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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
SDMG: Smoothing Your Diffusion Models for Powerful Graph Representation Learning | Accept (poster) | Summary: The paper presents a new diffusion-based approach for self-supervised graph representation learning (SDMG). Rather than reconstructing all graph details, which often leads to the overfitting of high-frequency noise, the authors provide both theoretical and empirical evidence that focusing on low-frequency comp... | Rebuttal 1:
Rebuttal: We appreciate your valuable input and positive comments, and we have carefully addressed your comments in our responses below.
**Question 1: Clarification on the Meaning of “–” Symbol on Table 1.**
In Table 1, the “–” symbol indicates that the corresponding method either exceeded the available... | Summary: The authors reveal that purely generation-oriented objectives can conflict with recognition goals, demonstrating that excessive non-smooth-frequency reconstruction can harm representation quality. Specifically, they systematically investigate how reconstructing different parts of the graph frequency spectrum a... | Rebuttal 1:
Rebuttal: Thank you for your constructive and positive feedback; we appreciate your insights and have addressed your comments below.
**Question 1 and Weakness 1: Clarification on Masking Strategy Motivation.**
The masking strategy has two primary motivations: enhancing robustness and promoting learning ... | Summary: The authors introduce a new diffusion-based self-supervised framework for graph representation. It addresses the issue that minimizing generation-based learning objectives can overfit high-frequency noise instead of capturing important global structures. To overcome this, the authors propose (1) learnable enco... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback. We appreciate your time and have addressed your comments below.
**Question 1: Clarification on noise addition in Figure 1.**
The primary goal of Figure 1(a) is to illustrate the misalignment between generative reconstruction objectives and representation l... | Summary: The authors study the problem of graph representation learning (GRL) via diffusion models. The authors argue that current graph diffusion models for representation learning are sub-optimal due to their focus on the high-frequency signals. However previous literature, including a preliminary study by the author... | Rebuttal 1:
Rebuttal: We thank you for taking the time to review our manuscript and for the suggestions that have helped clarify our original contributions. Due to its overarching importance, we would like to start with Weakness 2:
**Novelty**
Our work is not on GRL as such, but GRL for use in a downstream classific... | null | null | null | null | null | null |
Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models | Accept (poster) | Summary: This paper proposes a hierarchical framework for tackling the complex instruction following challenge in vision-language-action-based robotic control. The paper highlights challenges in existing methods that struggle with following intricate instructions. The proposed method, Hi Robot, addresses these issues b... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We address each point below and will incorporate these improvements in the revision.
> Limited contribution: hierarchical VLMs with synthetic data generation.
While individual components build on prior work, Hi Robot's novel synthesis enables critical real... | Summary: This paper, inspired by "System 1" and "System 2" cognitive processes, proposes a hierarchical VLM-based system to interpret high-level instructions and convert them into commands for a low-level VLA model. To train the model, the authors employ both human-labeled and synthetically generated interaction data. ... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We address your comments in detail below and will update our paper accordingly.
> For the comparison with GPT-4o high-level instruction decomposing experiments, what is the user prompt?
The user prompts for evaluation (e.g., "Hi robot, can you make me a ... | Summary: This paper presents Hi Robot, a hierarchical vision-language-action (VLA) model for open-ended instruction following. The system integrates a high-level vision-language model (VLM) that interprets complex prompts and user feedback with a low-level VLA policy that executes atomic actions. A synthetic data gener... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback. We address each concern below and will revise the paper accordingly.
> The hierarchical structure resembles prior work on dual-process systems.
While building on foundational ideas, Hi Robot introduces key innovations:
1. **Interactive Open-Ended Instructi... | Summary: In this work, the authors introduce Hi-Robot, a System-1/System-2 approach that leverages a Vision-Language Model (VLM) to interpret complex prompts and generate a more suitable sequence of instructions for a Vision-Language-Action Model (VLA) to complete a given task. The system also integrates feedback duri... | Rebuttal 1:
Rebuttal: Thank you for your positive review and constructive suggestions. We address each point below and will incorporate these improvements in the revision.
> Real-time inference timing analysis
We provide detailed latency measurements across components (tested on consumer-grade RTX 4090):
*Low-Level ... | null | null | null | null | null | null |
Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering | Accept (poster) | Summary: The authors propose Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering and conduct extensive experiments to analyze the proposed method.
Claims And Evidence: Yes.
Methods And Evaluation Criteria: Yes.
Theoretical Claims: Yes.
Experimental Designs Or Analyses: Yes.
Supplementary Material: Y... | Rebuttal 1:
Rebuttal: **Q: The two key problems addressed in the paper have already been extensively studied in previous works.**
**A:** Thank you for pointing out these issues. As you mentioned, these two problems have indeed been discussed in prior works, and we have acknowledged these discussions in our Introductio... | Summary: The paper introduces Hgformer, a novel Hyperbolic Graph Transformer framework designed to address two critical limitations in GNN-based collaborative filtering (CF): local structure bias caused by neighborhood aggregation and embedding distortion in Euclidean space. The proposed method combines a parameter-fre... | Rebuttal 1:
Rebuttal: **Q: Empirical validation of training/inference time on large-scale data is missing.**
**A:** Thank you for pointing this out. We have conducted empirical evaluations and reported the average computational time per epoch for several representative baselines and our proposed model on Amazon Book (... | Summary: The paper proposes a Hyperbolic Graph Transformer architecture, to tackle the long-tail problems in CF tasks, which leverages LHGCN for graph structure modeling and hyperbolic cross-attention for global information modeling
Claims And Evidence: Are the claims made in the submission supported by clear and conv... | Rebuttal 1:
Rebuttal: **Q: Inconsistency between two aggregation approaches**
**A:** Sorry for the misunderstanding. These two types of aggregation serve different purposes in our model. The aggregation in Section 2.2 indicates gathering neighbor information for each node to capture multi-hop relationships, which is p... | null | null | null | null | null | null | null | null |
Differential Privacy Guarantees of Markov Chain Monte Carlo Algorithms | Accept (poster) | Summary: This papers studies differential privacy
and R\'{e}nyi differential privacy
for Markov chain Monte Carlo (MCMC) algorithms
for the path and the final value of the algorithms.
The general results are then applied
to study two popular MCMC algorithms,
the unadjusted Langevin algorithm (ULA)
and stochastic grad... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for the positive review and for their comments. We will incorporate the reviewer's suggestions in the revised version of the paper. Below we reply to the questions that the reviewer raised:
- **In the very beginning of Section 3.2., do you need any assumption on $f... | Summary: This paper analyzes the differential privacy (DP) guarantees of Markov Chain Monte Carlo (MCMC) algorithms, focusing on both general MCMC methods and specific Langevin-based variants. It establishes that the DP properties of the posterior distribution are crucial for ensuring the privacy of MCMC samples, showi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their useful comments and remarks. We are going to incorporate them in the revised version of the paper. Below we address the questions from the "Questions for authors" section.
- **Proposition 4.6. Is "for any x, y, s..." missing?**
Indeed, we will add this in the r... | Summary: This is a theoretical work studying DP and MCMC algorithms, focusing on:
1. Connections between mixing, the privacy of the exact posterior, and the privacy of intermediate iterates.
1. The privacy of Markov chains based on Langevin diffusion, e.g. the Unadjusted Langevin algorithm (ULA).
I view the central ne... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful comments, and particularly for the suggestion of relevant connected work in the literature. In response to specific questions, we have the following responses:
- **The informal takeaways from Section 3 I view as obvious. It is not clear to me what value the... | Summary: This paper present theoretical results on differential privacy guarantees for Markov Chain Monte Carlo (MCMC) methods. They develop DP guarantees for both full chains and for the final state of the chain, and demonstrate how these results can be applied for specific instances of MCMC dynamics.
## Update after... | Rebuttal 1:
Rebuttal: We are grateful to the reviewer for the careful comments. In the revised version of the paper, we will incorporate all their comments and fix all the typos, notation clashes, and other issues that were mentioned by the reviewer. Below we reply to the questions that the reviewer raised:
- **Sorry ... | null | null | null | null | null | null |
An Adaptive Orthogonal Convolution Scheme for Efficient and Flexible CNN Architectures | Accept (poster) | Summary: The paper considers the problem of constructing convolutions which correspond to orthogonal operators. While the general approach is based on BCOP (Li et al. 2019), the authors extend this framework to certain variants of convolutions like stride, dilations, grouping and transposing and consider some aspects f... | Rebuttal 1:
Rebuttal: We sincerely appreciate the detailed proof verification, the interest you've shown in our paper, and the evaluation towards acceptance. We answer the main weaknesses raised in your review, and hope this can help you to increase your score.
About experiments
---------
> What is accurate and robu... | Summary: The paper proposed a new method to design scalable and versatile orthogonal convolutional layers.
This layer allows scaling further architecture composed of orthogonal layers; indeed, previous orthogonal layers lack common features of regular convolutional layers (strides, dilations, group convolutions, etc)... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for his comprehensive review. We sincerely appreciate the effort put into it and hope our response will be at the same level. We organized our answers into 4 sections that cover the various questions and remarks.
About the experiments:
-------------------------... | Summary: This paper explores research on orthogonal convolutional layers and introduces AOC, a scalable method for constructing orthogonal convolutions. The authors provide a detailed introduction and implementation of their methodology, including the construction of strided, transposed, grouped, and dilated orthogonal... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful evaluation and the interest you've shown in our paper, particularly regarding its practical contributions.
In addition to the implementation-oriented aspects that you seem to highly appreciate, we would like to highlight the theoretical novelty of our work:... | Summary: This paper purposes a orthogonal CNN structure. Orthogonal convolution has been shown success in BCOP and RKO. Utilizing the property of the product of orthogonal matrices are still orthogonal (prop 2.3), this paper invents AOC in Eq. 8, which is the product of RKO and BCOP. A parallel computing technique (fig... | Rebuttal 1:
Rebuttal: We want to thank the reviewer for his review. We will provide additional information that we hope you to find relevant.
**About 1.** While AOC does not improve the expressiveness of its original building blocks, its flexibility unlocks the construction of complex blocks as depicted in Fig. 5a. Gi... | null | null | null | null | null | null |
LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations | Accept (poster) | Summary: In this paper, an algorithm named LGDM (latent guidance in diffusion models) is proposed for NR-IQA. It utilizes the pretrained latent diffusion models for sampling process toward perceptually consistent regions on the manifold. Specifically, it extracts diffusion hyperfeatures, which are multi-scale and multi... | Rebuttal 1:
Rebuttal: Your feedback for our work is greatly appreciated. In the following, we will address your questions and concerns in detail.
1. **Missing References:** Thank you for pointing out the missing recent papers, GRepQ and QCN. We have already included both methods in our updated manuscript. See point-3 i... | Summary: This work proposes an NR-IQA model by leveraging latent diffusion models (LDMs). The core idea is that diffusion models inherently learn perceptually consistent regions within their data manifold. The authors introduce Perceptual Manifold Guidance (PMG), a technique that extracts multi-scale and multi-timestep... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback and your recognition of our work. We would like to answer your questions and address your concerns below.
1. **Reliance on Theoretical Assumptions:** We agree with the reviewer that our theoretical proofs rely on assumptions, which are idealizations of real... | Summary: This paper is the first to propose using a pre-trained latent diffusion model as a perceptual model to extract perception features aligned with human perception for NR-IQA. The authors introduce a novel sampling method that adjusts samples to align with human perception while preserving the manifold. Additiona... | Rebuttal 1:
Rebuttal: We appreciate your valuable insights and acknowledgment of our efforts. Below, we respond to your questions and discuss the issues you have raised.
1. **Clarification on the calculation of $\psi_p$:** We apologize for the lack of clarity regarding the calculation of the perceptual features $\psi_p... | Summary: This paper introduces Latent Guidance in Diffusion Models (LGDM) for No-Reference Image Quality Assessment (NR-IQA), leveraging the powerful representation capabilities of pretrained Latent Diffusion Models (LDMs). Specifically, the authors propose Perceptual Manifold Guidance (PMG) to steer the sampling proce... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. We would like to address your questions and concerns below.
1. **Multi-Timestep Features (Weakness 3):** Thank you for bringing up this question. In our method, Perceptual Manifold Guidance (PMG), operates iteratively during the DDIM sampling process. At ... | null | null | null | null | null | null |
Best of Both Worlds: Regret Minimization versus Minimax Play | Accept (poster) | Summary: This submission studies whether an algorithm can achieve O(1) regret compared to a specific fixed comparitor strategy while also guaranteeing O(T^0.5) regret compared to the best strategy in hindsight in symmetric two-player zero-sum games. The authors focus on bandit feedback settings, where the learner only ... | Rebuttal 1:
Rebuttal: Thank you for carefully reviewing our paper and providing valuable feedback!
## **Remark on non-symmetric games:**
> What are the barriers preventing you from extending your results to non-symmetric zero-sum games?
**Response:** We would like to point out that, in fact, the game in question d... | Summary: In this paper, the goal is to develop a bandit algorithm that guarantees simultaneously constant regret with respect to a given strategy and $\sqrt{T}$ regret with respect to the best strategy in hindsight.
Claims And Evidence: - An extension of the phased algorithm of Even-Dar to the bandit case
- An analysi... | Rebuttal 1:
Rebuttal: Thank you for carefully reviewing our paper and providing valuable feedback!
> The algorithm presented by the authors is really close to Even-Dar et al. (2008). Authors should focus on what are the differences beyond simply using Exp3 as a base algorithm.
**Response:** Indeed, the high-level di... | Summary: The paper proposes a bandit algorithm for two-player symmetric zero-sum games that guarantees $O(1)$ regret against the minimax strategy and $O(\sqrt{T})$ regret against any strategy.
## update after rebuttal
I keep my score, which remains positive.
Claims And Evidence: I do not see any problematic claims.
... | Rebuttal 1:
Rebuttal: Thank you for carefully reviewing our paper and providing valuable feedback!
> The related work and its comparisons with the manuscript can be more adequately discussed.
**Response:** We agree that further elaboration on the related work would increase the clarity of the paper. In the current v... | Summary: This paper derives regret upper bounds that vary depending on the comparator in the setting of online learning with bandit feedback. Specifically, the paper proposes an algorithm that simultaneously achieves an $O(1)$ regret upper bound when the comparator of the regret lies in the interior of the probability ... | Rebuttal 1:
Rebuttal: Thank you for carefully reviewing our paper and providing valuable feedback!
> It is possible that I have not fully understood the motivation behind this work. If I receive a convincing response on this point, I am willing to raise my score.
**Response:** From the viewpoint of learning in game... | null | null | null | null | null | null |
Ranked Entropy Minimization for Continual Test-Time Adaptation | Accept (poster) | Summary: This work studies continual test-time adaptation and proposes a method to tackle it. In particular, based on experimental motivation on how entropy minimization collapses under continual TTA to predicting the same class, they propose to optimize the cross entropy between the prediction of the model on two mask... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and insightful suggestions. Below, we provide detailed responses to your questions.
>**1. Computational intensive for mask**
A1. The token-wise attention used to compute the mask is applied only to the final self-attention layer and accounts for a very small ... | Summary: This paper addresses the model collapse issue in CTTA. Specifically, it aims to reconcile the trade-off between fast but unstable EM methods and stable yet computationally expensive CR methods. The authors propose REM, a novel EM-based approach incorporating a progressive masking strategy. This strategy gradua... | Rebuttal 1:
Rebuttal: We appreciate your constructive suggestions. We address your questions below.
>**1. Model collapse problem**
Model collapse in entropy minimization methods is a well-known issue, but we would like to emphasize that it remains an unresolved challenge in achieving stability during TTA process. Fig.... | Summary: This paper proposes a novel Ranked Entropy Minimization method for test-time adaptation. While leveraging entropy as a supervision signal may risk model prediction collapse, the authors address this challenge by first constructing an explicit masking chain with varying masking ratios on the original images. Th... | Rebuttal 1:
Rebuttal: We appreciate your constructive comments and positive reviews. We address your questions below in detail.
>**1. Sampling through image masking vs. sampling from a queue using active learning**
A1. The active approach of capturing accurate samples and adapting using selected samples is meaningful... | Summary: This paper proposes a masked consistency loss (MCL) and entropy ranked loss (ERL) based learning mechanism for continual test-time adaptation (CTTA).
The MCL involves incrementally masks the images for data augmentation, and the ERL contributes in ensuring that the entropy of predictions with a low masking rat... | Rebuttal 1:
Rebuttal: We appreciate your positive reviews and valuable suggestions. We address your concerns and questions below in detail.
>**1. Values of M and N involved in mask ratios and chains**
A1. We provide the ablation test regarding hyperparameters of the mask. Although the best accuracy was achieved when N... | Summary: This work introduces two novel loss functions for CTTA, which utilize different views of the samples to enforce consistency alignment while preserving their relative ranking. Experimental results demonstrate the effectiveness of the proposed approach.
Claims And Evidence: Yes
Methods And Evaluation Criteria:... | Rebuttal 1:
Rebuttal: We appreciate your constructive comments. We address all the concerns and provide new experiments to support our contributions.
>**1. Novelty**
Compared with [1,2], our novelty is summarized;
- [1] proposed a contrastive learning method based on strong-weak augmentations, but our method addres... | null | null | null | null |
Inverse Optimization via Learning Feasible Regions | Accept (poster) | Summary: This paper proposes an inverse optimization approach to learn parameters in an optimization problem using historical decision data. Specifically, the authors adapt the predictability and sub-optimality losses, previously used in the literature for learning objective functions, to the context of constraint lear... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments/questions. Here is our response:
**[Q1: Loss functions meaning]**
Table 2 presents the performance of different methods across four metrics: true predictability loss $(\ell^\text{p})$, true suboptimality loss $(\ell^\text{sub})$, estimated predictability l... | Summary: This paper investigates inverse optimization, with a particular focus on learning feasible regions in optimization problems with linear objectives. The authors introduce two loss functions, propose a hypothesis class for the constraint function, and develop reformulations and smoothing techniques to address th... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments/questions. Here is our response:
__[Contributions of loss functions:]__
We agree that introducing the loss functions is not a major contribution. However, their combination in (5a) and (5b) with the hypothesis class in (7) enables reformulation into finit... | Summary: The authors derive a method based on inverse optimization to learn the feasible region of an optimization model. To this end, they use two loss functions, incorporating infeasibility and suboptimality of the solutions over the hypothesized feasible region. Besides, they set a parametrized hypothesis class for ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments/questions. We will address all minor points in the revision. Please find all the code used for the paper in https://github.com/orRepo/InverseOptimization. Below is our response:
__[Q1: Main contributions]__
The primary contribution is the hypothesis class ... | Summary: The paper proposes a novel inverse optimization (IO) framework to learn feasible regions (constraints) from observed decisions. Key contributions:
Two Loss Functions:
Predictability Loss: Minimizes the perturbation required to make observed decisions feasible and near-optimal.
Suboptimality Loss: Penalizes vi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments/questions. Here is our response:
__[Q1: Implication of nonpositive loss]__
The reviewer is indeed correct that for the function $J_\theta$ defined in (4), the condition $J_{\theta}(\boldsymbol{x}, \boldsymbol{s}) \leq 0$ does not necessarily imply that the... | null | null | null | null | null | null |
Low-Rank Adapting Models for Sparse Autoencoders | Accept (poster) | Summary: This paper proposes improving the use of SAEs in LLMs by fitting a LORA to the LLM with the SAE inserted. The results indicate that LORAs are an effective countermeasure to the reduction in performance of inserted SAEs.
Claims And Evidence: The claims of the paper are well supported through a variety of exper... | Rebuttal 1:
Rebuttal: Thank you for your feedback and time! We are especially glad to hear you found our experiments and results to be thorough and convincing.
---
> Can you explain 5.2 in more detail?
Certainly! We apologize for not being clearer and hope the below will explain in more detail how we’re computing ou... | Summary: This paper trains some parts of a Transformer with LoRAs in order to make an SAE more accurate at reconstructing activations as a sparse linear sum. They achieve impressive upstream metric results at a low cost, and also achieve reasonable SAEBench downstream (ish!) results too.
Claims And Evidence: The claim... | Rebuttal 1:
Rebuttal: We are greatly thankful for your time and feedback, especially related to the limitations of our work. We were glad to hear you found the evaluations comprehensive.
---
> Additionally, I am concerned that far too much focus is on improving simple SAE metrics… At absolute minimum, the authors sho... | Summary: This paper introduces an approach for improving sparse autoencoders (SAEs) used in language model interpretability by using Low-Rank Adaptation (LoRA) to finetune the language model itself around a previously trained SAE. Their method freezes both the original model and the SAE while only training low-rank ada... | Rebuttal 1:
Rebuttal: We are grateful for your time and thorough feedback, especially with regards to our methodology. We were glad to hear you found the implementation to be well executed.
---
> How do you reconcile your approach of modifying the model with the traditional goal of mechanistic interpretability?
This... | Summary: This paper introduces a novel approach to improve the interpretability of SAE by using LoRA to fine-tune **LLMs** around previously trained SAEs. Unlike previous work that focused on optimizing SAE architectures, this approach optimizes the language model itself to work better with an existing SAE. Across vari... | Rebuttal 1:
Rebuttal: We thank you for your time and help, especially in regards to your suggestions for additional experiments, which we have now run. We are especially glad that you appreciated the novelty of modifying the model itself to better fit a sparse autoencoder, as we agree that this idea is the major impact... | null | null | null | null | null | null |
Focal-SAM: Focal Sharpness-Aware Minimization for Long-Tailed Classification | Accept (poster) | Summary: The paper proposes a new sharpness aware minimization (SAM) algorithm suited for long-tailed classification. It works by formulating SAM as a regularizer, and then applying different regularization strengths for each class. The authors show that this approach can be more computationally efficient than baseline... | Rebuttal 1:
Rebuttal: > Q1: Please include standard errors over multiple runs for numerical results tables. The performance of Focal-SAM is often quite close to baselines, so establishing this statistical significance will be important.
**A1**: Thanks for your valuable suggestion!
In the original paper, we report th... | Summary: The paper introduces Focal-SAM, a new variant of Sharpness-Aware Minimization (SAM) designed for long-tailed classification. It aims to improve generalization for both head and tail classes by integrating the focal mechanism with SAM. Compared with baselines like ImbSAM and CC-SAM, Focal-SAM efficiently achiev... | Rebuttal 1:
Rebuttal: > Q1: Why don't use $\tilde{L}_S^{FS} (w)$ (eq 5) as the main objective? Specifically, why manage $\tilde{L}_S^{FS} (w)$ as a penalty of standard loss as in eq 6, which will add an additional hyperparameter $\lambda$?
Thanks for your questions!
The motivation behind SAM-based methods is that th... | Summary: This paper proposes a learning mechanism named Focal Sharpness-Aware Minimization (Focal SAM), which is an exquisite extension of SAM theories over long-tailed classification tasks. Compared with existing methods, the proposed Focal SAM excels at keeping the flatness of landspaces of both head and tail classes... | Rebuttal 1:
Rebuttal: > Q1: (1) From Table. 2, I observe a significant performance enhancement with the original SAM, but the improvements from SAM to Focal-SAM are not obvious. (2) I notice that the running time of Focal-SAM is 50% more than the original SAM. This dilates the necessity of employing Focal SAM.
**A1**... | Summary: For long-tailed (imbalanced) classificaton, the authors propose
Focal-SAM that aims at class-wise SAM so that flatter minima are
found. They first show that imbSAM, which applies SAM only to tailed
classes, can increase sharpness in the head classes. They then show
that CC-CAM could be computational expensiv... | Rebuttal 1:
Rebuttal: > Q1: On computational overhead, how different are the perturbations between the two methods?
**A1:** Thanks for your question!
The perturbation in Focal-SAM is computed as:
$$
\hat{\epsilon}(w) = \rho \frac{\nabla_w L_S^\gamma(w)}{|| \nabla_w L_S^\gamma(w) ||_2}
$$
where $L_S^\gamma(w) = \s... | null | null | null | null | null | null |
Test-time Adaptation on Graphs via Adaptive Subgraph-based Selection and Regularized Prototypes | Accept (poster) | Summary: This paper find that existing graph neural network methods struggle with performance degradation when adapting to test-time domain shifts, and current test-time adaptation methods for graphs mainly focus on Euclidean data and face challenges with label scarcity and knowledge utilization. The authors propose AS... | Rebuttal 1:
Rebuttal: We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your concerns and provide additional clarification.
> Q1. However, I notice tha... | Summary: This paper investigates the problem of test-time adaptation on graphs, addressing the challenge of adapting a pre-trained graph neural network (GNN) to unseen test data without access to the original training set. The authors propose ASSESS (Adaptive Subgraph-based SElection and Regularized Prototype SuperviSi... | Rebuttal 1:
Rebuttal: We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your concerns and provide additional clarification.
> Q1. While the paper focu... | Summary: This paper proposes an algorithm ASSESS that handles graph test time domain adaptation for graph-level task. The algorithm mainly contain two components as the adaptive subgraph-based selection and regularized prototype supervision. The subgraph-based selection tends to select reliable test graphs by setting i... | Rebuttal 1:
Rebuttal: We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following.
> Q1. Lack of illustration and more concrete motivation behind the unique challenge appear for graph TTA in the introduction. For instance, what can be... | null | null | null | null | null | null | null | null |
Rethinking External Slow-Thinking: From Snowball Errors to Probability of Correct Reasoning | Accept (poster) | Summary: The paper provides theoretical analysis of the snowball error in reasoning models with external thinking mechanism. The paper unveil some interesting properties of reasoning models. However, the strong assumptions make most results somewhat obvious and/or unhelpful.
Claims And Evidence: - The setting in the p... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback. Below are our responses to the concerns raised.
## About the Probability that the Thoughts are Incorrect
Our theoretical analysis hinges on the framework presented in Fig.1, where we model the LLM reasoning process as a *planning-execution* mechani... | Summary: This paper discusses the issue of understanding and improving external slow-thinking methods in LLM reasoning. The authors (1) propose a theoretical framework based on information theory that analyzes snowball errors and (2) connects them to the probability of reasoning errors in LLMs. This method aims to expl... | Rebuttal 1:
Rebuttal: We sincerely appreciate the insightful and constructive feedback provided in your reviews. Below, we address each concern in detail.
## The General Linkage between MI Decay and Reasoning Errors
We formalize the reasoning process of LLMs as a *planning-execution* framework. Given a question, as il... | Summary: This paper analyzes the potential snowball error effect that may occur during the reasoning process of Large Language Models (LLMs), and connects it to the probability of correct reasoning using information theory. Within this theoretical framework, external slow thinking methods can be interpreted as strategi... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback. Below, we address your concerns point by point.
## Regarding Prior Work [1]
While our work is inspired by [1], and both employ information theory to analyze chain-of-thought (CoT) reasoning, **there are significant distinctions between our approache... | Summary: The paper focuses on analyzing ``snowball effects'', where the model's implicit thinking process is not well represented by the tokens they generate at each step, hence accumulating throughout the inference. They analyze this phenomenon through information-theoretic perspective to give lower bounds for correct... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback and constructive comments. Below, we provide point-by-point responses to address the raised concerns.
## The Snowball Error and the Length
We acknowledge your insightful observation that the length of responses could be correlated with question diffi... | Summary: This paper aims to provide rationales for the effectiveness of inference-time compute scaling, also known as slow thinking, particularly from an information-theoretic perspective.
First, it argues that as the length of a reasoning path increases, the probability of encountering an error along the path also gr... | Rebuttal 1:
Rebuttal: We sincerely appreciate the thorough and constructive reviews. Below we address each concern raised by the reviewers.
## Concerns about the Assumptions
We note your thoughtful concerns regarding our methodological assumptions, specifically concerning: **(1) the coverage of multiple valid reasonin... | Summary: The paper analyzes the mathematical mechanism behind the slow thinking of large language models. First, the authors define snowball errors using information theory. Then, they derive a lower bound on the probability of reasoning errors based on snowball errors. This bound indicates that the probability of erro... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable feedback. Below are our point-by-point responses:
## Lack of Empirical Verification for Proposition 4.3
Proposition 4.3 was designed to facilitate subsequent analyses (Sec 4.2, line 214) while maintaining readability. The negative exponential form provides a... | null | null |
Parameters vs FLOPs: Scaling Laws for Optimal Sparsity for Mixture-of-Experts Language Models | Accept (poster) | Summary: The paper investigates the optimal trade-off between the total number of parameters and FLOPs per example in Mixture-of-Experts (MoE) models under a fixed pretraining FLOPs budget. Experimental results show that increasing total model parameters (i.e., increasing sparsity and reducing active parameters per inp... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time reviewing our paper. We find it encouraging that the reviewer finds our work valuable and relevant as we systematically study optimal model size and FLOPS-per-token in MoEs. We address the concerns below. Please note that we partially quote or paraphrase review... | Summary: This paper explores parameter-FLOP trade-offs in sparse MoE LLMs. The author finds that:
1. Increasing sparsity during pretraining improves efficiency and performance under a fixed compute budget.
2. More parameters benefit pretraining, while FLOPs are crucial for inference, especially for reasoning tasks.
3. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful and thorough review and for their support. We thank the reviewer for pointing out the comprehensive nature of our empirical work and presentation and respond to the question(s) raised by them below:
### Response to Weaknesses
> Sparsity improves deploym... | Summary: This paper investigates the relationship between the number of model parameters and the compute per example, measured in Floating Point Operations (FLOPs), in the context of sparse Mixture-of-Experts (MoE) language models. The authors aim to understand how varying the sparsity level—defined as the fraction of ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful and thorough review and for their support.
> Essential References Not Discussed
The reference pointed to by the reviewer discusses techniques to sparsify MoEs after training whereas we discuss optimal sparsity during pretraining and its implications on... | Summary: The paper provides empirical scaling laws for MoE-based LLMs. The experimental setup is simply training MoE-LLMs on the RedPajama dataset and then evaluating by comparing the eval loss. With this setup the authors find:
1. For every sparsity level and FLOPs budget, there seems to be a unique optimal model siz... | Rebuttal 1:
Rebuttal: ### Response to Weaknesses (W):
> The novelty is lacking. There have already been many scaling papers for MoEs and it's not clear what is new here. The paper shows that for a given FLOPs budget and sparsity there is an optimal model size. This has been known for dense models for a long time, and ... | null | null | null | null | null | null |
Self-Discriminative Modeling for Anomalous Graph Detection | Accept (poster) | Summary: The paper presents self-discriminative modeling method for graph-level anomaly detection. By generating pseudo-anomalous graphs to interpolate between normal and anomalous samples, the method constructs a reliable decision boundary solely based on normal data. The claims are well supported by corresponding the... | Rebuttal 1:
Rebuttal: **To W1:** Thank you for the insightful question. The adversarial training of SDM-ATI and SDM-ATII utilizes a generator to produce pseudo-anomalous graphs, with a discriminator balancing this via an adversarial loss (Eq. 9). Therefore, the numbers of training epochs for the generator and discrimin... | Summary: This paper proposes a novel GLAD framework named Self-Discriminative Modeling (SDM). The key idea of SDM is to generate pseudo-anomalous graphs from normal graphs and train a classifier/discriminator to distinguish them. The generative model and discriminative model are jointly trained to learn a more robust d... | Rebuttal 1:
Rebuttal: **We thank the reviewer for recognizing our work. Below are our responses to your concerns:**
**To W1:** We would like to address your concern from the following two aspects:
1. From the embedding visualization (e.g., Figure 7), the embedding distributions of the pseudo-anomalous graphs are most... | Summary: The paper introduces a new framework called Self-Discriminative Modeling (SDM) for detecting anomalous graphs. The proposed method trains a deep neural network using only normal graphs, without access to real anomalous examples. To achieve this, the authors generate pseudo-anomalous graphs from normal graphs. ... | Rebuttal 1:
Rebuttal: **We thank the reviewer for the positive comments. Below are our responses to your concerns:**
**To Q1:** The statement "no overlap between $\mathscr{D}$ and $\tilde{\mathscr{D}}$" is intended to facilitate a more precise problem definition for our paper, as we would not be able to identify norma... | null | null | null | null | null | null | null | null |
Training High Performance Spiking Neural Network by Temporal Model Calibration | Accept (poster) | Summary: This paper systematically summarizes previous logit gradient calculation schemes, including SDT and TET, then proposes a new temporal gradient rescaling method to enhance the learning capability of SNNs.
Claims And Evidence: Yes
Methods And Evaluation Criteria: I tend to think the proposed method can provide... | Rebuttal 1:
Rebuttal: **Rebuttal Appendix:https://anonymous.4open.science/r/TMC-0262**
**1. I tend to think the proposed method can provide a new perspective for the SNN community to a certain extent. The research perspective of Proposition 3.3 seems interesting. The authors conduct research on temporal heterogeneity ... | Summary: This paper proposes a temporal confidence calibration method for SNNs, improving both the model's performance and heterogeneity, as demonstrated across several static classification tasks.
Claims And Evidence: The proposed method is simple yet efficient, with a clear motivation at increasing the model's tempo... | Rebuttal 1:
Rebuttal: **Rebuttal Appendix:https://anonymous.4open.science/r/TMC-0262**
**1.Provide more results on dynamic datasets to make the method more convincable, like action recognition datasets including DVS-Gesture, SL-Animals, etc.**
**1)DVS-Gesture**: We train VGGSNN with T=10 on DVS-Gesture and compare th... | Summary: this paper introduces a new training method for spiking neural networks called temporal model calibration . the goal is to improve the performance of snns by increasing their temporal heterogeneity, which is how much their outputs vary over time. the authors argue that existing training methods, like direct tr... | Rebuttal 1:
Rebuttal: **Rebuttal Appendix:https://anonymous.4open.science/r/TMC-0262**
**1.Visualization or mathematical proof of increased heterogeneity.In-depth analysis of how tmc affects neuron activations.**
**1)Temporal Heterogeneity Visualization**
We compare VGGSNNs trained by TMC and TET on DVSCIFAR10, visu... | Summary: This work finds that the logit gradients have insufficient diversity in the temporal dimension during SNN training. The authors then rescale the gradient in each time step to improve diversity, resulting in SOTA performance for image classification tasks.
Claims And Evidence: yes
Methods And Evaluation Crite... | Rebuttal 1:
Rebuttal: **Rebuttal Appendix:https://anonymous.4open.science/r/TMC-0262**
**1.The relationship between temporal heterogeneity and the performance of SNNs remains unclear. It is not clear whether the so-called logit gradient should be diverse.**
**1)Relationship between Temporal Heterogeneity and Performa... | null | null | null | null | null | null |
EvoPress: Accurate Dynamic Model Compression via Evolutionary Search | Accept (poster) | Summary: The paper proposes Evopress, a novel pruning approach for the dynamic compression of LLMs based on evolutionary computation. The authors identify a critical issue in current compression algorithm approaches: error monotonicity does not apply to LLM compression. Aiming to resolve such drawbacks, they propose a ... | Rebuttal 1:
Rebuttal: We thank the reviewer for a thorough and detailed view with a lot of insightful comments and suggestions. The concerns are addressed below:
**Stopping Criteria**
We ran the search for more generations than necessary to get insights about the convergence behavior. We agree that in practice one sh... | Summary: This paper proposes EvoPress, a general framework for LLM compression. The authors observe that error monotonicity does not hold for LLM, and proposes an evolutionary search approach to improve performance. Experimental results demonstrate that on three compression approaches, depth pruning, unstructured spars... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments. Below, we address the two weaknesses:
> The theoretical proof does not fully explain its effectiveness in nonlinear LLM compression. It is unclear whether EvoPress maintains optimization capability in nonlinear regions.
Fundamentally this is true – we ca... | Summary: This paper aimed at LLM compression and motivated by the observation that depth pruning LLM further may improve the performance. Evolutionary search algorithm is applied to search pruned model with compressed size and performance constraint. It also applied to layer/block-wise non-uniform unstructured pruning ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback. The concerns are addressed below:
**Experimental design and analyzes**
*Runtime of the method*
Similar as for unstructured pruning in the main text, we have indicated the runtime of EvoPress for block dropping and quantization within convergence plots (s... | Summary: The paper introduces EvoPress for dynamic LLM compression, which optimizes compression levels across different model components to minimize accuracy loss while satisfying a global compression threshold. By formulating dynamic compression as an optimization problem, EvoPress efficiently determines optimal compr... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback. The questions are addressed below:
> To quantify model degradation, KL divergence was used in Section 3. KL divergence is popular to do it, but there are other ways, for example max absolute value, or we can even use a small calibration dataset to measure p... | null | null | null | null | null | null |
Limitations of measure-first protocols in quantum machine learning | Accept (poster) | Summary: This work is motivated by randomized measurement protocols to analyze the separation in quantum machine learning when processing quantum data using fully quantum operations versus measuring the input data and utilizing classical information. It highlights the limitations of measure-first protocols and provides... | Rebuttal 1:
Rebuttal: **The paper discusses the impact of noise on the task but lacks numerical validation. Why do you not provide numerical validations?**
We appreciate the reviewer’s suggestion regarding numerical simulations and experimental validation. We would like to mention that we are already collaborating wit... | Summary: This paper establishes a theoretical framework contrasting two quantum machine learning approaches: "fully-quantum" protocols that adaptively measure quantum data versus "measure-first" protocols restricted to fixed initial measurements. The authors prove that certain learning problems can be efficiently solve... | Rebuttal 1:
Rebuttal: **The description of the task could be more accessible for the machine learning community.**
We regret that the reviewer finds the description insufficiently accessible to the ML audience and take their criticism to heart. We have made a concerted effort to present a clear explanation in Section ... | Summary: The paper compares two quantum learning paradigms: (i) the measure first protocols where the learner uses a priorly determined measurements to obtain some classical information about the training samples (shadow tomography), (ii) a fully quantum learner that is allowed to make measurements that depend on the o... | Rebuttal 1:
Rebuttal: We naturally find it disappointing that our work was evaluated with 1 out of 5. We are hopeful that the majority of this evaluation stems from misunderstandings, probably caused by our wording, which we can dispel and clarify. We will do our best to address the reviewer’s concerns and demonstrate ... | Summary: This paper details the difference of two approaches in the task of quantum state or quantum distribution identification. The first approach is the measure-first approach where form a given set of quantum samples, a randomized measurement is performed and then a classical representation is constructed. The seco... | Rebuttal 1:
Rebuttal: **My main concern in this work is that of concrete advancement. ... As such, is the novelty defensible?**
We appreciate the reviewer’s comment and would like to clarify that our contribution is both novel and nontrivial.
First, our work is motivated by practical ML scenarios, focusing on quantu... | null | null | null | null | null | null |
Does Low Rank Adaptation Lead to Lower Robustness against Training-Time Attacks? | Accept (poster) | Summary: This paper analyzes the training-time robustness of Low-Rank Adaptation (LoRA) when fine-tuning large language models, focusing specifically on (1) untargeted data poisoning attacks (e.g., label flips) and (2) backdoor attacks (trigger insertion). The core claim is that LoRA exhibits lower robustness than stan... | Rebuttal 1:
Rebuttal: We sincerely appreciate your enlightening reviews and the potential willingness of raising the score.
Below, we present a point-by-point response to address your concerns.
# Clarification of Misunderstandings
## "Seemingly" Discrepancy: "Would this finding contradict the theoretical results (Theo... | Summary: This paper investigates the impact of Low-Rank Adaptation (LoRA) on the training-time robustness (TTR) of LLMs, focusing on their resilience to data poisoning and backdoor attacks. The authors propose a theoretical framework for robustness analysis, leveraging tools from neural tangent kernel theory and inform... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful feedback and critical review of our work, especially the constructive suggestions regarding the organization and clarity of the paper.
While revising the submission is not permitted by ICML's rebuttal policy, we assure you that your suggestions will be care... | Summary: This paper makes a theoretical investigation into the security implications of LoRA’s low-rank structure during fine-tuning in the context of robustness against data poisoning attacks. The authors theoretical analysis shows that LoRA presents greater robustness compared to full-parameter fine-tuning (FFT), but... | Rebuttal 1:
Rebuttal: We sincerely appreciate your commendation on our work!
Regarding your suggestion on line 306, we acknowledge that the current phrasing lacks clarity. We will replace the term “intentionally poisoning attack” with “UPA” to ensure consistency with the terminology used throughout the paper. Thank yo... | Summary: This paper presents an extensive theoretical analysis of data poisoning attacks in the low rank adaptation phase, using neural tangent kernels as well as information theory to establish a link between LoRA structure and vulnerability to training attacks. The authors find that LoRA is more robust to backdoor at... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful feedback and constructive suggestions for improving our work. Below, we provide a point-by-point response to address your concerns:
# Response to Questions
1. *Clarifying Key Findings.* We will highlight our core conclusions in both the Introduction and Abst... | null | null | null | null | null | null |
LDMol: A Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Surpasses AR Models | Accept (poster) | Summary: The authors introduce LDMol, a latent diffusion model for generating molecules based on text inputs. By integrating a chemically informed autoencoder and utilizing contrastive learning, their proposed diffusion model ensures structural consistency in the latent space, addressing the problem of multiple SMILES ... | Rebuttal 1:
Rebuttal: We greatly appreciate Reviewer mgY4 for the review and thoughtful feedback. Below, we provide detailed point-by-point responses to address your remaining concerns.
**[mgY4] asked for additional metrics on text-to-molecule generation.**
As the reviewer suggested, we measured the uniqueness, novel... | Summary: This paper proposes a text-conditioned molecule (SMILES) generation model based on latent diffusion. LDMol learns a structurally informative latent space through contrastive learning, and surpasses AR models. The authors claim LDMol is the first diffusion model outperforms AR models in text-to-mol generation.
... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and for considering our work, and we appreciate your time and evaluation. Below we provide point-by-point responses on your questions and concerns.
**[qAD8] asked for visualization of the latent space and its structural information.**
To visualize the struc... | Summary: The paper proposes a latent diffusion model for text-conditioned molecule generation. The authors claim that their primary innovation lies in introducing a contrastive learning approach to capture molecular structural features from SMILES sequences. In tasks such as molecule-to-text retrieval and text-guided m... | Rebuttal 1:
Rebuttal: We appreciate the reviewer for your detailed comments and valuable suggestions. Below, we provide thorough point-by-point responses to address your concerns.
**[bRG5] asked how the proposed contrastive learning can learn the structural differences.**
We would like to remind the reviewer that in... | Summary: - This paper proposes a SMILE-based latent diffusion method for text-driven molecule generation task.
This paper augments the data by aligning enumerated SMILES with the traditional contrastive learning approach at the pretraining stage to ask the model to learn the invariant features from SMILES.
With the pre... | Rebuttal 1:
Rebuttal: We thank Reviewer YcAV for the constructive feedback. Below, we provide point-by-point responses on your questions and concerns.
**[YcAV] asked for the statistics of the enumerated SMILES.**
Please note that the SMILES enumeration we’ve introduced is done by different SMILES constructions of the... | null | null | null | null | null | null |
Scalable Non-Equivariant 3D Molecule Generation via Rotational Alignment | Accept (poster) | Summary: This paper proposes Aligned Latent Diffusion Model (ALDM) to improve 3D molecule generation by relaxing SE(3) equivariance constraints in diffusion models. Instead of enforcing equivariance, ALDM learns a sample-dependent SO(3) transformation using an autoencoder to align molecular representations, enabling th... | Rebuttal 1:
Rebuttal: We thank Reviewer bfzT for taking the time to review our paper and provide valuable feedback. We respond to the concerns as follows.
**Not SOTA on QM9**: Our paper mainly aims to improve non-equivariant diffusion models. On QM9, compared with the best non-equivariant baseline GraphLDM-aug, all th... | Summary: This paper proposes learning a roto-aligned latent space for 3d molecule generation. The alignment is achieved through learning sample-wise rotation with auto-encoder. More specifically, ALDM adopts an equivariant encoder and a non-equivariant decoder, largely alleviating the constraints in architecture design... | Rebuttal 1:
Rebuttal: We thank Reviewer Biv1 for taking the time to review our paper and provide valuable feedback. We are glad to see that they found our approach interesting and our results good enough to support our claims. We respond to the concerns as follows:
**Lack of stronger baselines**: We thank the reviewer... | Summary: The paper explores a non-equivariant alternative to protein generation. It uses an explicit rotation network to rotate zero-centered molecule coordinates and builds a latent space on top of the rotated coordinates. The latent space is learned via VAE objective and encodes aligned features. This allows one to l... | Rebuttal 1:
Rebuttal: We thank Reviewer uMCt for taking the time to review our paper and provide valuable feedback. We respond to the concerns as follows:
**Scalability of non-equivariant diffusion models**: In fact, we did experiments to show the scalability of our model. We kindly refer the reviewer to Table 3 on pa... | Summary: The paper suggests learning a sample-dependent rotational transformation during molecule generation. This approach aligns the molecules to specific directions, eliminating the necessity of employing equivariant models. The proposed method demonstrates promising performance and efficiency on benchmark datasets.... | Rebuttal 1:
Rebuttal: We thank Reviewer t7B8 for taking the time to review our paper and provide valuable feedback. We respond to the concerns as follows.
**Relation to canonicalization**: We thank the reviewer for pointing this out! We admit that we overlooked the literature on canonicalization, and we will add them ... | null | null | null | null | null | null |
Curvature Enhanced Data Augmentation for Regression | Accept (poster) | Summary: The paper proposed a new data augmentation approach called Curvature-Enhanced Manifold Sampling (CEMS) specifically for the regression task. It moved one step further by utilizing the second-order representation instead of a first-order approximation of the data manifold. Experiments are conducted on both in-d... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful feedback. We’re glad the paper was found clear and the core contribution appreciated. Your comments on method design and evaluation helped improve the revised manuscript. Below, we address each point.
[Additional tables](https://jmp.sh/GDqh987h)
1. **The contribution... | Summary: This paper targets the problem of data augmentation for regression tasks where data has some intrinsic manifold structure. Specifically, the goal is to capture this manifold structure and generate new data on this manifold. Local neighborhoods are formed through nearest neighbor algorithms. The tangent space i... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and constructive feedback. We are glad that the core methodology and experimental evaluations were found to be clear and well-executed. We appreciate the suggestions to improve the discussion of intrinsic dimension estimation and to better situa... | Summary: This paper proposes a data augmentation method tailored for regression problems, leveraging the manifold hypothesis in the joint input-output space. The method approximates the data manifold up to the second order and samples new data points that adhere to this approximation. The effectiveness of the approach ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback. We're glad the method’s empirical strength and relevance were recognized. We addressed the comments on hyperparameters, differentiability, and implementation through added clarifications and revisions. Below are our detailed responses.
[Additio... | Summary: Presents a data-augmentation method for regression problems, taking advantage of the manifold structure of the data. Defined on the concatenation of data and labels, the local neighbourhood of all points defines the assumed manifold structure, and the two first moments (mean and Hessian) are used to sample poi... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful and encouraging feedback. We appreciate the recognition of our method’s integration of manifold learning and regression, as well as the clarity of our experiments and supplementary material. Below, we address the comments and describe the corresponding revi... | null | null | null | null | null | null |
VersaPRM: Multi-Domain Process Reward Model via Synthetic Reasoning Data | Accept (oral) | Summary: The paper introduces VersaPRM, a multi-domain Process Reward Model (PRM) designed to improve reasoning abilities across diverse domains beyond mathematics. Traditional PRMs have been primarily trained on mathematical reasoning tasks and fail to generalize effectively to other disciplines such as Law, Philosoph... | Rebuttal 1:
Rebuttal: We thank the reviewer for your meaningful feedback and recognizing that (i) our paper presents strong empirical results and robust experiments, (ii) it addresses an important gap in PRM research, (iii) it provides thorough comparison between math PRMs and VersaPRM, and (iv) it facilitates open sci... | Summary: This paper introduces VersaPRM, a multi-domain Process Reward Model (PRM) designed to enhance reasoning capabilities across diverse domains beyond mathematics. The authors identify that current PRMs are predominantly trained on mathematical data and demonstrate poor generalization to non-mathematical domains. ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful feedback. We've addressed your concerns below.
> Multiple test-time computation methods (MV, WMV, BoN, Beam Search, MCTS) but computational costs across these methods are not controlled
We clarify the computational costs of test-time methods: WMV and BoN ... | Summary: Process Reward Models (PRMs) have been effective in improving mathematical reasoning for Large Language Models (LLMs) by utilizing increased inference-time computation. However, their generalizability to non-mathematical domains remains unproven. This work demonstrates that current PRMs perform poorly in non-m... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback, and for acknowledging that (i) our method fills a gap in the existing literature on PRMs, (ii) we provide comprehensive evaluation criteria for our model, and (iii) our open-sourced code facilitates future exploration of this field.
> Although the automatic... | Summary: This paper describes an automated pipeline for annotating chain-of-thought rationales with stepwise correctness labels. The authors generate a set of these annotated CoTs for MMLU-Pro, then train a model to predict these correctness labels. The result is VersaPRM, a process reward model for reasoning beyond ma... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful feedback on the paper, recognizing that (i) its claims are well supported and (ii) this work is the first attempt to broaden PRMs beyond the math domain. We will incorporate your suggested revisions into our final paper.
> The comparisons the authors carry... | null | null | null | null | null | null |
A Reductions Approach to Risk-Sensitive Reinforcement Learning with Optimized Certainty Equivalents | Accept (poster) | Summary: The paper propose two meta-algorithms for solving static risk sensitive MDPs by using the augmented formulation for OCE objectives. The first one uses models based oracles and the other one using policy optimization based oracles. Regret bounds are analyzed with the assumption that the oracles are reasonably w... | Rebuttal 1:
Rebuttal: Dear Reviewer 74co,
Thank you for your encouraging review and strong support! We truly appreciate the time and effort that have been invested in providing constructive comments.
Regarding your question about notation on the value function:
* $\hat V_{1,k}$ is an estimated augmented value functi... | Summary: In this paper, the authors study the risk-sensitive RL problem to optimize a risk-measure of cumulative rewards. A family of risks called Optimized Certainty Equivalents (OCEs) are considered, and this includes popular risk measures such as CVaR and entropic risk. Two algorithms have been proposed in the paper... | Rebuttal 1:
Rebuttal: Dear Reviewer t1og,
Thanks very much for providing helpful feedback. We truly appreciate the time and effort that have been invested in providing constructive comments. Please find our responses below.
1. Our assumption to normalize cumulative rewards is actually *more general* than normalizing ... | Summary: This work studies risk-sensitive reinforcement learning, where the target is to maximize $\max_{\pi} \max_{b} \\{b+E_{\pi}[u(\sum_{h=1}^H r_h - b)]\\}$ where $u$ is some utility function. For this problem, the authors prove that by augmentation, there exists a Markovian policy which reaches the optimality. Bas... | Rebuttal 1:
Rebuttal: Dear Reviewer ZPJ1,
Thanks very much for providing helpful feedback. We truly appreciate the time and effort that have been invested in providing constructive comments. Please find our responses below.
**Reviewer: The empirical results are too minor to support the claim that the augmentation met... | Summary: **I am very unfamiliar with this topic. I will maintain the lowest confidence level.**
This paper develops a study on risk-sensitive RL, which is formulated through OCE. Two meta algorithms are proposed with further analysis.
Claims And Evidence: The paper propose an augmented MDP for the OCE problem, which ... | Rebuttal 1:
Rebuttal: Dear Reviewer 5fhe,
Thanks for providing your helpful feedback. We truly appreciate the time and effort that have been invested in providing constructive comments. Please find our responses below.
**Reviewer: The paper could benefit from more explanation. For example, in eq (1) defining OCE, it ... | Summary: This paper studies risk-sensitive reinforcement learning (RSRL) with the goal of learning a history-dependent policy that optimizes Optimized Certainty Equivalents (OCE) risk measures of cumulative rewards. The authors propose two meta-algorithms, one based on optimism and another on policy gradients, that red... | Rebuttal 1:
Rebuttal: Dear Reviewer 1jGh,
Thanks very much for providing helpful feedback. We truly appreciate the time and effort that have been invested in providing constructive comments. Please find our responses below.
**Reviewer: the constant $V^{\max}_u$ is not defined formally**
**Authors' reply:** We kindly... | null | null | null | null |
Correlated Errors in Large Language Models | Accept (poster) | Summary: This paper investigates the extent of correlation across large language models (LLMs) and its implications for systemic bias and multi-model collaboration. Key factors influencing this correlation include shared base model architecture and development organization. The study highlights that larger, more accura... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and positive review. Given that you didn’t have any questions, we will be very brief in our response.
One thing you point out is that leaderboard metrics do not fully represent real-world performance. We appreciate this comment, and will make note of it more clearly... | Summary: This paper investigates the correlated errors of different LLMs, which I find interesting and novel. The authors conduct extensive experiments analyzing the performance of different LLMs, revealing a high correlation in model errors. The experimental results largely support the core claims of the paper, and th... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review and comments. We’re glad that you found the experiments to be extensive, and supportive of our primary claims, and that you found the paper “interesting” and “highly relevant to the existing LLM research.”
We appreciate the comments you give, which we discuss b... | Summary: This paper studies how correlated the mistakes of large language models (LLMs) are across two public leaderboards. This study considers a large range of LLMs and a large set of questions. The findings are that there is substantial correlation in model errors. A regression analysis suggests that high agreement ... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful and detailed review. We are happy to see that you found the paper to be a significant contribution to the LLM behavior and monoculture literature, and that you found the analyses to be interesting and thorough.
We now discuss several of the points you raise, which we ... | Summary: The paper studies agreement of LLMs on samples they make mistakes on, showing model pairs have well above chance error correlation. It shows that models from the same developer, and of higher accuracies, make more correlated errors. It then measures effects on candidates when LLMs are used to score CVs, simula... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review. We’re glad that you found our claims well substantiated. We appreciated the questions you raised, which we aim to address below:
**“How would figure 2 look if we replaced qwen 2.5 72b with other llms?”** Thanks for this question. We extended this analysis t... | null | null | null | null | null | null |
Explicit Discovery of Nonlinear Symmetries from Dynamic Data | Accept (poster) | Summary: The paper proposes a method for detecting non-linear symmetries in data. In a nutshell, the authors solve this by first defining a base for the Lie algebra space (which is the space of infinitesimal generators), then solve the matrix equation of which of the linear combinations of these generators align with t... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and valuable feedback! Below we will address each of your concerns point by point.
**Theoretical Claims**
Taking the discovery of linear symmetries in the Heat equation $u_t=u_{xx}$ as an example. We specify the function library as $\Theta=[t,x,u]^T\in R^{3\tim... | Summary: This paper introduces LieNLSD, a novel method for discovering nonlinear symmetries from trajectory data. It addresses the limitations of previous methods that primarily focus on linear symmetries. LieNLSD aims to determine the number of infinitesimal generators and their explicit expressions. The method involv... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and valuable feedback! Below we will address each of your concerns point by point.
**Methods And Evaluation Criteria**
(1) See point (1) in Methods And Evaluation Criteria section of Rebuttal to Reviewer 1vcx.
(2) According to the procedure in Section 2.4 of t... | Summary: The paper introduces a novel data-driven PDE symmetry discovery. In contrast to previous arts based on end-to-end symmetry discovery (Ko et al., Yang et al.,), the proposed methods applies a post-hoc nullspace analysis on the learned PDE operator to discover the subspace of infinitesimal symmetry generators th... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and valuable feedback! Below we will address each of your concerns point by point.
**Claims And Evidence**
Note that in our paper, the symmetry is defined on $X\times U$ (see the "Symmetries of differential equations" paragraph in Section 2, lines 82-88, column... | Summary: In this paper, the authors proposed LieNLSD, a method for explicitly discovering nonlinear Lie group symmetries that are not represented by $\mathrm{GL}(n)$, from data governed by a PDE $ F(x, u^{(n)}) = 0$. The proposed method parameterizes a nonlinear infinitesimal group action $\mathbf{v}$ acting on $(x, u)... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and valuable feedback! Below we will address each of your concerns point by point.
**Methods And Evaluation Criteria**
(1) See the Methods And Evaluation Criteria section of the Rebuttal to Reviewer 1vcx.
(2) We conduct ablation experiments using a limited $\T... | null | null | null | null | null | null |
A Generic Family of Graphical Models: Diversity, Efficiency, and Heterogeneity | Accept (poster) | Summary: To infer high dimensional graphical models with a variety of data types, this paper introduces a marginally recoverable parametric family. The family is very flexible, and it includes Gaussian, PLN, and latent Gaussian copula for count and binary data. Within this family, the joint distribution can be characte... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time in reviewing our paper and your insightful comments. In the following response, we would like to address your major concern and provide additional clarification.
For weakness:
> The author(s) write the family in a very general way, but the examples are more or l... | Summary: The paper introduces a novel family of graphical models, termed the marginally recoverable parametric family, to address diversity, efficiency, and heterogeneity in high-dimensional network inference. The paper proposes a Maximum Marginal Likelihood Estimator (MMLE) for efficient parameter estimation and exten... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time in reviewing our paper and your positive feedback. Here we address your comments as follows.
For weakness:
>Marginal recoverability may not extend to distributions with higher moments.
Thanks for your comment. To facilitate an intuitive network representation, w... | Summary: This article introduces a new class of graphical models, referred to as the marginally recoverable parametric family, which aims to tackle challenges related to efficiency and heterogeneity in high-dimensional network inference tasks. The proposed family is rather flexible, with the joint distribution characte... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time in reviewing our paper and your insightful comments. In the following response, we would like to address your major concern and provide additional clarification.
For theoretical claims:
> I believe that neither Theorem 4.2 nor Theorem 4.3 should be labeled as "T... | Summary: The paper proposes a new, unified class of graphical models—called the marginally recoverable parametric family—designed to handle diverse data types (e.g., Gaussian, Poisson log‑normal, and latent Gaussian copula models) and heterogeneous structures via mixture modeling. The authors introduce an efficient max... | Rebuttal 1:
Rebuttal: We sincerely appreciate your time in reviewing our paper and your insightful comments. In the following response, we would like to address your major concern and provide additional clarification.
For your major concern in the claims and evidence section, as well as the similar concern raised in t... | null | null | null | null | null | null |
Aligning Spoken Dialogue Models from User Interactions | Accept (poster) | Summary: This paper introduces an alignment framework for full-duplex spoken dialogue models (like Moshi). The authors construct preference data from real user interaction data and then fine-tune a spoken dialogue model (Moshi) using direct preference optimization (DPO). Preference pairs are constructed by using a mode... | Rebuttal 1:
Rebuttal: We are thankful for your time and careful feedback. We answer specific points below:
## Question
> In Equation (2), [...], a sequential relationship.
The model’s text and audio must have some dependency pattern, otherwise the two streams would quickly have different content.
Other patterns ha... | Summary: This paper integrates DPO and its variants into Moshi, a full-duplex voice interaction model, to enhance several aspects such as content and timing. To accomplish this, the authors collected dataset and used it for training and evaluation. Notably, aside from concurrent research, all published studies on end-t... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thorough feedback, and for acknowledging that our work presents a novel contribution for the alignment of voice interaction models.
## Questions & Concerns
> Loss of context-based prosody
We agree that discarding the original audio (for privacy) loses some audio in... | Summary: This work introduces a framework for aligning real-time, full-duplex spoken dialogue systems using user spoken interactions (building on Moshi system from kyute.ai). Unlike existing preference learning methods focused on text-based models, this approach addresses the complexities of dialog speech, such as inte... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive assessment and appreciate the insightful and careful feedback. We respond to the points below.
## Questions
> section 3.3: [...] how do you derive preference data from this? [...] unclear to me.
Thank you for the question. For **content-related** problems:... | Summary: The paper describes an approach to aligning the Moshi spoken dialog model to preference data automatically derived from human-model interactions. The preferences are elicited from transcripts of the spoken interactions via a textual LLM (Mistral Large 2). Context and responses are (re)synthesized via TTS and ... | Rebuttal 1:
Rebuttal: We appreciate that the reviewer finds our work valuable, and thank the reviewer for the constructive remarks.
We answer specific points below.
## Question
> 376: It wasn't clear to me why the preference data doesn't fully transfer between different voices [...] this problem.
The preference data... | null | null | null | null | null | null |
On the Role of Label Noise in the Feature Learning Process | Accept (poster) | Summary: This paper provides a theoretical analysis of how label noise impacts the feature learning process in deep neural networks. The authors prove a two-stage learning dynamic for networks trained with label noise:
1. Stage I – The model first learns from clean samples while ignoring noisy ones, leading to good ge... | Rebuttal 1:
Rebuttal: Thank you for your great efforts on the review of this paper and for recognizing the value of our contributions. We will try our best to address your questions.
**Q1: Suggestions for more explanations of why the model tends to prioritize learning from clean-labeled samples. “However, its impact c... | Summary: The paper is concerned with a theoretical analysis of gradient descent when the training samples contain label noise as well as features uncorrelated to the correct class. The chosen methodology is feature learning theory, where highly simplified learning problems are introduced in which the features carrying ... | Rebuttal 1:
Rebuttal: Thank you for your great efforts on the review of this paper and for appreciating our novelty and contributions! We will try our best to address your questions.
**Q1: Concerns about the presentation. “The formalism is well designed and it is possible to follow, although not too easily, eg the two... | Summary: This paper analyzes the training dynamics of a (custom) two-layer ConvNet under binary class-conditional label noise. The main result is a a two-stage characterization of "first fitting all clean samples, then overfitting to noisy samples".
Claims And Evidence: Yes, the claims were proved.
Methods And Evalua... | Rebuttal 1:
Rebuttal: Thank you for your great efforts on the review of this paper and for recognizing the value of our contributions. We will try our best to address your questions.
**Q1: Suggestion for acknowledgement in setup part. “The setup is built upon Kou et al. (2023), it would be good to indicate that in the... | null | null | null | null | null | null | null | null |
LV-XAttn: Distributed Cross-Attention for Long Visual Inputs in Multimodal Large Language Models | Accept (poster) | Summary: 1. The paper presents LV-XAttn, a distributed cross-attention mechanism designed to address the high memory demands and communication overheads in multimodal large language models (MLLMs) when processing large visual inputs.
2. LV-XAttn reduces communication volume by keeping large key-value blocks locally on... | Rebuttal 1:
Rebuttal: Thank you for your time and feedback.
**Q1: Limited model scope.**
A1: Please refer to Q2 of Reviewer zbCn.
**Q2: Lack of statistical significance testing.**
A2: All runtime results are averaged over five iterations, excluding the first two warm-up iterations.
**Q3: Limited discussion of trad... | Summary: Cross-attention layers takes a significant memory for applications involving large visual inputs, such as long video understanding, making scaling difficult due to high memory requirements. To address this, the authors present LV-XAttn, a distributed and exact cross-attention mechanism that leverages sequence-... | Rebuttal 1:
Rebuttal: Thank you for your detailed questions and useful feedback!
**Q1: Which datasets were used for training and evaluation?**
A1: For our experiments, we do not pretrain or finetune the models. Instead, we use the checkpoints from pretrained models and replace their cross-attention operations with LV... | Summary: The paper introduces LV-XAttn, a distributed cross-attention mechanism designed to handle large visual inputs in Multimodal Large Language Models (MLLMs) with minimal communication overhead. Cross-attention is commonly used in MLLMs to integrate visual information into the language backbone, but processing l... | Rebuttal 1:
Rebuttal: Thank you for your insightful questions and feedback!
**Q1: Details on implementation.**
A1: LV-XAttn is implemented using PyTorch and Triton. It uses `torch.distributed` for distributed communication, while the modified FlashAttention kernels to account for rescaling operations are implemented ... | Summary: This paper presents LV-XAttn, a distributed cross-attention mechanism with low communication overhead that can significantly reduce the inference time of MLLMs adopting cross-attn strategy. The authors also introduce the activation recomputation technique enabling supporting longer visual inputs. The proposed ... | Rebuttal 1:
Rebuttal: Thank you for the constructive feedback and questions!
**Q1: Llama-3V is not discussed in the paper.**
A1: Thank you for pointing this out! Llama-3V [1] also utilizes a cross-attention architecture, allowing LV-XAttn to be applied to it for large visual inputs. We conducted additional experiment... | null | null | null | null | null | null |
Rethinking Chain-of-Thought from the Perspective of Self-Training | Accept (poster) | Summary: This paper investigates CoT approach and proposes a novel CoT framework inspired by its similarity to self-training to enhance reasoning performance. The framework consists of two core modules: a task-specific prompt module that optimizes the initial reasoning process and an adaptive reasoning iteration module... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the valuable feedback and encouraging comments. We are motivated by the suggestions and have addressed each concern in detail as follows:
> **Q1.** There is a lack of discussion on negative results. While the paper primarily demonstrates the advantages of the m... | Summary: Inspired by self-training, this paper designs the CoT framework to improve reasoning performance. It contains two core elements: a task-specific prompt and an adaptive reasoning iteration. This paper finally conducted experiments on 10 reasoning datasets and achieved improved results.
Claims And Evidence: I t... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the valuable feedback and encouraging comments. We are motivated by the suggestions and have addressed each concern in detail as follows:
> **Q1.** Main experiments should compare both effectiveness and token overhead of each method.
**A1.** We have added toke... | Summary: This paper explores the conceptual similarity between CoT reasoning and self-training, highlighting their shared goal of minimizing predictive uncertainty by iteratively leveraging model-generated information. Based on this insight, the authors propose a novel CoT framework integrating a Task-Specific Prompt m... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the valuable feedback and encouraging comments. We are motivated by the suggestions and have addressed each concern in detail as follows:
> **Q1.** Although the authors propose a method for automatically searching for the "optimal prompt," they do not provide s... | Summary: This paper discusses the similarities between Chain-of-Thought (CoT) reasoning and self-training, and proposes how to reduce prediction uncertainty for both iteratively leveraging on model-generated information. In particular, this paper introduces a novel CoT framework with two main ingredients: (i) a task-sp... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the valuable feedback and encouraging comments. We are motivated by the suggestions and have addressed each concern in detail as follows:
> **Q1.** The paper lacks sensitivity analysis for the sampling number $N$.
**A1.** We evaluated the impact of $N$ on four... | null | null | null | null | null | null |
Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs | Accept (poster) | Summary: This paper presents Putnam-AXIOM, a benchmark of 522 problems from the Putnam competition along with their ground truth solutions. The paper also proposes some manual modification to the original problems to make evaluation less ambiguous, as well as manual variable and constant substitutions for 100 problems ... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and detailed response. We deeply appreciate the time and effort you invested in providing such valuable feedback.
> I am under the impression … not proof-based.
We'd like to clarify that Putnam-AXIOM explicitly addresses the challenge of evalu... | Summary: This paper introduces a new mathematical benchmark made of 522 questions from the William Lowell Putnam math competition from 1938 to 2023, among which 100 can be infinitely modified by changing variable names (100 of them) and constant values (37/100 of them) – called the Variation split.
This paper also pres... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and detailed response.
> “This is not the correct approach as you already have a metric that will correlate 100% of the time with accuracy, it is accuracy itself. The whole point of using proxy metrics is to evaluate something else than accurac... | Summary: The paper introduces a new dataset comprised of mathematical problems that have appeared at the Putnam contest for university students. To enable automatic evaluation, the problems are selected or have been reprhased to be such that the solution can be checked automatically as a boxed answer. This rephrasing o... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and detailed response. We deeply appreciate the time and effort you invested in providing such valuable feedback.
> ***how are the Putnam problems selected exactly?***
As outlined in Section 3.1, we selected problems based on two primary crite... | Summary: Putnam-AXIOM is a new benchmark designed to assess higher-level mathematical reasoning in large language models (LLMs), using 522 challenging problems from the William Lowell Putnam Mathematical Competition. To address data contamination, the authors introduce functional variations of 100 problems by altering ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful and detailed review, particularly for recognizing that our work addresses an important gap by introducing a benchmark capable of measuring higher-level mathematical reasoning, and appreciating the thoroughness of our experimental designs and comprehensive... | null | null | null | null | null | null |
Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks' Internal Representations | Accept (poster) | Summary: The paper addresses a critical challenge in concept-based explanation methodology—specifically, the generation of "concepts" for explanations. Traditionally, this required practitioners to manually guess and collect various candidate concept image sets. The paper introduces a novel approach, utilizing reinforc... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and for recognizing the technical soundness and novelty of our framework. We appreciate your critical insights and provide detailed clarifications below.
Please refer to this **[Anonymized GitHub Link](https://anonymous.4open.science/r/RLPO-9577/Rqpq/readme.md... | Summary: This work introduces an RL-based method to construct a vision-language concept-level preference dataset purely from synthesized images by taking TCAV score as the reward. It first prompts the trained MLLM to ground the common concepts represented as language phrases in the images. Then, it generate preference ... | Rebuttal 1:
Rebuttal: Please refer to this **[Anonymized GitHub Link](https://anonymous.4open.science/r/RLPO-9577/ADMa/readme.md)** where we have compiled detailed explanations for better understanding.
We are thankful to the reviewer for the feedback. While we appreciate the reviewer's recognition of the proposed fra... | Summary: The paper reframes the concept set creation as a concept generation problems. It proposes a method based on generative model to generate concept images. It aims to create to reliably generate diverse concepts that are challenging to craft manually. The process involves various components including Reinforcemen... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s constructive comments and recognition of the importance of concept generation in our work. Our method improves upon traditional automatic concept retrieval approaches. While conventional methods extract concepts directly from the dataset—risking semantic information le... | Summary: The authors proposed a method to discover and visualize the "concept" or hidden knowledge learnt by a neural network. They proposed a reinforcement learning framework to achieve this goal. A score (TCAV) was proposed to evaluate whether a hidden representation of NN forms a concept.
Claims And Evidence: The a... | Rebuttal 1:
Rebuttal: Thank you for identifying the novelty and appreciating the experiments. We hope the following explanations will clarify the queries for the reviewer.
## Q1: It seems that the users need to have a good understanding of the target concept in order to generate good concept seeds (i.e, proper VQA de... | null | null | null | null | null | null |
Feature learning from non-Gaussian inputs: the case of Independent Component Analysis in high dimensions | Accept (spotlight poster) | Summary: This paper investigates the unsupervised learning method ICA as a simplified framework for feature learning in (deep) CNNs. In particular, the authors derive sample complexity thresholds for escaping the search phase of FastICA and SGD, considering a toy model of a dataset sampled from an isotropic distributio... | Rebuttal 1:
Rebuttal: Thank you for your careful feedback and for having checked both the numerical experiments and the proofs. Your suggestions will definitely help to clarify the paper.
> One potential weakness of this work is the relative simplicity of the model. However, I do not consider this a major weakness, as... | Summary: The study quantifies the sample complexity of two learning algorithms: FastICA, Stochastic Gradient Descent (SGD). The key results are the following: FastICA requires at least $n \gtrsim d^4$ samples to recover a single non-Gaussian direction in high-dimensional inputs. SGD outperforms FastICA in feature learn... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback and your useful comments, which spurred us to run an additional experiment to link ICA with deep CNNs, and to provide two additional plots to provide intuition on the effect of smoothing, and on the intermediate regime of FastICA. We start with these points bef... | Summary: Motivated by empirical observations that features learned by deep convolutional networks resemble those recovered by independent components analysis (ICA), this paper presents a concrete algorithmic sample complexity bound for various algorithms for recovering a non-Gaussian direction from $d$-dimensional data... | Rebuttal 1:
Rebuttal: Thank you for your accurate comments and for the attention to the supplementary material, including the strategies of the proofs. Your suggestions offered valuable food for thought.
> I don't see any glaring weaknesses to the paper, beyond the possible restrictiveness of the single non-Gaussian ... | null | null | null | null | null | null | null | null |
SPRI: Aligning Large Language Models with Context-Situated Principles | Accept (poster) | Summary: Large Language Models (LLMs) often require guiding principles to ensure their responses are well-aligned and contextually appropriate. While prior work has leveraged predefined principles or constitutions for synthetic data generation, these approaches often fail to adapt to situation-specific needs.
SPRI is a... | Rebuttal 1:
Rebuttal: Thank you for your positive comments on the innovation of SPRI, which dynamically generates context-adaptive principles to align LLMs while relying on minimal-to-no human supervision. We are also grateful for your acknowledgment of the experimental results — a strength that other reviewers also ap... | Summary: The proposed SPRI framework automates real-time generation of context-specific guiding principles for LLM alignment, minimizing reliance on human expertise while addressing the limitations of generic predefined rules. SPRI achieves performance on par with expert-crafted principles in domain-specific tasks.
Cl... | Rebuttal 1:
Rebuttal: We are grateful for your valuable feedback! We appreciate your recognition of the novelty and importance of SPRI in automating the alignment guidance per query to enhance the safety and reliability of LLMs with minimal human supervision, which other reviewers concurred with. Thank you also for poi... | Summary: This paper proposes a novel framework named SPRI for aligning LLMs with human preferences. The framework operates through a two-stage collaborative process between models:
1. A base model dynamically generates context-specific principles tailored to each input query, iteratively refined through feedback from ... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of the meaningful contribution SPRI makes toward aligning LLMs with context-situated principles while relying on little-to-no human effort. We are also grateful for your recognition of SPRI’s robustness, which is supported by detailed experimental results — a key st... | null | null | null | null | null | null | null | null |
B-score: Detecting biases in large language models using response history | Accept (poster) | Summary: This paper discusses the potential of multi-turn interaction with LLMs to quantify the bias in LLM's response better. Specifically, the proposed framework calculates the multi-turn appearance probability of the answers by repeating the same question multiple times in a single conversation. The difference betwe... | Rebuttal 1:
Rebuttal: Thank you for your suggestions!
**Summary:** We experimented with confidence baselines, highlighted our focus on detecting bias, and provided empirical evidence showing that LLMs can self-calibrate in multi-turn due to their inherent ability to do so.
> The experiments lack meaningful baselines.... | Summary: This paper proposes a new score (B-score) for estimating the degree of bias in a preferred LLM response. The key is to not rely on only a single sample output from the model with a self-reported confidence, but rather probe the model multiple times and estimate the preference for a particular response. The B... | Rebuttal 1:
Rebuttal: Thank you for your detailed and constructive feedback!
**Summary**: We've carefully extended the experiment on the BBQ bias benchmark to address concerns about dataset size, clarifying our results in `Tab. 3` vs `Tab. 4` and number of queries in single-turn/multi-turn, revised our writing (e.g.,... | Summary: This paper investigates biases in large language models (LLMs) and introduces a metric, the B-score, to quantify bias by comparing single-turn and multi-turn interactions. The authors identify that LLMs exhibit biases across various dimensions (e.g., gender, race, numbers, names) when repeatedly asked the same... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback!
**Summary**: We've extended our experiments to the BBQ benchmark to address the reviewer's concern about test size and clarified our rationale for using four distinct question categories to capture different aspects of bias.
> This is a relatively small d... | null | null | null | null | null | null | null | null |
Generalized Venn and Venn-Abers Calibration with Applications in Conformal Prediction | Accept (poster) | Summary: This paper proposes Venn and Venn-Abers Calibration defined as connected to the loss function. First, the isotonic regression and quantile losses are examined for the marginal calibration. Furthermore, conditional calibration is discussed. There are a series of theoretical results for the proposed algorithms.... | Rebuttal 1:
Rebuttal: Q1: By augmenting the dataset with all possible imputed outcomes for the test point whose outcome we wish to predict we are able to convert a point prediction into set prediction, where the width of this set captures epistemic uncertainty in the calibration process. This augmentation approach is t... | Summary: This paper introduces a unified framework for Venn and Venn-Abers calibration, generalizing Venn calibration to hold with respect to arbitrary given loss functions. Unlike point calibrators (e.g., histogram binning, isotonic regression), which map predictions to a single calibrated value, Venn calibration cons... | Rebuttal 1:
Rebuttal: Thank you for the helpful comments.
- Thank you for these thoughtful suggestions. We will add clarification on how the Sherman–Morrison approach can be used to efficiently implement the algorithm. We appreciate the request for more detail on the case of infinite $\mathcal{Y}$, and will revise th... | Summary: The authors propose a unified framework for Venn and Venn-Abers calibration that leverages binning calibrators to construct prediction sets that contain at least one marginally perfect calibrated prediction. Furthermore, they propose a novel method for Venn calibration technique across subpopulations. Their me... | Rebuttal 1:
Rebuttal: Thank you very much for your detailed comments and suggestions. We will incorporate them in the revised version of the paper.
We note our primary contribution is the generalization of Venn and Venn–Abers calibration to arbitrary loss functions. As a secondary contribution, we show that the same t... | Summary: Conformal Prediction arrises as a techniques for turning a point predictor to a set predictor, allowing for a guarantee that the set contains the ground-truth target with high probability. Venn-Predictor operates similarly but output a set of probabilistic prediction with the guarantee that at least one of the... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful questions and suggestions. We will take incorporate your suggestions in the revised version of our manuscript. We clarify the efficiency and practicality of our method below.
1. **Efficient Approximation and Scalability**
The algorithm can be efficiently approxim... | null | null | null | null | null | null |
LIMEFLDL: A Local Interpretable Model-Agnostic Explanations Approach for Label Distribution Learning | Accept (poster) | Summary: Existing interpretability models are designed for single-label paradigm and struggle to directly interpret label distribution learning (LDL) models. To solve this, the paper proposes an improved LIME algorithm capable of effectively interpreting black-box models in LDL. The authors also provide analysis on ana... | Rebuttal 1:
Rebuttal: Thank you for your positive review of our paper; we greatly appreciate your comments and questions.
[Comment 1] Steps 3, 4 in Algorithm 1 lack sufficient rigor.
A: We will introduce Steps 3, 4 of Algorithm 1 in more detail in the revised version. Step 3 of Algorithm 1 generates $m$ samples b... | Summary: This paper proposes an interpretability model for label distribution learning (LDL). The classification LIME approach is adapted to handle LDL. An optimization objective is proposed to estimate the parameters of the interpretability model. Theoretical analysis is performed, including convergence, stability, an... | Rebuttal 1:
Rebuttal: Thank you for your constructive review of our paper.
[Comment 1]About the structure of the proofs.
A:We will restructure proofs in the revised version. The proof of Theorem 3.2 is appeared in Appendix B.2 (lines 1025~1040).
[Comment 2]The method's novelty appears limited as it resembles a mu... | Summary: In order to mitigate the interpretability challenge inherent in most label distribution learning (LDL) algorithms when applied to risk-sensitive decision-making scenarios, this paper introduces a novel local interpretable model-agnostic explanation framework specifically tailored for LDL. This approach takes i... | Rebuttal 1:
Rebuttal: Thank you for your positive review of our paper; we greatly appreciate your comments and questions.
[Comment 1] This paper also has some limitations. For example, the writing of this paper needs improvement, and the experimental results in appendix need more discussions.
A: We will follow up... | Summary: To address the local interpretability issues of label distribution learning (LDL), this paper proposes an improved LIME algorithm, namely LIMEFLDL. The algorithm is mainly manifested in three aspects: first, by introducing a feature attribution matrix to address the label dependency issue in LDL tasks; second,... | Rebuttal 1:
Rebuttal: Thank you for your detailed review of our paper; we greatly appreciate your comments and questions.
[Comment 1]About the errors and unclear mathematical symbols.
A:
1. Equation 23 should be modified: $\sum_{k=0}^{f} e^{\frac{(k-f)}{\sigma^{2}}}\frac{k}{f} \frac{f!}{2^{f}(f-k)!k!}$, this e... | null | null | null | null | null | null |
Exponential Family Variational Flow Matching for Tabular Data Generation | Accept (poster) | Summary: The paper introduces the application of Variational Flow Matching to tabular data generation. To extend VFM, the authors propose to represent the variational distribution in VFM as an exponential family. The motivation behind this proposal stems from the heterogeneous nature of tabular data thus, the claim is ... | Rebuttal 1:
Rebuttal: Dear reviewer FzAE,
We thank you for your effort to review our work. Moreover, we appreciate you mentioning the originality of our exponential-family formulation and the value of Bregman divergence connections. We will reply to the points mentioned in the review here:
- Regarding **TabDiff and T... | Summary: In this work, the authors propose a new method called Exponential Family Variational Flow Matching which adds a variational formulation on top of VFM which helps them leverage sufficient statistics/moment matching procedure to obtain a probabilistic generative modelling framework. The exponential family perspe... | Rebuttal 1:
Rebuttal: Dear reviewer RH9Z,
First, we'd like to express our gratitude for the thorough and extensive feedback on the paper. Our response to the points raised is as follows:
- We agree that the **theoretical assumptions/limitations** should be made more explicit, especially as this seems to be a recurrin... | Summary: This paper proposes a new method that introduces variational flow matching to table generation. Specifically, it incorporates a function family, exponential family, for mapping the table data to the prior, which is a more general form for flow starting from widely used priors.
Claims And Evidence: The main cl... | Rebuttal 1:
Rebuttal: Dear reviewer NBdr,
Thank you for carefully examining and giving feedback on our work. We will reply to the points raised one by one:
- Regarding the **typos** in our work, that is poor validation on our end, thank you for pointing this out. We made sure that not only the bold fonts are now corr... | Summary: They propose TabbyFlow, a variational flow-matching method for generating mixed tabular data. An advantage over previous methods is that the exponential-family version allow modelling mixed data (continuous and categorical), and contrary to other methods even other types of data such as Poisson counts. The the... | Rebuttal 1:
Rebuttal: Dear reviewer rhRe,
Thank you for your thoughtful and constructive review. We are glad that the theoretical contributions and generality of the method came across clearly. We will reply to the points raised in the review pointwise:
- We fully agree that including **flow-matching baselines** is ... | null | null | null | null | null | null |
Online Detection of LLM-Generated Texts via Sequential Hypothesis Testing by Betting | Accept (poster) | Summary: The paper focuses on developing an algorithm for online detection of texts generated by
large language models (LLMs). The main contribution is an algorithm based on sequential
hypothesis testing techniques, which allows for quick and accurate identification of LLMgenerated texts in a streaming setting. The al... | Rebuttal 1:
Rebuttal: We first would like to thank the reviewer for the positive feedback and good suggestions. Here are our responses.
**(Supplementary material.)**
We would like to clarify that the supplementary material includes a detailed README.md file, which provides a clear overview of the codebase, experiment ... | Summary: The paper studies the problem of detecting whether a series of texts is LLM/machine-generated sequentially. To this end, they build on existing work on offline detection of LLM-generated text which proposes a variety of scoare functions, as well on recent work in sequential hypothesis testing. More concretely,... | Rebuttal 1:
Rebuttal: We first thank the reviewer for your positive feedback and helpful suggestions. Below are our responses.
**(Assumption on the score function.)**
We agree that the assumption of the existence of a score function that produces distinguishable means for human-written texts and LLM-generated texts is... | Summary: This work has studied an online detection method for AI-generated texts, and it can identify texts from unknown source models. The proposed method mainly makes use of sequential hypothesis testing and has an advantage of non-parametric property. Comparision experiments with several baseline methods (e.g. Detec... | Rebuttal 1:
Rebuttal: We thank the reviewer for your careful reading and useful comments. Here are our responses.
**(Necessity.)** Our emphasis is not on the frequency of adjustments, but rather on the fact that **most** offline detectors detect by comparing the text score with a **pre-determined** threshold, chosen ... | null | null | null | null | null | null | null | null |
Hyperspherical Normalization for Scalable Deep Reinforcement Learning | Accept (spotlight poster) | Summary: The main claim of the paper is that a novel architecture (SimbaV2) can improve the scaling of RL algorithms. The benefits of using SAC and SimbaV2 are demonstrated across a variety of domains.
Claims And Evidence: I find that the authors do an excellent job at demonstrating the empirical performance benefits ... | Rebuttal 1:
Rebuttal: Dear reviewer fmAi,
Thank you for your thoughtful and constructive feedback. We address your concerns below and would be happy to clarify further.
> **Question 4.1**
The scalability claims are not fully justified:
(1) Only SAC is tested, (2) only width scaling is shown, (3) UTD scaling shows lim... | Summary: This paper introduces SimbaV2, an RL architecture that improves scalability and stability in deep RL. The authors use hyperspherical normalization to control weight and feature norm growth, alongside distributional value estimation with reward scaling to maintain stable gradients. Using SAC as the base algorit... | Rebuttal 1:
Rebuttal: Dear Reviewer 4564,
Thank you for suggesting future research direction! We respond to each of your comments below and are happy to clarify further if needed.
> **Question 3.1**
The claim that this leads to 'scalable' RL remains incorrect, as scaling limits are reached very quickly (See Fig. 5).... | Summary: This paper proposes SimbaV2, an improved version of Simba, by replacing several key components of Simba with a scale-preserved l2-normalization (i.e., hyperspherical normalization), distributional value function approximation, reward scaling, etc. The authors presents a comprehensive experimental study with 57... | Rebuttal 1:
Rebuttal: Dear Reviewer ZJka,
Thank you for your constructive feedback and positive support! We respond to each of your points below and would be happy to clarify further if needed.
> **Q2.1:**
> The necessity or effects of the design changes “Linear → Linear + Scaler” and “Residual Connection → LERP” ar... | Summary: The paper introduces SimbaV2, an RL architecture that stabilizes training and improves scalability through hyperspherical normalization and distributional value estimation with reward scaling. Built on Soft Actor-Critic (SAC), it achieves state-of-the-art performance across 57 continuous control tasks and scal... | Rebuttal 1:
Rebuttal: Dear reviewer R7VW,
Thank you for your thoughtful and constructive feedback. We address your concerns in detail below and would be happy to clarify any remaining questions.
> **Question 1.1**
It would be helpful if the authors could provide more theoretical justification or intuitive explanat... | null | null | null | null | null | null |
Hardware and Software Platform Inference | Accept (poster) | Summary: This manuscript proposed a method called Hardware and Software Platform Inference (HSPI), aiming at identifying the GPU and software stack for machine learning models. The authors introduced 2 methods:
1. HSPI with Border Inputs (HSPI-BI), which is building inputs that are at the decision boundary of the mode... | Rebuttal 1:
Rebuttal: Thank you for offering valuable questions. We will address them one by one.
> # Claims And Evidence
> No, the proposed method is called "hardware and software platform inference", which I doubt is kind of big regarding the manuscripts, which is mostly focusing on GPU types and data types (like in... | Summary: The paper introduces Hardware and Software Platform Inference (HSPI), a novel method for identifying the underlying GPU architecture and software stack of machine learning models based on their input-output behavior. HSPI uses computational differences across various GPUs and software environments to detect th... | Rebuttal 1:
Rebuttal: Thank you for valuable suggestions and questions. We would like to address them one by one.
> # W1. Large models and systems
> The paper states that limited GPU memory would constrain the ability to scale to larger language models; even with HSPI-LD the experiments were conducted on rather small... | Summary: The paper presents an interesting idea where a client can infer the hardware and software platform that was used for model inference based on its input and output behavior. It banks on the observation that there are inherent differences between different GPUs and software stack. The idea has potential to allow... | Rebuttal 1:
Rebuttal: Thank you offering valuable suggestions and questions. We will address them one by one. Due to word limits, we uploaded **three new result tables** here: https://imgur.com/a/VbONCoU
> # E1. Essential References Not Discussed
Our work is different from these two papers in terms of goals, threat m... | Summary: This paper introduced Hardware and Software Platform Inference (HSPI), which is a method for identifying the hardware and software stack based on the input-output behavior of machine learning models. The proposed method leverages the inherent differences of various GPU architectures and compilers to distinguis... | Rebuttal 1:
Rebuttal: Thank you for offering valuable suggestions and questions. We would like to address them one by one.
> S1. The authors shared their code to training the classifier and run inference. We don't have the logit dataset to run but it looks legit.
- We know collecting the logits from various combinat... | null | null | null | null | null | null |
Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting | Accept (poster) | Summary: The paper introduces a Conditional Diffusion Model (CDM) for generative modeling, leveraging denoising diffusion probabilistic models (DDPMs) to generate high-quality samples conditioned on specific inputs. The key contribution is a conditioning mechanism that guides the diffusion process, allowing for control... | Rebuttal 1:
Rebuttal: ### **Methods And Evaluation Criteria:**
**Q1: But the paper lacks efficiency comparisons to assess computational cost**
**A1:** We now present the computational cost analysis to compare our model with other diffusion-based time series forecasting methods. Refer to the tables below for training... | Summary: This paper proposes an approach for time series forecasting using conditional diffusion models. The approach utilizes a learnable forward process, making the transition operator and ending points of the forward process learnable and controlled by the conditioning observations. It derives a non-Markovian revers... | Rebuttal 1:
Rebuttal: ### **Claims and Evidence:**
**Q1: In the first Paragraph $\cdots$ . I would $\cdots$ straight reverses processes.**
**A1:**
**In the first $\cdots$ untrainable.**
Our claim follows the well-known observation that incorporating a flexible forward process can enhance performance and address th... | Summary: This paper introduces CN-Diff, a novel conditional diffusion model tailored for time-series forecasting. The core idea is to integrate a nonlinear data transformation and a learnable condition within the forward process of diffusion, as opposed to more conventional diffusion-based approaches that use only a fi... | Rebuttal 1:
Rebuttal: ### **Other Strengths And Weaknesses:**
**Q1: It would be better if the authors can also compared the computational and memory complexity of the CN-Diff with other methods.**
**A1:** We now present the computational cost analysis to compare our model with other diffusion-based time series fo... | Summary: The paper presents a conditional diffussion method for time series forecasting. The proposed method CN-Diff adds a non-linear transformation in the forward process. This transformation is a learnable paramter and results in non-Markovian series of latent variables. These latent variables are learned in the re... | Rebuttal 1:
Rebuttal: ### **Questions For Authors:**
**Q1: The proposed method is claimed to do a better job of learning the patterns in the time series.Will not the use of synthetic data will demonstrate better the effectiveness of the method?**
**A1:** We agree with your reasoning that the structured nature of synth... | null | null | null | null | null | null |
A Unified Framework for Generalization Error Analysis of Learning with Arbitrary Discrete Weak Features | Accept (poster) | Summary: The paper presents a unified formalization and theoretical analysis of discrete Weak Features Learning (WFL) to handle learning with arbitrary discrete weak features (WFs). It introduces a set of algorithms that jointly learn the estimation model for WFs and the predictive model for a downstream task while con... | Rebuttal 1:
Rebuttal: Thank you for your careful evaluation and valuable suggestions to improve our paper.
**Comparison with Benchmark Methods:**
The most relevant benchmark is using WFs directly as input features, because our framework is designed to improve upon this approach by accommodating various learning metho... | Summary: In this paper, the authors propose a unified framework called weak feature learning which accommodates arbitrary discrete weak features and a broad range of learning algorithms. The authors also introduce a class of algorithms that learn both the estimation model for weak features and the predictive model for ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and thorough review of our manuscript, and for pointing out areas for further clarification and improvement.
**Necessity of Theorem 3.1 Given Lemma 4.1 (about Question 1.):**
We agree that Lemma 4.1 provides a tighter bound and that Theorem 3.1 can be derived from i... | Summary: The paper presents a unified framework called WFL for analyzing generalization error in learning tasks involving arbitrary discrete WFs. The authors propose a risk-based formulation that accommodates various types of WFs and a broad range of learning algorithms.
## update after rebuttal
The author's response... | Rebuttal 1:
Rebuttal: We sincerely appreciate the time and effort the reviewer dedicated to evaluating our work and are grateful for your insightful feedback and constructive suggestions.
**The Impact of Different Types of WFs on the Learning Process and Experimental Datasets:**
Our framework and analysis hold regard... | Summary: This paper introduces a unified framework for Weak Feature Learning (WFL), which aims to address the challenge of learning with arbitrary discrete weak features (WFs)—features that are incomplete, erroneous, or ambiguous due to various real-world constraints. The authors propose a risk-based formulation that j... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and constructive feedback on our work.
**Extending the Framework to Handle Continuous Weak Features:**
Extending weak features learning (WFL) to continuous weak features (WFs) is an important research direction. We have investigated this and found that repla... | null | null | null | null | null | null |
Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications | Accept (poster) | Summary: This paper proposes Poly2Vec, a Fourier-transform approach to encoding shapes (points, lines, polygons) for geospatial tasks. The authors find that Poly2Vec outperforms baselines in preserving shape topology, direction, and distance over OSM datasets for two cities, New York and Singapore.
## Update after reb... | Rebuttal 1:
Rebuttal: Thank you for these detailed, thought-provoking comments.
___
***Reviewer (paraphrased)***: Existing baselines use OSM data, including LLMs[1]. ResNet-18 is weak given [2,3,4].
***Authors:*** Poly2Vec is designed to handle arbitrary geometric data described by coordinates (vector spatial data), r... | Summary: The authors introduce a method of encoding representations of geospatial objects which they call Poly2Vec. This method is capable of representing points (e.g. points of interest), polylines (e.g. roads), and polygons (e.g. buildings) while competing methods struggle to represent all of these different formats.... | Rebuttal 1:
Rebuttal: Thank you for your insightful comments.
___
***Reviewer:*** (Note we grouped these comments, because our answer is essentially the same for all of them.) The method involves learning how to encode objects in a way that seems to allow it to be fine tuned to specific tasks which might not be helpf... | Summary: The paper presents a Fourier based encoding strategy for geospatical principles, i.e., points, lines, and polygons, into a deep-learning compatible vector format. The methodology is well-explained and founded in signal processing theory. It well-explains Fourier analysis and inherent properties (Affine Transfo... | Rebuttal 1:
Rebuttal: Thank you for the thorough review and encouraging comments.
___
***Reviewer:*** The experiments are well-structured according to 4 research questions. However, all results seem to be based on the classification of point-to-polyline etc relationships, which is reasonable for this experimental set... | Summary: This work considers the problem of encoding points, polylines, and polygons for the purposes of prediction tasks that require understanding of spatial relationships such as topology, direction, and distance. Similar to previous point and polygon encoding approaches (Space2vec, NUFTspec), Fourier transformation... | Rebuttal 1:
Rebuttal: We thank the reviewer for the insightful comments.
___
***Reviewer:*** Similar to previous point and polygon encoding approaches (Space2vec, NUFTspec), Fourier transformation is used to transform vector space data to fixed length presentations.
***Authors:*** We note that our Poly2Vec differs fr... | null | null | null | null | null | null |
Generalization Analysis for Supervised Contrastive Representation Learning under Non-IID Settings | Accept (poster) | Summary: This paper proposes a modified framework where ERM is performed using a small subset of input tuples assembled from a fixed pool of label data points. It derives generalization bounds for the empirical minimizer of U-Statistics and a sub-sampled risk. The results are applied to obtain bounds for common classes... | Rebuttal 1:
Rebuttal: We truly value your thoughtful feedback. Below, we do our best to allay your concerns within the character limit, and stand ready to answer any questions in further detail during the rolling discussion.
Firstly, we clarify that our work assumes $N$ **labeled** examples $x_{1:N}$ are used to form... | Summary: This paper provides the first statistical generalization theory of Contrastive Representation Learning (CRL) where data does not follow the strong assumption of independently and identically distributed (i.i.d.). Previous theoretical research on CRL assumes that data points used for training are drawn independ... | Rebuttal 1:
Rebuttal: Dear reviewer, we truly appreciate your **recognition of the strength and novelty** of our contributions. We are also particularly grateful for your detailed review and the effort you put into carefully **checking our technical proofs** in Appendices B and C.
To answer your question - yes - the ... | Summary: For the generalization analysis of contrastive representation learning (CRL) in non-IID settings, several new theory bounds are proposed in this paper. First, this paper proposes a revised theoretical framework for CRL. Then, a U-Statistics formulation for the population unsupervised risk is proposed, and boun... | Rebuttal 1:
Rebuttal: Many thanks for your thorough review. We especially appreciate your efforts in checking the correctness of our theoretical analysis as well as raising an interesting question regarding the proof techniques. Please refer to our response below regarding the difference between the technicality of our... | Summary: This paper revisits contrastive representation learning by relaxing the traditional i.i.d. assumption on training tuples. Instead of assuming independent tuples, the authors analyze a practical setting where a fixed pool of labeled data is recycled to form multiple tuples. They derive generalization bounds usi... | Rebuttal 1:
Rebuttal: Many thanks for the thoughtful and detailed response. We sincerely appreciate your recognition of the novelty in our decoupling techniques. Below, we provide our responses to address your concerns.
**Re (Remark 4.2)**: To clarify your concern, suppose we have an one-sample, $m$-order U-Statistic ... | null | null | null | null | null | null |
How Do Transformers Learn Variable Binding in Symbolic Programs? | Accept (poster) | Summary: This paper studies how transformers can learn to implement variable binding through training. In particular, the authors focus on a specific task where the transformer is given a synthetic program where variables are assigned either numerical values or other variables, and the target is to correctly return the... | Rebuttal 1:
Rebuttal: Thank you for your assessment that we "carefully designed the experiments to empirically validate [our] claims" and for your thoughtful suggestions. Your questions about generalization inspired several new experiments that strengthen our findings.
## Does the model actually learn a systematic mec... | Summary: In this paper, the "variable binding" ability of Transformer model is studied, which is to autonomously assign correct values to symbolic variables. The paper focuses on controlled experimental design. A symbolic program dataset is constructed, which consists of programs that involve value passing among variab... | Rebuttal 1:
Rebuttal: Thank you for your feedback that our paper "proposes an interesting paradigm in analyzing the symbolic learning ability of Transformer." We appreciate your critical assessment and have conducted new experiments to address your concerns.
## Addressing key concerns:
### 1. "The model does not disti... | Summary: The authors investigate how variable binding is learned by decoder-only transformer models. In particular, they use causal interventions to understand how transformers propagate binding in tasks where they are asked for the value of a variable, e.g., "c" where ("c=b, p=f, b=a, f=2, a=1") is given. Their key fi... | Rebuttal 1:
Rebuttal: Thank you for your positive assessment that our paper is "well-written and clear" with claims that are "well-supported." We're especially pleased you highlighted our most interesting finding: that relevant circuits form with "behavior itself being fairly uninteresting."
## Response to your questi... | Summary: This paper performs a mechanistic interpretability study on a transformer trained on a synthetic task that requires tracking values assigned to variables in a mock programming language. Programs consist of 16 variable assignments, and the transformer has 12 layers. The authors use interchange interventions to ... | Rebuttal 1:
Rebuttal: We thank you for your thorough review and positive feedback that our paper is "straightforward and generally well-written" with experiments that "validate the authors' claims convincingly."
## Responses to your questions:
### Q1&Q2: Why exactly 17 lines? Did you try testing on longer/shorter pr... | null | null | null | null | null | null |
Nonlinearly Preconditioned Gradient Methods under Generalized Smoothness | Accept (oral) | Summary: This work studies nonlinearly preconditioned gradient methods. After comparison with existing methods and discussing the existing connections, an extension of anisotropic smoothness was proposed. For specific choices of the reference function (1- isotropic reference function and 2- separable reference function... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback.
Regarding the experimental design, our paper is focused on theoretical analysis and thus we only provide preliminary experimental results on problems that are associated with generalized smoothness in the related literature (Chen et al., 2023; Gorbunov ... | Summary: The paper studies non-linearly preconditioned gradient methods, analyzing algorithms of the form $x^{k+1} = x^k - \gamma\nabla\phi^*(\lambda \nabla f(x^k))$, where where $\phi^*$ is a convex dual reference function, and $f$ is continuously differentiable but potentially non-convex.
This general framework encom... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback.
Our answers for **Questions 1 and 2** are summarized in the following. Regarding the $(L, \bar L)$-anisotropic smoothness condition, our contribution is threefold:
- First and foremost, we describe a general setting where the second-order condition prese... | Summary: 1. This paper introduces a unified preconditioning gradient descent scheme for some of the popular optimization algorithms including Adam, Adagrad, gradient clipping.
2. The paper introduces the concept of anisotropic smoothness, which generalizes the relative smoothness. It gives 2nd order sufficient conditio... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s time and effort in providing feedback.
Regarding the stated **Weaknesses**:
1. Our results cover both the convex and nonconvex setting. The convergence results in the nonconvex setting are presented in Section 3.1.
2. The scope of the current manuscript is to lay the ... | Summary: The studies a non-linearly preconditioned gradient descent method for optimizing some scalar functions $f:R^n\to R$.
The iterations of the method are of
$$x_{k+1}=x^k-\gamma \nabla \psi^*(\lambda \nabla f(x^k)),$$
which is similar to the approach of "Dual Space Preconditioning for Gradient Descent" by Maddiso... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback.
Answer to **Question 1**:
There are three major differences in the conditions considered in the current paper compared to (Laude & Patrinos, 2022):
1. We consider reference functions $\phi$ of possibly not full domain. This leads to a more gener... | null | null | null | null | null | null |
Stray Intrusive Outliers-Based Feature Selection on Intra-Class Asymmetric Instance Distribution or Multiple High-Density Clusters | Accept (poster) | Summary: This paper proposes a supervised FS method, Stray Intrusive Outliers-based FS (SIOFS), for data classification with intra-class ADMHC. By focusing on Stray Intrusive Outliers (SIOs), SIOFS modifies the skewness coefficient and fuses the threshold in the 3σ principle to identify the class body, scoring features... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer 5k2j for the constructive and valuable comments. The concerns are addressed as follows.
## Q1: In Fig. 1b, class "2" has two high-density clusters, but there is only class "1" and "3" in the figure.
Many thanks for the comment. We revise this typo from class "2" to clas... | Summary: This paper proposes a supervised FS method, Stray Intrusive Outliers based FS (SIOFS), for data classification with intra-class ADMHC. By focusing on Stray Intrusive Outliers (SIOs), SIOFS modifies the skewness coefficient and fuses the threshold in the 3σ principle to identify the class body, scoring features... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer’s feedback. Below, we address the concerns in detail.
## Q1: Some ambiguous descriptions about the theorems. Inconsistency between theory and experiment about α.
As addressed in our response to **Reviewer 52hZ, Q1**, we clarify that the condition for in Theorem... | Summary: For the problem of high-dimensional data classification with intra-class asymmetric instance distribution or multiple high-density clusters (ADMHC), a novel supervised feature selection (FS) method named Stray Intrusive Outliers-based FS (SIOFS) is proposed. The proposed method uses the RDM center to character... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer 52hZ for the recognition of our work and for providing constructive comments.
## Q1: Explain the inconsistency between Theorem 2 and results about $\alpha$.
Sorry for the incomplete statement about Theorem 2. We clarify that the condition for $\alpha$ in Theorem 2 does... | null | null | null | null | null | null | null | null |
The Elicitation Game: Evaluating Capability Elicitation Techniques | Accept (poster) | Summary: This paper aims to test capability elicitation techniques for evaluating language models. The motivation is that models are often deployed with undiscovered harmful capabilities that later resurface in deployment. Building on prior work for password-locked models, they introduce models with more sneaky passwor... | Rebuttal 1:
Rebuttal: Thanks for your help and feedback! We’re glad that you found the motivation and contributions compelling.
> The main thing that this paper does not do is show that its model organisms are particularly good models of real-world capabilities. One could red-team models that had undergone unlearning... | Summary: This paper evaluates techniques for eliciting hidden capabilities in language models. The authors create two types of "model organisms" (password-locked and circuit-broken models) with intentionally hidden capabilities, then test various elicitation methods including prompting, activation steering, and fine-tu... | Rebuttal 1:
Rebuttal: Thanks for your feedback. We're glad you appreciated our novel model organism approach, clear results, and practical insights.
> The model organism approach is nice, but a rehashed idea that is evaluated in this work
Our paper focuses on using model organisms to evaluate techniques to elicit hid... | Summary: In this paper, the authors seek to evaluate different capability elicitation methods from the literature. Since there are too many to reasonably evaluate, they choose a subset of techniques by prioritizing their cost, simplicity and ease of use. The authors propose to use *model organisms* that mimic real-worl... | Rebuttal 1:
Rebuttal: Thanks for taking the time to review our paper and offer high-quality feedback! We’re glad that you found the paper clear and enjoyable to read, and that you appreciated the strength of our novel model organism based on circuit-breaking.
Note: we abbreviated some of your comments to meet the char... | Summary: This paper introduces Model organisms which are language models that are trained to imitate AI systems with human capabilities via circuit breaking. This method of training is more robust to elicitations than simple password-locked training approach. On the tasks of MCQA and code generation, they are able to s... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our paper. We’re glad that you appreciated our contribution to the capability evaluation and sandbagging literature, with our evaluation of techniques for eliciting hidden capabilities, and the introduction of a more robust model organism.
> Why did you cho... | null | null | null | null | null | null |
FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks | Accept (poster) | Summary: The paper describes the method for accelerating the computation of concept activation vectors, which uses, instead of the SVM linear classifier, the Fisher's linear discriminant analysis (LDA).
## After the rebuttal
Many thanks to the authors for the work and a good rebuttal discussion. The authors address m... | Rebuttal 1:
Rebuttal: We thank the reviewer for their effort in checking our theoretical contributions and providing insightful comments. Particularly, we appreciate the reviewer for highlighting the correctness and soundness of our analysis.
## Implemented Changes
We agree with and follow the suggestion of the review... | Summary: The paper introduces FastCAV, a method to compute Concept Activation Vectors (CAVs) up to 63.6× faster than traditional SVM-based approaches by leveraging simple mean vector computations under theoretical assumptions. FastCAV achieves comparable accuracy and interpretability to existing methods while significa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive feedback and appreciate that they see our contribution well-situated within prior work. We are glad that our empirical comparison to SVM-based CAV computation was found to be fairly comprehensive. Additionally, we agree with the reviewer that more applicati... | Summary: The paper introduces a faster approximation to compute concept activation vectors. They do so through computing the mean vector for representations for the concept of interest and a random reference concept. The authors demonstrate that their approach is the same as the standard approach for computing concept ... | Rebuttal 1:
Rebuttal: We appreciate the thoughtful feedback of the reviewer. We are glad that the logic in Sections 3.3 and 3.4 was judged as reasonable and intuitive. Similarly, we welcome the feedback regarding our experiments in Section 4. Particularly, the reviewer asked about our choice of pairwise cosine similari... | Summary: The paper introduces fastcav, extending tcav to identify concept activation vectors by computing the mean of vectors and find the approximate direction of concept activation vectors to the concept space, which is assumed to be nearly orthogonal. The usage of mean reduces the dimensional complexity of using svm... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments and thoughtful feedback. Specifically, we appreciate that the reviewer found our comparison between FastCAV and SVM-based computation valid and reliable. Concerning the lack of comparison to other CAV calculation methods, we note that we focused on SVMs fol... | null | null | null | null | null | null |
Regression for the Mean: Auto-Evaluation and Inference with Few Labels through Post-hoc Regression | Accept (poster) | Summary: This paper works on statistical inference and its applications in model evaluation. The paper analyzes the pitfalls of a previous method, Prediction Powered Inference (PPI), and proposes two variants to improve performance. The first proposed method is to use ridge regression and add an additional hyperparamet... | Rebuttal 1:
Rebuttal: We greatly appreciate your review and your suggestions for improving the work. We especially appreciate your thoroughness in reading the paper, and will be sure to fix any typos currently in the draft for the camera ready version of the paper. Below we respond to your individual points.
**Model E... | Summary: This paper makes the observation that the variance-minimizing parameter $\lambda$ in PPI++ can be seen (in the scenario where $N$ is large) as a linear regression coefficient when regressing $Y$ on $f(X)$. Motivated by this observation, this paper proposes an analog of ridge regression, where $\lambda$ is reg... | Rebuttal 1:
Rebuttal: Thank you very much for your insightful review. Your questions and comments have helped to make the draft much stronger and clearer. In the camera ready version of the paper, we will improve the related works section and better describe the LLM refusal dataset. We address your individual concerns ... | Summary: The paper focuses on the specific scenario where PPI is used for mean estimation. The authors note an equivalence between the tuning parameter estimation for PPI++ and linear regression, and examine several modifications that can be derived. Results are validated through empirical evidence on several benchmark... | Rebuttal 1:
Rebuttal: We would like to express our appreciation for your thoughtful review and constructive critique. We look forward to making the writing more clear following your points. Specifically, we will amend the abstract and make the description of the LLM refusal dataset more clear. We respond to individual ... | Summary: The paper provides a method to improve over standard prediction-powered inference by weighting the correction factor in the same way that is done in control variates. The methods are simple and improve over existing baselines, and the authors provide an illuminating discussion.
Claims And Evidence: Claims wer... | Rebuttal 1:
Rebuttal: First of all, we thank you for your constructive feedback and positive reception of the work. We look forward to improving the manuscript based on your review. We will be sure to improve readability of the figures for the camera ready draft. Below, we respond to individual points you raised.
**Op... | null | null | null | null | null | null |
Reflection System for the Abstraction and Reasoning Corpus | Reject | Summary: The authors propose an augmented ARC dataset, AugARC, that can be used for finetuning LLMs to solve ARC tasks. They also propose a two-stage system for solving ARC tasks that runs multiple ARC solvers in parallel for a solution followed by a reflection LLM that chooses the best solution. They demonstrate that ... | Summary: The paper introduces a reflection system and data generation techniques for ARC. The authors evaluate baseline LLMs on ARC and find sort of unsatisfactory results, and try to address this in two ways: AugARC and the reflection system. They create AugARC, which consistently improves LLM performance through task... | Summary: In this work, the authors examine the performance of large language models (LLMs) when paired with a program synthesis solver to tackle the ARC challenge. They also explore augmentation strategies to expand the training data size, which was limited to 400 tasks in the original version. Fine-tuning the LLM with... | Summary: This paper proposes a “Reflection System” for solving the ARC challenge, combining multiple ARC solvers (notably DSL-based program synthesis and LLM-based solvers) and then using an additional “reflection” model to select among the candidate predictions. The authors also introduce AugARC, an augmented version ... | null | null | null | null | null | null | ||||
What makes an Ensemble (Un) Interpretable? | Accept (poster) | Summary: The paper provides a complexity-theoretic investigation into the interpretability of ensemble models, focusing on three major types of explanation queries (sufficient reasons, contrastive reasons, Shapley values, etc.) and three common base-model classes (decision trees, linear models, neural networks). The ma... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback, and for recognizing the significance of our work.
**Discussion on more specialized ensemble schemes**
We thank the reviewer for raising this point. While our complexity results are established for both majority voting and weighted voting infer... | Summary: The paper presents complexity results for a variety of explanation
queries on different machine learning models with the main aim of
comparing the behaviour of single models with ensembles. In
particular, the paper considers the following explanation queries:
Minimum Sufficient Reason (MSR), Minimum Contrastiv... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback and for acknowledging the significance of many of our results.
**Relation to Ordyniak et al. (KR, November 2024)**
We appreciate the reviewer for highlighting the important connection to the recent and highly significant work by Ordyniak et al., pu... | Summary: This paper studies the interpretability of ensemble of models from the computational complexity theory. Authors show both negative results (e.g., intractability even for constant-size models) and positive results (e.g., tractability for small decision tree ensembles).
Claims And Evidence: Yes, there are proof... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive feedback and for recognizing the importance of the results provided in our work.
**The importance of complexity separations between base-models and ensembles**
While the reviewer acknowledges the importance of some of our unexpected results, such as th... | null | null | null | null | null | null | null | null |
Zero Shot Generalization of Vision-Based RL Without Data Augmentation | Accept (poster) | Summary: The authors propose a method for generalizing vision-based RL policies in the presence of image distribution changes or distractors, such as modifying the color scheme of the image. There method aims to learn a representation that disentangles the different components of an image, specifically into two types: ... | Rebuttal 1:
Rebuttal: Thank you for your in-depth feedback and for providing references to work on bi-simulation for visual generalization. We will add bi-simulation to the related works section and fix the reward / RL objective notation issues for the final iteration of the manuscript. We respond to individual comment... | Summary: This paper proposes Associative Latent DisentAnglement (ALDA) that builds on standard off-policy RL towards zero-shot generalization. It learns a disentangled representation from the training data and then uses an associative memory model to recover data points in the original training distribution given OOD d... | Rebuttal 1:
Rebuttal: Thank you for your in-depth feedback and for providing the additional reference. We will add it to the related works and discussion sections for the camera-ready version of our manuscript. We respond to individual comments and concerns below.
**However, I have some concerns in its novelty...**
A... | Summary: The paper introduces Associative Latent Disentanglement (ALDA), an approach to zero-shot generalization in vision-based reinforcement learning (RL) without relying on data augmentation. ALDA leverages disentangled representations and associative memory mechanisms to enable RL agents to generalize to novel envi... | Rebuttal 1:
Rebuttal: Thank you for your feedback and suggestions on how we could improve the paper's clarity. We respond to individual comments and questions below.
**Some theoretical discussions, particularly regarding disentanglement and associative memory mechanisms, could be more clearly explained for a broader a... | Summary: The authors present ALDA - an approach for training disentangled representations along with off-policy learning for OOD generalization.
They build upon existing QLAE-based latent model, which is SOTA disentanglement method, which uses latent space dimensions, each having their own codebook.
The prove that d... | Rebuttal 1:
Rebuttal: Thank you for your comments and feedback on the paper. We will fix the grammar errors for the camera-ready version of the manuscript. Regarding your questions and concerns, we respond to each one individually below.
**I think it would be nice to discuss this for SVEA vs ALDA since that would be ... | null | null | null | null | null | null |
What Limits Bidirectional Model's Generative Capabilities? A Uni-Bi-Directional Mixture-of-Expert Method For Bidirectional Fine-tuning | Accept (poster) | Summary: This paper explores the impact of bidirectional fine-tuning on the performance of unidirectional language models, particularly focusing on the decline in generative ability caused by bidirectional fine-tuning. The authors attribute this decline to subsequent dependence and perform an in-depth analysis to suppo... | Rebuttal 1:
Rebuttal: > Weakness 1&2: Test on More Models: The authors have only tested the Qwen 1.5 series of models. The experiments on more models are yet to be conducted. Test on Recent Models: It is beneficial to conduct experiments on recent LLMs.
Thank you for your advice. We evaluate the effectiveness of our p... | Summary: There exists a common belief that causal language models (i.e., unidirectional models) perform better in generation tasks while bidirectional models perform better in embedding tasks. However, bidirectional finetuning of unidirectional models usually leads to significantly inferior generation performance, whic... | Rebuttal 1:
Rebuttal: > Weakness 1: The authors claim that the inferior performance ... this conclusion more convincing.
As we replied to Reviewer bS72, Table 1 shows that there is a consistent trend between subsequent dependence and the general capability of the model. This indicates a correlation between subsequent... | Summary: The paper investigates the impact of bidirectional fine-tuning on unidirectional language models. The authors argue that bidirectional attention mechanisms, while enhancing embedding tasks, degrade generative performance. To address this, they integrate unidirectional FNN layers with bidirectional ones trained... | Rebuttal 1:
Rebuttal: > Weakness 1: Limited discussion of potential trade-offs in training complexity.
Thanks for the helpful comments. Our training consists of two main parts: bidirectional expert training and UBMoE-LLM model training. For bidirectional expert training, compared to previous bidirectional approaches, ... | Summary: Due to the unidirectional attention mechanism, current LLMs underperform in embedding tasks. Some studies have modified the unidirectional attention to bidirectional attention in LLMs and fine-tuned them using contrastive learning, resulting in models better suited for embedding tasks. However, this modificati... | Rebuttal 1:
Rebuttal: > W1: The new model developed ... computational demands for generation tasks.
UBMoE activates only one expert for each token, adding computational overhead during inference solely for token routing. Compared to the computational cost of the causal language model itself, this additional overhead ... | null | null | null | null | null | null |
Simple and Critical Iterative Denoising: A Recasting of Discrete Diffusion in Graph Generation | Accept (poster) | Summary: This paper introduces Iterative Denoising, a novel framework to improve discrete diffusion and flow matching models for graph generation. Traditional discrete diffusion models suffer from error accumulation and propagation due to time dependencies in the noising process, particularly in mask diffusion. The pro... | Rebuttal 1:
Rebuttal: We would like to thank you for the review.
## Claim and Evidence
**Paragraph 1**
The justification for our work is as follows:
- We identify a limitation of discrete diffusion models—namely, the issue of compounding denoising errors.
- We propose a theoretically grounded method that directly... | Summary: This manuscript observes that choices taken at the early steps of the generation process may, in hindsight, turn out to be "errors", and empowering the model with a mean to fix these errors could prevent their accumulation and improve performances. Iterative Denoising (ID) is introduced to this end, and furthe... | Rebuttal 1:
Rebuttal: We would like to thank you for your review, your careful reading and questions. We believe they will contribute to significantly improve our submission.
## Answer to the questions
### Question 1:
We broadly agree with your summary in **Claim and Evidence** and specifically with your assessmen... | Summary: The paper aims to address *Compounding Denoising Error* in discrete diffusion models by removing time dependency in the noising process. To enhance performance, it introduces a Critic that aligns generated samples with the data distribution. The authors provide relevant experimental results as supporting evide... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for their careful reading of our submission and for the thoughtful questions and comments. We believe these insights will substantially help us improve the quality and clarity of our paper.
### 1. Time Dependency in the Noising Process
We at time used *time de... | null | null | null | null | null | null | null | null |
Controllable Data Generation with Hierarchical Neural Representations | Accept (poster) | Summary: This paper proposes a hierarchical Implicit Neural Representations (INR) framework that aims to provide better control over hierarchical representations during the generation process. In the first stage of the framework, a Layer-of-Experts (LoE) model is trained, and a latent variable is learned for each layer... | Rebuttal 1:
Rebuttal: We thank reviewer ZbmH for the valuable feedback.
### **W1. Qualitative results.**
Thank you for the valuable feedback. We agree that earlier layers exhibit more visible influence in Figure 4. This is expected in our hierarchical design—deeper layers naturally introduce finer, subtler variations... | Summary: The paper proposes a framework for controllable data generation using hierarchical implicit neural representations (INRs).
It models conditional dependencies across layers in the parameter space to improve control over the generation process.
Claims And Evidence: The paper presents clear evidence to support ... | Rebuttal 1:
Rebuttal: We thank reviewer aYYY for the constructive feedback.
### **W1. Lack of predefined interpretability for each layer.**
Thank you for the thoughtful comment. CHINR is designed to align the hierarchical structure of INRs with semantic abstraction, allowing each layer to control different levels of ... | Summary: The paper introduces CHINR, a framework for controllable data generation using hierarchical neural representations. It addresses limitations of existing generative INR approaches that fail to capture hierarchical structures in data, leading to limited control over generation. The method consists of two stages:... | Rebuttal 1:
Rebuttal: We thank reviewer N8d9 for the valuable comments.
### **W1. Scalability to larger datasets.**
Thank you for pointing this out. As discussed in the conclusion, scaling CHINR to larger datasets is a known challenge. While CHINR demonstrates the core idea of hierarchical control through INR paramet... | Summary: The paper presents a novel method, CHINR, for controllable generative INR by exploiting the hierarchy structure in parameters. The authors employ a Layers-of-Experts (LoE) network to encode data with layer-wise latents and propose a Hierarchical Conditional Diffusion Model (HCDM) to learn conditional dependenc... | Rebuttal 1:
Rebuttal: We thank reviewer T7ed for the thoughtful comments.
### **W1. Out-of-distribution samples in conditional chain.**
Thank you for your insightful feedback. We address error accumulation during both the training and inference phases.
In the training phase, we begin by teaching the model to capture... | Summary: The paper introduces a novel framework to capture hierarchical data semantics with implicit neural representations, enabling improved control over data generations. The framework is structured in two stages. For the first stage, a layer-of-expert (LOE) architecture is employed to capture general semantics with... | Rebuttal 1:
Rebuttal: We thank reviewer XvvS for the valuable feedback.
### **W1. Scalability to larger datasets**
Thanks for raising this scalability concern. The CHINR framework focuses on establishing the connection between INR parameters and data semantics for controllable generation. While our experiments use sm... | null | null | null | null |
Generalization in Federated Learning: A Conditional Mutual Information Framework | Accept (poster) | Summary: The paper introduces a novel information-theoretic framework to analyze generalization in federated learning (FL). By extending the supersample-based conditional mutual information (CMI) framework with a “superclient” construction, the authors decompose the generalization error into two components: the partici... | Rebuttal 1:
Rebuttal: We thank you sincerely for your valuable feedback on our paper. Our responses follow.
>- While the mathematical methods are solid, including more straightforward explanations or diagrams could help clarify the concepts for readers who may not be experts in the field.
>- Another suggestion is to... | Summary: The paper studies the generalization error of the federated learning algorithm using the CMI framework. The goal is to capture the out-of-sample gap and the participation gap using this framework. To do so, a federated learning setup is considered where each user observes data distributed according to a distr... | Rebuttal 1:
Rebuttal: We thank you sincerely for your valuable feedback on our paper. Our responses follow.
>-... the bounds established do not capture any characteristic of FL ...
**Response.** We respectfully disagree with the reviewer's claim that our bounds fail to capture key characteristics of FL, and the argum... | Summary: The paper proves mutual information-based generalization bounds for a federated learning setting, that bounds both the out-of-sample gap (between the empirical and population distributions of the participating clients) and the participation gap (between the participating clients and the underlying meta distrib... | Rebuttal 1:
Rebuttal: We thank you sincerely for your valuable feedback on our paper. Our responses follow.
>- ... why the MI-based ... will be useful and how ... lead to regularization ...
>- ... uncertain about the practical role ...
**Response.** We note that our bounds do have practical implications, e.g., as men... | Summary: This work studies the question of generalisation in federated learning (FL), where $K$ users aims to share some benefits of their learning phase via a central server without sharing their data. Authors propose novel generalisation bounds tailored to FL involving the Conditional Mutual Information (CMI) framewo... | Rebuttal 1:
Rebuttal: We thank you sincerely for your valuable feedback on our paper. Our responses follow.
>- you said in l.122-123 left column that your CMI framework is inspired by the meta-learning one of Hellstrom \&Durisi 2022. Is it possible to precise what are the specificities of your derived results (e.g. th... | null | null | null | null | null | null |
Falsification of Unconfoundedness by Testing Independence of Causal Mechanisms | Accept (poster) | Summary: Addressing the common assumption of causal sufficiency, this work proposes an approach to falsify this assumption in causal effect estimation settings. It relies on datasets from multiple environments to test for violations of the Independence of Causal Mechanisms (ICM) principle. The authors study a causal mo... | Rebuttal 1:
Rebuttal: We thank the reviewer for their invaluable feedback, suggestions for improvement and questions. Below are answers to your questions.
# 1: Fixed feature representations with linear parameter shifts
We agree that it is important to consider what it entails to assume a model class with fixed featur... | Summary: The paper proposes a test for unconfoundedness based on the usual assumptions of causality from the potential outcomes perspective, but more critically assumptions about the independence of causal mechanisms and a specific functional form (including specifying functional form well). The test requires observing... | Rebuttal 1:
Rebuttal: We thank the reviewer for the response, in particular their excitement for our proposal for extending our method to implicit feature representations. Whereas it felt as to be outside the scope of the current manuscript, we agree this is one of the more promising directions of our work and we also ... | Summary: This paper addresses the problem of unmeasured confounding in observational data. Most causal estimation methods assume that there are no unmeasured confounders, an assumption that is hard to test in practice. The authors propose a method for testing this assumption in situations where data from multiple env... | Rebuttal 1:
Rebuttal: We thank the reviewer for the invaluable feedback and questions. We have provided answers and proposed changes related to your questions below. In addition, we also do the following changes for the camera-ready version:
- update Section 4 with more explicit assumptions on the feature representatio... | Summary: This manuscript presents an algorithm for falsifying the assumption of no unmeasured confounding in a setting of observational data from multiple environments. To this end, the authors employ the Rubin potential outcome causal model, assume positivity, consistency and no unmeasured confounding; subsequently th... | Rebuttal 1:
Rebuttal: We thank the reviewer for the valuable feedback and for spotting a number of typos which now have been addressed. We have provided answers and proposed changes related to your questions below.
# 1: Connection between our theory and that of of Karlsson and Krijthe (2023)
While the possibility of ... | null | null | null | null | null | null |
Overshoot: Taking advantage of future gradients in momentum-based stochastic optimization | Reject | Summary: Proposes a variant of Nesterov's accelerated gradient method and presents some experiments
Claims And Evidence: As this paper doesn't present any theory, the strength of the paper needs to the experimental evaluation. I don't see these experiments as strongly convincing for a number of reasons:
- The biggest... | Rebuttal 1:
Rebuttal: Thank you for your evaluation and insights. We will incorporate the suggestions therein into future incarnations of our work. | Summary: This paper proposes an overshooting technique for optimization, which evaluates the gradient at point that extrapolate the standard optimizer update.
Claims And Evidence: The presentation of the Algorithm is somewhat confusing: for the general Algorithm 1 it remains unclear what is allowed for the update rule... | Rebuttal 1:
Rebuttal: Thank you for putting your time into reviewing our work and for your insights. We will incorporate the suggestions therein into future incarnations of our work.
> and how differs $\phi^{\prime}$ from $\phi$.
$\phi$ and $\phi^{\prime}$ represents the same optimization methods, each for each weigh... | Summary: The submission presents Overshoot, a momentum optimizer that can be used with adaptive algorithms like Adam and SGD. Unlike Nesterov's Accelerated Gradient (NAG) or classical momentum (CM), Overshoot updates model weights in advance in anticipation of the upcoming momentum update even before gradients are calc... | Rebuttal 1:
Rebuttal: Thank you for putting your time into reviewing our work and for your insights. We will incorporate the suggestions therein into future incarnations of our work.
> while the claim of a 15% reduction in steps is based on the Steps-to-95% Loss Reduction" metric. However, it’s not clearly defined, li... | Summary: This paper draws inspiration from Nesterov’s Accelerated Gradient (NAG) and introduces a novel method called Overshoot. The Overshoot method calculates the gradient at model weights shifted in the direction of the current momentum, thereby leveraging information from the surrounding landscape more effectively.... | Rebuttal 1:
Rebuttal: Thank you for putting your time into reviewing our work and for your insights. We will incorporate the suggestions therein into future incarnations of our work.
> Unlike NAG, Overshoot employs a specialized reformulation that aims to reduce memory overhead.
Overshoot does not aim to reduce memor... | null | null | null | null | null | null |
Debiased Orthogonal Boundary-driven Efficient Noise Mitigation | Reject | Summary: This paper exploits the properties of high-dimensional orthogonality to identify a robust and effective boundary in cone space for separating clean and noisy samples. They propose One-Step Anti-noise (OSA), a model-agnostic noisy label mitigation paradigm that employs an estimator model and a scoring function ... | Rebuttal 1:
Rebuttal: We thank the reviewer DHCR for the positive, patient and professional review, as well as the valuable suggestions for improvement. Our responses to the reviewer’s questions are below:
***W1 : The claim of "table2 show that OSA outperforms all previous approaches on all metrics with a huge gap" is... | Summary: This paper proposes One-Step Anti-noise (OSA), a model-agnostic noise mitigation paradigm leveraging high-dimensional orthogonality and cone effects in pre-trained models (e.g., CLIP) to distinguish noisy and clean samples. Key contributions include: 1) It identifies a shifted orthogonal boundary in cone space... | Rebuttal 1:
Rebuttal: We thank the reviewer Ho5e for the positive, patient and professional review, as well as the valuable suggestions for improvement. Our responses to the reviewer’s questions are below:
***W1 : It is more meaningful when validated on large-scale datasets.***
**A1:** Thank you for your insightful r... | Summary: This paper proposes One-Step Anti-noise (OSA), a model-agnostic noise mitigation method, addresses label noise in large-scale pre-training tasks. It leverages pre-trained models’ high-dimensional orthogonality and the cone effect, which shifts the orthogonal boundary in embedding space, intersecting clean and ... | Rebuttal 1:
Rebuttal: We thank the reviewer PeYE for the positive, patient and professional review, as well as the valuable suggestions for improvement. Our responses to the reviewer’s questions are below:
***W1 : Although the evaluation datasets cover diverse image and text content, they may not fully represent all p... | null | null | null | null | null | null | null | null |
A Generalization Theory for Zero-Shot Prediction | Accept (oral) | Summary: The papers takes a theoretical approach towards understanding key quantities driving zero-shot prediction. Introducing and deriving bound for the aforementioned problem setting, the paper analyzes translation between modalities and effectiveness of prompt engineering strategies a multi-modal learning setting.
... | Rebuttal 1:
Rebuttal: Thank you for your hard work in verifying the paper. The authors are happy to hear that you found the narrative insightful and well-written. We address your concerns below.
> **“[The Methods And Evaluation Criteria] make sense, even though they can be extended considerably… [The Experimental Desi... | Summary: This paper provides a formal modeling of the two-stage learning procedure, known as CuPL: (1) pretraining on multimodal labeled data and (2) zero-shot prediction (ZSP) on the pre-trained model with natural language prompts. The goal is to offer a theoretical explanation of the success of CuPL. To achieve this,... | Rebuttal 1:
Rebuttal: Thank you for your thorough review. We address your comments below.
> **“I hope the authors could discuss more on the relationship between FSL, ZSP and MAML.”**
We discuss model agnostic meta-learning (MAML) in its [offline](https://proceedings.mlr.press/v70/finn17a.html) variant, i.e., as a too... | Summary: This paper explores the theoretical foundations of zero-shot prediction (ZSP) in foundation models, establishing a formal statistical framework to analyze how pretraining on large-scale, multimodal, unlabeled datasets transitions into downstream zero-shot inference via prompting. The authors identify key facto... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read our manuscript critically. This is undoubtedly one of the most comprehensive reviews we have ever received. Please see your comments addressed below.
> **“What are the implications of prompt bias and residual dependence?... Clarifying their role... would stre... | Summary: The paper proposes a theoretical framework for zero-shot prediction linking pre-training to prompting and also introduces residual dependence (information loss between modalities) and prompt complexity (sample/prompt trade-offs). Risk bounds show ZSP needs huge pre-training data but few prompts.
## update af... | Rebuttal 1:
Rebuttal: Thank you for your review—we address your main comment below.
> **"The theoretical analysis is limited to CLIP-like multi-modal models."**
Our analysis broadly describes multimodal encoder + prompting strategies, where the encoders could be learned by a variety of objectives (not only CLIP... | null | null | null | null | null | null |
Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training | Accept (poster) | Summary: The document introduces DOMAIN2VEC, a technique for optimizing data mixtures in training large language models by decomposing datasets into linear combinations of "Meta-Domains" to enable efficient identification of optimal data mixture ratios.
DOMAIN2VEC uses a meta-Domain classifier to classify any dataset ... | Rebuttal 1:
Rebuttal: Dear Reviewer 75ct,
Thanks for your very valuable review and recognition of our work! We will address your questions point by point.
## Q1 :Why did you use K-means for finding the domains? Why did you choose to represent these domains with 240 Meta-domains? Also, it is unclear whether these cl... | Summary: This paper presents a method for determining the optimal data mixture weights for combining different pre-training datasets to train language models. The authors formulate this as an optimization problem, where the goal is to find the appropriate weights over a set of meta-domains. These meta-domains are const... | Rebuttal 1:
Rebuttal: Dear Reviewer H5YS,
Thanks for your insightful review and suggestions! We will respond to your questions one by one.
## Q1: The paper should provide a more thorough discussion of its connections to prior related work, i.e, Task2Vec [1].
A1: Similar to [2] and [3] cited in lines 407–412 (righ... | Summary: Authors propose a sampling method for training LLMs on multiple sources. Their core idea is as follows: They construct a universal set of real-valued vectors from a large textual corpus--using K-means and doc embeddings. Each vector approximately represents a topical domain of the corpus. Then they take random... | Rebuttal 1:
Rebuttal: Deal Reviewer HtfL,
Thanks for your careful review! We will reply to your questions one by one and hope to solve your concern.
## Part1: Clarification of Our Method
### Q1: Line 88 (right column). what does this mean: (where each element vj of v represents the projection (weight) of the datase... | Summary: Domain2Vec introduces a method for vectorizing datasets by decomposing them into linear combinations of Meta-Domains, which enables efficient identification of optimal dataset mixtures for LLM pretraining. They sampled and embedded texts from which predetermined clusters/labels, which they then clustered and t... | Rebuttal 1:
Rebuttal: Dear Revierwer ZTKc,
Thank you for your recognition of the value of our work, as well as for your valuable comments. We will respond to your questions one by one.
## Q1: Domain2Vec involves much more modelling primitivesand other hyperparameters, as such, demonstrating the method’s robustness ... | null | null | null | null | null | null |
A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle | Accept (poster) | Summary: The paper investigates the effectiveness of Transformer-based models for time series forecasting, focusing on why simpler Transformers outperform more complex ones. These findings include that intra-variate dependencies dominate the performance of existing forecasting benchmarks, tokenization/channel independe... | Rebuttal 1:
Rebuttal: Thank you very much for your comments. We truly appreciate your effort and time for reviewing our work.
*Response to the comment challenging the assumption that the model's output is deterministic:*
While MSE is linked to the assumption of Gaussian noise during traini... | Summary: The paper focuses on a recently heated topic and an important topic in time series forecasting -- which is to conduct further evaluation on the previously proposed models, not only on how a model can improve the forecasting performance, but also consider the performance changes are potentially related to the c... | Rebuttal 1:
Rebuttal: Thank you very much for your thoughtful comments. We truly appreciate your effort and time for reviewing our work.
*Response to comments regarding additional metrics for evaluating forecasting accuracy:*
In this work, our primary focus is on understanding how differen... | Summary: There has been many proposals of transformer architectures for time series forecasting, and some of them are simpler some are more sophisticated, some work well, and some struggle. This work, examines why some of them work better and some struggle: In doing so, it uses an existing classification (Wang et al 20... | Rebuttal 1:
Rebuttal: Thank you very much for your encouraging comments. We truly appreciate your effort and time for reviewing our work.
*Response to Claims And Evidence:*
1. In Table 3, removing the skip connection leads to a clear performance drop across all benchmark datasets, with particularly significant degrad... | Summary: The authors survey the literature on time-series forecasting with transformers, and divide previously published approaches into 3 categories. Given multiple time-varying signals (variates), the signals can be chopped into tokens along the time axis, along the variate axis (one token per variate), or both.
... | Rebuttal 1:
Rebuttal: Thank you very much for your encouraging comments. We truly appreciate your effort and time for reviewing our work.
*Response to questions:*
1. Why use transformers?
In previous work on developing new transformer architectures for time series forecasting, there has been little validation o... | null | null | null | null | null | null |
Alpha-SQL: Zero-Shot Text-to-SQL using Monte Carlo Tree Search | Accept (poster) | Summary: This paper proposes a novel approach based MCTS to enhance the zero-shot performance of LLMs in the text to SQL domain. Authors designed a set of task specific actions such as question rephrasing, schema selection, SQL generation, column value identification, column function identification, and SQL revision. A... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your comprehensive review and for providing a clear summary of our proposed Alpha-SQL method.
We understand the need for comparison against a "Best-of-N" baseline, the omission of base model performance for larger Qwen models in Table 4, and the lack of detailed ana... | Summary: The paper presents a novel approach to Text-to-SQL that eliminates the need for fine-tuning by leveraging the reasoning capabilities of large language models (LLMs). Alpha-SQL employs a Monte Carlo Tree Search (MCTS) framework to progressively construct SQL queries by breaking them down into smaller, manageabl... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your detailed feedback. We appreciate the opportunity to clarify our approach, particularly regarding the core concerns about the MCTS framework's validity, the reward mechanism, experimental fairness, and novelty, which we believe may stem from some misunderstandings... | Summary: This paper introduces a novel zero-shot Monte Carlo Tree Search (MCTS)-based Text-to-SQL approach that constructs SQL queries progressively, enhancing the Text-to-SQL capabilities of Qwen2.5-Coder-32B. The proposed method achieves an execution accuracy of 69.7% on the BIRD dev set and 87.0% on spider dev set, ... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your detailed and insightful review of our paper.
Below, we address each point in detail:
**1. Fair Comparison, Model Choice, and Performance Validation (Addressing Concerns on Experiments, Q1, Q2a, and Rationale for Qwen)**
- **Our Response**: We acknowledge the cr... | Summary: The authors propose a Monte Carlo tree search framework for zero-shot text-to-SQL with LLMs. The action space is a set of sub-tasks whose composition (subject to ordering rules) defines a reasoning path that terminates in a SQL output given a question and database. They generate candidate SQL queries by MCTS r... | Rebuttal 1:
Rebuttal: Dear Reviewer,
Thank you for your thorough review and constructive feedback. We appreciate your positive assessment and valuable suggestions, which will help improve our paper's clarity.
**1. Terminology ("Partial SQL Query States")**
- **Our Response**: Thank you for highlighting this ambiguity... | null | null | null | null | null | null |
Learning With Multi-Group Guarantees For Clusterable Subpopulations | Accept (poster) | Summary: This paper focuses on providing *multigroup* guarantees (with a focus on multicalibration, though the techniques are more general) in a stochastic online prediction game in the novel setting where the groups of interest are not provided beforehand as functions of the feature values, but, rather, as unknown end... | Rebuttal 1:
Rebuttal: Thanks for your detailed review and suggestions! To briefly address your questions:
**Q1: Online-to-batch.**
> This is not a huge weakness, but it might make the paper more complete to have a corresponding batch setup, with results for the batch case. I wonder if the authors have already consi... | Summary: The paper considers a multi-group online learning problem with instances $(x_t, y_t)$ arriving in sequence. In contrast to prior work in online multi-group learning, the groups themselves are not known at each instance. Instead, the paper assumes that there is an endogenous unknown subgroup model, such as, e.g... | Rebuttal 1:
Rebuttal: Thanks for your review and suggestions! We will incorporate your presentation recommendations in the next versions.
**Q1: Covering of H**
> To obtain theorem 4.4, we need to run algorithm 2 with G=H, correct?
For Theorem 4.4, we run Algorithm 2 with $\mathcal{G}$ defined according to eq. 1 (fo... | Summary: This paper focuses on evaluating prediction performance on meaningful subpopulations rather than the overall population in a clustering problem. It proposes two levels of guarantees for capturing performance per subgroup: (1) evaluating the performance if assigning an individual to the most likely cluster, and... | Rebuttal 1:
Rebuttal: Thanks for your review and questions!
**Q1: On empirical results**
> I don't see any empirical studies nor experiments to justify your theory and your algorithm in both main text and supplementary. Can you provide empirical results to justify the effectiveness of your method?
We emphasize that... | null | null | null | null | null | null | null | null |
On Differential Privacy for Adaptively Solving Search Problems via Sketching | Accept (oral) | Summary: First the authors consider the differentially private ANN problem. They reduce this problem of differentially private ANN to differentially private selection problem, and then solve the differentially private selection problem via differentialy private selection methods. Next, the authors consider the problem... | Rebuttal 1:
Rebuttal: We thank the reviewer for their question about prior work, which we address below:
1) Prior works using differential privacy for adaptivity focus on numerical estimates [Hassidim et al., 2022; Beimel et al., 2022; Song et al., 2023; Cherapanamjeri et al., 2023]. Our work instead tackles the *sear... | Summary: The paper addresses the problem of hiding internal randomness of data structures using differential privacy. Specifically, the authors focus on nearest neighbor search and regression problem and aim to protect the internal randomness against adaptive adversary using differentially private techniques.
## Updat... | Rebuttal 1:
Rebuttal: We thank you for your very helpful comments. We would also like to address your questions and comments regarding differential privacy.
Let us start with a high-level overview of the framework, first introduced in [Hassidim et al., 2022; Beimel et al., 2022]. The rough idea is to treat the random ... | Summary: The paper studies the problem of adaptive algorithms: the algorithms where the adversary interacts with a randomized algorithm and the goal of the adversary is to increase the error probability.
It is known that differential privacy could be used to design such algorithms, but the existing applications were f... | Rebuttal 1:
Rebuttal: We appreciate your encouraging and helpful comments. Regarding your comments:
- **Line 40:**
We will clarify this sentence. Our intention is to express both ANN and regression in a unified way, and we will revise it to clearly explain the query and update models in each problem.
- **Line 94:... | Summary: The paper explores using differential privacy for answering adaptively chosen queries (potentially by and adversary) for search problems. The authors consider approximate nearest neighbor and regression as the main problems in this work. The main contributions are showing that under reasonable assumptions, it ... | Rebuttal 1:
Rebuttal: We thank you for your very helpful comments and we would like to address your comments on the quadratic dependence on $\kappa$ and linear dependence on $T$ for the alternative.
For the $\kappa^2$ dependence, the main reason is that by using the $\ell_\infty$ guarantee together with the standard r... | null | null | null | null | null | null |
Optimal transport-based conformal prediction | Accept (poster) | Summary: The paper addresses multivariate conformal prediction and proposes an approach based on multivariate quantiles derived from optimal transport (OT), using a notion of multivariate statistical depth (MK depth) to define ranks. The application to both multivariate regression and classification is discussed, inclu... | Rebuttal 1:
Rebuttal: We thank the reviewer for the careful reading and evaluation. We agree that our numerical experiments primarily compare against simple baselines. Our goal, however, is not to claim superiority over all methods but rather to motivate the use of transport-based quantiles in fundamental settings. We ... | Summary: The paper introduces OT-CP, a novel conformal prediction method for multi-output tasks based on optimal transport. The method constructs quantile regions for multivariate conformity scores while ensuring finite-sample coverage and achieving asymptotic conditional coverage. It introduces a new nonconformity sco... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive evaluation and address their comments with additional discussion and empirical results to illustrate the relevance and computational properties of OT-CP.
*(Region size in regression)*. In Fig. 2, we reported the volumes to demonstrate that OT-CP achieves th... | Summary: In the paper Optimal Transport-based Conformal Prediction, the authors propose a new conformal prediction framework leveraging optimal transport to produce multivariate score functions. They prove that such a framework also achieves distribution-free coverage. They validate the method for multi-output regressi... | Rebuttal 1:
Rebuttal: Thank you for the positive and detailed feedback.
*(OT-CP+ and the use of k-nearest neighbor)*. We agree that OT-CP+ is not the only way to make OT-CP adaptive: our proposed methodology aims to demonstrate that conditional coverage can be achieved with only a slight modification of the generic OT... | Summary: The authors tackle the problem of conformal prediction in regression and classification settings when the target random variable is multivariate. To do so, they use the quantiles definition as in [1], that defines the quantile function of a r.v. $Y \in R^n$ as the euclidean optimal transport map between a unif... | Rebuttal 1:
Rebuttal: ## General comment
We would like to thank all the reviewers for their time and feedback. We have revised the paper accordingly and provide detailed responses below.
1. We emphasize that integrating multivariate quantiles into the conformal prediction framework while ensuring theoretical coverage... | null | null | null | null | null | null |
How to Evaluate and Mitigate IP Infringement in Visual Generative AI? | Accept (poster) | Summary: This paper explores the intellectual property (IP) infringement risks posed by state-of-the-art visual generative AI models, such as DALL-E 3, Stable Diffusion XL. The study shows that these models can generate content resembling IP-protected characters (e.g., Spider-Man, Iron Man, Superman) even when given pr... | Rebuttal 1:
Rebuttal: Thank you very much for your insightful comments. We hope the following results
and clarifications can address your concerns. Please let us know if anything is
still unclear. We are more than willing to provide further clarification and
conduct more experiments if needed.
**Q1:** The proposed met... | Summary: The paper presents a method for creating prompts that may cause T2I/T2V models to generate images infringing on IP rights and show that IP infringement issues are widespread across different visual generative models based on their constructed
prompts. They then develop a defensive method which combines detecti... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable comments. We hope the following results
and clarifications can address your concerns. Please let us know if anything is
still unclear. We are more than willing to provide further clarification and
conduct more experiments if needed.
**Q1:** As the proposed de... | Summary: This paper illustrates that IP Infringement often happens under both ‘Name-based Lure Prompt’ and ‘Description-based Lure Prompt’ situations. Then, this paper proposes a defensive method, named TRIM, to mitigate the infringement. TRIM blocks the targeted name and detect the infringement to regenerate the image... | Rebuttal 1:
Rebuttal: Thank you very much for your helpful comments. We hope the following results
and clarifications can address your concerns. Please let us know if anything is
still unclear. We are more than willing to provide further clarification and
conduct more experiments if needed.
**Q1:** Recall rate of VLM-... | Summary: This paper discovers that SOTA diffusion models tend to generate content that very highly resembles data that could be protected by IP rights, for example, Marvel characters. Their human evaluation studies show that the risk with these characters/concepts is very high. They also propose a mitigation method tha... | Rebuttal 1:
Rebuttal: Thank you very much for your thoughtful comments.
**Q1:** Corpus of selected characters/content.
**A1:** Thank you very much for your constructive comments. Besides the results
in our main paper, we also have the results on different types of non-human IP
contents in Table 6 in the Appendix. We ... | null | null | null | null | null | null |
A Reasoning-Based Approach to Cryptic Crossword Clue Solving | Accept (poster) | Summary: This paper presents a multi-stage pipeline approach to solving cryptic crossword clues, focusing on reasoning through the wordplay mechanisms that make these puzzles challenging. The system consists of:
1. A candidate answer generator (fine-tuned Gemma2-9B model)
2. A wordplay suggestion generator (fine-tuned... | Rebuttal 1:
Rebuttal: ### Statistical Significance of Improvement over GPT-4o
We acknowledge the reviewer's point regarding the margin of error. Since the improvement over GPT-4o on our sampled test set was within the simple margin of error, we performed the Bayesian IRT analysis presented in the paper that suggests ... | Summary: This paper proposes a reasoning-based approach to solving cryptic crossword puzzles, integrating large language models (LLMs) with Python formal verification.
The system generates answer candidates, derives wordplay explanations, and translates them into verifiable Python code for validation.
It achieves a new... | Rebuttal 1:
Rebuttal: We appreciate the reviewer highlighting the identified limitations of our Python-based verification system. To clarify, these shortcomings were not overlooked, but rather explicitly discussed in Section 4.5 ("Known Limitations of the System") to ensure transparency. Presenting these potential 'sh... | Summary: The paper presents a reasoning-based system for solving cryptic crossword clues using open-licensed LLMs. It follows a three-step pipeline: (1) generating answer candidates, (2) proposing wordplay explanations, and (3) verifying solutions via a Python-based formalizer. The system outperforms prior methods on t... | Rebuttal 1:
Rebuttal: In Section 4.5 ("Known Limitations of the System"), we chose to explicitly discuss the limitations of our Python-based verification system to ensure transparency. However, these potential 'shortcuts' are mainly a cautionary note regarding the potential for a Reinforcement Learning loop (in future... | Summary: This paper presents a system for solving cryptic crossword clues using a collection of fine-tuned and ICL-prompted open-weight language models, as well as a custom Python-based domain-specific interpreter. The proposed system first samples a set of candidate solution words from a fine-tuned proposal model. Ano... | Rebuttal 1:
Rebuttal: ### Answer to "How many wordplay/definition suggestions are generated per answer candidate?"
For each clue, we generate 20 candidate answers (the cumulative probability graph for the upper bound of success after this is given in Figure 7a). These are then deduplicated, and we create 5 definition... | null | null | null | null | null | null |
Learning Single Index Models with Diffusion Priors | Accept (poster) | Summary: The work addresses the problem of signal reconstruction in semi-parametric single-index models, where the link function is unknown. They propose a new method relying on parametrizing the signal prior by a diffusion model. Building on the observation that the measurements may be related to noisy versions of the... | Rebuttal 1:
Rebuttal: Thanks for your recognition of this paper and the helpful comments and suggestions. Our responses to the main concerns are as follows. All citations refer to the reference list in the main document.
(**To strengthen the paper, it would be beneficial to include a version of Theorem 2 for the SIM-D... | Summary: The authors propose a diffusion model sampling scheme to reconstruct an unknown signal from measurements, assuming a single-index model with a known compressed sensing matrix and noise distribution but unknown and potentially nondifferentiable link function. The approach leverages a property of the link functi... | Rebuttal 1:
Rebuttal: Thanks for your helpful comments and suggestions. Our responses to the main concerns are as follows. All citations refer to the reference list in the main document.
(**It is strange that DPS and DAPS with knowledge of the link function (i.e., DPS-N and DAPS-N) perform worse than they do without k... | Summary: In this manuscript, the authors address a notable shortcoming in current signal recovery techniques based on diffusion models: most existing methods either concentrate on narrowly defined reconstruction tasks or fail to handle nonlinear measurement models with discontinuous or unknown link functions. To tackle... | Rebuttal 1:
Rebuttal: Thanks for your recognition of this paper and the helpful comments and suggestions. Our responses to the main concerns are as follows.
(**FID is widely recognized as a key evaluation metric in diffusion models. Why wasn’t it employed here?**)
We carry out additional experiments to report the FID... | Summary: Summary
This paper proposes a novel method for reconstructing images from measurements obtained through a nonlinear compressed sensing model. The degradation model consists of a measurement matrix, an unknown and potentially discontinuous nonlinear element-wise link function, and additive Gaussian noise.
The... | Rebuttal 1:
Rebuttal: We thank the reviewer for the feedback. We are pleased that the claims in our submission have been recognized as "supported by evidence" and our experiments have been recognized as "sound". Regarding your major concern about the practical relevance of Theorem 2, our responses are as follows:
- Fo... | null | null | null | null | null | null |
Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing | Accept (poster) | Summary: This paper extends the *locate-then-edit* approach of model editing proposed by [Meng et al, 2023](https://arxiv.org/pdf/2202.05262) to multi-hop factual recall tasks. During the localization experiments the authors find that: on multi-hop factual recall tasks, the LM retrieves implicit subject information in ... | Rebuttal 1:
Rebuttal: Thank you very much for your recognition of our work. Here is our response.
> Response to Q1.a in Questions For Authors
In the past exploration process, we also used patchscope to perform the experimental effects shown in figure 2, and judged whether the corresponding information was contained b... | Summary: This paper investigates how LLMs handle multi-hop factual recall under knowledge editing. Using various interpretability methods, the authors uncover a critical insight: for multi-hop questions, LLMs rely primarily on implicit subject information encoded in deeper MLP layers to derive final answers. This mecha... | Rebuttal 1:
Rebuttal: Thank you very much for your recognition of our work and for your suggestion to extend the framework to other editing methods (such as **MEMIT**) to demonstrate the generalization of the method itself. We conducted detailed experiments to verify this point. Due to time constraints, we tested it on... | Summary: This paper introduces a method called IFMET to perform multi-hop factual edits to language models. It first conducts an analysis of LM factual recall in the presence of multi-hop edits and finds that LMs integrate hops beyond the second in deeper MLP layers, compared to single-hop facts which are retrieved in ... | Rebuttal 1:
Rebuttal: > Response to Q1.1 in Claims And Evidence.
In our actual experiments, we conducted a simple exploration, and found that the probability of articles, e.g. *the*, *a* will increase significantly in the last few layers, which will squeeze the absolute probability of the answer to a certain extent. B... | Summary: This paper identifies significant limitations in the existing knowledge editing methods based on the "locate-then-edit" paradigm, particularly when applied to multi-hop factual recall tasks. To explore these limitations, the authors first employed the LogitLens technique and discovered that multi-hop queries t... | Rebuttal 1:
Rebuttal: Thank you very much for your recognition of our work. Here is our response.
> Response to Q1:How is the dimension of K0 determined in Equation 3?
K0 represents the knowledge we try to preserve when modifying specific fact. Throughout the calculation process, we do not calculate K0 separately, but... | null | null | null | null | null | null |
On Efficient Estimation of Distributional Treatment Effects under Covariate-Adaptive Randomization | Accept (poster) | Summary: This paper proposes a method of estimating the distributional treatment effect, in a setting that uses randomised experiments using covariate-adaptive randomisation. The distribution is captured through the cumulative distribution function, and estimation is done via regression-adjustment.
Claims And Evidence... | Rebuttal 1:
Rebuttal: We are deeply grateful for your detailed and thoughtful review and appreciate your positive assessment of our paper.
**Theoretical Presentation**
Thank you for the valuable suggestions to improve our mathematical presentation:
1. We will provide definitions of VC-type function classes in the Ap... | Summary: The authors propose an estimator and inference method based on asymptotic normality for distributional treatment effects under a covariate-adaptive randomization. The primary estimand considered in this work is the difference between cumulative distribution functions for a fixed value $y$ between treatments. ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your detailed review of our manuscript. Your insightful feedback has greatly helped us identify key areas to strengthen in our work.
**Distinguishing Our Work from Existing Literature**
You kindly raised a question about positioning our work uniquely beyond existing appro... | Summary: The paper proposes a method to estimate distributional treatment effects in randomized experiments that leverages additional covariates, beyond stratum indicators, to improve precision. The authors posit a regression adjustment based on Neyman-orthogonal moment conditions to flexibly estimate the nuissance par... | Rebuttal 1:
Rebuttal: We are grateful for your thoughtful and constructive feedback on our paper. Your positive assessment regarding clarity and theoretical development is encouraging.
Following your comments, we will highlight three key distinctions from Jiang et al. (2023) in the revision:
1. **Discrete Random Vari... | Summary: The paper develops a regression‐adjusted estimator for distributional treatment effects (DTEs) under covariate‐adaptive randomization (CAR). It presents (i) a derivation of the estimator’s limit distribution under CA, (ii) a semiparametric efficiency bound result, and (iii) empirical/simulation demonstrations ... | Rebuttal 1:
Rebuttal: We are grateful to your detailed review of our manuscript with constructive criticism, which has helped us identify important areas for improvement.
**Discrete Outcomes and Assumption 5.1**
You raised an important point regarding our claim about handling discrete outcomes versus Assumption 5.1 r... | null | null | null | null | null | null |
IO-LVM: Inverse Optimization Latent Variable Models with Graph-based Planning Applications | Reject | Summary: This paper introduces IO-LVM (Inverse Optimization Latent Variable Models), and this method learns latent representations of constrained optimization problem (COP) costs based on observed solutions, with applications to graph-based planning problems. The authors come up with a new variate of VAE training loss ... | Rebuttal 1:
Rebuttal: The main points from the reviewer are:
* Q1) How the baseline VAE is trained to output y
* Q2) Marginal improvement only compared to VAEs (Table 1)
* Q3) RL-based and other baselines (also to reviewer 2TrQ)
* Q4) Problems other than graphs
.
* Answer to Q1 and Q2.
This is a key question. Actu... | Summary: This paper develops a latent variable model for constrained optimization problems like path planning in graphs. The key insight is to modify the traditional Variational Auto-Encoder (VAE) with two distinct stages of reconstruction - from the latent space to an unconstrained space, and from the unconstrained sp... | Rebuttal 1:
Rebuttal: The main points from the reviewer are:
* Q1) Results showing that the latent representations learned are distengled / interpretable compared to standard VAEs
* Q2) Concern of the graphs in the experiments being only in 2D
* Q3) Writing
* Q4) In what type of applications the proposed approach migh... | Summary: The paper proposes a representation learning based method for generating feasible solutions for constrained optimization problems. Given a black-box solver and a dataset of constraints and their solutions, the proposed IO-LVM algorithm learned a latent representation model which is used to generate feasible so... | Rebuttal 1:
Rebuttal: The main points from the reviewer are:
* Q1) Requirement of a blackbox solver in the framework.
* Q2) No evidence supporting if minimizing the fenchel yong loss can lead to a good latent representation.
* Q3) Baselines / Related work
.
* Answer to Q1.
The requirement of a blackbox solver in th... | Summary: The paper provides a method to learn the representations of the underlying "cost" functions of trajectories in a structured domain. They achieve it by using a VAE like method that uses a "fenchel young" loss function instead of the reconstruction error. The representations are supposed to encode the cost that ... | Rebuttal 1:
Rebuttal: The main points from the reviewer are:
* Q1) Y space restriction and inner product formulation.
* Q2) Equation 6 (gradient estimation)
* Q3) Limited-size problems compared to standard Deep Learning
* Q4) VAE has low number of latent dimensions
* Q5) Related work and baselines
* Q6) Tables 1 and ... | null | null | null | null | null | null |
Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas | Accept (poster) | Summary: The authors explore how vision-language models’ struggle with spatial reasoning, focussed on how misdirected attention (i.e. to irrelevant parts of image) within transformer blocks contributes to such behavior. They analyze attention patterns and report how attention prioritizes text tokens over image tokens. ... | Rebuttal 1:
Rebuttal: We thank Reviewer 3qjz for the comments. However, we are concerned that some comments suggest a disconnect or oversight of our paper. We respond to each point below, and respectfully encourage the reviewer to revisit our work.
> Essential References Not Discussed
Among the four references mention... | Summary: This paper investigates the Visual Attention Distribution in VLM and finds that it affects vision-centric understanding. Based on such observation, paired with the confidence score of the model when generating tokens, the authors propose ADAPTVIS, a temperature scaling mechanism for the attention scores that e... | Rebuttal 1:
Rebuttal: We thank Reviewer p97N for the encouraging comments and thoughtful feedback. Below, we address the concerns raised in detail.
> I'm wondering how well the proposed method works for other VLMs.
- Experiments on Qwen2-VL, a SOTA VLM with different architecture. We intervene in the image attention ... | Summary: The paper examines the attention patterns in the visual-language models, and finds patterns which might explain why spatial reasoning can be hard for VLMs. Specifically, they first find that a large chunk of the attention is focussed on the text stream, even though the number of visual tokens are more. However... | Rebuttal 1:
Rebuttal: We appreciate your thorough review and detailed comments! Your suggestions will be helpful in improving the paper. We address your concerns below.
Q1:
>whether the claims of this paper are generally valid for VLMs or is it specifically valid for LLAVA models that this paper tests on.
- The ... | Summary: This paper introduces ADAPTVIS, an adaptive attention mechanism designed to enhance spatial reasoning in vision-language models (VLMs). By analyzing attention distributions, the authors identify that errors often arise when models focus on irrelevant image regions and that attention patterns differ between fam... | Rebuttal 1:
Rebuttal: Thanks for your comments and advice, We address your concerns below.
> it is unclear whether the findings would generalize to other VLMs.
- We agree generalizability is important so we have especially included different variances of LLaVA-series models, with LLaVA-1.5 (224×224 visual encoder, ML... | null | null | null | null | null | null |
Enhancing Rating-Based Reinforcement Learning to Effectively Leverage Feedback from Large Vision-Language Models | Accept (poster) | Summary: This paper introduces ERL-VLM, a straightforward yet effective method for learning reward functions by leveraging feedback from VLMs. By querying a VLM for absolute ratings of trajectory segments rather than relying on pairwise comparisons, the method aims to improve sample efficiency and expressiveness of the... | Rebuttal 1:
Rebuttal: We thank the reviewer for their in-depth review, and for recognizing the simplicity and effectiveness of our method, the strength of our results, and our efforts in real-world experiments. We address your comments in detail below:
- **Q1**. Could you elaborate on the criteria for selecting traject... | Summary: The paper introduces ERL-VLM, a method that efficiently utilizes feedback from large VLMs like Gemini to generate reward functions for training RL agents. Instead of pairwise comparisons, it queries VLMs for absolute evaluations of individual trajectories on a Likert scale. This approach allows for more expres... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed and thoughtful review, and for recognizing the novelty of our method, the extensiveness of our experiments, and the strength of our results across a wide range of tasks and domains. We respond to your comments in detail below:
- **Q1**. The method is highly ... | Summary: This paper studies the problem of automated reward generation for RL policy training via VLMs. While prior work has shown that rewards can be extracted from VLMs either by preference (relative comparison) between two or more trajectory segments, or by using the representation itself as a distance metric, these... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful and constructive feedback, and for recognizing the clear motivation behind ERL-VLM, the clarity of our writing, and the strength of our experimental results. We address your comments in detail below:
- **Q1**. What do trajectories labeled with different ... | null | null | null | null | null | null | null | null |
Nemotron-CORTEXA: Enhancing LLM Agents for Software Engineering Tasks via Improved Localization and Solution Diversity | Accept (poster) | Summary: This paper proposes CORTEXA, a coding agent for bug-fixing based on Agentless v1.5. CORTEXA differs from Agentless in two key ways: 1) it uses an embedding-based file-retrieval (using a finetuned embedding model) combined with an agentic entity retrieval approach instead of direct prompting for localization, 2... | Rebuttal 1:
Rebuttal: We truly appreciate the time and effort you have put into reviewing our work and are grateful for your insightful comments that help improve our paper. For comparisons to SOTA works, please refer to our response to Reviewer WZSG. We would like to highlight that Cortexa outperforms Agentless in sev... | Summary: - This paper introduces CORTEXA, a software agent that involves training a model specifically for localizing the right files, building a localization agent to identify the right entities within a file, and a workflow for diverse patch generation and selection.
- CORTEXA outperforms Agentless and achieves simil... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our work and for offering constructive feedback to help enhance our paper. Regarding the comparison to the SOTA works, please refer to our response to Reviewer WZSG. Additionally, our sample efficiency is demonstrated by achieving higher resolution rates tha... | Summary: The work proposes an agentic system around LLM to solve GitHub issues (Swe-bench tasks).
They mainly proposed a code embedding model used for file retrieval and built a localization, diverse patch generation and filtering mechanism around it.
Finally, they demonstrated good performance of Cortexa while being ... | Rebuttal 1:
Rebuttal: Thank you very much for your valuable feedback. We appreciate your thoughtful comments.
**Comparison to SOTA:** In response to your comment on Cortexa's performance compared to other works, we would like to emphasize that our primary goal was to identify ways to increase the efficiency of coding ... | null | null | null | null | null | null | null | null |
Dynamical Modeling of Behaviorally Relevant Spatiotemporal Patterns in Neural Imaging Data | Accept (poster) | Summary: This work proposed a novel learning framework, SBIND, to model spatiotemporal neural imaging. SBIND is able to model the spatiotemporal neural dynamics of neural activity, at the same time SBIND can disentangle the behavioral relevant neural dynamics. The authors experimented with both calcium and ultrasound i... | Rebuttal 1:
Rebuttal: ### [1]: motivation for ConvRNNs
We thank the reviewer for this insightful question. There are two reasons we need to pass the ConvRNN1 states to ConvRNN2. First, ConvRNN2 aims to capture **residual neural dynamics** not explained by behaviorally relevant states, $X_k^{(1)}$. To ensure $X_k^{(2)}$... | Summary: The authors propose SBIND, which learns behaviorally relevant and irrelevant neural dynamics directly from high-dimensional imaging data without preprocessing. The authors apply SBIND to widefield imaging datasets and functional ultrasound, and find that SBIND outperforms existing methods that involve preproce... | Rebuttal 1:
Rebuttal: ### [1]: Comments & Suggestions
Thank you for these helpful suggestions for improving the clarity of our work.
* We will fix the typo on Line 87 and ensure all symbols are clearly defined upon first use. Specifically, $n_y$ is the number of neural images in $Y_k$ at time k, and H and W are the he... | Summary: This work propose SBIND, a data-driven deep learning framework to model the spatiotemporal dependencies in the neural image data and behavior data. Existing methods fail to model the dependencies of behaviors and neural dynamics. This work allows modeling the complex local and global spatial temporal patterns,... | Rebuttal 1:
Rebuttal: ### [1]: New Baselines
We thank the reviewer for their constructive feedback. We agree that comparing SBIND with more baselines strengthens our contribution. In response, we have added new baseline comparisons.
**Comparison with STNDT:**
STNDT uses a Transformer architecture for spatiotemporal ... | Summary: This work proposes SBIND, a dynamical model for neural imaging data designed to extract behaviorally relevant spatiotemporal patterns. The model mainly uses a double-RNN technique to disentangle behaviorally relevant neural dynamics from other covariates of high-dimensional neural activity. The first RNN captu... | Rebuttal 1:
Rebuttal: ### [1]: Neuroscientific meaning
We thank the reviewer for raising this key point regarding the scientific utility of neural imaging data compared to electrophysiology (EP) data, which we will now clarify in the manuscript. Brain function relies on diverse spatial and temporal scales from single n... | null | null | null | null | null | null |
Revisiting Instance-Optimal Cluster Recovery in the Labeled Stochastic Block Model | Accept (poster) | Summary: The authors study the labeled stochastic block model, which is similar to the regular SBM, but each edge is additionally assigned one label from a candidate set. 0 corresponds to no edge existing, and the authors study the sparse regime where the probability of a label not being 0 is small ($O(\log n / n)$).
... | Rebuttal 1:
Rebuttal: Thank you very much for your positive feedback and for carefully reading our draft.
> Additional data such as the running time would have been interesting to get a fuller picture on the practical properties of the algorithms.
Thank you for this suggestion. We have conducted additional experimen... | Summary: The authors propose a new tractable algorithm for cluster recovery in the labeled stochastic block model. An upper bound for the asymptotic error rate is derived and is shown to match known lower bounds.
## Update after rebuttal
I maintain my score, thank you.
Claims And Evidence: The proposed algorithm foll... | Rebuttal 1:
Rebuttal: Thank you very much for your positive feedback and for carefully reading our draft. Please let us know if you have any further questions. | Summary: This paper considers the problem of community detection in the Labeled SBM, a generalization of the standard SBM in which each edge is associated with one of L+1 labels (where the zero label is most frequent). The authors study the case of a growing number of communities and propose an algorithm for achieving ... | Rebuttal 1:
Rebuttal: Thank you very much for your positive feedback and for carefully reading our draft.
> it would be nice to see an empirical validation of the threshold in Theorem 1.2.
We will include a comparison plot of the empirical error rates and the lower bound by varying $n$ in the appendix of the paper.
... | Summary: This paper provides a detailed algorithm for clustering under the LSBM model, and theoretically shows that it achieves the known asymptotic lower bound on the number of miss-classifications (YP2016). Furthermore, the algorithm does not need to know the LSBM parameters, essentially showing that the bound in (YP... | Rebuttal 1:
Rebuttal: Thank you very much for your positive feedback and for carefully reading our draft.
> The authors could move some of the discussion to the supplementary material and instead present the numerical studies in the main body of the paper.
Thank you for the suggestion. We will take this into consider... | null | null | null | null | null | null |
Textual Unlearning Gives a False Sense of Unlearning | Accept (poster) | Summary: The paper investigates the effectiveness of machine unlearning (MU) in LMs and introduces new auditing and attacking methods to evaluate its reliability and privacy risks. They propose U-LiRA+, which uses mislabeled samples to rigorously audit unlearning effectiveness. The results reveal that over 70% of unlea... | Rebuttal 1:
Rebuttal: # Response to Reviewer PnWC
We sincerely thank the reviewer PnWC for your valuable and constructive feedback!
## Q1: Concerns about our assumptions and their practicality.
We would like to provide further explanations on our assumptions:
(1) In the **black-box scenario**, we assume that an advers... | Summary: This paper critically demonstrates that current machine unlearning mechanisms give a false sense of effective unlearning. First, they propose U-LiRA+, a rigorous textual unlearning auditing method, and find that the unlearned texts can still be detected with very high confidence after the unlearning process. F... | Rebuttal 1:
Rebuttal: # Response to Reviewer 9rYC
We sincerely thank the reviewer 9rYC for your valuable and constructive feedback!
## Q1: How about the proposed reconstruction attack (TULA-DR) against the exact unlearning method (including retraining)?
We mainly focus on TULA-DR against inexact unlearning because: (... | Summary: The authors demonstrate that current unlearning methods fail to adequately protect the privacy of unlearned texts in language models. To address this, they propose a robust unlearning auditing method, U-LiRA+, which utilizes membership inference attacks and deliberately introduces mislabeled samples to reveal ... | Rebuttal 1:
Rebuttal: # Response to Reviewer tLdQ
We sincerely thank the reviewer tLdQ for your valuable and constructive feedback!
## Q1: How to determine the convergence of TULA-DR?
Our proposed TULA-DR is an **optimization-based attack**. Empirically, the optimized candidates converge gradually as the number of ite... | Summary: The authors demonstrate that the textual unlearning mechanism can not ensure privacy as expected. They propose a rigorous unlearning auditing method (U-LiRA+) and and investigate privacy attacks in both black-box and white-box scenarios. Through empirical evaluations on large language models and synthetic data... | Rebuttal 1:
Rebuttal: # Response to Reviewer c3T7
We sincerely thank the reviewer c3T7 for your valuable and constructive feedback!
## Q1: Is it possible for the proposed U-LiRA+ to be adopted to other MIAs?
Yes, our approach is **MIA-agnostic** (membership inference attack), as its core idea is to **properly construct... | Summary: The paper proposes a new auditing method to check whether unlearning text from a model is completely unlearned. The auditing method called U-LiRA+ is based on U-LiRA and checks whether it is possible to differentiate between unlearned and not seen samples. Additionally two methods for investigating privacy ris... | Rebuttal 1:
Rebuttal: # Response to Reviewer p1Cm
We sincerely thank the reviewer p1Cm for the valuable and constructive feedback!
## Q1: Clear the differences between the proposed TULA and [1].
Here are the key differences:
(1) **Broader and more realistic assumptions.** ***[1] considers only the relaxed black-box s... | null | null | null | null |
Assessing Safety Risks and Quantization-aware Safety Patching for Quantized Large Language Models | Accept (poster) | Summary: The paper studies an important but relatively underexplored problem. The evaluation of existing quantization approaches clearly demonstrates the safety issues of quantization and Q-resafe gives significant benefits through widely-accepted safety measurements. Experimental results show that Q-resafe outperforms... | Rebuttal 1:
Rebuttal: Thank you for the positive and detailed comments. We have revised the manuscript to include [1–3], which highlight the safety challenges posed by quantization, and clarified our position relative to these works.
**Responses to Weaknesses**
1.Thank you for raising this important concern. To red... | Summary: This paper measures the safety of quantized methods and proposes Q-resafe, a method that restores the safety capabilities of quantized LLMs by adding a LoRA module.
## Update after rebuttal
Thank you for providing the additional results. I will raise my score to 2. However, I still have some confusion regard... | Rebuttal 1:
Rebuttal: We are sincerely thanks the reviewer for their valuable feedback. Due to limited words, we given summary response. We are eager to have more profound discussion.
**Respouse to Questions:**
**Q1. Supplementary evaluation results**
We updated the quantization baselines with state-of-the-art metho... | Summary: The paper presents a comprehensive safety evaluation of quantized LLMs. Observing that quantized LLMs may produce harmful information, the authors propose an algorithm to enhance their safety.
Claims And Evidence: The claims in the paper are supported by clear and convincing evidence.
Methods And Evaluation ... | Rebuttal 1:
Rebuttal: We greatly value your feedback and appreciate your insightful suggestions. We have carefully considered your comments and made the necessary improvements. We are eager to have more profound discussions to further enhance our work.
**For Weaknesses: Add more quantization methods without fine-tun... | null | null | null | null | null | null | null | null |
Federated Oriented Learning: A Practical One-Shot Personalized Federated Learning Framework | Accept (poster) | Summary: The paper introduces Federated Oriented Learning (FOL), a novel one-shot personalized federated learning (OPFL) framework designed for communication-constrained environments such as LEO satellite networks. FOL integrates multi-stage processes—fine-tuning, structured pruning with alignment regularization, ensem... | Rebuttal 1:
Rebuttal: **1. Concern About the Local Validation Set Has the Same Distribution as the Final Test Set.**
**Response:** We would like to clarify that, in personalized learning, it is standard practice to assume that the local validation set and the test set follow the same distribution. In real-world scenar... | Summary: • In order to address the situation of limited client communication in federated learning, this paper introduces a novel federated learning paradigm - OPFL and presents a four-stage one-shot PFL algorithm FOL (Federated Oriented Learning). FOL can learn a personalized model for each client without the need of ... | Rebuttal 1:
Rebuttal: **1. Fairness Issue in the One-Shot Communication Setting.**
**Response:** Please note that for personalized learning, fairness does not mean equal accuracy across individual users, but instead it means every user has comparable opportunity to improve its accuracy (i.e., opportunity of learning).... | Summary: This paper first introduces an important limitation of existing Personalized Federated Learning methods, which is the need of multiple communication rounds to update models. This will lead to massive communication costs and impracticable for the real-world scenarios. Moreover, the authors argue that personaliz... | Rebuttal 1:
Rebuttal: **1. Optimal Weighted Ensemble Clarification.** *Optimal Weighted Ensemble described in Section 3.3 appears to be misleading...; Additionally, the statement that "the ensemble outcome (a.k.a. the teacher model) is K times larger...*
**Response:** We respectfully clarify that the "Optimal Weighted... | Summary: This paper proposes Federated Oriented Learning, a novel framework for One-Shot Personalized Federated Learning designed for environments with constrained or infrequent communication or limited contact windows. The authors further provide two theoretical guarantees on empirical risk discrepancy between student... | Rebuttal 1:
Rebuttal: **1. Regarding Evaluation on 1–3 Clients per Setting.**
**Response:** The accuracy of 6 additional clients are given in following Table. Similar performance trends as those 3 shown in our original paper can be observed in this table (i.e., those 3 shown in our original table indeed are representa... | null | null | null | null | null | null |
LongRoPE2: Near-Lossless LLM Context Window Scaling | Accept (poster) | Summary: This paper mainly introduces LongRoPE2, aiming to achieve an effective long context window while preserving short-context performance by context extension. Based on LongRoPE, LongRoPE2 introduces a new needle-PPL guided evolutionary search method for settling the rescaling factors, and proves it to be more eff... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and for recognizing our contributions. We greatly appreciate your acknowledgment of our New RoPE OOD Hypothesis and the role of needle-PPL-guided search in validating this hypothesis through empirical results. We are also glad that you found our extensive exper... | Summary: This paper proposed LongRoPE2, a RoPE scaling method to extend the context window of LLMs. The primary extension compared to LongRoPE1 is that LongRoPE2 utilizes a needle-based search rather than perplexity-based one for various rope dimension scaling. The experimental results demonstrate the superior performa... | Rebuttal 1:
Rebuttal: >Q1: Clarification on LLaMA3.1-8B long-context evaluation numbers, and the "overclaim" comments
**Response**: We appreciate your feedback and would like to clarify the following points:
1. **65.1 is the En.MC score of the instruct version, not LLaMA3.1-8B.**: As noted in Table 2 of the LLaMA3.1... | Summary: Maintaining the performance on both long and short benchmarks are a critical challenge for existing long context extension methods. LongRoPE2 is a new approach that extends the effective context window of pre-trained large language models to the target length, while preserving the performance on the original s... | Rebuttal 1:
Rebuttal: **Response**: Thank you for your valuable feedback and for recognizing the strengths of our work. We appreciate the opportunity to address your concerns.
1) **Affordable evolutionary search computational cost**: we acknowledge that evolutionary search introduces additional costs. To further clari... | null | null | null | null | null | null | null | null |
Better to Teach than to Give: Domain Generalized Semantic Segmentation via Agent Queries with Diffusion Model Guidance | Accept (spotlight poster) | Summary: This paper proposes QueryDiff, an agent query-driven learning framework based on diffusion model guidance for DGSS, which utilizes the scene distribution priors embedded in diffusion models to enhance semantic segmentation generalization. Various experiments show the model’s effectiveness and reach the sota pe... | Rebuttal 1:
Rebuttal: **Q**: It is suggested that an ablation experiment of AQ + $L_{dist}$ (without $L_{sup}$) be included in Table 3.
**A**: The purpose of $L_{dist}$ is to suppress the visual texture details within the matrix $S_j^{t_w}$. Subsequently, $S_j^{t_w}$ is utilized in Equation (10) to derive optimized qu... | Summary: This paper proposes an agent Query-driven learning framework based on Diffusion model guidance for DGSS. The method employs agent queries to learn scene distribution knowledge from the diffusion model, capitalizing on the inherent consistency of this distribution across domains to improve segmentation model ge... | Rebuttal 1:
Rebuttal: **Q**: The proposed method should be compared with previous methods in terms of computational complexity, time consumption, and memory usage.
**A**: Thank you for the suggestion. We have included a detailed comparison of computational resources between our proposed method and recent diffusion-bas... | Summary: The authors leverage on refined features of diffusion models to stabilize the features of vision transformers and other backbones when feeding them into the mask2former decoders for semantic segmentation. In this way, the authors achieve considerable domain generalization capabilities for their network. In the... | Rebuttal 1:
Rebuttal: **Q**:Reported results for Rein differ from those in its original paper.
**A**:The performance difference arises because we reproduced Rein at a 512×512 resolution on the ACDC validation set using the official code—as indicated in Table 2 (line 351)—rather than the original 1024×1024 resolution o... | Summary: The paper presents a novel framework for utilizing diffusion models for domain-generalized semantic segmentation. While previous works often struggle to generate reasonable scenes for semantic segmentation, this paper introduces agent queries from segmentation features and incorporates additional pretrained kn... | Rebuttal 1:
Rebuttal: **Q**: Have you conducted experiments in in-domain settings, such as few-shot or fully-supervised learning on the Cityscapes dataset? This approach could potentially achieve state-of-the-art performance even beyond domain-generalized settings.
**A**: Thank you for this valuable suggestion. We con... | null | null | null | null | null | null |
Unnatural Languages Are Not Bugs but Features for LLMs | Accept (poster) | Summary: This paper studies unnatural prompts, strings that seem unintelligible to humans yet able to make Language Models produce a specific target output. The paper claims that unnatural prompts contain latent features that LMs respond to. Using a gradient-based method, the authors find the unnatural versions of exam... | Rebuttal 1:
Rebuttal: Thank you for your insightful reviews and comments. We appreciate the time and effort you have put into providing valuable feedback. We would like to address your concerns as follows:
---
> **Concern #1 Unnatural language contains keywords**
We acknowledge that the unnatural language contains ke... | Summary: This paper offers an interesting investigation into whether and how LLMs interpret unnatural contexts on various tasks. It proposes a heuristic optimization algorithm to search for the optimal unnatural tokens based on the log probabilities. Two synthetic datasets are also curated for fine-tuning and evaluatio... | Rebuttal 1:
Rebuttal: Thank you for your insightful reviews and comments. We appreciate the time and effort you have put into providing valuable feedback. We would like to address your concerns as follows:
---
> **Concern #1 Improved searching algorithm**
In the current implementation, we use the GCG algorithm [1], w... | Summary: This study posits that LLMs are highly effective at picking up on latent patterns in non-human-readable strings. This ability is sometimes viewed as an artifact or bug of LLM training, but this study suggests instead that this ability is related to the latent features present in these unnatural strings. To dem... | Rebuttal 1:
Rebuttal: Thank you for your insightful reviews and comments. We appreciate the time and effort you have put into providing valuable feedback. We would like to address your concerns as follows:
---
> **Concern # 1 Length of input**
Yes, we completely agree with your point. This is precisely why we employ ... | Summary: This work argues that there exist versions of human language that is not human readable (unnatural), but maintains a semantic meaning for large language models (LLMs). The authors propose a gradient-based sampling procedure to translate natural to unnatural language for a given LLM, or use GPT-4 to perform the... | Rebuttal 1:
Rebuttal: Thank you for your insightful reviews and comments. We would like to address your concerns as follows:
---
> **Concern #1: Experiment details**
1. Uncertainty
In Table 2, we do not report uncertainty because the decoding temperature is set to 0, eliminating any randomness. However, to further a... | null | null | null | null | null | null |
Instruction-Following Pruning for Large Language Models | Accept (poster) | Summary: This article tackles the problem of letting LLMs select the most suited parameters for each prompted task and proposes a novel instruction-following pruning paradigm called IFPruning. Specifically, IFPruning uses a sparse mask predictor to predict a input-dependent mask for each context input. To train the pre... | Rebuttal 1:
Rebuttal: We thank Reviewer bo9C for the support and the valuable suggestions. Below we address each point raised.
**Q1: Clarification on actual model speedup.**
Since inference speed is concerned, we would like to first clarify that our method is motivated and designed for on-device models (e.g. on smart... | Summary: The paper introduces Instruction-Following Pruning (IFPruning), a dynamic structured pruning method for large language models (LLMs). Instead of using a fixed sparsity mask, IFPruning employs a sparse mask predictor that selects the most relevant model weights (specifically, rows/columns of transformer feed-fo... | Rebuttal 1:
Rebuttal: We thank Reviewer Kh2x for the support and the valuable suggestions. Please see our response below.
**Q1: Why does continued pre-training help and what if we remove it?**
Thank you for the question. We first explain our motivation followed by the ablation study on continued pre-training.
- Intu... | Summary: The paper proposes "Instruction-Following Pruning" (IFPRUNING), a novel approach to dynamic structured pruning of large language models (LLMs). Unlike traditional static pruning methods that determine a fixed pruning mask for a model, this approach generates input-dependent pruning masks that adapt based on th... | Rebuttal 1:
Rebuttal: We thank Reviewer epS3 for the support and the valuable feedback. Please see our response below.
**Q1: What is the reasoning behind designing a method that requires task-specific pruning for general-purpose LLMs?**
We totally agree with the reviewer that having “general-purpose LLMs and adaptabi... | Summary: The paper proposes a dynamic pruning method in which a router determines the pruning strategy of the FFN layers in an LLM model using the input instruction. The sparse mask predictor and LLM weights are jointly trained using instruction-following data and the pre-training corpus. Experiments on different targe... | Rebuttal 1:
Rebuttal: We thank Reviewer Lv3b for the support and the valuable suggestions for our paper. We will address the writing feedback, such as adding statistics of the dataset and discussing additional related work, in the next version. Please see our response to the questions and/or major comments below.
**... | null | null | null | null | null | null |
Risk-aware Direct Preference Optimization under Nested Risk Measure | Reject | Summary: This paper tackles token-level preference optimization for LLM alignment by making it “risk-aware.” It modifies the usual Bradley-Terry setup to include a nested risk measure that accounts for potential variability in model updates. They define a token-level advantage function that uses this risk measure, lead... | Rebuttal 1:
Rebuttal: **We sincerely appreciate the valuable comments from the reviewer. We hope our responses below provide further clarity.**
## Response to Other Comments Or Suggestions: The running …… "Theorem 3.6 Restated"?
We apologize for the confusion caused by our oversight. We will correct these errors in the... | Summary: This paper presents a risk-aware version of direct preference optimization (DPO) algorithm. The key innovation is to employ a risk-aware objective that operates at the token level (which results in a different algorithm due to the presence of KL divergence). The risk is calculated sequentially in terms of the ... | Rebuttal 1:
Rebuttal: **We sincerely appreciate the valuable comments from the reviewer. We hope our responses below provide further clarity.**
## Response to Claims And Evidence:
_**Risk awareness:**_ In this paper, risk awareness refers to the sensitivity to risks arising from deviations from the reference model. I... | Summary: The paper introduces Risk-aware Direct Preference Optimization (Ra-DPO), a new method for fine-tuning token-level large language models (LLMs) with higher-order nested risk measures. The moderation of path dependency utilizes Bellman equation. Comprehensive theoretical remarks and justification are provided re... | Rebuttal 1:
Rebuttal: **We sincerely appreciate the valuable comments from the reviewer. We hope our responses below provide further clarity.**
## Response to Methods And Evaluation Criteria: Yes, the …… different datasets.
In Appendix Figures 6 and 7, **we present experimental results conducted on the Anthropic HH da... | Summary: This paper introduces a risk-aware direct preference optimization method that incorporates a nested risk measure into a token-level objective function. The ultimate objective function maximizes the likelihood of the policy while suppressing the deviation between a training model and the reference model using a... | Rebuttal 1:
Rebuttal: **We sincerely appreciate the valuable comments from the reviewer. We hope our responses below provide further clarity.**
## Response to Weaknesses 1:
We apologize for the confusion caused by failing to give a clear explanation and would like to re-clarify our motivation.
Before restating our m... | null | null | null | null | null | null |
Procurement Auctions via Approximately Optimal Submodular Optimization | Accept (spotlight poster) | Summary: The paper studies the design of procurement auctions with submodular welfare. The problem involves an auctioneer and $n$ sellers, each possessing an item for sale with a private cost $c_i$, representing the minimum price at which they are willing to sell. The auctioneer's valuation over items is given by a mon... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to read our paper and for their valuable feedback. We will re-organize section 4 and name the assumptions we use for the submodular optimization algorithm. | Summary: This paper focuses on procurement auctions where an auctioneer aims to acquire services from strategic sellers with private costs. The quality of services is represented by a submodular function, and the goal is to design efficient mechanisms that maximize the difference between service quality and total selle... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to read our paper and for their valuable feedback.
>I hope the author can analyze the complexity of the algorithm in more detail.
Thanks for the comment, notice that all of our algorithms run in polynomial time. For example, in algorithm 2... | Summary: In this paper, they develop a framework to convert a family of greedy algorithms for submodular maximization to a mechanism for procurement actions.
Moreover, they provide an improved analysis of the Distorted Greedy algorithm.
Finally, they consider the case of Descending auctions where they design a mechanis... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to read our paper and for their valuable feedback. We respond to each of the points they raised below.
>Information on the specs of the machine and the MIP solver.
Thanks for the suggestion, we will add more details about the specs of the m... | Summary: The paper studies a procurement mechanism with objective function f(S) - \sum_{i \in S} p(i) that is truthful, individually rational, and has nonnegative surplus, and provide a bi-criteria approximation.
To the best of my knowledge, this is the first to study such objective function in the procurement auction ... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to read our paper and for their valuable feedback. We agree with the suggestions and we will do the appropriate reorganization of our content based on their feedback. If our manuscript gets accepted, we will also make sure to utilize the extr... | null | null | null | null | null | null |
SEMU: Singular Value Decomposition for Efficient Machine Unlearning | Accept (poster) | Summary: The paper proposed a machine unlearning method that only fine-tunes by the subspace of the gradient orthogonal to the weight. It claims to effectively “unlearn” forgetting sets while eliminating the dependency on the original training dataset.
Claims And Evidence: see Methods And Evaluation Criteria
Methods ... | Rebuttal 1:
Rebuttal: **Referencing other works** Thank you for highlighting these works. We will discuss the differences between SEMU and these methods and include this analysis in the camera-ready version of our work. Regarding [1], our method does not rely on samples or gradients from the remaining dataset, nor do w... | Summary: This paper proposes a machine unlearning (MU) method named Singular Value Decomposition for Efficient Machine Unlearning (SEMU). The authors disentangle the gradients of parameter weights with Singular Value Decomposition (SVD) to identify the important proportion for MU. They keep all original weight matrices... | Rebuttal 1:
Rebuttal: We appreciate your feedback. Below we address the concerns.
**Previous works R1 and R2** Regarding R1, we would like to highlight that it requires an additional surrogate dataset $\mathcal{D}^{sur}$, which is not required for SEMU. This means our approach does not rely on extra datasets to maint... | Summary: The paper proposes SEMU, a machine unlearning method using Singular Value Decomposition (SVD) to efficiently erase specific data influences from trained models. SEMU leverages SVD to project model gradients into a low-dimensional subspace, identifying critical weights linked to unwanted data. By updating a sma... | Rebuttal 1:
Rebuttal: **Time consumption**
In the Table below we show a comparison of time needed to unlearn DDPM model:
|Method|Preprocessing time|1000 iters time|
|:-|:-:|:-:|
|SEMU|44.18s|308s|
|SEMU_retrain|44.18s|530s|
|SalUn|50.69s|1170s|
Also, we show the time of unlearning of ResNet-18:
|Method|Dataset|Prepro... | Summary: The authors proposed a Singular Value Decomposition for Efficient Machine Unlearning method which solve two problems 1) the need remaining dataset for unlearning process and 2) changes too many parameters during unlearning process
Claims And Evidence: This article show good evidence to support its claim that ... | Rebuttal 1:
Rebuttal: **Why SVD?** Our objective was to minimize the number of altered weights during the unlearning process to maintain the model's behavior. To achieve this, we looked for an effective selection mechanism. SVD, our first choice, proved to be successful, so we did not explore other parameter selection ... | Summary: The paper "SEMU: Singular Value Decomposition for Efficient Machine Unlearning" introduces a new method for machine unlearning (MU). The goal is to remove specific data from AI models without damaging overall performance. Traditional unlearning methods require modifying large portions of the model or retrainin... | Rebuttal 1:
Rebuttal: Thank you for your thorough review. We would like to address some of your concerns below:
**Time needed for SEMU when compared to SalUn** In the Table below we show a comparison of time needed to unlearn DDPM model:
|Method|Preprocessing time|1000 iters time|
|:-|:-:|:-:|
|SEMU|44.18s|308s|
|SEM... | Summary: The paper performs an SVD decomposition for machine unlearning, which enables them for efficient unlearning. They also propose a dataset-free scenario, addressing data privacy concerns. Experiments show their superiority over other methods
At the core, SEMU aims to change a minimal number of model parameters... | Rebuttal 1:
Rebuttal: To address the concerns and questions raised by the Reviewer, we would like to point out the following:
**More on theoretical aspects of projection**. In practice, some directions are more important than others for all weights. Observe that the weights are roughly proportional to the averaged gr... | null | null |
scSSL-Bench: Benchmarking Self-Supervised Learning for Single-Cell Data | Accept (spotlight poster) | Summary: This paper proposed a benchmarking analysis for SSL method's application in single-cell data analysis.
## update after rebuttal
I raised my score.
Claims And Evidence: Yes.
Methods And Evaluation Criteria: Yes.
Theoretical Claims: No, they do not have theoretical claims in the manuscript.
Experimental De... | Rebuttal 1:
Rebuttal: We express sincere gratitude to the reviewer for providing feedback and raising several points about the validity of the work, which we address below and extend our evaluation accordingly. We hope that the reviewer will consider updating their review score if they find our comments and new results... | Summary: This paper proposes a self-supervised learning (SSL) benchmark for single-cell data. The authors tried twelve representative SSL methods and conducted comprehensive evaluations on eight datasets across three downstream tasks. The experimental designs are technically sound, and the paper is well-organized and w... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback. We appreciate your recognition that our paper could be a helpful reference for researchers interested in single-cell representation learning and that we conducted comprehensive evaluations. We address your questions and suggestions, which we h... | Summary: The authors present scSSL-Bench, a single-cell data benchmark that integrates 12 different approaches and 8 different datasets. The authors run extensive experimentations to answer three critical questions and provide invaluable insights and takeaways - This is no easy feat considering there are many moving pa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive feedback and highlighting the relevance of our benchmark and the quality of our experiments. In the following, we address your questions and suggestions, which have improved the quality of the paper and the benchmark.
**Single-cell FM:** We acknowledge... | null | null | null | null | null | null | null | null |
Parametric Scaling Law of Tuning Bias in Conformal Prediction | Accept (poster) | Summary: The manuscript explores the phenomenon of tuning bias in the field of conformal prediction, which is a statistical method used to ensure that prediction intervals or sets cover the true value with a specified probability. The focus is on how the tuning of parameters, when done on the same dataset used for cali... | Rebuttal 1:
Rebuttal: Thanks for your positive review and insightful feedback.
**1. Exchangeability definition/discussion**
Thank you for the suggestion. We agree that introducing the exact definition earlier can improve the clarity. In the current version, we defaultly introduce Exchangeability as a well-known assum... | Summary: In this paper, the authors focus on the tuning bias produced by parameter tuning in many conformal preidction methods. First, they reveal that the tuning bias is negligible for simple parameter tuning in many conformal prediction methods. Then, the authors establish a parametric scaling law, showing that tuni... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback. Below, we address your concerns point by point.
**1. C-Adapter vs. Vector Scaling**
> Please explain why C-adapter achieves much smaller tuning bias but vector scaling cannot? It seems C-adapter tunes more parameters than VS.
Thank you for the i... | Summary: This paper points out the problem that the exchangeability assumption of conformal prediction does not hold if the holdout set (applied for parameter tuning) and calibration set are identical. A parametric scaling law is proposed such that the tunning bias increases with parameter space complexity and decrease... | Rebuttal 1:
Rebuttal: We thank the reviewer for the nuanced and constructive feedback. Below, we address your concerns point by point.
**1. Lack of a sophisticated solution**
We want to clarify that the primary objective of this work is to provide a comprehensive understanding of tuning bias in conformal prediction r... | Summary: This paper finds out that the coverage gap of using same dataset for tuning and calibration is negligible in most of the conformal prediction methods. Also, this paper observes a scaling law about how parameter space complexity and calibration set size influence the tuning bias. Then this paper proposes a theo... | Rebuttal 1:
Rebuttal: Thanks for your positive and valuable feedback.
**1. Results of reducing parameter numbers:**
Thank you for the suggestion. In the manuscript, we presented two empirical evidences to validate the effect of reducing the parameter number:
1. TS vs. VS (Table 1, Fig. 1 and 3): we present a pilot s... | null | null | null | null | null | null |
Domain-Adapted Diffusion Model for PROTAC Linker Design Through the Lens of Density Ratio in Chemical Space | Accept (poster) | Summary: In this work the authors explore a domain-adapted diffusion model for unconditional molecular generation, focusing on PROTAC linker design. While this is an interesting application, the novelty is somewhat limited as it does not explore one of the most compelling aspects of molecular design: conditional molecu... | Rebuttal 1:
Rebuttal: Thanks for the constructive comments!
# 1. Experimental Details
Regarding the training/test split, since the proposed model is a pretrain-finetuning model, we use different datasets for the two phases. We use the ZINC dataset (438610 small molecule samples as the training set) to pretrain the mod... | Summary: This paper introduces DAD-PROTAC, a domain-adapted diffusion model for designing linkers in Proteolysis-targeting chimeras (PROTACs). The main algorithmic idea is the efficient fine-tuning strategy via density ratio estimation, avoiding full retraining of the diffusion model.
Claims And Evidence: The claims m... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments!
# 1. Figure 3
We will make Figure 3 clearer in the final version with fewer annotations and highlight more on how the score correction term is obtained. We will also explicitly annotate the input and output of two phases and each model component.
# 2. Pr... | Summary: This study focuses on domain adaptation in diffusion models for biology. The authors pretrain a diffusion model on the ZINC dataset and attempt to use the model on the PROTAC domain. For finetuning, they use density ratio estimation techniques to correct the score function on the ZINC dataset. The method looks... | Rebuttal 1:
Rebuttal: Thank you for the constructive comments!
# 1. Errors for large $t$
In Eq.(16), we train a time-dependent classifier $W(X_t^L,t)$ with samples from potentially all steps $t$ jointly, instead of training different classifiers $W_t(X_t^L)$ for each step $t$ individually. This means that the **train... | null | null | null | null | null | null | null | null |
Three-Dimensional Trajectory Prediction with 3DMoTraj Dataset | Accept (poster) | Summary: In this paper, the authors address the challenge of predicting 3D trajectories, which is more complex than 2D trajectory prediction. To achieve this goal, the authors first introduce the 3DMoTraj dataset, collected from unmanned underwater vehicles (UUVs) in oceanic environments. Then, they propose a new metho... | Rebuttal 1:
Rebuttal: **Q1: Demonstrate the generalization capability to 2D datasets on more baselines.**
**A1**: We evaluated our prediction strategy on four additional baselines using two widely adopted 2D datasets: ETH&UCY and SDD. All models were trained and tested on the same machine for fair comparison.
|Method... | Summary: The paper addresses the problem of 3D trajectory prediction by introducing a novel dataset and an innovative prediction framework. Building upon the 3DMoTraj dataset, they propose a dual-component prediction method that decomposes the 3D trajectory prediction task into two stages. The first stage, decoupled tr... | Rebuttal 1:
Rebuttal: **Q1: I look forward to seeing future validation on the trajectory dataset of unmanned aerial vehicles.**
**A1**: As outlined in our conclusion, future work will involve collecting a large-scale 3D trajectory dataset from unmanned aerial vehicles (UAVs) to validate our proposed methodology furthe... | Summary: Firstly, this paper introduces the 3DMoTraj dataset, a novel 3D trajectory dataset collected from unmanned underwater vehicles (UUVs) in oceanic environments. The dataset includes annotations for both static (endpoint octant) and motion (velocity change) intentions. Secondly, to address the increased complexit... | Rebuttal 1:
Rebuttal: **Q1: The dataset is simulation data and has not been validated in real-world scenarios.**
**A1**: While our dataset is based on predefined formation trajectories, real-world disturbances naturally cause deviations from planned paths and formation shifts. These deviations reflect real challenges ... | Summary: This paper proposes a 3D trajectory dataset named 3DMoTraj collected from unmanned underwater vehicles (UUVs) in ocean environments, which fills the research gap in this field.
Regarding the setting of 3D trajectory prediction, the paper highlights the challenge of computational complexity and provides theoret... | Rebuttal 1:
Rebuttal: **Q1: The paper tries to mitigate the computational complexity but has not provided ablations or method comparisons on inference cost.**
**A1**: At first, we clarify that the complexity our paper aims to mitigate is prediction complexity, which is crucial for optimizing 3D trajectory prediction. ... | null | null | null | null | null | null |
Enforcing Idempotency in Neural Networks | Accept (poster) | Summary: Idempotent Generative Networks (IGN) require an operator $f$ to satisfy $f = f \circ f$. This paper addresses such idempotency by
analysis from perturbation theory, identifying the polynomial $3K^2 - 2K^3$ as one that projects matrices to the idempotent matricse manifold. The authors then adapt this approach... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and valuable feedback. We are pleased that the reviewer finds this to be an important contribution with much potential, and with a strongly interdisciplinary character. We appreciate the concerns raised, and we aim to address these below.
**(W2) Runtime and me... | Summary: This paper introduces a novel approach for training idempotent neural networks.
Leveraging techniques from perturbation theory on idempotent matrices, the authors propose a new method for projecting matrices onto the idempotent manifold.
They further extend this approach to nonlinear neural networks.
Finally, ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and valuable feedback. We are happy to respond to the questions that have been raised.
**Extended experiments with CelebA.** We agree with the reviewer that extending experiments to cover other large datasets is interesting. In the graph (https://imgur.com/a/w... | Summary: The paper presents a new approach to enforcing idempotency in neural networks through a modification of the backpropagation algorithm, termed Modified Backpropagation. The key idea is the derivation of an idempotent corrector function $g(K) = 3K^3 - 2K^2$, which iteratively projects a real-valued matrix onto t... | Rebuttal 1:
Rebuttal: We thank the reviewer for their time and valuable feedback. We appreciate the concerns raised, and we aim to address these below.
**Theoretical development.** We employ a novel theoretical framework which allows gradient-free training of an idempotent property. For us it is interesting that there... | null | null | null | null | null | null | null | null |
Continuously Updating Digital Twins using Large Language Models | Accept (poster) | Summary: This paper proposes CALM-DT (Context-Adaptive Language Model-based Digital Twin), a novel digital twin framework that leverages large language models (LLMs) for simulation of dynamical systems.
Claims And Evidence: Key claim: LLMs can serve as digital twins that continuously update without re-design or retrai... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and suggestions. We give answers to the following:
- (A) Is the context window exceeded?
---
**(A) Is the context window exceeded?**
Thank you for raising this point. You are correct in saying that the initial cystic fibrosis (CF) dataset we investigated d... | Summary: This paper presents CALM-DT, a framework using large language models to create digital twins that can update continuously without redesign or retraining. Unlike traditional approaches, CALM-DT handles new variables and incorporates new information through in-context learning. Testing on cystic fibrosis patient... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and suggestions. We address the following:
- (A) Additional settings
- (B) Time-series summary ablation
- (C) Prompts
- (D) World models
- (E) LLM thoughts on 'drug X'
---
**(A) Additional settings**
We evaluate CALM-DT on three additional datasets:
1) Non... | Summary: The paper addresses the challenge of maintaining the relevance of digital twins in dynamic environments where state/action variables and relevant information constantly change. The authors frame digital twinning as an in-context learning problem using LLMs. They propose CALM-DT, which uses fine-tuned encoders ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and suggestions. We address the following:
- (A) LLM reliance
- (B) Further experiments
- (C) Application-specific metrics
- (D) LLM sample selection
- (E) Actions
---
**(A) LLM reliance**
We agree it is important to elaborate on hallucinations, bias, and f... | Summary: This paper proposes a way to use a frozen LLM (e.g. GPT-4o) to construct an auto-regressive model of the temporal evolution of a few variables, like a medical patient's height, weight, and lung function measurement, in response to certain interventions (administration of medications). The main insight is that ... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and suggestions. We address the following:
- (A) Ivacaftor experiments
- (B) Uncertainty
- (C) Simple baselines
- (D) Domain-specific model
- (E) Tokenization
---
**(A) Ivacaftor experiments**
We agree that more robust evidence is necessary for demonstrati... | null | null | null | null | null | null |
HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport | Accept (poster) | Summary: The paper proposes a self-supervised heterogeneous graph neural network (HGNN) coined ``HGOT''. It first incorporates optimal transport (OT) into heterogeneous graphs to better facilitate the learning of a more semantic accurate similarity measure between graph instances and structure. The method introduces th... | Rebuttal 1:
Rebuttal: R(1): Different from other self-supervised learning methods, optimal transport (OT) can capture the matching information from the original graph space to the representation space, obtaining node representations that exhibit consistency with the optimal transport plans. Second, reconstruction-based... | Summary: This paper presents HGOT, a self-supervised heterogeneous graph neural network that harnesses optimal transport theory to establish an optimal transport plan between the meta-path and aggregated views. By compelling the model to learn node representations that faithfully preserve the intrinsic matching relatio... | Rebuttal 1:
Rebuttal: Other Strengths And Weaknesses:
Q: However, the complexity of Gromov-Wasserstein optimization raises scalability concerns for large graphs. The paper lacks a discussion on runtime and memory overhead, which are crucial for practical deployment. Furthermore, interpretability could be improved with... | Summary: This paper proposes a novel self-supervised heterogeneous graph neural network (HGOT), which aims at addressing the limitations of existing contrastive learning methods on heterogeneous graphs by combining Optimal Transport method. HGOT avoids data augmentation and the construction of positive and negative sam... | Rebuttal 1:
Rebuttal: Experimental Designs Or Analyses:
Q1: The paper does not provide sufficient explanation or analysis for the selection of certain parameters (such as the σ and ρ parameters in optimal transport), particularly lacking a comprehensive discussion on the impact of different parameter values on model p... | Summary: This paper proposes a novel self-supervised learning framework for heterogeneous graphs that leverages optimal transport theory to align meta-path views with an aggregated central view, eliminating the need for graph augmentation or explicit positive/negative sampling. The method achieves state-of-the-art perf... | Rebuttal 1:
Rebuttal: Theoretical Claims:
Q: A concern is why the fused Gromov-Wasserstein distance is the optimal choice for heterogeneous graphs? A theoretical or empirical comparison with alternative OT metrics (e.g., Wasserstein barycenters) is missing.
R: Thank for your comments. The fused Gromov-Wasserstein dist... | null | null | null | null | null | null |
Text-Image Dual Consistency-Guided OOD Detection with Pretrained Vision-Language Models | Reject | Summary: This paper introduces DualCnst, a novel text-image dual consistency framework for zero-shot Out-of-Distribution (OOD) detection using pretrained Vision-Language Models (VLMs) like CLIP. The core idea is to leverage both semantic similarity (text-based) and visual similarity (image-based) by generating syntheti... | Rebuttal 1:
Rebuttal: Response to Reviewer aTAh
We thank the reviewer aTAh for the valuable feedback. We addressed all the comments. Please find the point-to-point responses below. Any further comments and discussions are welcomed\!
**W1:** The paper assumes that Stable Diffusion can reliably supplement visual info... | Summary: This paper propose DualCnst for CLIP-based zero-shot OOD detection. It enhances zero-shot OOD detection by combining text-image dual consistency, leveraging both semantic similarity to textual labels and visual similarity to synthesized images. This unified framework achieves state-of-the-art performance acros... | Rebuttal 1:
Rebuttal: Response to Reviewer Ci5u
We thank the reviewer Ci5u for the valuable feedback. We addressed all the comments. Please find the point-to-point responses below. Any further comments and discussions are welcomed\!
**W1:** Why ID images are not accessible (fig. 1\) ? Please justify. I think most t... | Summary: This paper presents a simple and effective method to enhance the performance of OOD detection. In addition to utilizing the similarity between test images and text features, it also introduces images through a diffusion model, thereby leveraging the similarity between test images and generated images to furthe... | Rebuttal 1:
Rebuttal: Response to Reviewer yyQ7
We thank the reviewer yyQ7 for the valuable feedback. We addressed all the comments. Please find the point-to-point responses below. Any further comments and discussions are welcomed\!
**W1:** While the method’s simplicity is a strength, it risks underselling technica... | Summary: This paper proposed a novel OOD approach named DualCnst, based on text-image dual consistency. In addition to detecting OOD samples by assessing the similarity between test images and ID/OOD label texts, this paper synthesizes OOD images using text-to-image models and incorporates the visual similarity between... | Rebuttal 1:
Rebuttal: Response to Reviewer ViPC
We thank the reviewer ViPC for the valuable feedback. We addressed all the comments. Please find the point-to-point responses below. Any further comments and discussions are welcomed\!
**W1:** The paper uses OOD-specific α (tuned per OOD dataset) , violating OOD agnos... | null | null | null | null | null | null |
Enabling Optimal Decisions in Rehearsal Learning under CARE Condition | Accept (poster) | Summary: The paper introduces a CAnonical REctangle (CARE) condition for the Avoiding Undesired Future (AUF) problem. Under this CARE condition, along with additional assumptions on the problem structure and the noise term, the AUF problem can be reformulated as a convex optimization problem. The authors propose a proj... | Rebuttal 1:
Rebuttal: Thanks for the valuable feedback! We hope our responses can address your concerns.
**Q1.** Extension for non-Gaussian noise.
**A1.** Thanks for your insightful question. We would like to clarify that the Gaussian noise assumption is primarily used to establish theoretical guarantees. For cases w... | Summary: This paper addresses the AUF (Avoiding Undesired Future) problem in machine learning decision-making, where the goal is to identify actions that prevent undesirable outcomes predicted by ML models. It introduces the CARE condition (CAnonical REctangle), a novel assumption under which the AUF probability—i.e.,... | Rebuttal 1:
Rebuttal: Thanks for your valuable feedback and appreciation of our work! We hope that our responses can address your concerns.
**W1.** Further Discussion on the CARE Condition.
**A1.** Thanks for your insightful question. In practice, the dimensions of $\mathbf{Y}$ are often dependent, meaning that the c... | Summary: The paper proposes an algorithm for decision making that helps avoid undesirable future (AUF), i.e., increasing the AUF probability. The new algorithm is shown to reduce time complexity compared to prior work and has showed performance improvement compared to a few baselines.
Claims And Evidence: The theoreti... | Rebuttal 1:
Rebuttal: Thanks for your detailed feedback! We hope our responses address your concerns.
**Q1.** Relevance to RL research.
**A1.** Thanks for your insightful question. Below, we clarify the connection between RL and AUF problem and explain distinctions.
- **Connection between RL and AUF.** When interact... | null | null | null | null | null | null | null | null |
Global Context-aware Representation Learning for Spatially Resolved Transcriptomics | Accept (poster) | Summary: The paper introduces Spotscape, a novel framework for representation learning in Spatially Resolved Transcriptomics (SRT) data. The key contribution of the paper is the Similarity Telescope module, which captures global relationships between spots, addressing the limitations of existing graph-based methods tha... | Rebuttal 1:
Rebuttal: Thank you for taking the time to provide constructive feedback on our paper. To address your concerns, we have added tables and figures in this [external link](https://anonymous.4open.science/r/Spotscape-31B6/Rebuttal.pdf)
**Q1) Theoretical Justification**
The similarity consistency loss in _Equ... | Summary: This paper proposed a new computational method, known as Spotscape, to integrate different spatial transcriptomics data. This model is improved by graph neural networks.
## update after rebuttal
I keep my score.
Claims And Evidence: Yes.
Methods And Evaluation Criteria: Yes.
Theoretical Claims: Yes, I hav... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback. To address your concerns, we upload additional results in the [external link](https://anonymous.4open.science/r/Spotscape-31B6/Rebuttal.pdf).
**Q1-1) Presentation of Figure 1**
_Figure 1_ highlights the limitations of previous methods that address the is... | Summary: The paper introduces Spotscape, a novel framework for representation learning in Spatially Resolved Transcriptomics (SRT) data. Spotscape incorporates a Similarity Telescope module to capture global similarity relationships and integrates Prototypical Contrastive Learning (PCL) and a similarity scaling strateg... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback regarding our manuscript. To address your concerns, we have added tables and figures in this [external link](https://anonymous.4open.science/r/Spotscape-31B6/Rebuttal.pdf).
**W1) The scalability analysis is incomplete**
In our _manuscript_, we initially focus... | null | null | 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.