divide-and-conq... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments. Below, we respond to your concerns:
### 1. How does the DQBubble protocol work under the encrypted distance, if all the encrypted distance use the same seed, will there be risk to leak seed such that disclose the original distance?
Our proposed DQBubble pr... | Summary: The paper introduces Kona, an efficient privacy-preserving framework for KNN classification, and optimizes communication overhead through two key methods: (1) It designs novel Euclidean triples to eliminate the need for online communication during secure computations of Euclidean squared distances. (2) It prop... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments. Below, we respond to your concerns:
### 1. Can the proposed methods be applied using distance metrics other than Euclidean distance?
Yes, both of our proposed optimizations can be applied to other distance metrics.
First, our proposed divide-and-conquer... | null | null | null | null | null | null | null | null |
Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding | Accept (poster) | Summary: The paper empirically observed that LLM with RoPE would have massive QK values, which has positive correlation to the knowledge understanding capability.
## updated after rebuttal
Claims And Evidence: LLM with RoPE would have massive QA value pattern, which is responsible for knowledge understanding. The auth... | Rebuttal 1:
Rebuttal: > D1: Allen-Zhu and Li also highlighted the importance of RoPE. Please provide a discussion.
Thank you for the excellent suggestion. We acknowledge that our current version does not yet include a discussion of Allen-Zhu and Li [1]. We appreciate the reviewer’s input and will include proper citati... | Summary: This paper investigates the emergence of massive values in self-attention within LLMs and their impact on contextual knowledge understanding. Through extensive experiment, the authors demonstrate that these massive values play a crucial role in interpreting contextually provided knowledge rather than retrievin... | Rebuttal 1:
Rebuttal: > W1: Lacks a deep analysis of attention pattern changes due to massive values.
We thank the reviewer for pointing this out. We agree that a deeper analysis of how massive values affect attention patterns would strengthen the paper, and we will include such an analysis in the revision. Specifical... | Summary: This paper investigates the massive value elements in output vectors of internal layers of Transformer from the viewpoint of their influences on task-solving performance and the reason for them. The empirical analysis of various large language models (LLMs) reveals that the massive values influence tasks requi... | Rebuttal 1:
Rebuttal: > S1: subsection 3.5: (Yu et al.) -> (Yu et al. 2023)
Thank you for your suggestion. We will correct the citation in subsection 3.5 from (Yu et al.) to (Yu et al. 2023) to ensure proper citation formatting.
> Q1: How the massive value occurs is still uncertain.
Thank you for raising this critic... | Summary: The authors study the case of Massive values K,Q values which appear in LLMs.
The authors conclude that the primary reason for this ROPE embeddings.
Another interesting finding the authors proposed was that the massive K,Q values have different impact in different types of task.
Claims And Evidence: The autho... | Rebuttal 1:
Rebuttal: > W1: Several prior works like Cachegen StreamingLLMs have similar insights. My primary concern is that some of these findings are not new. I think the study is quite interesting, however the impact is quite unclear. For example- Existing optimization uses similar insights.
Thank you for pointing... | null | null | null | null | null | null |
Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect | Accept (poster) | Summary: This paper addresses the challenges of the exploration-exploitation tradeoff in the adaptive estimation of the Average Treatment Effect (ATE), proposing a novel approach based on Neyman regret. The authors clearly outline the motivation and theoretical foundation of the algorithm, and conduct simple experiment... | Rebuttal 1:
Rebuttal: > It would be beneficial to compare the proposed method with some of the many related works, if possible.
We did not compare our algorithm with Kato et al. (2020) because Cook et al. (2022) showed that ClipSDT significantly outperforms Kato’s algorithm, even under weaker assumptions (see Figure 2... | Summary: The paper designs an adaptive algorithm, OPTrack, to estimate the average treatment effect (ATE) by integrating the augmented inverse probability weighting (AIPW) estimator with the optimism principle from multi-armed bandits. The paper claims that OPTrack adaptively balances the exploration of minimizing the ... | Rebuttal 1:
Rebuttal: > On line 144, what is the definition of $R(a)$?
We meant that $\Delta = \mathbb{E}[R_t(1) - R_t(0)]$ for any $t$ and we had omitted the subscript for brevity since the choice of $t$ is immaterial. However, we agree that this could be confusing. To clarify, we have revised the definition to $\Del... | Summary: This study proposes optimistic algorithms for the adaptive estimation of the average treatment effect. While existing studies focus on clipping-based algorithms to stabilize performance, the present work develops a method by setting the treatment-assignment probability $\pi_t$ close to $1/2$. In addition, the ... | Rebuttal 1:
Rebuttal: Thank you for your positive and thoughtful review. We appreciate your feedback and are glad that you found our work worthy of acceptance.
In response to your suggestion, we have incorporated a discussion of the concurrent work by Noarov et al. (“Stronger Neyman Regret Guarantees for Adaptive Expe... | Summary: This paper proposes an adaptive algorithm for estimating the Average Treatment Effect (ATE) in a two-arm setup. The authors introduce an Optimistic Policy Tracking (OPTrack) method that uses “optimism in the face of uncertainty” to allocate subjects so as to reduce the variance of their final ATE estimator. Th... | Rebuttal 1:
Rebuttal: > The claims are substantiated, although clarity and exposition could be greatly improved (e.g., Section 3 should be more clear and define the variables accordingly, the definition of Neyman regret seems off in eq. (5), result of Lemma B.1 is central to the algorithm and should be moved to the mai... | null | null | null | null | null | null |
Learning to Reuse Policies in State Evolvable Environments | Accept (poster) | Summary: This work introduces introduces a framework to handle non-stationary state distributions during deployment of an RL agent. The authors are motivated by the deterioration or maintenance of sensors on real-world systems that do not have a constant fidelity. The proposed framework is structured around a State Evo... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments and suggestions. We provide our responses below and present the additional experimental results in https://anonymous.4open.science/r/Lapse-7846.
### Q1. Performance of baselines
The baselines underperform because they cannot anticipate all possible state e... | Summary: The authors propose the framework Learning to reuse policies in state evolution (Lapse) to solve the issue of model degradation when there is an explicit change in the environment representation that is received by the agent in the form of input. They formalize this scenario as State Evolvable Reinforcement Le... | Rebuttal 1:
Rebuttal: Thank you for careful review our paper and providing constructive comments and suggestions. We offer some clarification to your questions here, and we would appreciate any further comments you might have.
### Q1 Changes are directly signaled to Lapse
Yes, the change in the environment should be ... | Summary: This paper introduces the problem of state evolvable reinforcement learning (SERL) and proposes a method called Lapse to address it.
The key contributions are as follows:
* Formalization of the SERL problem, where the state space of an environment evolves over time due to sensor changes.
* Lapse takes a two-p... | Rebuttal 1:
Rebuttal: Thank you for your careful review and thoughtful comments. We hope our responses can relieve your concern. We present the additional experimental results in https://anonymous.4open.science/r/Lapse-13F8.
### Q1: Prior knowledge about sensor changes
While this assumption does not hold universally,... | null | null | null | null | null | null | null | null |
PEINR: A Physics-enhanced Implicit Neural Representation for High-Fidelity Flow Field Reconstruction | Accept (poster) | Summary: The paper introduces two contributions:
1. Multi-Fidelity Large-Scale CFD Dataset: A comprehensive dataset generated using the WENO solver, comprising five different configurations of 2D and 3D flows that include shock phenomena. Each configuration is provided at two resolutions: a low-resolution version (app... | Rebuttal 1:
Rebuttal: Thanks for your reviews. Here are our explanations about the weaknesses and problems.
1.Response to comments on transferability and generalization of PEINR
We appreciate your query regarding the generalization of our model. Since the physical phenomena in the flow field vary greatly from problem... | Summary: This paper introduces HFR-Bench, a large-scale CFD dataset, with 33,600 unsteady 2D and 3D vector fields across various grid resolutions and numerical precisions, providing a benchmark for flow field reconstruction. It also proposes PEINR, a novel physics-enhanced INR model that improves flow field reconstruct... | Rebuttal 1:
Rebuttal: Thanks for your reviews. Here are our explanations about the weaknesses and problems.
1.Response to comments on the missing citation CoordNet(Line 246)
Thanks for your suggestions and we will correct it in our revised paper. ResuMLP is proposed by CoordNet, and combines residual connections with... | Summary: This paper introduces PEINR, a physics-enhanced implicit neural representation framework for high-fidelity flow field reconstruction. The authors address three key limitations of existing INR methods: 1) invalid grid independence assumption, 2) temporal-spatial complexity disparity, and 3) spectral bias. The w... | Rebuttal 1:
Rebuttal: Thanks for your reviews. Here are our explanations about the weaknesses and problems.
1.Response to comments on the single 3D dataset
The SV case is a classical benchmark in CFD that incorporates both shock waves and vortical structures, which can effectively validate our method's capability in ... | Summary: The paper introduces PEINR (Physics-enhanced Implicit Neural Representation), a framework designed for high-fidelity flow field reconstruction. The authors highlight the limitations of current implicit neural representation (INR) methods in handling complex spatiotemporal dynamics and accurately capturing fine... | Rebuttal 1:
Rebuttal: Thanks for your reviews. Here are our explanations about the weaknesses and problems.
1.Response to comments on missing citations
[1] de Vito et al. (2024) Implicit Neural Representation For Accurate CFD Flow Field Prediction.
[2] Du et al. (2024) Conditional neural field latent diffusion mode... | null | null | null | null | null | null |
Liger: Linearizing Large Language Models to Gated Recurrent Structures | Accept (poster) | Summary: Transformer Language models based on linear recurrent structures (Katharopoulos et al., 2020; Yang et al., 2023; Qin et al., 2024b) are substantially more efficient than regular transformers due to their linear dependency on sequence length and constant memory requirements. Prior work - SUPRA (Mercat et al., ... | Rebuttal 1:
Rebuttal: Dear Reviewer 5t3M,
Thanks for your thorough reviews and we appreciate the attention to detail and would fix the writing errors in the revision. We give point-by-point responses to your comments based on your questions.
> `Q1`: Is pooling performed across the time dimension?
No, the pooling op... | Summary: This paper proposed a novel method, Liger, for transforming pretrained Transformer-based LLMs into gated linear recurrent models, repurposing key matrix weights to create gating mechanisms.
It utilizes Liger Attention, an intra-layer hybrid attention mechanism involving Gated Recurrent Modeling (GRM) and Slid... | Rebuttal 1:
Rebuttal: Dear Reviewer 2WYJ,
We are truly grateful for the time you have taken to review our paper and your insightful comment. We address your questions in the following.
> `Q1`: Discussion on intra-layer hybrid attention with similar designs
We carefully read these works for further comparison and dis... | Summary: This paper studies the problem of how to convert pretrained standard attention-based LLMs into more efficient hybrid models, a research area that recently emerges along with significant progress on sub-quadratic model architectures. It proposes a methodology to convert current Mistral/Llama models to constant-... | Rebuttal 1:
Rebuttal: Dear Reviewer xf9Q,
We sincerely thank you for your careful comments and thorough understanding of our paper! Here we give point-by-point responses to your comments and questions.
> `Q1`: The baseline that only uses sliding window attention is missing.
Thanks for your careful comment. We have s... | null | null | null | null | null | null | null | null |
How Compositional Generalization and Creativity Improve as Diffusion Models are Trained | Accept (poster) | Summary: This paper shows empirically that diffusion models learn compositional rules in increasing depth. They then provide an analytic computation of the sample complexities of learning different levels of rules and show that the empirical sample complexities align with these predictions. Finally they define local an... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and address their specific concerns below.
**Scope of the analysis.** Our theoretical results are limited to the RHM, which is a form of PCFG. As suggested, we will modify the writing to stress this point in the abstract, introduction, and conclusion.
**M... | Summary: This paper studies the sample complexity of learning hierarchical structures, mainly through theoretical analysis of RHMs accompanied by some experiments on synthetic RHMs as well as simple text and and image experiments. For the RHM, their theoretical analysis predicts that the production rules at a given lev... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and address their specific concerns below.
**Role of branching factor $s$.** The number of samples required to learn the full hierarchy in the absence of weight sharing is $m^{L+1}$, which is independent of $s$. In the presence of weight sharing, we expec... | Summary: In the paper, the authors use probabilistic context-free grammars and random hierarchy models in order to analyze compositional generalization in diffusion models. The papers begin by training a D3PM, analyzing how the model learns the rules of RHM at different hierarchical levels. They empirically observe a l... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and address their specific concerns below.
**Paper claims**
1. **Creativity.** By *creativity*, we refer to the well-established phenomenon in linguistics (e.g., Chomsky 1976) in which a finite set of rules can be learned by an infant so as to generate a... | Summary: The paper proposes a model for understanding the evolution of learning feature dependencies when training diffusion models, specifically in terms of modeling the sample complexity required to learn certain structures. To do this, they adopt the adopt the perspective that data is generated by some ground truth ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and address their specific concerns below.
**Models of images.** We agree that the relation between pixel and latent space in images is subtle. While images are continuous, they admit high-level abstract representations that can be described with discrete ... | null | null | null | null | null | null |
E-LDA: Toward Interpretable LDA Topic Models with Strong Guarantees in Logarithmic Parallel Time | Accept (poster) | Summary: The paper introduces a novel approach to Latent Dirichlet Allocation (LDA) topic modeling by developing a combinatorial method to efficiently infer topic assignments with strong interpretability guarantees. The proposed method, Exemplar-LDA (E-LDA), departs from traditional gradient-based approaches and instea... | Rebuttal 1:
Rebuttal: **Thank you for your overall positive review, and for remarking that our algorithm provides "strong provable guarantees" that "address a longstanding challenge in topic modeling", and that it improves interpretability and achieves strong empirical performance.**
**We recognize that your remaining... | Summary: The authors present a novel non-gradient-based, combinatorial approach for estimating topic model parameters. The model converges to an algorithm that achieves near-optimal posterior probability in logarithmic parallel computation time, significantly faster than known algorithms for LDA. Furthermore, this manu... | Rebuttal 1:
Rebuttal: **Thank you for your thoughtful and detailed review, and your comment that "The approach proposed in this paper enhances the LDA model from a novel perspective and offers theoretical guarantees...demonstrating high originality and inspiration for future work in the field."**
We recognize that yo... | Summary: The paper introduces new algorithms for Latent Dirichlet Allocation (LDA). The algorithms address the problem of inferring the topics assigned to each document in an LDA topic model. The approach is non-gradient-based and combinatorial.
The algorithms converge to near-optimal posterior probability in loga... | Rebuttal 1:
Rebuttal: **We thank you for your positive review, and we are sincerely grateful for your many detailed and thoughtful comments. We respond below to your specific points about comparing our computational complexity with alternatives, interpretability, causal inference, and the additional references.**
**Re... | Summary: The paper introduces Exemplar-LDA (E-LDA), a novel combinatorial approach to Latent Dirichlet Allocation (LDA) topic modeling, addressing the NP-hard problem of inferring document-topic assignments. Unlike traditional gradient-based methods, E-LDA uses a non-gradient, submodular optimization framework, achievi... | Rebuttal 1:
Rebuttal: **Thank you for your overall positive review, and for your positive comments about our "rigorous theory" and "solid empirical results." We also thank you for your positive comments that the paper "is interesting and gives insight to the community."**
**We believe we address your remaining reserva... | null | null | null | null | null | null |
Causal Logistic Bandits with Counterfactual Fairness Constraints | Accept (poster) | Summary: This paper introduces a framework for causal logistic bandits with counterfactual fairness constraints, addressing sequential decision-making under fairness requirements and offering theoretical guarantees and practical algorithms for balancing performance and fairness.
## update after rebuttal
The authors ad... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful feedback. Below, we address the main comments provided in the review.
### _**Claim and Evidence**_
> _1. "performance guarantee" This paper has only theoretical effectiveness analysis, no experimental validation of performance._
We will add empirical results using a... | Summary: This paper studies the problem of imposing counterfactual fairness constraints in causal logistic bandit problems. The authors propose an algorithm that leverages primal-dual optimization for constrained causal logistic bandits, where the non-linear constraints are a priori unknown and must be learned over tim... | Rebuttal 1:
Rebuttal: We appreciate your time and feedback. Below, we address the main comments provided in the review.
### _**Weaknesses**_
> _1. The paper could benefit from extensive numerical studies._
We have done some empirical experiments using a synthetic dataset, evaluating both cumulative regret and constr... | Summary: This paper forwards an online decision-making framework constrained using counterfactual fairness constraints. The paper considers a sequential decision-making setup where a set of applicants arrive at each time step. The goal is to make labeling decisions that maximize the cumulative outcome reward associated... | Rebuttal 1:
Rebuttal: We are grateful for your detailed and constructive suggestions on our submission. We address the main comments provided as following.
### _**Claims and Evidence**_
> _1. Intuitively, if I understand correctly, the larger the $\delta$ value the easier it likely is to find a "fair" policy, and the... | Summary: The paper introduces a new problem formulation: Constrained Causal Logistic Bandits (CCLB), which addresses online decision-making under counterfactual farness constraints. This paper proposes an algorithm utilizing primal-dual optimization for constrained causal logistic bandits with non-linear constraints an... | Rebuttal 1:
Rebuttal: We appreciate your time and feedback. Below, we address the main comments provided in the review.
### _**Weaknesses**_
> _1. My primary concern is the lack of empirical validation._
We will add empirical results using a synthetic dataset, evaluating both cumulative regret and constraint violat... | null | null | null | null | null | null |
$Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training | Reject | Summary: This work proposes Q♯, a distributional value-based algorithm for KL-regularized RL aimed at improving the post-training alignment of LLMs. The key idea is to learn the optimal Q function via a supervised distributional learning approach. The authors provide theoretical guarantees and validate their approach o... | Rebuttal 1:
Rebuttal: Dear Reviewer gBqj,
Thank you so much for your detailed review and suggestions for us to improve the paper. We respond to individual points below.
**Could the authors conduct experiments with more models to enhance the generality of the approach?**
**Authors' reply:** Thank you for the suggesti... | Summary: The paper introduces a novel value-based reinforcement learning algorithm designed for post-training LLM's. Unlike traditional policy-based methods like PPO and DPO, which may struggle with inherited pre-training biases, Q♯ leverages distributional RL to optimize a KL-regularized objective. It learns an opti... | Rebuttal 1:
Rebuttal: Dear Reviewer B1zk,
Thank you for your encouraging review and strong support! We address individual questions below.
**On the right hand side of the first equation should be $Q^{\star,\eta}_h$.**
**Authors' reply:** Thanks for catching this typo! We will make sure to fix it in the final version... | Summary: The paper introduces a value-based algorithm for KL-regularized RL in deterministic MDP, where the authors leverage distributional RL to estimate the optimal regularized $Q$-function. Theoretically, the authors reduce KL-regularized RL to no-regret online learning, proving variance-dependent PAC bounds under r... | Rebuttal 1:
Rebuttal: Dear Reviewer GoFf,
Thank you for the detailed review and constructive feedback! We truly appreciate the time and effort that have been invested in providing constructive comments. Please find our responses below.
**Star-graph results for claiming policy-based method shortcomings in planning tas... | null | null | null | null | null | null | null | null |
A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models | Accept (poster) | Summary: The paper investigates a likelihood-based framework for regression in a view of conditional distribution estimation. The authors propose a Sieve Maximum Likelihood Estimator (MLE) to estimate the conditional density of the response given predictors. Importantly, they derive convergence rates (under Hellinger a... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful and constructive feedback, and we sincerely appreciate their recognition of the strengths in our likelihood‐based framework and its theoretical underpinnings.
**Claims and Evidence:** We thank the Reviewer for highlighting these key aspects of our theor... | Summary: In this paper, the authors explore the theoretical properties of deep neural network-based generative models for conditional density estimation. They tackle the specific problem of estimating a conditional distribution on a lower dimensional manifold, and find the convergence rate to be dependent on the intrin... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and constructive feedback. While our theoretical contributions are inherently technical, we have made every effort to present them in a manner accessible to a broader audience. Below, we address each of the reviewer’s points in detail:
**Claims... | Summary: This submitted manuscript works on the statistical aspects of conditional generative models, which is an important topic in the broad area of deep generative models. The analysis focuses on the likelihood-based models and, under the assumption that the high dimensional response variables have low-dimensional d... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and constructive feedback, and we appreciate the positive remarks on the main results and the comprehensibility of our proofs.
**Our derivations beyond Theorems 1 and 2** follow a unified strategy controlling two sources of error. First, we bound the stati... | null | null | null | null | null | null | null | null |
Wolfpack Adversarial Attack for Robust Multi-Agent Reinforcement Learning | Accept (poster) | Summary: This paper propose Wolfpack adversarial attack for robust MARL, which targets not only the initial perturbed agent, but also the agents that follow-up to help this perturbed agent. Based on wolfpack attack, the author further propose the wolfpack-adversarial learning for MARL, which learns robust policy based ... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and constructive comments. We greatly appreciate the opportunity to address the main concerns regarding the scope of our method, the perturbation budget, and the theoretical contributions. We provide detailed responses to each of these points be... | Summary: This paper proposes a Wolfpack adversarial attack to train robust MARL systems. Specifically, it modifies the attack process to attack an agent and its follow-up teammates. Experiments on SMAC benchmarks demonstrate the effectiveness of the proposed method.
Claims And Evidence: Yes
Methods And Evaluation Cri... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and constructive comments. We greatly appreciate the opportunity to address the main concerns regarding computational cost and environmental diversity, and we provide detailed responses below.
**Extra computational cost:**
Thank you for pointi... | Summary: This paper introduces a new kind of adversarial attack, Wolfpack, and trains policies against those attacks. They find that trained models are not only more robust against Wolfpack attacks but are also more robust to other adversarial attacks (EGA) and noise injection (Random Attack).
Wolfpack builds upon pre... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and constructive comments. We greatly appreciate the opportunity to address the main concerns regarding general robustness and environmental diversity, and we provide detailed responses below.
**General robustness beyond action perturbations:**... | null | null | null | null | null | null | null | null |
Do Not Mimic My Voice : Speaker Identity Unlearning for Zero-Shot Text-to-Speech | Accept (poster) | Summary: This paper proposes a method to exclude specific speakers from zero-shot text-to-speech to preserve voice privacy. It utilizes a guided learning approach in which some other speakers' voices are used in mask-based learning, especially a teacher-guided learning method in which the target guidance is generated f... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer 87b9 for the thoughtful and constructive comments.
## **1. Human Subjective Evaluation**
We deeply agree with the reviewer on that speaker similarity metric may not perfectly reflect human perception. Speech consists of pitch, prosody, and unique speaker character... | Summary: The paper proposes Teacher-Guided Unlearning (TGU), a novel framework for removing specific speaker identities from Zero-Shot Text-to-Speech (ZS-TTS) models to mitigate privacy and ethical concerns related to unauthorized voice cloning. The method leverages controlled randomness guided by a teacher model durin... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer d6q4 for the constructive feedback. Below we address each concern and clarify the raised points:
## **1. Similar Voice and Robustness**
Please refer to Section 7 in our response to reviewer (bkoi)
## **2. Computational Resources**
Please refer to Section 1 in ou... | Summary: The paper introduces Teacher-Guided Unlearning (TGU), a method to remove specific speaker identities from zero-shot text-to-speech (ZS-TTS) models to address privacy concerns. TGU guides the model to generate random voice styles for "forgotten" speakers while retaining synthesis quality for others. A new metri... | Rebuttal 1:
Rebuttal: We thank the reviewer bkoi for the constructive feedback and valuable insights. We address the primary concerns as follows:
## **1. Practical Use Case**
The 145K steps used during unlearning processes were chosen based on the common fine-tuning practices when adapting Flow-Matching-based pre-trai... | Summary: The paper presents two machine unlearning methods for zero-short TTS, with the main one called teacher-guided unlearning (TGU). This is an under-explored topic/task, which is of interest and importance. It further proposes a new evaluation metric for this task. A set of experiments were conducted to validate t... | Rebuttal 1:
Rebuttal: We deeply appreciate the reviewer k36c's insightful comments and thorough review. Below, we address each comment clearly and concisely.
## **1. Clarification and Typos**
As you correctly indicated, Exact Unlearning is an exact method rather than approximate. We will correct this with other typos... | null | null | null | null | null | null |
ArrayDPS: Unsupervised Blind Speech Separation with a Diffusion Prior | Accept (poster) | Summary: This paper presents a diffusion based method for speech separation using microphone arrays. The core idea is to use a DPS based approach where they alternative between diffusion steps, and an estimation of the reverberation/mixing matrix A. The matrix A is necessary to calculate the conditional likelihood upda... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful and encouraging comments.
>Q: I think the method is very good, but the authors could have done a lot more to strengthen the evaluation. The authors only test with 2 speakers from the WSJ mix, never showing a result with more speakers.
We sincerely apprecia... | Summary: This paper proposes ArrayDPS for multi-channel blind speech separation. The main contributions claimed by the authors are as follows:
1. Proposal of an array-agnostic unsupervised learning-based blind speech separation model that is robust to microphone array variations.
2. Introduction of the first method uti... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed and constructive comments. New results from real-world experiments and a 3-speaker dataset are discussed in the rebuttal to reviewer Go2n. Corresponding demos are updated in **https://arraydps.github.io/ArrayDPSDemo/**.
>Q: GET_SCORE(.) in Algorithm 2 contai... | Summary: The paper introduces ArrayDPS, an unsupervised, generative, and array-agnostic method for blind speech separation (BSS). ArrayDPS leverages a diffusion-based generative model as a speech prior, combined with a novel approximation method for the intractable likelihood arising due to unknown array geometries and... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed and constructive comments.
>Q: The paper could strengthen its evidence by evaluating more diverse real-world scenarios or noisy acoustic conditions to demonstrate true generalizability. Have you considered evaluating ArrayDPS on realistic datasets involving m... | Summary: The paper proposes a method for blind source separation in scenarios where the microphone array and RIR are unknown. The authors adopt a virtual source model, mapping all sources to an imaginary reference microphone.
They formulate the separation problem as an inverse problem, where: (i) a single-source diffus... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed and constructive comments.
> Q: The advantage of using a virtual source model is unclear and questionable.
Yes, we understand this concern now. We had tried to explain this but the explanations are scattered in the paper (lines 147-152 (column 1), line 259 (... | null | null | null | null | null | null |
Machine learning on rigid classes of Euclidean clouds of unordered points | Reject | Summary: The paper aims to identify a set of invariants for point clouds that can uniquely determine the coordinates of all points up to rigid motions. It also introduces a new measure to evaluate the continuity of the proposed invariants.
Claims And Evidence: Overall, the desired properties are demonstrated both theo... | Rebuttal 1:
Rebuttal: Dear Reviewer fCa6
>The paper aims to identify a set of invariants for point clouds that can uniquely determine the coordinates of all points up to rigid motions. It also introduces a new measure to evaluate the continuity of the proposed invariants. the desired properties are demonstrated both t... | Summary: This paper introduces a novel approach for machine learning on Euclidean clouds of unordered points, focusing on achieving complete and bi-continuous invariants under rigid motion. The authors propose a new invariant called Nested Distributed Projection (NDP) and a corresponding Nested Bottleneck Metric (NBM).... | Rebuttal 1:
Rebuttal: Thank you for the detailed review.
>a novel approach for machine learning on Euclidean clouds of unordered points
Yes
>demonstrate the effectiveness of their approach on molecular datasets (QM9, GD), reporting high accuracy
Yes
>claim of "over 98% accuracy" in predicting chemical elements is ... | Summary: This paper focuses on preserving any output, e.g., class labels or predicted property, when the same object is represented differently, and introduces the so-called complete invariants to satisfy all metric axioms. For validating the proposed method, experiments have been conducted on the "world's largest coll... | Rebuttal 1:
Rebuttal: Dear Reviewer Je3e, thank you for the detailed review.
>This paper focuses on preserving any output, e.g., class labels or predicted property, when the same object is represented differently, and introduces the so-called complete invariants to satisfy all metric axioms.
Yes, the distance metric ... | Summary: This work proposes a representation of unordered point clouds that is invariant under rigid motions and complete (point clouds have the same representation if and only if they are related under a rigid motion). First, the authors provide a detailed specification of the properties required from an invariant poi... | Rebuttal 1:
Rebuttal: Dear Reviewer V9mS, thank you for your helpful review.
>This work proposes a representation of unordered point clouds that is invariant under rigid motions and complete(point clouds have the same representation if and only if they are related under a rigid motion).
Thank you for correctly summar... | null | null | null | null | null | null |
Scaling Video-Language Models to 10K Frames via Hierarchical Differential Distillation | Accept (poster) | Summary: Long video processing needs to exploit spatio-temporal redundancies. The core idea of this paper is hierarchical differentiation by key-frame selection and non-keyframe compression leading to processing abiility upto 10K frames on a single GPU and shows competitive performance on long-video benchmarks.
Claims... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. Here is our reply to your main concern:
> Can the authors shed some lights of its possible extension to a streaming framework?
We sincerely appreciate the reviewer’s insightful question. While our proposed DKS and DFM processes significantly reduce token cou... | Summary: This paper presents a novel video token compression strategy named Hierarchical Differential Distillation, designed for long video understanding. Specifically, the hierarchical differential distillation process is achieved through two key mechanisms: Differential Keyframe Selection (DKS) and Differential Featu... | Rebuttal 1:
Rebuttal: Thank you for your time and effort, here are our replies:
> Q1: Do the authors consider comparing LongVU with some training-free token compression methods[1-4]?
Since VisionZip[1], FasterV[2], and FastV[4] are based on the LLaVA-NeXT model, which is optimized for single image scenarios, they can... | Summary: This paper introduces a new long-video LLM framework called ViLaMP to incorporate a differential distillation mechanism to compress the redundant video frame sequence's information into a more compact representation for better trade-off between performance and computational cost. Specifically, this work mainly... | Rebuttal 1:
Rebuttal: > Although this method's effectiveness has been verified on video question answering benchmarks for multimodal large language models, it is still unclear whether this method can truly preserve necessary information for more temporal-sensitive tasks like action recognition, video grounding or tempo... | null | null | null | null | null | null | null | null |
ATA: Adaptive Task Allocation for Efficient Resource Management in Distributed Machine Learning | Accept (poster) | Summary: In distributed machine learning, traditional greedy methods waste resources by assigning tasks to every worker, even when only a few results are needed. This paper tells a story of efficiency: it introduces ATA, a smart task allocator that learns the speed of each worker on the fly. Instead of needing prior kn... | Rebuttal 1:
Rebuttal: Thank you for the review
> they lack real-world experiments on real-world benchmarks such as MLPerf.'}
There might be a misunderstanding on the nature of our contribution: We propose an allocation strategy across workers that adapts to their computation times and that is *agnostic* to the specif... | Summary: This paper proposes a new task allocation algorithm for efficiently managing computation resources in distributed SGD.
Claims And Evidence: Yes.
Methods And Evaluation Criteria: Yes.
Theoretical Claims: I have read the theoretical claims in the main part. The proofs in the appendix are not checked.
Experim... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their valuable feedback.
> The paper only shows the simulation results, but there is no real dataset experiment (e.g. distributed deep learning training allocation) to show the performance of their approach.
There might be a misunderstanding on the natur... | Summary: This paper suggests an adaptive method for task allocation, Adaptive Task Allocation (ATA). In the context of machine learning, this approach addresses federated learning, where in each round, $n$ workers collaboratively perform minibatch SGD with total batch size of $B$. In contrast to the standard approach, ... | Rebuttal 1:
Rebuttal: Thank you for the review
> The paper provides theoretical guarantees in the framework of online convex bandit
Since the action set is discrete, the proposed reduction does not fall under the framework of online convex bandits but rather corresponds to a combinatorial bandit problem.
> however, ... | Summary: This paper introduces ATA (Adaptive Task Allocation), a method that dynamically adapts to heterogeneous and random worker computation times without prior knowledge of their distributions. Theoretical analysis shows that ATA achieves optimal task allocation, performing as well as methods that have prior knowled... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their valuable feedback.
> missing ablation study (e.g., impact of data-dependent concentration inequality)
We are not sure what the reviewer means here: We did study the impact of the data-dependent concentration inequality. Indeed, we test both ATA (the ... | null | null | null | null | null | null |
Navigating the Social Welfare Frontier: Portfolios for Multi-objective Reinforcement Learning | Accept (poster) | Summary: This paper proposes a method to create a portfolio of multi-objective reinforcement learning policies that aim to preserve the social welfare among agents; in other words, a portfolio that ensures that different stakeholders fulfill their individual preferences. The presented method (*p-MeanPortfolio*) revolve... | Rebuttal 1:
Rebuttal: We thank you for your incredibly detailed and insightful comments! Additional results are available in the anonymous link [here](https://anonymous.4open.science/r/approximation-portfolios-for-rl-ICML)
**1.** *On the overstimation of oracle complexity in Theorem 4.1, and how this impacts our curre... | Summary: The paper proposes algorithms for computing a portfolio of policies for multi-objective MDPs. The individual reward functions are scalarized by an aggregation function called p-means, which is dependent on a parameter $p \in [-\infty,1]$. The authors propose an algorithm, which generates an $\alpha$-approxima... | Rebuttal 1:
Rebuttal: We thank you for the insightful comments! Additional results are available in the anonymous link [here](https://anonymous.4open.science/r/approximation-portfolios-for-rl-ICML)
**1.** *The selection of baselines seems somewhat unfair ... much more Oracle calls.*
- Thank you for your comment. How... | Summary: The paper investigates an interesting problem of learning fair policies in multi-objective reinforcement learning (MORL) by proposing an $\alpha$-approximate portfolio which is a finite set of policies that are approximately optimal across the family of generalized p-means social welfare functions (e.g., Nash,... | Rebuttal 1:
Rebuttal: We thank you for the insightful comments! Additional results are available in the anonymous link [here](https://anonymous.4open.science/r/approximation-portfolios-for-rl-ICML)
**1.** *Including one or more realistic environment with higher number of objectives.*
- We respectfully note that our e... | Summary: In this paper, the authors proposed a novel setting in optimizing social welfare while considering different preferences. The authors proposed an algorithm which can trade off among approximation factor, portfolio size, and computation efficiency. The proposed p-MEANPORTFOLIO algorithm chooses several p-value,... | Rebuttal 1:
Rebuttal: We thank you for the insightful comments! Additional results are available in the anonymous link [here](https://anonymous.4open.science/r/approximation-portfolios-for-rl-ICML)
**1.** *Why there is a need to find $\alpha$-approximation for p-value. Why not use previous MORL algorithms?*
- We than... | null | null | null | null | null | null |
Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models | Accept (poster) | Summary: This paper introduces a way to decide the optimal proportion of different domain datasets for LLM fine-tuning. The Scaling law is utilized to learn the optimal proportions, improving fine-tuning performance in various settings.
Claims And Evidence: Yes.
Methods And Evaluation Criteria: Yes.
Theoretical Clai... | Rebuttal 1:
Rebuttal: Thank you so much for your feedback! We will discuss your comments as follows.
---
Q: Relatively marginal improvement of "Ours" compared to "Original"
A: We have noticed this issue and, over the past month, we conducted additional experiments to understand it. Our responses are as follows: In o... | Summary: This paper addresses the problem of data mixing for supervised fine-tuning (SFT) of large language models (LLMs). The authors frame data mixing as an optimization problem and introduce a method to minimize validation loss. Their approach involves parameterizing the loss by modeling effective data transferred a... | Rebuttal 1:
Rebuttal: Thank you! We will address your comments as follows.
---
Q: The generalization of the method to different data domains (excluding IF, Math and Code) could be explored further.
A: In Section 5.2, we conducted experiments beyond the three domains (IF, Math, and Code), exploring additional domains... | Summary: This paper proposes a data mixing optimization problem for supervised fine-tuning of LLMs. The problem is based on a scaling law determined by neural scaling laws for SFT. The paper proves convexity properties about the optimization problem and applies it to optimize mixtures for several finetuning settings.
... | Rebuttal 1:
Rebuttal: Thank you! We will address your comments as follows.
---
Q1 : Improvements appear marginal
A1 : We have noticed this issue and, over the past month, we conducted additional experiments to understand it. Our responses are as follows: In our paper, we mentioned using repetitive sampling for domain... | Summary: This paper proposes an algorithm to optimize proportions of data sources for the supervised fine-tuning (SFT) stage in LLM training by minimizing validation loss. Given that jointly optimizing the mixing weights and model parameters is expensive, the paper makes an approximation. It makes uses of data scaling ... | Rebuttal 1:
Rebuttal: Thank you! We will address your comments as follows.
---
Q: Validation loss approximation in MSE
A: The MSE of the validation loss estimation in the experiments from Section 5.1 is presented in Table 0.1.1 (https://anonymous.4open.science/r/Data_Mixture-5D78/src/Table0.1.1.png). Additionally, t... | null | null | null | null | null | null |
Attention Mechanisms Perspective: Exploring LLM Processing of Graph-Structured Data | Accept (poster) | Summary: As an experimental analysis work, the article examines the limitations of LLMs in processing graph-structured data from the perspective of attention mechanisms. The experimental design is relatively comprehensive, covering changes in attention distribution, structural information interference, comparisons of d... | Rebuttal 1:
Rebuttal: >the bimodal trend mentioned in the text focuses on mean differences with t-tests while KS tests emphasize differences between overall distributions ("Given that the KS test focuses on the overall distribution and the t-test emphasizes the mean value:"). It would be better if relevant materials we... | Summary: This article explores the attention mechanisms of Large Language Models (LLMs) when processing graph-structured data. It first observes that while LLMs are aware of graph-structured data, they fail to properly utilize link information. The analysis and explanation focus on two broad aspects: attention distribu... | Rebuttal 1:
Rebuttal: >In some of the experimental demonstrations (Figure 1,2,3,5), three datasetswere selected and showed different forms of performance. Could you comprehensively explain how these different datasets performed across multiple experiments?
We are happy to address your question. In our experimental dem... | Summary: The paper empirically studies the behavior of the attention mechanism on graph inputs. First, they show that the attention distribution changes after finetuning, therefore concluding that LLMs can recognize graphs. Second, they show that altering the connectivity does not affect significantly the performance, ... | Rebuttal 1:
Rebuttal: > Please discuss the prompts used and their impact on your findings. I see that Appendix J presents the prompt, but does not analyze the impact of changing it, especially because the assistant part can be significantly modified.
Thank you for your suggestions regarding the experimental section of... | Summary: This paper does an empirical investigation of attention patterns when an LLM is applied to graph problems. They find that their treatment of graphs in LLMs is not very sensitive to graph connectivity information. Some improvement is gained with attention masking.
Claims And Evidence: One fundamental problem... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. However, there seems to be some misunderstanding regarding our study's objectives. The content you claimed is missing is mentioned in the appendices. Below are our responses to clarify the points raised.
>What does "node sampling configured as 8x8" mean?
No... | null | null | null | null | null | null |
Optimal Task Order for Continual Learning of Multiple Tasks | Accept (poster) | Summary: This paper analyzed the impact of task order and task similarity on continual learning. With a linear teacher-student model, the authors derived an analytical expression that connects task order and task similarity to the learning performance. The theoretical results further derive two principles (the peripher... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. Please find our replies below.
> However, I'm not sure the practical benefits of using these insights to improve continual learning
To illustrate the practical applicability of our work, let us first note that, our method requires only a small number of data ... | Summary: This paper investigates the task order problem in continual learning. Specifically, it first establishes a theoretical connection between task order and final performance, which is then empirically validated through toy examples. Additionally, the paper examines the impact of task order from two key perspectiv... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments. Please find our replies below (we quoted your comments only partially due to space constraint).
> Does this imply that we have access to all tasks when adjusting task order?
While we need some information on all tasks, we do not need to have access to all da... | Summary: This paper investigates the impact of task order on learning performance in continual learning. It first mathematically models the correlations between tasks and then derives two rules to enhance learning effectiveness: (1) learning non-representative tasks before typical ones, referred to as the *periphery-to... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments. Please find our replies below.
> it lacks a clear table summarizing the performance impact of applying or not applying the two proposed rules.
The table below summarizes the results from Figs. 5a-f.
| | Fashion MNIST| CIFAR-10 | CIFAR-100 |
|:-... | Summary: The paper considers task orders in continual learning. It formulates a linear teacher-student model and derives an explicit expression for the errors in the asymptotic case. Then the paper analyzes how the change of task order would affect the error.
Claims And Evidence: Some claims made in the submission are... | Rebuttal 1:
Rebuttal: Thank you for your detailed comments on our work. Please find the replies below (we quoted your comments only partially due to space constraint).
**Claims And Evidence**
> For example, the title is...
Please note that, both in the abstract and main text, we noted that these principles are base... | null | null | null | null | null | null |
Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting | Accept (spotlight poster) | Summary: This paper proposes a sample weighting scheme for fine-tuning data based solely on the loss of the pre-trained model.
The proposed method weights samples with low loss of the pre-trained model (i.e., easy samples) to suppress the divergence from the pre-trained model.
Existing methods operate in the parameter ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed evaluation and encouraging feedback. We respond to your questions and feedback below:
**Performance for larger/different models and larger datasets:** We believe that our language experiments are indeed large-scale, and the size of our vision datasets is c... | Summary: This paper examines how we can mitigate catastrophic forgetting during fine-tuning. They propose that prioritizing "easy" examples (i.e. those that the base model already has low loss on). Concretely, this allows mitigating forgetting without significant access to the pretraining dataset -- as would be require... | Rebuttal 1:
Rebuttal: We thank the reviewer for recognizing the novelties of the simple design of our sample-wise weighting scheme. We appreciate the constructive and encouraging feedback. Please find our responses to your major comments below.
**Wider range of fine-tuning tasks, especially for language modeling:** Wh... | Summary: The work relates to the issue of catastrophic forgetting in finetuned models wherein the pre-trained knowledge of the model is wiped out substantially after finetuning, and proposes a method that up-weighs the samples in the finetuning dataset that incur a low loss value (in contrast with existing approaches),... | Rebuttal 1:
Rebuttal: Thanks for the positive assessment of our work! We address your questions below.
**Ablation for $\tau$:** In the table below, we show the pre-training and fine-tuning accuracies as a function of $\tau$ when the model is ResNet-50, the pre-training dataset is ImageNet-1K, and the fine-tuning data... | Summary: The paper contributes the following:
- a weighting strategy (called FLOW) for finetuning examples in the data-oblivious setting where there is no access to pre-training examples, that upweights examples with small loss, and aimed at mitigating forgetting of the pretraining task.
- a benchmark against other s... | Rebuttal 1:
Rebuttal: Thanks for the positive assessment of our work! We address your questions below.
**Number of fine-tuning epochs:** For the language experiments, this was chosen based on existing literature recommendations for fine-tuning epochs. Following the observation of [1], we trained for 2 epochs due to c... | null | null | null | null | null | null |
Revisiting Differentially Private Algorithms for Decentralized Online Learning | Accept (poster) | Summary: Two types of algorithms for differentially private online learning (PD-FTGL and PD-OCG) and their theoretical regret bounds are presented. While the conventional D-FTGL (Wan et al., 2024) employs a standard gossip step, the proposed PD-FTGL effectively exploits the accelerated gossip step (Liu & Morse, 2011), ... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews!
---
Q1: No experiments were conducted … demonstrating how the theoretical regret bound improvements manifest empirically ... would have made the paper more compelling. This absence has been a factor in my relatively lower score.
A1: We agree with the si... | Summary: This paper explores the problem of decentralized online learning in a private setting. Previous works fail to achieve $(0, \delta)$-DP over $T$ iterations and do not attain an optimal regret bound comparable to the non-private setting. This paper introduces a blocking update mechanism, an accelerated gossip st... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We hope the reviewer could reevaluate our paper, and are very happy to respond more questions during the reviewer-author discussion period.
---
Q1: No experimental results ...
A1: Please check **our response to Q1 of Reviewer J5zj**, which presents exper... | Summary: This paper proposes differentially private decentralized online learning algorithms that achieve regret bounds nearly matching those in the non-private setting. The algorithms can be adapted to be projection-free using Frank-Wolfe methods, resulting in a degradation of the regret bounds. However, these bounds ... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews!
---
Q1: ... Frank-Wolfe methods on KK can be just as challenging to solve as the projection operator ... it is inaccurate to claim that ... is a projection-free algorithm.
A1: There may be some misunderstandings about our projection-free algorithm. We ag... | Summary: This paper revists the regret bound for differentially private onilne learning. Based on the block-decoupled accelerated gossip strategy, the authors achieved a refined uppper bound $O(n\rho^{1/4}/\epsilon\sqrt{T})$ under $(\epsilon,0)$-DP. Theoretical results are derived under certain assumptions.
Claims And... | Rebuttal 1:
Rebuttal: Many thanks for the constructive reviews! We hope the reviewer could reevaluate our paper, and are very happy to respond more questions during the reviewer-author discussion period.
---
Q1: ... the theoretical results rely on restrictive assumptions ... an empirical evaluation against existing me... | null | null | null | null | null | null |
DLP: Dynamic Layerwise Pruning in Large Language Models | Accept (poster) | Summary: The paper relates to the weight pruning of LLMs and introduces a method to prune individual layers non-homogeneously, i.e. instead of applying a fixed sparsity ratio across all layers, every layer is sparsified according to its importance.
The method extends, and is compatible with previous importance estimat... | Rebuttal 1:
Rebuttal: Dear Reviewer zWSE:
We sincerely appreciate your positive assessment of DLP's non-parametric layer importance scoring and the rigor of our experimental design. We are also grateful for your acknowledgment of the paper’s detailed literature review and its clear presentation, which we hope will fa... | Summary: The paper focused on finding important layers to perform non-uniform pruning a.k.a to prune each layer in different rates. More important layers are pruned less aggressively, less important weights pruned more. The paper can be easily extended to any sparsification approach with just determining the pruning r... | Rebuttal 1:
Rebuttal: Dear Reviewer EVqz:
We sincerely appreciate your time and thoughtful review of our work. Please find our detailed, point-by-point responses below.
>Q1: Clarify LOD metric and its significance (Line 253) with a formula and explanation.
Thank you for your valuable comment. Followed by OWL, weight... | Summary: This paper aims to improve LLM pruning by introducing a new layerwise sparsity ratio inspired by OWL. Unlike OWL, which assigns layerwise sparsity based on predefined threshold of outlier ratio, DLP measures layerwise importance based on the median of outlier ratios. This simple modification not only avoids h... | Rebuttal 1:
Rebuttal: Dear Reviewer 9Taq:
We sincerely appreciate your recognition of DLP’s median-based importance scoring and its extensive validation across multiple compression paradigms (pruning, quantization, etc.). Below, we address your comments point by point:
> Q1: The paper presentation can be significantly... | Summary: The paper proposes a new method for pruning LLMs by determining the importance of each layer for determining the layer-wise sparsity ratio. The proposed method, DLP, adaptively determines the importance of each layer by combining model weights with input activation information. DLP uses a median to detect outl... | Rebuttal 1:
Rebuttal: Dear Reviewer AtTG:
Thank you for taking the time to read and review our paper! In the following, we would like to address your comments point by point.
>Q1: Limited technical novelty and direct extension using median-based outlier filtering.
Thank you for your insightful comments. We would li... | null | null | null | null | null | null |
SAFE: Finding Sparse and Flat Minima to Improve Pruning | Accept (spotlight poster) | Summary: This work tried to learn sparse and flat minima during training, so such sparse and flat minima can be more suitable for network pruning. The idea is simple but effective. This work formulates pruning as a sparsity-constrained optimization problem where flatness is encouraged as an objective. This optimization... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer’s recognition and constructive feedback. We have addressed the reviewer’s specific comments below, while we remain open to any additional suggestions.
---
**Convergence for nonconvex case?**
We appreciate the reviewer’s insight. We believe that the converge... | Summary: This paper introduces SAFE and SAFE+ algorithms for finding sparse subnetworks with flat loss landscapes. The methods utilize an augmented Lagrange dual approach, with SAFE+ extending the base algorithm through a generalized projection operation. The authors evaluate their approaches on image classification an... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer for taking the time to review our work and providing us with constructive feedback. While we respond to the reviewer’s specific comments below, we would be keen to engage in any further discussion.
---
**No sparsity ratio in our convergence analysis?**
We ap... | Summary: The authors tackle the problem of performance degradation in sparse neural networks at very high sparsity levels due to reduced capacity. The proposed method is motivated by the recent discoveries connecting the flatness of the solution (loss) landscape and the improved generalization performance of neural net... | Rebuttal 1:
Rebuttal: We really appreciate the reviewer’s insightful comments and constructive feedback. While we address specific comments below, we would be keen to engage in any further discussions.
---
**Extended sharpness measurements**
Thank you for the insightful suggestion. We extend our sharpness analysis b... | Summary: This paper presents an algorithm to produce networks that are both flat and sparse. Recent work has demonstrated that flat networks have better generalization and have proposed some approaches for producing flat and sparse networks. This paper introduces a theoretically principled approach for producing sparse... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for finding our work interesting and giving us constructive suggestions. We make clarifications to specific comments below.
---
**Robustness to different types of noisy data**
Thank you for the insightful suggestion. To test the robustness of SAFE against image c... | null | null | null | null | null | null |
Beyond Self-Repellent Kernels: History-Driven Target Towards Efficient Nonlinear MCMC on General Graphs | Accept (oral) | Summary: The paper introduces a new "history-driven target" (HDT) framework to improve Markov Chain Monte Carlo (MCMC) sampling on graphs. HDT modifies the target distribution based on the history of visited states, improving efficiency. This helps overcome the computational limitations of the recently-proposed class o... | Rebuttal 1:
Rebuttal: ## Q1. Downstream benefits in ML algorithms
Our HDT framework enables more efficient sampling of large or complex graphs for arbitrary target measures, reducing estimator variance and cost. Key impacts include:
1. **Estimating Global Properties of Large-Scale Graphs:** In social networks, e-commer... | Summary: This paper introduces the History-Driven Target (HDT) framework, a novel approach to improving Markov Chain Monte Carlo (MCMC) sampling on discrete state spaces, such as general undirected graphs. The HDT framework replaces the original target distribution with a history-dependent target distribution $ \boldsy... | Rebuttal 1:
Rebuttal: Thank you for your constructive and practical feedback. We agree that demonstrating robustness to initialization and quantifying uncertainty in our results will enhance the paper. Below is the setting of additional experiments. Due to space constraints, we defer the simulation results on more real... | Summary: - Propose a history-driven target (HDT) framework, extending the self-repellent random walk (SRRW) framework on general graphs [Doshi et al., 2023] to allow for construction from a non-reversible sampler, whereas the original framework in [Doshi et al., 2023] required a reversible sampler.
- On the practical ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and detailed review. We appreciate your positive remarks on our paper’s theoretical depth and are grateful for the questions you raised. Below, we address each of your points in turn.
---
## Response to Q1
We agree that this point warrants further explanation. Here... | null | null | null | null | null | null | null | null |
Emergence and Effectiveness of Task Vectors in In-Context Learning: An Encoder Decoder Perspective | Accept (spotlight poster) | Summary: This work studies the mechanisms underlying the emergence of in-context learning.
The authors show that, on a synthetic dataset composed of a mixture of sparse linear regression tasks, (a) the first layers of a transformer progressively learn to map each task into a separable latent space and (b) the followi... | Rebuttal 1:
Rebuttal: Thank you for your detailed and thoughtful review. We greatly appreciate that you find “the authors make great work at demonstrating the emergence of the task encoding and task decoding mechanisms” and that you consider “the synthetic experiments … and the real LLM experiments to be well-designed ... | Summary: This paper studies the task vector behavior in the trained-from-scratch and pretrained transformer, demonstrating the following conclusions:
1. When the transformer develops the task encoding (that generates the task vector), it simultaneously generates the corresponding decoding functions.
2. In pretrained LL... | Rebuttal 1:
Rebuttal: We are grateful for your careful review. You highlight a key point that “the different encoding progress for different tasks is really interesting,” which resonates strongly with our primary objective of illustrating how task encodings develop and shape downstream ICL performance.
Below, we addre... | Summary: This paper investigates the learning mechanisms of autoregressive Transformers in In-Context Learning (ICL), with a particular focus on the emergence of task vectors and their impact on ICL task performance. Building on the Bayesian view of ICL, the authors propose an encoder-decoder analytical framework and e... | Rebuttal 1:
Rebuttal: Thank you for your positive and thoughtful review. We appreciate that you see our “theoretical logic as comprehensive, extending from experimental observations in regression tasks to natural language experiments,” and that you find our encoder-decoder framework provides “new theoretical foundation... | Summary: This paper studies how transformers learn task vectors for various tasks during pretraining. Task vectors refer to the intermediate representations from a middle layer of a transformer network, given an in-context learning (ICL) task. The paper investigates how these task encodings separate and cluster based o... | Rebuttal 1:
Rebuttal: Thank you for your positive and thoughtful review. We appreciate your recognition of Task Decodability as “a compelling new measure” and that our activation patching analysis “provides convincing evidence [of transformers’] encoder-decoder structure.” We also value your suggestion to generalize to... | null | null | null | null | null | null |
Complete-Tree Space Favors Data-Efficient Link Prediction | Accept (poster) | Summary: This paper studies the link prediction task under the scarce data scenario. In the real-world application, the observed links in the network are far fewer compared to the unobserved ones. However, current LP methods usually focus on scenarios where observed links are abundant. To solve the problem, this study ... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer for spending valuable time reviewing our manuscript and providing insightful comments. We have improved our paper accordingly and our responses are as below.
Q1: **Hierarchical Modularity Detectability**: ‘the claim that "hierarchically modular structure can s... | Summary: This paper addresses the understudied problem of link prediction when only sparse links are available. It proposes studying the complete tree space (CT), which offers a significantly lower bound on sample complexity compared to the commonly used Euclidean space. Building on this, the authors introduce leaf mat... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer for spending valuable time reviewing our manuscript and providing insightful comments. We have improved our paper accordingly and our responses are as below.
Q1: **Practical Use Case**: ‘it would be beneficial to include practical use cases illustrating scenar... | Summary: This paper proposes a Complete-Tree based approach to explore the hierarchical structure of graphs such that the sample complexity for link prediction can be improved. The key idea is to develop a (hierarchical) complete tree to model the node distance in graphs, and utilize leaf matching to embed nodes into t... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer’s time and insightful comments. Our justifications and improvements are as below.
W1: **Hierarchical Evidence**: **Numerous existing studies have empirically verified** hierarchical modularity in many important real-world networks: metabolic [1], customer purc... | Summary: This paper explores the challenge of link prediction in data-scarce networks by introducing the CT space, a discrete metric space that formalizes hierarchical modularity in networks. The authors leverage group theory to prove that the CT space provides a lower bound on sample complexity compared to Euclidean s... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer for spending valuable time reviewing our manuscript and providing insightful comments. We have improved our paper accordingly and our responses are as below.
Q1: **Scalability** ‘…does not provide detailed runtime comparison baselines; explanations and analyse... | null | null | null | null | null | null |
OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable? | Accept (poster) | Summary: The paper introduces OOD-Chameleon, a framework for automatically selecting training algorithms to improve out-of-distribution (OOD) generalization. Instead of trial-and-error, it learns to predict the best algorithm based on dataset characteristics. The authors create a meta-dataset by re-sampling datasets wi... | Rebuttal 1:
Rebuttal: Many thanks for the thoughtful review. We are happy that you found our work novel and the analysis insightful.
---
**`Comment 1 (Theoretical analysis to explain the success of OOD-Chameleon):`** The bottom line is that our meta-learning-like approach turns algorithm selection into a *supervised*... | Summary: This paper presents _OOD-Chameleon_, a principled and bottom-up approach for selecting training algorithms in out-of-distribution (OOD) generalization tasks. By framing algorithm selection as a multi-label classification problem, the method leverages a meta-dataset of datasets representing diverse distribution... | Rebuttal 1:
Rebuttal: Thanks for the constructive feedback. We are glad that you found our work timely, valuable, and appreciated its clarity and insights.
---
**`W1 (Would a predictor trained on a meta-dataset of small datasets generalize?):`** We performed additional experiments, reported below. The algorithm selec... | Summary: This paper introduces an interesting task: choosing the right training algorithm for the right dataset. The authors propose an approach called OOD-CHAMELEON, which predicts whether a given algorithm can generalize to unseen OOD data based on its past performance on the current dataset. I find this work suffici... | Rebuttal 1:
Rebuttal: Thanks for the thoughtful feedback. We are glad that you found our work novel and well presented.
---
**`Q1 (algorithm selection at test time):`** We simply pick the method with the *highest logit* (L159 right column), which corresponds to the one the model is the most confident in.
An alternat... | Summary: This paper tackles the interesting problem of algorithm selection for OOD generalization. In particular, the paper proposes a meta-learning-like algorithm which takes in summary statistics from a dataset as well as a distributionally robust learning algorithm. This tuple is then mapped to the performance of tr... | Rebuttal 1:
Rebuttal: Thanks for the insightful review. We are glad that you found the studied problem and insights interesting. We address the questions below and propose several improvements to the paper.
---
**`W1: No guarantee beyond the support of the training distribution:`**
Correct, this is true of any learni... | null | null | null | null | null | null |
Overcoming Spurious Solutions in Semi-Dual Neural Optimal Transport: A Smoothing Approach for Learning the Optimal Transport Plan | Accept (poster) | Summary: Semidual Neural OT (SNOT) is one of the NN-based formulation for computing transport maps for general OT problems. The paper seeks to address the issue of spurious solutions for SNOT formulation, a problem that has been known before. For example, when the source distribution places mass on low-dimensional sets... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for carefully reading our manuscript and providing valuable feedback. We are especially grateful for the reviewer's recognition that our work is the **first to identify a reasonable sufficient condition for the avoidance of spurious solutions** for the minimax formu... | Summary: The manuscript presents a minimax algorithm used to find solutions to the semi-dual optimal transport plan problem. A primary drawback of this method is that not all solutions of the semi-dual formulation yield optimal transport maps and optimal potential functions; in other words, the formulation can lead to ... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for carefully reading our manuscript and providing valuable feedback. We greatly appreciate the reviewer's acknowledgment that our method is **theoretically founded**. Moreover, we’re pleased to hear that the reviewer **appreciated reading the examples, as they prov... | Summary: This paper investigates convergence issues that arise when learning Optimal Transport (OT) maps using the Semi-Dual formulation. Specifically, the authors:
- Identify a sufficient condition under which the standard Semi-dual Neural OT approach converges to the correct OT map, resolving issues with inaccurate (... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for carefully reading our manuscript and providing valuable feedback. We greatly appreciate the reviewer's acknowledgment that "this paper presents a nice contribution that would be relevant for the conference". We hope our responses to be helpful in addressing the ... | null | null | null | null | null | null | null | null |
Enhancing Graph Invariant Learning from a Negative Inference Perspective | Accept (poster) | Summary: The paper presents NeGo, a negative inference graph OOD architecture designed to address complex environmental shifts in OOD scenarios. Extensive experiments demonstrate the effectiveness of proposed architecture.
Claims And Evidence: Yes, the Yes, this work brings a new perspective and method to graph OOD le... | Rebuttal 1:
Rebuttal: Dear Reviewer vzxf,
Thank you so much for taking your valuable time to review our manuscript. We greatly appreciate the positive feedback on our work. Here we will carefully address your concerns.
**W1&Q1. Intuitive analysis**
Prompt learning and negative learning collaboratively contribute to ... | Summary: This paper introduces a negative inference graph OOD framework (NeGo) to broaden the inference space for environment factors. Technically, this paper is inspired by the concept of prompt learning and, from the perspective of negative learning, propose an innovative approach called negative prompt. Experimental... | Rebuttal 1:
Rebuttal: Dear Reviewer wCN4,
Thank you for taking valuable time to review this work. We have carefully considered your comments, and the following are our detailed responses. We hope these details could address your concerns.
**W1. Parameterized networks**
${f_\phi }$ represents the negative prompter to... | Summary: In this work, the authors tackle the complex environment shift challenge in graph learning. Through theoretical analysis and an in-depth discussion of the challenge, the authors propose the negative prompt graph learning framework NeGo. Extensive experiments on real-world datasets across domains and synthetic ... | Rebuttal 1:
Rebuttal: Dear Reviewer sxw9,
Thank you for taking valuable time to review our work. We have carefully considered your comments, and the following are our detailed responses. We hope these details could address your concerns.
**W1. The environment in the graph data.**
The environment factors in graph da... | Summary: Through the experimental exploration, this paper observe that the existing approaches lack the ability to decouple causal subgraphs from complex environments. To tackle this problem, a negative prompt environment inference framework is proposed. The theoretical analysis, along with the excellent experimental p... | Rebuttal 1:
Rebuttal: Dear Reviewer cMu6,
We appreciate your time on reviewing our work. We will provide further clarification to address your concerns.
**W1 and Q1. Complex environments.**
Our proposed NeGo remains effective in other complex environment scenarios, not limited to the size-dependent environment cas... | null | null | null | null | null | null |
Federated Disentangled Tuning with Textual Prior Decoupling and Visual Dynamic Adaptation | Accept (poster) | Summary: This paper proposes FedDDA, a federated framework designed to enhance the adaptation of Vision-Language Models under data heterogeneity. Toward existing works, the authors highlight two key limitations in both textual and visual modality: Textual Property Loss and Visual Feature Diversity. To address these cha... | Rebuttal 1:
Rebuttal: Dear Reviewer UBUx:
Thank you for your positive response! We value the constructive feedback and have carefully considered the concerns raised. Below we provide detailed responses to address the reviewer's concerns:
**Weakness & Question**
**W1 & Q1: The rationale for employing dual adapters sh... | Summary: The paper proposes a novel FedDDA approach to improve federated learning for fine-tuning Vision-Language Models (VLMs) under data heterogeneity. The focus is on addressing two primary challenges in both language and vision modality. On the textual side, the method integrates two prompts (global and local) in a... | Rebuttal 1:
Rebuttal: Dear Reviewer 287K:
We sincerely appreciate your constructive and insightful comments. We will explain your concerns point by point.
**Weakness**
**W1: Strange decoupling.**
**A1:** A key challenge in decoupling prompts lies in achieving efficient collaboration between prompts with different ... | Summary: This paper addresses the challenge of data heterogeneity in Federated Parameter-Efficient Fine-Tuning. To tackle this issue, the authors propose FedDDA, which is designed from two modalities. In the text modality, the proposed approach decouples prompts into global and local components, which are connected thr... | Rebuttal 1:
Rebuttal: Dear Reviewer oran:
Thank you very much for your recognition of our work and for raising thoughtful questions and concerns about our work. We have carefully considered each comment and provided responses.
**Weakness**
**W1: Analysis of t-SNE visualization.**
**A1:** Thank you for your insight... | Summary: This paper considers the collaborative effectiveness caused by data heterogeneity during Federated Parameter-Efficient Fine-Tuning process. A simple yet effective algorithm is proposed to address the above challenges via Textual Prior Decoupling and Visual Dynamic Adaptation. In this case, both modalities can ... | Rebuttal 1:
Rebuttal: Dear Reviewer Zexp:
Thanks for raising thoughtful questions and concerns. We sincerely hope this point-to-point response allows the reviewer to update the score. We will add a more detailed comparison or analysis of the existing decoupled works in the final version.
**Weakness**
**W1: Analysis ... | null | null | null | null | null | null |
Density Ratio Estimation-based Bayesian Optimization with Semi-Supervised Learning | Accept (poster) | Summary: In this article, the authors consider using a classifier to perform Bayesian optimization rather than regression models. To avoid issues with overconfident predictors, the proposition is to rely on unobserved inputs (unlabeled data), propagate the labels to this set before selecting the next query point among ... | Rebuttal 1:
Rebuttal: We thank the reviewer for your constructive feedback to improve our work. We answer your questions and concerns in the following.
> Given observation values, the regression problem is supposedly simpler than a classification (skipped)
We think that this question is regarding the general advantag... | Summary: This paper observes the overconfidence, or to say, over-exploit behavior of supervised learning methods in DRE-BO. With unlabeled points available, semi-supervised learning is combined with DRE-BO to tackle the issue. The proposed method, DRE-BO-SSL, are empirically evaluated at two cases including unlabeled p... | Rebuttal 1:
Rebuttal: We thank the reviewer for your constructive feedback to improve our work. We answer your questions and concerns in the following.
> I am sort of confused about the relationship between experiments and the motivation. DRE-BO-SSL is motivated by the overconfidence of DRE-BO-SL, giving the impressio... | Summary: This paper proposes an algorithm for black-box optimisation which is inspired by Bayesian Optimisation by density-Ratio Estimation (BORE, Tiao et al., 2021), a BO algorithm that transforms the usual BO regression problem into classification, and is combined with semi-supervised learning to better balance the e... | Rebuttal 1:
Rebuttal: We thank the reviewer for your constructive feedback to improve our work. We answer your questions and concerns in the following.
> One my main concerns with this paper in its current form is that, to me, it seems that semi-supervised learning would only make sense if the data had a meaningful un... | null | null | null | null | null | null | null | null |
Widening the Network Mitigates the Impact of Data Heterogeneity on FedAvg | Accept (poster) | Summary: This paper provides an NTK based analysis of FedAvg for training deep neural networks under the MSE loss, showing that the effect of heterogeneity disappears in the infinite-width limit. Essentially, the model divergence (distance between local model parameters) is inversely proportional to the square root of ... | Rebuttal 1:
Rebuttal: Thanks for your constructive comments. We will clarify your questions and concerns below.
**1. Non-rigorous statement:** We will revise Assumption 1 and Theorem 3 as:
> **Assumption 1.** The terms of order $\mathcal{O}(n^{-1})$ and higher are neglected for $n \to \infty$.
> **Theorem 3.** Unde... | Summary: The paper provides theoretical analysis showcasing that increased neural network width provably reduces the effects of data heterogeneity, and approaches the convergence of centralized gradient descent. However, it is also proven that data heterogeneity slows convergence.
Claims And Evidence: The theory clear... | Rebuttal 1:
Rebuttal: Thanks for your insightful comments. We will clarify your questions and concerns below.
**1. Novelty compared with Lee (2019) and Shi (2024):** The main novelty of our work compared to Lee (2019) and Shi (2024) lies in addressing *data heterogeneity*, which was not considered in those earlier an... | Summary: This paper mainly studies that in the federated learning scenario, as the width of the neural network continuously increases, the negative impact of non-IID data on the FedAvg algorithm will gradually weaken and finally disappear completely when the width is infinite. Based on the theory of neural tangent kern... | Rebuttal 1:
Rebuttal: Thanks for your constructive comments. We will clarify your questions and concerns below.
**1. References regarding the convergence behavior of multi-layer NTK in real-world applications**:
After carefully reviewing the existing literature again, we were unable to find further studies regarding... | Summary: This paper analyzes the performance of FedAvg in infinite-width fully connected neural networks trained with gradient descent. It provides theoretical results on both convergence and generalization, revealing new insights—such as the diminishing impact of data heterogeneity as network width increases.
Claims ... | Rebuttal 1:
Rebuttal: Thank you for acknowledging our contributions. Our response to your question is provided below.
**What are the main technical challenges in extending the NTK analysis from centralized learning (e.g., [Jacot et al., 2018]) to FedAvg?**
The main technical challenges in extending NTK analysis from ... | null | null | null | null | null | null |
Gap-Dependent Bounds for Federated $Q$-Learning | Accept (poster) | Summary: This work studies instance-dependent regret guarantees in tabular Markov Decision Processes within a federated learning framework, where multiple agents collaborate to learn the environment. Specifically, the author focuses on the minimal sub-optimality gap structure and provides the first logarithmic regret g... | Rebuttal 1:
Rebuttal: We thank you for your careful reading and thoughtful comments. Below are our point-by-point responses to your questions, and we hope our responses address your concerns.
**Theoretical Claim**: Specific factors lead to the continuous improvement
**Response**: Thanks for this insightful comment. W... | Summary: The paper presents a gap-dependent analysis of regret and communication cost for on-policy Federated Q-Learning method. It improves prior worst-case regret bounds that scale as $\sqrt{T}$ by $logT$. The paper introduces a refined communication cost bound that distinguishes between exploration and exploitation... | Rebuttal 1:
Rebuttal: We thank you for your careful reading and thoughtful comments. Below are our point-by-point responses, and we hope our responses address your concerns.
**Weakness 1**: Improve readability of technical details
**Response**: Thanks for this helpful advice. To improve readability, we now provide a ... | Summary: This paper studies the gap-dependent bounds for online federated Q-learning where multiple agents collaboratively explore and learn a global Q-function under tabular MDP setting. The authors provide analysis of both performance regret and communication cost of the online federated Q-learning algorithm (FedQ-Ho... | Rebuttal 1:
Rebuttal: We thank you for your careful reading and thoughtful comments. Below are our point-by-point responses, and we hope our responses address your concerns.
**Weakness**: Contribution of the work
**Response**: We respectfully disagree that our main result is merely an application of existing gap-depe... | Summary: In the last few years, there has been an increase of interest in the topic of federated reinforcement learning (FRL), where agents interacting with similar (or exactly identical) MDPs communicate to achieve some speedup in performance. The authors consider such a formulation in the context of episodic finite-h... | Rebuttal 1:
Rebuttal: We provide point-by-point responses to your thoughtful comments.
**W1: A direct extension.** We respectfully disagree that extending to FRL is trivial. FRL introduces unique challenges and requires new techniques beyond those used in single-agent RL, especially in controlling the estimation error... | null | null | null | null | null | null |
sciLaMA: A Single-Cell Representation Learning Framework to Leverage Prior Knowledge from Large Language Models | Accept (poster) | Summary: The paper introduces a VAE approach to integrate external gene knowledge from pretrained large language models (LLMs) with single‐cell RNA sequencing (scRNA‐seq) data. By adapting static gene embeddings (derived from multiple modalities such as textual descriptions, protein sequences, and single‐cell models) i... | Rebuttal 1:
Rebuttal: We address each of the concerns below with response figures and tables: https://github.com/anonymous-ICML2025/rebuttal_April1st/tree/main/H2Qr:
**Interpretability**:
The reviewer is correct that interpretability primarily relies on implicit gene module detection and clustering in the current man... | Summary: The authors propose a new method that combines gene embeddings from LLMs with scRNA-seq data to generate context-aware representations for cells and genes. This approach outperforms existing methods in batch effect correction, cell clustering, and gene module discovery, while remaining computationally efficien... | Rebuttal 1:
Rebuttal: The reviewer mentioned that we did not provide enough evidence to illustrate how text information contributes to cell representations. In the manuscript we showed text-based gene information helped with removing batch effects (while preserving biologically meaningful information through sciLaMA fr... | Summary: This paper proposes a new method for single cell embedding that leverages external textual information (or internal knowledge representation from large language models). The method works by considering a learnable matrix decomposition of the single cell data, where cell embeddings are produced from the transcr... | Rebuttal 1:
Rebuttal: Below are our responses to the questions:
(1) “Why do you not use siVAE for comparison?”
We now clarify this more explicitly in the manuscript. siVAE is functionally equivalent to sciLaMA-self-informed (sciLaMA-s.i.), which instead of using LLM-based external gene embeddings as the input to the ... | Summary: The paper introduces sciLaMA, a novel framework for single-cell RNA sequencing analysis that integrates gene embeddings from large language models (LLMs) with scRNA-seq data using a Variational Autoencoder based architecture. This approach allows for context-aware representations of both cells and genes, enhan... | Rebuttal 1:
Rebuttal: Below, we address the three major points raised:
(1) “How does sciLaMA perform if the gene embeddings are random vectors?”
Using random gene vectors will prevent the model from generalizing to unseen genes. To further clarify this point, we conducted additional experiments: (i) using random gen... | null | null | null | null | null | null |
MITIGATING OVER-EXPLORATION IN LATENT SPACE OPTIMIZATION USING LES | Accept (poster) | Summary: This paper proposes a new score for optimizing the encoded latent space from pre-trained VAEs. The proposed new score employs the pre-trained decoder as a constraint to mitigate the over-exploration problem.
Claims And Evidence: The author is motivated to use a pre-trained decoder (Sec.3, why use the sequence... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort in reviewing our work.
### Experimental Design and Analysis
We calculate Equation 13 for a given **Z** because we are interested in a score that can be attributed to a single latent vector, rather than to a distribution over the output space \( p(x... | Summary: The paper develops Latent Exploration Score (LES) to mitigate the over-exporation in Latent Space Optimization (LSO). LES can be applied to any VAE decoder, and the paper develops a numerical procedure to incorporate LES as a constraint in LSO. The paper evaluates LES on multiple benchmark tests and demonstrat... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort in reviewing our work.
### Comments and Suggestions
We thank the reviewer for their comments and suggestions, which we will incorporate into the camera-ready version of our work.
### Question
**All code and trained VAEs will be released if the pa... | Summary: This work targets the problem of over exploration in latent space optimization. The overarching problem is that when we optimize over discrete data in the latent space of a generative model e.g. a VAE, we often select points that decode poorly or are otherwise invalid. This work proposes a score (LES) to guide... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort in reviewing our work. We are glad to see that the reviewer finds our work to address an appropriate hole in the research landscape.
We would like to clarify that we did not claim that LES is faster than all alternative approaches, but rather that i... | Summary: The paper proposes a latent exploration score method to mitigate over-exploration in latent space optimization, utilizing VAEs' properties for continuous optimization. It also demonstrates cases where solutions may become impractical.
Claims And Evidence: The paper begins by examining cases of over-exploratio... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort in reviewing our work.
### Typos
Thank you for pointing this out—we will update the text to replace "quadratic formula" with "quadratic function."
### Additional Visualizations
We thank the reviewer for their great suggestion to add more visualiz... | Summary: The paper proposes LES, a novel method for mitigating the over-exploration phenomenon in latent space optimization for discrete black-box optimization. LES can be easily integrated as a regularizer for acquisition function optimization. Experiment results validate that LES can improve the performance of latent... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort in reviewing our work.
**Deep Kernel**
The deep kernel affects only the computation of the kernel in the GP surrogate model—it does not reduce the dimensionality of the latent space on which LES is applied. Based on our results with the SELFIES VAE... | Summary: This paper addresses the problem of latent space optimization (LSO) for combinatorial problems. LSO uses an encoder-decoder architecture to map from a continuous latent space to a combinatorial space. The optimization can be carried out in a continuous space which has the advantage that it allows for continuou... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort in reviewing our work.
We address the reviewer's questions and comments chronologically.
Thank you for suggesting the addition of the important work by Tripp et al. We would like to point out that we refer to this work in line 155:
> “The frequent... | Summary: The authors propose a scoring method, Latent Exploration Score (LES), for Latent Space Optimization (LSO), to mitigate overexploration, which often results in unrealistic solutions in discrete optimization problems. LES uses the VAE decoder’s data distribution without additional training or architectural chang... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and effort in reviewing our work, and we are particularly glad to see that the reviewer agrees that our claims regarding the effectiveness of LES are supported by our empirical evidence, which the reviewer described as “extensive”.
First, thank you for catchin... |
Fair Class-Incremental Learning using Sample Weighting | Reject | Summary: This paper addresses the issue of ensuring fairness in class incremental learning. The authors identify a problem of unfair catastrophic forgetting, where previously learned information about certain sensitive groups (including classes) are disproportionately forgotten when new data is introduced. This work pr... | Rebuttal 1:
Rebuttal: Thanks for your thoughtful review.
To save space, we use the following shorthand terms:
* C&E: Claims And Evidence
* E&A: Experimental Designs Or Analyses
* M&E: Methods And Evaluation Criteria
* OCS: Other Comments Or Suggestions
>C&E1: Scope limited to class
Please refer to [W1] of Reviewer qr... | Summary: This paper investigates fairness in class-incremental learning (CIL). The authors argue that when learning new knowledge negatively impacts old tasks whose gradients oppose the new one while minimally affecting other tasks, unfairness arises. To address this, the authors introduce three fairness measures: equa... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and constructive feedback.
To save space, we use the following shorthand terms:
* C&E: Claims And Evidence
* E&A: Experimental Designs Or Analyses
* RL: Relation To Broader Scientific Literature
* M&E: Methods And Evaluation Criteria
>C&E: Each measure's effe... | Summary: This submission proposes a fairness-aware sample weighting (FSW) algorithm to address unfair catastrophic forgetting in class-incremental learning. A key contribution of this work is the detailed analysis of the forgetting mechanism, which demonstrates that catastrophic forgetting occurs when the gradient vect... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and constructive feedback.
>W1, Q1: Quadratic computational complexity
Clustering current data and assigning cluster-wise weight reduce LP variables, lowering computational complexity. Mini-batch K-means[1] is efficient for extremely large datasets where the c... | Summary: This paper addresses the fairness issue in class-incremental learning and proposes the Fairness-aware Sample Weighting (FSW) algorithm. The authors theoretically analyze the causes of unfair forgetting and optimize gradient direction by adjusting training sample weights to improve fairness. Furthermore, FSW fo... | Rebuttal 1:
Rebuttal: Thanks for your thoughtful review.
To save space, we use the following shorthand terms:
* C&E: Claims And Evidence
* E&A: Experimental Designs Or Analyses
* M&E: Methods And Evaluation Criteria
* OCS: Other Comments Or Suggestions
>C&E: More Baselines
While methods like parameter isolation and ... | null | null | null | null | null | null |
Linear convergence of Sinkhorn's algorithm for generalized static Schrödinger bridge | Accept (poster) | Summary: The authors leverage the process of Lyu and Mukherjee'24 to identify the linear convergence of the generalized Sinkhorn algorithm for a general class of metrics and solve the problem suffered by the KL divergence.
Claims And Evidence: NA
Methods And Evaluation Criteria: NA
Theoretical Claims: The existing l... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive comments and for raising some great questions.
**Theoretical Claims**
>The existing linear results assume a bounded domain or bounded cost functions, while yours do not explicitly specify that in the introduction or assumption section? Am I missing somet... | Summary: This paper presents a generalized SSB problem for any divergence and provides a new Sinkhorn-type algorithm to solve this problem. The paper focuses on the formalization of SB in a matrix optimization problem (1). The authors first construct the generalized Kantorovich duality and propose the generalized Sinkh... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful and helpful comments.
**Claims and Evidence**
While our work was motivated and inspired by the works the reviewer mentioned, our primary motivation was to identify the core mathematical structures that enable the elegant theory of SA/EOT in a significa... | Summary: This paper studies the generalized static Schrödinger bridge (SSB) problem, which extends the classical SSB by replacing the entropy divergence with any strictly convex function. Then it derives the Sinkhorn-type algorithm for solving the generalized SSB, and proves the linear convergence of the algorithm. Mor... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive evaluation of our work and further suggestions.
**Other Comments or Suggestions**
1. Thank you for pointing out this typo, we shall make the change to refer to iteration count by $k$ instead of $t$.
2. In the revision we will maintain $t$ for the upper bo... | null | null | null | null | null | null | null | null |
Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding | Accept (poster) | Summary: This paper explores PETL for 3D point cloud understanding to address the issues of full fine-tuning, such as forgetting pretrained knowledge and high storage costs. The authors propose a novel fine-tuning paradigm that integrates 2D vision semantic prompts from frozen pretrained models to improve the performan... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the detailed review and valuable suggestions.
**Mitigation of local and global ambiguity:** We understand the reviewer's concern and appreciate the opportunity to clarify. During the classification of 100 airplane and chair samples, we recorded which view t... | Summary: This paper proposes to leverage the rich semantic cue inherent in pretrained 2D foundation models as the semantic prompt for efficient point cloud learning. The authors identify the limitation of existing 3DPEFT research: they fail to align semantic relationships of features required by downstream tasks. To ad... | Rebuttal 1:
Rebuttal: We greatly appreciate the reviewer's thoughtful feedback and sincerely thank the reviewer for the positive recognition of our work.
**Feature inconsistency affects downstream task performance**：We appreciate the reviewer's question regarding the impact of feature inconsistency on downstream perfo... | Summary: This paper introduces a parameter-sparse fine-tuning strategy for fine-tuning pre-trained 3D point cloud models with vision semantic prompts of a frozen vision transformer in the 2D domain. It introduces a new Hybrid Attention Adapter (HAA) which aggregates the features of different views of the depth images i... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer's for the constructive comments and valuable suggestions.
**Performance gain on segmentation:** We understand the reviewer's concern and appreciate the opportunity to clarify. The ShapeNetPart is a nearly saturated dataset, where multiple methods have achieved simi... | Summary: This paper introduces a novel fine-tuning paradigm for 3D pretrained models by leveraging 2D semantic prompts from frozen 2D pretrained models to enhance point cloud understanding, while maximally preserving the pretrained knowledge. The proposed method has three main contributions:
1. Introduction of the Hyb... | Rebuttal 1:
Rebuttal: Thanks for the reviewer's careful reading and detailed comments. First, we respectfully clarify that all the pretrained models in our paper are trained on ShapeNet, these works aim to explore effective approaches for building 3D foundation models. ModelNet40, ScanObjectNN, and ShapeNetPart are use... | null | null | null | null | null | null |
Kernel-based Unsupervised Embedding Alignment for Enhanced Visual Representation in Vision-language Models | Accept (poster) | Summary: In this paper, the author proposed a kernel-based method to tune CLIP models for fine-grained visual information. Specifically, they propose a kernel based objective to align features from the original CLIP and another visual model that is capable of focusing on fine-grained details. Additionally, they introdu... | Rebuttal 1:
Rebuttal: We thank Reviewer u8xU for the thoughtful and constructive feedback on our work. The following is our response to the comments and questions in the review.
___
> **Preserving the original CLIP space in the alignment with DINOv2**
We would like to clarify that our objective function for the alignm... | Summary: The paper aims to enhance the fine-grained perception capability of CLIP by aligning with DINOv2 in the kernel space. It finds that such a practice can resolve the feature space conflict problem when aligning two different forms of features. Improvements are observed in image-text retrieval benchmarks and MLLM... | Rebuttal 1:
Rebuttal: We thank Reviewer 3gYu for the thoughtful and constructive feedback on our work. The following is our response to the comments and questions in the review.
___
> **Significance of the improvements**
We would like to highlight that our alignment fine-tuning achieves consistent improvements on seve... | Summary: This paper employs a kernel-based method to align CLIP's visual representation with DINOv2's representation. The goal is to enhance the CLIP vision encoder's fine-grained perception capabilities. The resulting aligned visual encoder demonstrates improvements in zero-shot object recognition, fine-grained spatia... | Rebuttal 1:
Rebuttal: We thank Reviewer ue7L for the thoughtful and constructive feedback on our work. The following is our response to the comments and questions in the review.
___
> **Comparison with DINOv2 encoder**
We would like to clarify that our numerical evaluation of the fine-tuned CLIP using alignment with D... | null | null | null | null | null | null | null | null |
MoEQuant: Enhancing Quantization for Mixture-of-Experts Large Language Models via Expert-Balanced Sampling and Affinity Guidance | Accept (poster) | Summary: This paper proposes a customized quantization method for MoE-LLMs, targeting two key challenges: inter-expert and intra-expert imbalance. To address these issues, the authors design a self-sampling approach to construct a balanced calibrated dataset and introduce an affinity-guided quantization loss to ensure ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive feedback. Below, we provide point-by-point responses, with all revisions incorporated accordingly.
---
>The proposed method shows marginal improvements against GPTQ
>When applying both proposed modules to QWENMOE-14B, the performance does not show a clea... | Summary: This paper proposes the MoEQuant framework, which aims to address the key challenges in the quantization of large language models (LLMs) with mixture of experts (MoE): inter-expert imbalance and intra-expert imbalance. Through two innovative methods, expert balanced self-sampling (EBSS) and affinity-guided qua... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive feedback. Below, we provide point-by-point responses, with all revisions incorporated accordingly.
---
>paper does not provide direct experimental or theoretical proof that the poor MoE quantization effect is caused by load imbalance
We provide additiona... | Summary: MoEQuant is a novel quantization framework designed for Mixture-of-Experts (MoE) large language models (LLMs). It addresses the challenges of accuracy degradation encountered in traditional post-training quantization (PTQ) methods. The framework introduces two techniques: Expert-Balanced Self-Sampling (EBSS) a... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive feedback. Below, we provide point-by-point responses, with all revisions incorporated accordingly.
---
>The study focuses solely on weight quantization and does not address activation quantization.
Memory constraints are the decisive factor for deployin... | Summary: This paper introduces a framework named MoEQuant which designed to efficiently quantize MoE large language models while addressing calibration imbalances. It tackles two key challenges: the unequal distribution of calibration samples across different experts, and the varying affinities between tokens and their... | Rebuttal 1:
Rebuttal: We sincerely appreciate your constructive feedback. Below, we provide point-by-point responses, with all revisions incorporated accordingly.
---
>Why do some tasks (like HumanEval) benefit significantly from MoEQuant, while others show minimal or no improvement?
Thanks for your insightful observ... | null | null | null | null | null | null |
SatFlow: Generative model based framework for producing High Resolution Gap Free Remote Sensing Imagery. | Reject | Summary: The paper proposes a model and an algorithm for generating high spatial res satellite images, given a single low spatial res satellite image together with past high spatial res satellite images. The generated high res image corresponds to the same day as the low res image. The low res images come from the MODI... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed review and appreciate the opportunity to clarify our work. Below, we address the specific points raised:
**Evaluation in Regions of Significant Change**
**comment:** *The metrics used in the paper compare differences between the generated image and the grou... | Summary: The authors present a generative model using Conditional Flow Matching (CFM) to fuse Landsat and MODIS images to produce frequent, gap-free images. The primary objective of their approach is to generate gap-free Landsat images conditioned on MODIS images and gap-free composite Landsat images. For example, thro... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thorough evaluation of our submission and appreciate the constructive feedback. We address each of the concerns below:
**Clarifying MODIS Downscaling**
**comment:** *Can your model generate high-resolution Landsat-like imagery using only MODIS input, or does it alwa... | Summary: - The paper presents SatFlow, a novel generative model-based framework designed to fuse low-resolution MODIS imagery with higher-resolution Landsat data using Conditional Flow Matching.
- The method generates gap-free, high-resolution remote sensing imagery, addressing limitations due to cloud cover and infreq... | Rebuttal 1:
Rebuttal: We thank the reviewer for their suggestions and appreciate the opportunity to clarify our work. Below, we address each of the concerns:
**Utility for Downstream Applications**
**comment:** *Claims regarding the advantages of high-resolution generated data for downstream tasks are not substantiat... | Summary: This paper addresses the challenge of the low temporal frequency of 30m resolution images collected by the Landsat satellite. It introduces SatFlow, a generative model based framework that integrates low-resolution MODIS imagery and Landsat observations to generate frequent, high-resolution, and gap-free surfa... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's thoughtful feedback and acknowledgement of the novelty and significance of our problem statement and methods introduced.
**Model Validation**
**comments:** *Could a dataset be constructed by using a previous Landsat image and a current MODIS image to predict the cur... | null | null | null | null | null | null |
Test-time Correlation Alignment | Accept (poster) | Summary: This paper proposes Test-time Correlation Alignment (TCA), a novel method to address the challenges of test-time adaptation (TTA) in deep neural networks. TCA aims to enhance model performance on out-of-distribution test data by aligning feature correlations between high-certainty test instances and the test d... | Rebuttal 1:
Rebuttal: We thank Reviewer Knrt for the valuable comments. (**Click [https://anonymous.4open.science/r/2025ICML_TCA_Rebuttal/README.md](https://anonymous.4open.science/r/2025ICML_TCA_Rebuttal/README.md) for Rebuttal Figures and Tables.**)
---
## Q1. Try to use more layers for transformation.
Following t... | Summary: The paper investigates the feasibility of performing correlation alignment during test time. Specifically, the authors:
1. Identifies the limitation of current TTA methods, and proposes Correlation Alignment (CORAL) as a potential soluton.
2. Introduces a strategy to construct a pseudo-source domain by sampli... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer cY7r’s insightful comments and constructive feedback. Below, we address each concern in detail. (**Click [https://anonymous.4open.science/r/2025ICML_TCA_Rebuttal/README.md](https://anonymous.4open.science/r/2025ICML_TCA_Rebuttal/README.md) for Rebuttal Figures and ... | Summary: The paper introduces TCA, which addresses some limitations of existing test-time adaptation (TTA) methods, such as neglecting feature correlation alignment and relying on backpropagation. TCA aligns feature correlations between test data and pseudo-source domains without accessing the source data. Two methods ... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer SZVF for the constructive and valuable comments. The concerns are addressed as follows. (**Click [https://anonymous.4open.science/r/2025ICML_TCA_Rebuttal/README.md](https://anonymous.4open.science/r/2025ICML_TCA_Rebuttal/README.md) for Rebuttal Figures and Tables.**)
-... | null | null | null | null | null | null | null | null |
Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal Link Prediction | Accept (poster) | Summary: This paper proposes using multiple geometries to analyze the problem of dynamic graphs. The authors propose to check for hyperbolicity locally and use hyperbolic geometry to update the node embeddings. They propose a framework consisting of two primary modules - Dynamic Geometric Modeling, Temporal State Aggre... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the time and effort in reviewing our paper. We take all comments seriously and try our best to address every raised concerns. Please feel free to ask any follow-up questions.
---
# S1
The typos on line 652 will be corrected in the next version, and we will care... | Summary: This paper introduces a novel framework for dynamic graph learning that integrates both Euclidean and hyperbolic geometric spaces. The authors propose a unified cross-geometric learning framework designed to capture the evolving nature of real-world systems, which are often dynamic. This framework includes a t... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the time and effort in reviewing our paper. We hope that our response can resolve your concerns. Please feel free to ask any follow-up questions.
---
# W1
We fully agree that MRR is an important metric for evaluating model performance in dynamic link prediction... | Summary: The paper proposes a unified cross-geometric learning framework that integrates both Euclidean and hyperbolic spaces to address the dynamic nature of graph data. The key idea is to analyze the evolving local structures via k‑hop ego-graphs and select the embedding space that best matches the local geometric ch... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the time and effort in reviewing our paper. We take all comments seriously and try our best to address every raised concerns. Please feel free to ask any follow-up questions.
---
# No Supplementary Material
Our appendices include notations, geometric background... | null | null | null | null | null | null | null | null |
Taming Knowledge Conflicts in Language Models | Accept (spotlight poster) | Summary: This paper conducts a study of knowledge conflicts in LMs. The authors reveal that some attention heads contribute to both contextual information and parametric memory, instead of only contribute to one of them. They call this phenomenon "superposition of contextual information and parametric memory."
Besid... | Rebuttal 1:
Rebuttal: **Claim and Evidence 1: The Effect of Scaling JuNE**
Thank you for the great suggestion. JuNE is an ablated variant of JuICE that only skips the second run but still incorporates and tunes its own scaling factor, as noted in Line 257 and Appendix E. We acknowledge that Proposition 5.5 may cause c... | Summary: This paper proposes JuICE, a test-time intervention method to address knowledge conflicts in language models (LMs) by steering models toward either parametric or contextual knowledge. The method first identifies two sets of attention heads that consistently achieve the desired effect with either positive or ne... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the time and thoughtful feedback. Before addressing specific points, we emphasize that our primary goal is to understand the interplay between parametric (internal) and contextual (external) knowledge in LLMs. Our methodology primarily aims to validate the novel... | Summary: The author proposed JUICE (Just Run Twice), a test-time intervention method to address knowledge conflicts in language models (LMs), where parametric memory (internal knowledge) contradicts contextual information. Extensive experiments across 11 datasets and 6 model architectures demonstrate that JUICE achieve... | Rebuttal 1:
Rebuttal: **Q1: Effectiveness of JuICE on different architectures**
Thank you for the insightful question. Our main paper evaluates JuICE on a range of widely used dense attention-based open-source LMs. Following your suggestion, we additionally test JuICE on Mistral-v0.1-7B, a leading sparse attention-bas... | Summary: The authors contribute a new method and analysis framework for understanding knowledge conflicts in RAG applications. Specifically, they define the usual setting of knowledge conflict where the contextual (retrieved) knowledge may differ from the parametric (internal) knowledge of the LLM. In these cases, and ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and for recognizing the significance of our work. We agree with the reviewer that many of the ideas presented in our paper extend beyond the scope of knowledge conflicts, as superposition appears to be a widespread phenomenon. Gaining a deeper understanding o... | null | null | null | null | null | null |
Scale Invariance of Graph Neural Network for Node Classification | Reject | Summary: The paper claims to prove "scale invariance" in GNNs and proposes a model that uses "multi-scaled" graphs.
## update after rebuttal
My original concerns regarding the technical correctness and impact of the work remain after the rebuttal. The notations can be defined much better and the presentation also nee... | Rebuttal 1:
Rebuttal: Dear Reviewer LdRA,
Thank you for your feedback on our paper. We would like to address two key misunderstandings in your review.
Misinterpretation of the Linear Model
You commented that our paper presents "exactly a linear model." However, this is a misunderstanding of our approach. In our paper... | Summary: The paper proposes a new approach for learning on graphs based on a multi-scale perspective inspired by CNN-like scale invariance and an adaptive self-loop addition strategy. The authors apply their method to four homophilic and two heterophilic benchmark datasets, reporting notable results in node classificat... | Rebuttal 1:
Rebuttal: Dear Reviewer YrK7,
Thank you for your thoughtful review and valuable feedback. We appreciate the opportunity to clarify several points and address any misunderstandings regarding our work.
First, we would like to clarify that our focus is on scale invariance, not scalability. Scale invariance i... | Summary: The work analyzes GCNs and demonstrates their scale invariance—i.e., a GCN with $k$ layers using the standard normalized adjacency matrix (SNA) is equivalent to a GCN with $k/2$ layers using the squared SNA, and other similar results. Leveraging this scale invariance property, the authors introduce GCN variant... | Rebuttal 1:
Rebuttal: Dear Reviewer 7DTZ,
Thank you for your valuable feedback on our paper. We appreciate the opportunity to clarify our contributions and address your concerns.
1. Lack of a "Related Work" section:
We acknowledge your concern regarding the absence of a dedicated "Related Work" section. However, we i... | Summary: The paper introduces the concept of scale invariance, which is the ability of a GNN to produce consistent classifications for a node regardless of the neighborhood radius used for its embedding. In this context, the paper proposes ScaleNet, a GNN designed to adapt to directed graphs (considering undirected gra... | Rebuttal 1:
Rebuttal: Dear Reviewer hJ6W,
We appreciate your thoughtful feedback and would like to address the key points you raised.
1. Equivalence of Hermitian Laplacian-Based Methods
Thank you for recognizing our contribution in uncovering the equivalence of Hermitian Laplacian-based methods.
By incorporating Mag... | null | null | null | null | null | null |
Joint Metric Space Embedding by Unbalanced Optimal Transport with Gromov–Wasserstein Marginal Penalization | Accept (poster) | Summary: In this paper the authors propose a novel alignment procedure procedure across different spaces by optimizing a Wasserstein distance in a joint embedding space between two distributions that are close in the GW sens to two distributions. The idea is similar to the Sturm variant of GW and the authors express... | Rebuttal 1:
Rebuttal: Many thanks for your positive report and the constructive comments.
We revised our manuscript accordingly,
especially with respect to the limiting cases
and additional references.
Please find our detailed answers below.
**Limiting Cases:**
Thanks for pointing this out,
which extends the limiting ... | Summary: The authors formulate an unbalanced OT problem where the marginals are penalized with GW distances and prove that their function have the minimizer. The authors show that if the regularization parameter goes to infinity, the solution their optimization problem is the embedded Wasserstein distance. To approxima... | Rebuttal 1:
Rebuttal: Many thanks for the positive evaluation
and your constructive feedback!
We have revised our manuscript according to your suggestions in particular
with respect to quantitative experiments.
Please find our detailed answers to your questions below.
**Quantitative FAUST:**
We agree that
the resul... | Summary: This paper proposes a framework for jointly embedding and aligning two metric measure spaces into a common metric space. The authors build their formulation from well-known prior works in this area, namely the Sturms and Memolis variants of Gromov Wasserstein distance. The main contribution of this paper is fo... | Rebuttal 1:
Rebuttal: Many thanks for your positive evaluation of our manuscript as an 'interesting submission' that 'is theoretically comprehensive and conceptually enriches many prior OT-based distances'
and for your constructive feedback!
We revised our manuscript according to your suggestions.
Please find our answe... | Summary: The paper proposes an unbalanced optimal transport problem with Gromov-Wasserstein marginal penalization, which can map data from two different domains, where no known correspondences exist, into a common metric space. The problem can be resolved using block-coordinate descent based on the Sinkhorn algorithm. ... | Rebuttal 1:
Rebuttal: Many thanks for your positive and constructive feedback!
We have revised our manuscript according to your suggestions.
Please find our detailed answers to your questions below.
**GW Comparison:**
If we consider the experiment behind Fig 1,
then we are looking for rotations (isometries)
which alig... | null | null | null | null | null | null |
MimicMotion: High-Quality Human Motion Video Generation with Confidence-aware Pose Guidance | Accept (poster) | Summary: This paper presents MimicMotion, a generative approach for animating a human in a single image using a 2D pose sequence. It employs a fixed pre-trained video diffusion model while training select layers, including PoseNet and U-Net modules. The PoseNet processes the 2D pose sequence, and its output is integrat... | Rebuttal 1:
Rebuttal: ### **1. Is confidence-aware pose guidance the optimal approach?**
We evaluated depth information as an additional input and found it does not significantly improve the resolution of front-back ambiguity or incorrect pose issues beyond what our confidence-aware approach already achieves. Depth es... | Summary: This paper proposed an SVD-based human video animation framework. The major contributions of this work are 1) a confidence-aware strategy to reduce the negative impact of inaccurate pose skeleton detection 2) a local hand region enhancement strategy to alleviate hand motion blur 3) a long video generation stra... | Rebuttal 1:
Rebuttal: ### **1. Out-of-Domain Generalization**
The model is trained exclusively on human dancing videos. There are two reasons for this generalizability:
First, the SVD backbone can generate cartoon and animal videos. Our fine-tuning enhances the model with pose controllability, and since the pose feat... | Summary: Based on the confidence-aware strategy, this work propose hand region enhancement to alleviate hand distortion, which improves video generation performance. Additionally, the supplementary materials provide generation results in other scenes (such as animals and other styles), verifying the generalization abil... | Rebuttal 1:
Rebuttal: ### **1. Essential References Not Discussed:** StableAnimator, ControlNeXt
Thank you for your effort and valuable feedback during the review process. We will add a discussion of these two related works and revise our manuscript accordingly.
### **2. The influence of hand region enhancement.**
> ... | Summary: This research focuses on the task of video generation, aiming to produce videos where the subject replicates a reference pose based on a single input image and specified pose conditions.
The paper introduces three significant advancements over existing baseline methods:
1. Integrating confidence awareness t... | Rebuttal 1:
Rebuttal: ### **1. Essential References Not Discussed: DiT**
Thank you for your effort in reviewing our paper. We will revise our manuscript and add a discussion about DiT. Compared with U-Net, DiT is considered a more novel and promising foundation model, especially in video generation tasks, where it demo... | null | null | null | null | null | null |
Socialized Coevolution: Advancing a Better World through Cross-Task Collaboration | Accept (poster) | Summary: This paper introduces a practical learning paradigm of Socialized Coevolution (SC), to overcome the limitations of methods that rely heavily on data-driven, single-model approaches and often fall short in leveraging inter-model knowledge transfer. The paper argues that existing methods like Knowledge Distillat... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We provide detailed responses to the issues. Please refer to the anonymous link (https://anonymous.4open.science/r/Supplementary-materials-for-SC/) for the supplementary tables and proofs.
>Q1. Specialization-generalization trade-off
R1. DISC balances spec... | Summary: This paper proposes a practical learning framework, DISC, to improve model performance across two task domains. The proposed method allows two expert models to exchange intermediate states and weights in a structured manner to achieve superior performance in both task domains. The authors evaluated DISC on ima... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We provide detailed responses to the issues. For the supplementary tables and proofs, please refer to the anonymous link (https://anonymous.4open.science/r/Supplementary-materials-for-SC/).
>Q1. Connection with social learning
R1. Human and machine societies... | Summary: The authors introduce a learning paradigm called Socialized Coevolution (SC), designed to enhance the performance of existing tasks with the support of an auxiliary model. Specifically, the main model progressively integrates hierarchical auxiliary information through a dynamically weighted communication mecha... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback. We provide detailed responses to the issues. Please refer to the link (https://anonymous.4open.science/r/Supplementary-materials-for-SC/) for the Additional tables and proofs.
>Q1. MTL setting
R1. The MTL setting is introduced to validate Sociability Infor... | null | null | null | null | null | null | null | null |
TabICL: A Tabular Foundation Model for In-Context Learning on Large Data | Accept (poster) | Summary: This paper propose a Tabular Foundation Model (TabICL) for classification tasks with new Distribution-aware Column-wise Embedding then Context-aware Row-wise Interaction design. It is a scalable and efficient model, it can handle up to 500K samples and 500 features with around 20GB of GPU memory. This model d... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful review and recognition of our work.
---
1. *More comparison results against the more recent models based on TabPFN ?*
Thanks for these relevant works. Due to limited computational resources, we prioritized methods based on code availability.
- TuneTable... | Summary: The paper proposes TabICL a tabular foundation model able to fit larger datasets than TabPFN (v1) and TabPFN (v2).
The main difference lies in the embedding strategy which leverages Set-Transformer as well as later in the model a positional rotary embedding layer. Similar to TabPFN v2 using FlashAttention, mak... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful review, and for mentioning that “the results are really great”.
Next, we respond to the reviewer's questions.
---
1. *A major drawback is ... the code or the model was released upon submission*
We **confirm that all of TabICL will be made fully public,... | Summary: This paper introduces a tabular foundation model, TabICL, to deal with the classification of larger number of table samples at affordable resources. Compred to the recent state-of-the-art TabPFNv2, the proposed TabICL has competitive performance while enjoys up to 10 times faster speed. The proposed architectu... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's valuable feedback. We now respond to the reviewer's questions.
---
1. *Why should the training be invariant to permutation of columns?*
In relational data, column order is typically arbitrary. The very visible publication [Why Tabular Foundation Models Should Be a ... | Summary: The paper introduces TabICL, a tabular foundation model designed to scale in-context learning (ICL) for tabular data with a two-stage architecture. This design allows TabICL to handle datasets with up to 500K samples using affordable resources. On the TALENT benchmark with 200 datasets, TabICL achieves compara... | Rebuttal 1:
Rebuttal: We appreciate the reviewer's valuable feedback.
**We kindly invite the reviewer to refer to the figures available at https://limewire.com/d/FhJbU#LPxKd0hlro, which may help better illustrate our responses. The link only contains figures.**
We realize that there may be a misunderstanding of how c... | null | null | null | null | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.