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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
AdaWorld: Learning Adaptable World Models with Latent Actions | Accept (poster) | Summary: This paper focuses on learning world models from general videos. Unlike previous approaches that solely rely on video-based learning, this work extracts latent actions in a self-supervised manner and leverages action information for large-scale world model pretraining. With the aid of latent actions, the model... | Rebuttal 1:
Rebuttal: Thanks for the constructive feedback. We answer each question below and will include all results and discussions in the revision.
> A visualization or feature similarity analysis of the action latents.
**R4-1**: As suggested, we randomly collect 1000 samples for each action from three environmen... | Summary: The authors propose a method to incorporate latent actions into the pre-training stage of World-models allowing for more efficient adaptation to downstream tasks. The authors curate a dataset spanning from ego perspectives and third-person views to virtual games and real-world environments. Finally, they evalu... | Rebuttal 1:
Rebuttal: Thanks for the helpful feedback. We answer each question below and will include all results in the revision.
> The authors mention using two action tokens $a_{t:t+1}$ during the latent action autoencoding but they mention that they use $a_{t+1}$ to approximate the posterior. What do the authors d... | Summary: This paper proposes a pretraining framework for learning world models that can generalize to various contexts. AdaWorld first learns a latent action representation using an unsupervised forward prediction objective. Subsequently, AdaWorld learns an autoregressive world model that conditions on latent actions ... | Rebuttal 1:
Rebuttal: Thanks for the thoughtful feedback. We answer each question below and will include all results in the revision.
> Compare to state-of-the-art baselines.
**R2-1**: To demonstrate the generality of our method, we use iVideoGPT as a state-of-the-art baseline. iVideoGPT is an action-controlled world... | Summary: This paper introduces AdaWorld, a world model learning approach that leverages self-supervised latent action extraction from videos to capture key transitions. It also introduces an autoregressive world model (single-frame SVD) conditioned on these latent actions and historical frames, enabling transfer and le... | Rebuttal 1:
Rebuttal: Thanks for the insightful feedback. We answer each question below and will include all results in the revision.
> The model requires a reference video to guide actions. Clarify how the model can achieve similar controllability to Genie or OASIS at inference time.
**R1-1**: Our world model does n... | null | null | null | null | null | null |
RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation | Accept (poster) | Summary: This paper introduces R E -I MAGINE: a framework to characterize a hierarchy of reasoning ability in LLMs, alongside an automated pipeline to generate problem variations across all the levels of the hierarchy. By altering problems in an intermediate symbolic representation, RE-IMAGINE generates arbitrarily man... | Rebuttal 1:
Rebuttal: # Response to Claims And Evidence: Some datasets also have a systematic hierarchy such as GPQA
The hierarchy in GPQA represents the difficulty of the **original problems**, whereas the hierarchy we introduce in Re-Imagine defines reasoning complexity through **variations of problems** from existi... | Summary: To identify whether the performance improvement of LLMs on public benchmarks such as GSM8K indeed comes from the stronger reasoning capabilities or results from mere memorization of training cases, the authors propose RE-IMAGINE to automatically make multi-level modifications to questions in the existing bench... | Rebuttal 1:
Rebuttal: We thank the reviewer for the detailed feedback.
---
# Response to weakness 1: Novelty and the Influence of the Question Difficulty
## (1) Novelty and contribution:
In summary,
* We present the reasoning ladder for LLMs, which systematically defines different levels of reasoning difficulty.... | Summary: The paper mainly introduces a benchmark synthesis pipeline for math and coding reasoning problem. The proposed pipeline can make modifications of the original benchmark (question, answer) pairs to make it a different (potentially more challenging) instance. The main motivation is to evaluate the true reasoning... | Rebuttal 1:
Rebuttal: # Response to Claims And Evidence: Robustness to stochastic synthesis
We believe there may be a misunderstanding about the experimental setup. To clarify:
* **All models in the paper are tested on the same generated benchmark instantiation**, ensuring fair comparison between models.
* We also ... | Summary: This work creates a framework for LLM reasoning evaluationexpands and scales up LLMs reasoning evaluation by means of an automated pipeline that converts benchmark problems into symbolic representations and then back again, and a 'mutations' step to create variations in pre-existing questions to further test r... | Rebuttal 1:
Rebuttal: # Response to Experimental Designs Or Analyses and Weaknesses: should directly compare against related works methods to show concrete improvements
We highlight that **the goal of Re-Imagine is to establish a unified reasoning hierarchy that integrates both previously studied mutations and the new... | null | null | null | null | null | null |
Improving the Effective Receptive Field of Message-Passing Neural Networks | Accept (poster) | Summary: This paper introduces an architecture called Interleaved Multiscale Message-Passing Neural Networks (IM-MPNN) to address limitations in traditional Message-Passing Neural Networks (MPNNs), particularly the problem of over-squashing and the limited Effective Receptive Field (ERF). The key issue identified is th... | Rebuttal 1:
Rebuttal: We thank Reviewer UcQT for their thoughtful and constructive feedback, and for acknowledging our work being *well-motivated*, *well-presented*, *distinct from existing solutions*, and *addressing a fundamental GNN limitation*. We are pleased to address your comments in detail below.
1. **Regardi... | Summary: This work proposes a hierarchical coarsening method during GNN message passing in order to increase the effective receptive field while reducing over-squashing. The method is compared against datasets within the Long Range Graph Benchmark.
Claims And Evidence: The experimental results well demonstrate the cla... | Rebuttal 1:
Rebuttal: We thank Reviewer 1PKF for their thoughtful and constructive feedback. We are happy to read that you found our *claims well demonstrated* and that the *paper tackles a problem important for the field*, and we are pleased to address your comments in detail below.
1. **Regarding related hierarchica... | Summary: The paper addresses the challenges faced in capturing long-range interactions in GNNs due to limited effective receptive field of the message passing mechanism and proposes a novel architecture based on a hierarchical coarsening of graph to improve communication between distant nodes.
## update after rebuttal... | Rebuttal 1:
Rebuttal: We thank Reviewer EHuP for the thoughtful and constructive feedback and for finding our paper *interesting*, *well-written*, valuable to the current GNN literature, and offering a *different perspective on enabling long-range interactions*. We are pleased to address your comments in detail below.
... | Summary: This paper proposes a new messages passing strategy to expand the receptive field of GNNs by transmitting information between graphs at multiple scales, and theoretical analysizes the influence decay of message passing along the path between nodes. The experiments are conducted on long-range graph benchmarks ... | Rebuttal 1:
Rebuttal: We thank Reviewer U4Ms for their thoughtful and constructive feedback. We appreciate you finding our paper *well-written and easy to follow*, and that the proposed method is *meaningful* with *impressive results*, and we are pleased to address your comments in detail below.
1. **Regarding SeBot... | null | null | null | null | null | null |
Nonconvex Theory of $M$-estimators with Decomposable Regularizers | Accept (poster) | Summary: This paper challenges the results of Section 9 of Martin's textbook "High-dimensional statistics". Surprisingly, this paper is able to recover the results of Proposition 9.13 and Theorem 9.19 of "High-dimensional statistics" for nonconvex loss functions. Moreover, Theorem 3.3 in this paper extends the results ... | Rebuttal 1:
Rebuttal: -Can we consider the other distributions for corrected linear regression?
Thanks for the question. (Rosenbaum & Tsybakov, 2010; Loh & Wainwright, 2012) have already studied the corrected linear regression, and they consider the sub-Gaussian parameters. This paper just follow them to use the exiti... | Summary: This paper studies the theoretical properties of regularized M-estimators with decomposable regularizers under nonconvex loss functions. The authors extend existing results on convex regularized M-estimators to the nonconvex case, demonstrating that estimation errors remain within a restricted set and that con... | Rebuttal 1:
Rebuttal: Can you comment on whether your results extend to more general forms of nonconvexity beyond those studied in the examples?
Thanks for the question. Our framework applies to a broad class of nonconvex loss functions; however, we require that the nonconvex loss satifies the dual norm bound.
---
R... | Summary: This paper develops a theoretical framework for analyzing regularized M-estimators with decomposable regularizers. Extending prior work in convex settings, the authors establish that estimation errors remain in a restricted set and that convergence rates can be recovered despite the loss function's nonconvexit... | Rebuttal 1:
Rebuttal: 1, The paper can benefit from a brief discussion on potential algorithmic implementations based on the theoretical results.
Thanks for the question. Our theoretical results show that the decomposable regularizers play the key role in facilitating convergence and improving generalization. So, it i... | Summary: The paper studies the high dimensional M-estimators for non-convex loss functions. The previous classical results only consider the convex cases. It is natural to consider the non-convex loss function in high dimensions. The motivation is strong. The central theoretical questions studied in this paper are whet... | Rebuttal 1:
Rebuttal: 1, What is the main difference between the proof of Theorem 2.7 in (Wainwright, 2019) and Theorem 3.3 in this paper? I think it is better to clarify this question in the paper, then the reader can quickly understand theorems and proofs.
Thanks for the question. The main difference between the pro... | null | null | null | null | null | null |
Handling Imbalanced Pseudolabels for Vision-Language Models with Concept Alignment and Confusion-Aware Calibrated Margin | Accept (poster) | Summary: This paper proposes a novel approach to improving pseudolabels generated by vision-language models (VLMS). The authors identify two key errors which contribute to the degradation of pseudolabel quality: concept mismatch and concept confusion. Concept mismatch occurs when the class text name provides a descript... | Rebuttal 1:
Rebuttal: Thanks for your time and encouraging comments! We address each of your concerns below.
> **Q1:** Ablations of hyperparameters are not included
Following your suggestion, we have conducted ablations of several hyperparameters in CAP.
**Ablation of $t$**
In CAP, we use $t$ to identify concept mi... | Summary: This paper proposes a concept-adaptive pseudo-labeling framework to generate balanced pseudo-labels for fine-tuning vision-language models (VLMs) on downstream tasks.
In the first stage, the paper introduces concept alignment to address the issue of concept mismatch by assigning precise pseudo-labels to miscl... | Rebuttal 1:
Rebuttal: Thanks for your time and helpful comments! We address each of your concerns below.
> **Q1:** The number of baselines is limited
In our paper, we primarily compare against state-of-the-art methods **specifically designed for leveraging CLIP’s zero-shot capabilities** through pseudolabel generatio... | Summary: This paper proposes a novel framework, CAP (Concept-Adaptive Pseudolabeling), to address the problem of imbalanced pseudolabels when fine-tuning Vision-Language Models (VLMs) like CLIP for downstream tasks using unlabeled data. The authors identify two key causes of imbalance: concept mismatch (where text feat... | Rebuttal 1:
Rebuttal: Thanks for your time and helpful reviews! We address each of your concerns below.
> **Q1:** A more formal or quantitative definition of concept mismatch and concept confusion
Generally, concept mismatch arises from a severe form of the semantic gap while concept confusion is more commonly observ... | Summary: The paper addresses concept mismatch and confusion when adapting VLMs to downstream tasks using pseudo-labeled data. The authors argue that this issue primarily arises from an imbalance in the pseudo-labels generated by VLMs. To mitigate this, they propose a concept alignment mechanism and a confusion-aware ca... | Rebuttal 1:
Rebuttal: Thanks for your time and helpful comments! We address each of your concerns below.
> **Q1:** The ablation studies could be more comprehensive
Following your feedback, we have conducted additional ablation studies on the components you mentioned.
**Ablation of gamma Selection**
We conducted exp... | null | null | null | null | null | null |
The Double-Ellipsoid Geometry of CLIP | Accept (poster) | Summary: The work analyses the geometric properties of CLIP embeddings. It builds upon previous work that studied the modality gap in embeddings. The main finding is that image and text embeddings both live within separate ellipsoid thin shells in high dimensional embedding space. The authors demonstrate that this p... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback. The applications section is indeed small since we devoted our paper to the geometric and statistical analysis, which we found the most important to delve in, explain and support experimentally. See additional comments embedded in our answers to the other two r... | Summary: This paper investigates the geometric properties of the CLIP embedding space, proposing that image and text modalities form independent double-ellipsoid structures displaced from the origin. The authors argue that this structure improves the performance of contrastive learning and provides explanations for pre... | Rebuttal 1:
Rebuttal: - Limited novelty: as other reviewers pointed out, we study well known phenomena using a different lens, i.e. a geometric perspective. As far as we know, we are the first to analyze the raw features prior to the normalization phase. The normalized features are forming a unit hypersphere by definit... | Summary: This paper investigates the geometry of the pre-normalized CLIP embedding space. The main finding is that image and text embeddings reside on linearly separable ellipsoid shells, which are not centered at the origin. This non-origin-centered, double-ellipsoid structure is proposed as a key factor in controllin... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback, intriguing comments and thought-provoking questions.
- The geometry explains the modality gap and the narrow cone effect:
We agree that the geometry does not fully explain the reasons to all observed phenomena. We have a strong evidence that the offset of the... | null | null | null | null | null | null | null | null |
Model Steering: Learning with a Reference Model Improves Generalization Bounds and Scaling Laws | Accept (spotlight poster) | Summary: This paper theoretically studies the mechanism behind a learning paradigm, called Learning with a Reference Model (LEAR) and proposes a new learning algorithm that achieves better scaling than the naive approach. They first relate the RHO loss with DRO and show how using the RHO loss can improve the generaliza... | Rebuttal 1:
Rebuttal: We thank the reviewer for appreciating the technical merits of this paper. We will follow the reviewer's suggestions to improve the paper, which should be an easy task.
> **Q1**: Would this method work for other tasks? Is there a reason to focus only on CLIP?
**A**: Yes! We have discussed that ... | Summary: The paper establishes a theoretical framework for RHO-based learning with a reference model using DRO as the perspective and introduces a novel DRRho risk. It further applies DRRho-based LEAR to CLIP, achieving good and data-efficient performance.
## update after rebuttal
My overall evaluation remains unchang... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable comments and suggestions.
> **Q1**: More powerful reference model does not always yield superior target model performance. But Corollaries 4.2 and 4.3 show that a better reference model should lead to greater improvements in target model training.
**A**: ... | Summary: The paper proposed DRRho risk minimization with a reference model and provided a theoretical analysis of it. It also applied this approach to training the CLIP model. Experiments show that the proposed method achieves better performance than the baselines.
Claims And Evidence: The claims made in the submissio... | Rebuttal 1:
Rebuttal: We thank the reviewer for their constructive and positive evaluation of our work. We are happy to address any concerns the reviewer may have in later stage. | Summary: Authors present a framework for using available open weights model to improve model training on given dataset (learning with a reference model - LEAR). The framework is based on distributionally robust optimization (DRO). DRO makes use of available data empirical distribution to create perturbed data distribut... | Rebuttal 1:
Rebuttal: We thank the reviewer for the suggestion on experiments. We believe that the comments raised by the reviewer are not critical drawbacks of this paper. We request the reviewer to consider our contribution in terms of theoretical framework and analysis, and our experiments comparison with multiple b... | null | null | null | null | null | null |
Edge-Colored Clustering in Hypergraphs: Beyond Minimizing Unsatisfied Edges | Accept (poster) | Summary: The submission provides a range of algorithmic and complexity-theoretic contributions to the Edge-Colored Clustering problem, which has been shown to have several applications in the general area of ML.
Claims And Evidence: Yes, all claims are supported by evidence (primarily proofs), although most of this is... | Rebuttal 1:
Rebuttal: Thanks for the helpful feedback on our manuscript. With regards to the motivation please see our response to the reviewer Tzam for details on our motivation for these variants.
Furthermore, we agree that some sections are quite dense on results and sparse on details. We wished to provide more de... | Summary: The authors study the edge-colored clustering (ECC) problem in hypergraphs. They generalize MaxECC from graphs to hypergraphs, and present an approximation algorithm of factor $(2/e)^r (r+1)^{-1}$, where $r$ is the maximum number of nodes that are allowed in a single hyperedge. A slight modification of this al... | Rebuttal 1:
Rebuttal: Thanks for your detailed review and feedback on our manuscript.
Regarding the “significance of the variant problems for ECC”, please see our response to reviewer Tzam, where we provide more details on motivation we could have been clearer about in our submission.
Thanks for your question about ... | Summary: This paper studies variants of the Edge Colored Clustering problem. The input consists of a hypergraph where the hyperedges are equipped with a color each. The goal is to find a coloring of the vertices. A hyperedge e is satisfied if all its vertices have the same color as e; otherwise e is unsatisfied.
An imp... | Rebuttal 1:
Rebuttal: Thanks for the detailed review! We agree the FPT algorithms are simple, and intended only as small supporting results. Also, please see our answer below regarding ETH bounds.
Thank you also for your line-based comments and suggestions. We will revise the manuscript accordingly. We’ll focus the ma... | Summary: This submission studies a clustering problem where one wants to cluster vertices of a hypergraph by coloring them while approximately maximizing the number of hyperedges all whose vertices get a hyperedge-specific color (those hyperedges are called satisfied). It gives the first algorithm for this problem on g... | Rebuttal 1:
Rebuttal: Thanks for your feedback and questions!
**Motivation for ECC variants.** We highlight 2 points we could have done better addressing in the main text.
(1) Correlation Clustering variants
One key motivation for our fair/protected ECC variants is that these are directly analogous to questions that... | null | null | null | null | null | null |
NExtLong: Toward Effective Long-Context Training without Long Documents | Accept (poster) | Summary: This paper introduces NExtLong, an effective framework that improves long-context modeling in LLMs through negative document extension. It first divides a document into meta-chunks and then inserts hard negative distractors to force LLMs learn the dependency between long documents. Experimental results have il... | Rebuttal 1:
Rebuttal: We sincerely appreciate your response! Your detailed and insightful feedback plays a crucial role in improving our article. The following text further clarifies some questions.
---
**Q1: "Regarding Table 1, it seems that NExtLong is using extra negatives when pre-training on the two datasets com... | Summary: The paper introduces the NExtLong framework, which aims to alleviate issues arising from the scarcity of high-quality long documents in long-context training. Traditional methods that concatenate shorter documents do not effectively capture necessary long-range dependencies, leading to problems with coherence ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your response! Your detailed and insightful feedback plays a crucial role in improving our article. The following text further clarifies some questions.
---
**Q1: "I would like to suggest to include the above work in the related section. It is the first work try to use in... | Summary: The paper introduces NExtLong, a new framework designed to address the challenge of training LLMs with extended context windows, particularly in the face of limited availability of long documents. The key contributions of the paper are the usage of hard negative documents mining for the construction of long do... | Rebuttal 1:
Rebuttal: We sincerely appreciate your response! Your detailed and insightful feedback plays a crucial role in improving our article. The following text further clarifies some questions.
---
**Q1: Dependency on FAISS retrieval quality**
**A1:** The quality of FAISS retrieval depends significantly on th... | Summary: Most LLMs require long documents to handle long-context processing, but in practice, high-quality long documents are scarce. Existing methods typically concatenate short documents either randomly or based on similarity, which is not effective for learning long-range dependencies. This paper proposes splitting ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your response! Your detailed and insightful feedback plays a crucial role in improving our article. The following text further clarifies some questions.
---
**Q1: "It would be better to explain why improving the long-context ability of LLMs can decrease the performance on... | null | null | null | null | null | null |
Revisiting the Predictability of Performative, Social Events | Accept (poster) | Summary: The authors consider a classic problem in social science--how can we make accurate predictions about the world if our predictions affect the world--from a learning theory perspective. In this setting, there are features $x$, a binary outcome $y$, and we wish to make probabilistic predictions $f(x)$ to predict... | Rebuttal 1:
Rebuttal: Thank you for all of the insightful comments on our work! We’re glad you found the paper interesting.
We will certainly add intuition behind the construction in section 5 to the introduction. Your comments in the summary of your review are very helpful in this regard. We appreciate it!
And yes,... | Summary: This paper formalizes the question of whether predictions can remain accurate when the act of predicting affects the state of the world. The authors address this question theoretically and show that a predictor can maintain some bounded level of calibration/validity. The paper provides a bound on how far a pre... | Rebuttal 1:
Rebuttal: Thank you for taking the time to carefully read our manuscript and provide comments. We’re delighted you find the direction interesting and look forward to more work in this area.
We will happily add more clarification and intuition on the distinctions between accuracy and calibration, the notat... | Summary: * This paper investigates multicalibration problems in performative settings.
* The models assume the performative prediction framework, where data distribution depends on the deployed model $(x,y)\\sim \\mathcal{D}(f)$.
* The main result is a convergence bound for the performative multicalibration loss, ach... | Rebuttal 1:
Rebuttal: We appreciate you taking the time to read and carefully critique our work. These are great questions.
Stateful world. We had not considered this possibility. We don’t believe it applies directly to the stateful case, but this is an interesting, open question for future work.
Also, given the ad... | Summary: The paper at hand claims to explore the predictability of socil events. Social predictions do not merely describe the future. They also influence it. Such predictions can affect market prices, voter behavior, and policy outcomes. This interaction complicates the ability to forecast accurately. Early theorists,... | Rebuttal 1:
Rebuttal: Thank you for taking the time to read and provide detailed feedback on our manuscript. It is very appreciated.
Re: Brown et al. 2022. This is an excellent paper. We mistakenly did not include it but will happily discuss it in the revision.
Re: Conceptual question. Thanks for raising this. Lipsch... | null | null | null | null | null | null |
Optimizing Social Network Interventions via Hypergradient-Based Recommender System Design | Accept (poster) | Summary: The paper studies interventions in the Friedkin-Johnsen opinion formation model. This model is defined by a graph in which each node has a fixed innate opinion and a time-dependent expressed opinion; the model is often used to study polarization and disagreement in networks from a mathematical angle. Many rece... | Rebuttal 1:
Rebuttal: We thank the reviewer for their feedback and constructive comments.
First, we would like to clarify one important point.
Our algorithm optimizes only over the network topology, i.e., the network weights $w$, but not the _innate_ opinions $s$.
The main optimization problem that we aim to solve, eq... | Summary: This paper investigates opinion polarization in social networks using a large-scale optimization approach to modify network interactions based on the Friedkin-Johnsen model. The authors propose a gradient-based algorithm designed to address this problem in a scalable and computationally efficient manner. Some ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive comments and their careful reading.
**Methods and Evaluation Criteria** The concern raised on "Methods and Evaluation Criteria" is a valid and important one.
Indeed, for problems/networks of huge dimensions our algorithm might face computational bottlenecks.... | Summary: This paper proposes a novel method, Best Intervention for Recommender Systems (BeeRS), which provides a method of using the hypergradient to optimize connection weights in a social network for certain objectives (for example, reducing polarization). The paper models the social network via the Friedkin-Johnson ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their positive evaluation and their constructive comments.
Concerning the weaknesses and questions raised by the reviewer:
**Question/Weakness 1** Thanks for pointing out this aspect. We foresee using the weights as a penalty (/boosting)
factor in a Weighted PageRank (... | Summary: This paper proposes a gradient descent based method for modifying network weights towards an optimal (general) downstream performance metric, under the framework of Friedkin-Johnsons opinion dynamics. Experiments show significant improvement on computation time on large-scale real-world dataset.
## update aft... | Rebuttal 1:
Rebuttal: We thank the reviewer for their very positive evaluation and their kind suggestions.
We will make sure to stress the advantage of considering directed networks in the paper and include the references suggested by the reviewer. | null | null | null | null | null | null |
Holes in Latent Space: Topological Signatures Under Adversarial Influence | Reject | Summary: The authors analyzes latent representations of several large language models (LLMs) under two main adversarial conditions: Extended Prompt Injection (XPIA) and backdoor sandbagging fine-tuning. To do this, the authors use persistent homology (PH) in order to capture the shape of data at multiple distance scale... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their thoughtful and supportive feedback on our work.
We recognize that the **computational challenges** of persistent homology remain an inherent limitation, as noted by the reviewer. In response to a similar comment from Reviewer mgPH, we have conducted addit... | Summary: This paper conducts a detailed analysis of representations of LLMs using a “topological data analysis” tool. The analysis shows statistical differences between benign natural inputs and adversarial inputs in two scenarios: indirect prompt injection and “sandbagging” (fine-tuning backdoor).
Claims And Evidence... | Rebuttal 1:
Rebuttal: We appreciate the careful review of our work and would like to address several of the concerns raised.
## Motivation of the methodology (PH)
We acknowledge that our motivation for the use of PH in this context could have been more accessible. We address now *why such representation of the data i... | Summary: In this paper, the authors propose a method to analyze the internal representations of LLMs using tools from Topological Data Analysis. They use Persistent Homology (PH) to show that a clear difference in the topology of the activations in an adversarial setting. They perform two sets of qualitative analyses -... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive evaluation of our work and constructive feedback, which we now address.
## Clarification of the local analysis
**Element-wise analysis:** Indeed, our method differs from conventional approaches that analyze full activation vectors using cosine similarity or... | null | null | null | null | null | null | null | null |
SynEVO: A neuro-inspired spatiotemporal evolutional framework for cross-domain adaptation | Accept (spotlight poster) | Summary: This work mainly focuses on observation, i.e., deep learning can imitate the neuroscience mechanism to expand the information boundaries through theoretically and empirically cross-domain collective intelligence learning. Then, drawing from neuroscience, this research introduces a synapse-inspired evolutional ... | Rebuttal 1:
Rebuttal: Dear reviewer ci9Q,
We deeply appreciate the time and efforts you have invested in reviewing our paper. We are honored to receive your recognition of the novelty and theoretical contribution. Your comprehensive feedback is helpful in guiding our revisions.
**W1. Why skip spatial adaptation**
Ur... | Summary: This paper, theoretically examines strategies for increasing information boundaries through cross-domain collective intelligence learning and introduces SynEVO, a synaptic evolutionary spatiotemporal network designed to enable cross-domain knowledge sharing and aggregation by addressing model independence cons... | Rebuttal 1:
Rebuttal: Dear reviewer n3PZ,
Thanks for your constructive feedback of our research. Your valuable advice contributes a lot to our work.
**W1&Q1(1). Explanations of 'harmony with diversity'.**
The concept of 'harmony with diversity' in our paper means that in order to build a generic spatiotemporal learn... | Summary: Drawing from neuroscience, this paper presents a theoretical investigation into methodologies for expanding information boundaries via cross-domain collective intelligence learning. The authors propose SynEVO, a synaptic evolutionary spatiotemporal network architecture. The framework employs a sample order reo... | Rebuttal 1:
Rebuttal: Dear reviewer esVG,
Thanks for your constructive comments on our research.
**W1. Why not use LLM?**
LLM is popular to empower diverse applications but it is more specific to language processing and generation tasks. The reasons for not using LLM in this research are two-fold.
(1) The core of t... | Summary: This paper introduces SynEVO, an interesting neuro-inspired spatiotemporal evolutional framework designed for cross-domain adaptation in spatiotemporal learning. The core idea is to enhance knowledge transfer and model evolution by mimicking synaptic plasticity and neurotransmitter mechanisms from neuroscience... | Rebuttal 1:
Rebuttal: Dear reviewer Nb7e,
Thanks for your encouraging comments!
**Relations between SynEVO and transfer learning, optimizer design, evolutionary algorithms**
**Transfer learning** is also a freeze-finetune mechanism where finetune is varied and specific to the problem itself. It does not involve acti... | null | null | null | null | null | null |
GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation | Accept (poster) | Summary: In this work, the authors tackle semi-supervised medical image segmentation (SSMIS) by proposing GraphCL. This is the first work to model data in a graph network for SSMIS. The authors propose a graph clustering loss function for optimization.
Claims And Evidence: Yes
Methods And Evaluation Criteria: - The a... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and constructive feedback on our manuscript. We are grateful for your positive remarks and are pleased that you found our work to be a valuable contribution to the field of semi-supervised medical image segmentation.
>Q1: The standard deviation of the performanc... | Summary: This paper introduces a graph-based clustering for semi-supervised medical image segmentation by modeling data structure in a unified network. A graph clustering loss function was proposed to optimize the correlation clustering task in SSMIS.
Claims And Evidence: The authors claim that 1) previous methods neg... | Rebuttal 1:
Rebuttal: >Q1: About the novelty
A1: For the importance of graph structural information, our method leverages two types of graph structural information: spatial relationships between voxels/pixels and semantic relationships based on feature similarity. Specifically, we construct dense instance graphs to ca... | Summary: The paper proposes GraphCL, a novel graph-based clustering framework for semi-supervised medical image segmentation (SSMIS). The key contribution is integrating graph data structures into deep learning models which leverages both labeled and unlabeled data, leading to better segmentation performance. The autho... | Rebuttal 1:
Rebuttal: >Q1: Essential References Not Discussed
A1: We acknowledge the relevance of works like GraphSAGE (neighborhood aggregation)[1], GAT (attention mechanisms)[2], Graph U-Nets (hierarchical pooling)[3] and MixMatch (unifiy dominant approaches)[4]. Different from this methods, GraphL uniquely address ... | null | null | null | null | null | null | null | null |
Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection | Accept (poster) | Summary: This paper presents a jailbreak attack against judge LLM detection.
**After rebuttal:** I read the author's rebuttal and most of my concerns are addressed. I am actively participating in reviewer-AC discussion to champion this paper.
Claims And Evidence: Yes
Methods And Evaluation Criteria: Yes, but could b... | Rebuttal 1:
Rebuttal: We thank the reviewer for all the insightful comments.
**Q1: Virus.**
We will cite the Virus paper. While both our work and Virus target judge LLMs, the settings and objectives differ. Virus attacks judge LLMs during the data filtering stage to preserve harmful content, which is subsequently us... | Summary: This paper proposes an Emoji attack to fool the judge LLM and thus enhance the attack power of jailbreaking. Emoji attack finds the position to insert the Emoji that can achieve the maximized segmentation bias. Empirical results show that Emoji can successfully bypass the judge LLM.
Claims And Evidence: Yes. ... | Rebuttal 1:
Rebuttal: We evaluated two types of potential defenses: (1) LLM-based filtering and (2) adversarial training of Judge LLMs (as suggested by Reviewer RkAj). Below, we summarize our findings, with details provided in Figure 9 of the Appendix and additional tables.
**1. LLM-Based Filtering.**
We use `gpt-3.5... | Summary: This paper introduces "Emoji Attack," a technique that exploits token segmentation bias to enhance jailbreak attacks against Judge LLMs. The authors demonstrate that inserting emojis into text can disrupt the tokenization process, causing embedding distortions that lead Judge LLMs to misclassify harmful conten... | Rebuttal 1:
Rebuttal: We thank the reviewer for all the insightful comments. We have addressed your questions and comments below.
**Q1: Limited Description of Datasets.**
Thank you for pointing this out. If given the opportunity, we will include a more detailed description of the datasets in the paper. Below, we outl... | Summary: This paper introduces "Emoji Attack," a novel technique exploiting token segmentation bias in Judge LLMs to bypass harmful content detection. The authors demonstrate that inserting emojis into text disrupts tokenization patterns and creates embedding distortions that significantly reduce the ability of safety ... | Rebuttal 1:
Rebuttal: We thank the reviewer for all the insightful comments. We have addressed your questions and comments below.
**Q1: Defense Mechanisms.**
Please see our response to Reviewer i3cs.
**Q2: Emoji Semantics and Its Impact on Attack Effectiveness.**
We agree that understanding the semantics of emoj... | null | null | null | null | null | null |
Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints | Accept (poster) | Summary: This paper presents CP-Composer, a zero-shot cyclic peptide design framework using composable geometric constraints. The key innovation lies in decomposing complex cyclization strategies into type constraints and distance constraints, integrated into a geometric graph diffusion model via conditional encoding. ... | Rebuttal 1:
Rebuttal: > Q1: Limited MD Validation: Only two test cases are simulated. A larger sample size and error analysis (e.g., standard deviations across replicates) would improve confidence.
Sorry for the limitation, due to the time and resource limitation, we adopted the rhotheta score as an auxiliary metric t... | Summary: Cyclic peptides exhibit superior biochemical properties and can be used to address emerging medical needs. However, due to the limited availability of training data, research on cyclic peptide design remains scarce. This paper introduces a novel generative model that employs a composable constraint approach, e... | Rebuttal 1:
Rebuttal: Thanks for the insightful comments which help improve the quality of our paper!
> Q1: The type and distance constraints may not be independent but rather intertwined, which the authors have not discussed.
Sorry for the confusion. We ensure that the constraints are jointly feasible by considering... | Summary: The paper proposes CP-Composer, a novel diffusion-based generative framework for zero-shot cyclic peptide design. The authors motivate their approach by highlighting the data scarcity problem in cyclic peptide design, where obtaining experimental data for diverse cyclization patterns is challenging. The key in... | Rebuttal 1:
Rebuttal: Thank you for your suggestions!
> Q1:Comparing against any existing methods specifically designed for cyclic peptide generation is important for contextualizing the performance gains.
We include CADS, an advanced diffusion conditional sampler, and DiffPepBuilder, a model specifically designed fo... | null | null | null | null | null | null | null | null |
RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning | Accept (spotlight poster) | Summary: The paper proposes an end-to-end multi-turn RL framework to teach LLM self-repair/refine based on execution feedback, particularly focused on the code generation domain, where the unit tests and execution feedback is easy to obtain. The main algorithm is PPO with turn-level value function to calculate the adva... | Rebuttal 1:
Rebuttal: We deeply appreciate the reviewer's comments and feedback on our manuscript.
> the generalization performance to HumanEval and MBPP, can we report CI for the numbers in Table 2 as well? The generalization performance seems worse compared to CC. Test, is the improvement statistically significant?
... | Summary: This paper finetunes LLMs for multi-turn code generation with PPO. The action is code generation/refinement using LLMs. The model is given public test cases for code evaluation and then refinement. The epsidoe ends either when reaching the maximum turn limit or when the generated codes pass the public tests. T... | Rebuttal 1:
Rebuttal: We thank the reviewer for their valuable feedback and comments and provide the following responses:
> Data Contamination
We build off the Llama 3.1 models which were originally evaluated on benchmarks like HumanEval and MBPP as well. On top, we add training data from CodeContests exclusively; we... | Summary: This paper introduces RLEF - a reinforcement learning method for improving natural language to code generation in an iterative setting. The method treats code generation as a multi-turn conversation, where a language model first produces a program then receives and interprets textual execution feedback to refi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful review and valuable feedback.
Based upon their feedback, we investigated the performance of DeepSeek-R1-Distill-Llama-70B in our exact same setting. With a single generation (temp=0.6, prompted with "<think>\n"), we obtained solve rates of 38.5 and 33.9 ... | Summary: This paper proposes an RL training strategy for Code LLMs to enable them to refine generated code using execution feedback besides the capability of following instructions. They present an exhaustive analysis on different aspects of their RLEF-trained models including their inference time behavior, performance... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thorough analysis of our paper, their valuable suggestions and stimulating questions.
## Updates
Re. GRPO: We will add the following to the end of our related work section:
More recently, DeepSeek-AI et al. (2025) observe emerging reasoning capabilities with a larg... | null | null | null | null | null | null |
Trajectory World Models for Heterogeneous Environments | Accept (poster) | Summary: This manuscript has 2 contributions:
1. A trajectory dataset UniTraj, a large-scale dataset including over one million trajectories collected from various distributions from 80 heterogeneous environments.
2. A Transformer-based architecture TrajWorld integrates interleaved variate and temporal attention mecha... | Rebuttal 1:
Rebuttal: We sincerely appreciate Reviewer bd1u's strongly positive feedback on our work. Your recognition of our clear writing, well-motivated approach, and the effectiveness of both our UniTraj dataset and TrajWorld architecture is truly encouraging. We are grateful for your support and share your belief ... | Summary: This paper introduces the UniTraj dataset, which contains a large set of trajectories collected from 80 heterogeneous environments. It also presents a world model, TrajWorld, pretrained on this dataset. The pretrained world model demonstrates positive transferability to new environments in zero or few-shot set... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer 8E9V for the thorough review and valuable questions.
## Q1: MPC evaluation
Following Schubert et al., we have added **online MPC experiments**. In this setting, TrajWorld outperforms both baselines and its counterpart trained from scratch (see [anonymous figure](http... | Summary: The paper aims to tackle the heterogeneity issue in world model pretraining. To achieve this, the authors curate a unified trajectory dataset from 80 control environments. Based on this dataset, they introduce TrajWorld, a world model architecture that naturally accommodates varying sensors and actuators, ther... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer puvo for the thorough review, insightful questions, and a positive evaluation of our work.
## W1: Model-predictive control (MPC) evaluation
We have conducted MPC experiments comparing different world models.
**Setup**: Following Schubert et al, we first attempted *M... | Summary: This paper presents a trajectory world model that handles varying sensor and actuator information across different environments. To support the generalization of the world model, this work composes a large dataset, UniTraj, comprising over one million trajectories from 80 environments. The key ingredient of th... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer 2GyG for the thoughtful review and valuable comments, especially the recognition of our idea of unifying world modeling across heterogeneous environments.
## Q1: Dataset contribution
**Dataset construction**: We respectfully disagree with the assessment that our datas... | null | null | null | null | null | null |
M2PDE: Compositional Generative Multiphysics and Multi-component PDE Simulation | Accept (poster) | Summary: The paper proposes a compositional multiphysics and multicomponent simulation model that uses diffusion. The approach, MultiSimDiff, consists of learning conditional distributions of individual components conditional on the other physical processes. At test-time, the approach samples from the conditional distr... | Rebuttal 1:
Rebuttal: > Re1: Relation to Gibbs Sampling.
Gibbs sampling splits a problem into conditional distributions and samples them sequentially. In our setting, this is essentially our baseline (named **surrogate model** in manuscript), where we iteratively update each physical field using a surrogate model (See... | Summary: The paper proposed a compositional diffusion model framework to handle multi-physics multicomponent surrogate model for physics systems. It leverages the existing diffusion backbone and demonstrated it effects on three multi-physics/multicomponent PDE systems. The paper claims the contribution are: 1. Introduc... | Rebuttal 1:
Rebuttal: > Re1: What is the novelty for your proposed diffusion framework?
Our study is application-driven rather than aiming to improve diffusion models directly. Our contribution is not diffusion model itself, but the higher-level algorithm on top of diffusion models for multiphysics and multi-component... | Summary: This paper proposes MultiSimDiff, a novel compositional generative model for multiphysics and multi-component simulations. The core idea is to use diffusion models to learn energy functions representing the conditional probability distributions of different physical processes or components. During inference, M... | Rebuttal 1:
Rebuttal: > Re1: the line 145, 197, right column, "z=(z1,z2,...,zn)" and "V=v1∪v2∪...∪vn", should use capital "N" instead of "n" ?
**Answer**: Thank you for pointing this out. We wll make the correction.
> Re2: The statement of z\_{i}^{e} in eq.9 is not clear enough. For readers who are familiar with diff... | Summary: This paper introduces MultiSimDiff a new method for solving multi-physics/multi-compnennts simulations efficiently by learning the conditional score of each component's solution given its parameters and solutions of other components. Experiments demonstrate that MultiSimDiff outperforms largely a simpler surro... | Rebuttal 1:
Rebuttal: > Re1: The end of the abstract (3 last sentences) appears a bit narrow and verbose.
Thank you for the suggestion! We will revise the ending of the abstract as follows:
We demonstrate the effectiveness of MultiSimDiff through two multiphysics tasks—reaction-diffusion and nuclear thermal coupling... | null | null | null | null | null | null |
G-Sim: Generative Simulations with Large Language Models and Gradient-Free Calibration | Accept (poster) | Summary: The paper proposes a G-Sim, LLM-guided simulator with expert domain knowledge with gradient-free optimization while introducing a new problem with environment building. With experiments on three environments, the paper verified the flexibility of G-Sim.
Claims And Evidence: With limited experiments and no the... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer for the valuable and constructive feedback. Below, we address each concern and outline key improvements in the revised manuscript.
---
### **1. Experimental Validation and Benchmark Fairness**
We acknowledge the concerns regarding validation and benchmark selecti... | Summary: This paper attempts to generate simulators via LLMs coupled with a gradient-free optimisation process to choose parameters. An LLM-guided search loop identified the simulator's structural components and a gradient-free optimisation procedure sets their parameters. The method relies on the generalisation abilit... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer for their detailed and constructive feedback. We fully agree on the importance of clearly presenting our methodology within the manuscript itself. To comprehensively address these concerns, we will allocate the additional page in the camera-ready version specificall... | Summary: This paper introduces G-Sim, a framework for automatically constructing simulators by combining Large Language Models (LLMs) and gradient-free optimization (GFO). The LLM is used to generate the structural components of the simulator (submodules, causal relationships), based on provided domain knowledge. GFO i... | Rebuttal 1:
Rebuttal: We sincerely thank Reviewer for their insightful and constructive feedback, which significantly helps strengthen our paper. Below, we address each major concern explicitly and outline concrete improvements for the camera-ready version.
---
**Baseline Tuning**
We appreciate the reviewer’s concer... | null | null | null | null | null | null | null | null |
Conformity Score Averaging for Classification | Accept (poster) | Summary: This paper proposes to improve conformal prediction by optimally averaging multiple conformity score functions. This papers explores various data splitting methods and optimal weights for aggregating the score functions.
Claims And Evidence: The main claim of this paper is that by optimally weighting multiple... | Rebuttal 1:
Rebuttal: Thank you for your detailed feedback and constructive suggestions.
---
### **1. About why our method can improve performance**
In our setting, there exist $d$ score functions corresponding to weights $w = e_1, \dots, e_d$, where $e_i$ is the $i$-th standard basis in $\mathbb{R}^d$. The optimal... | Summary: The paper proposes a method for improving prediction set efficiency in classification tasks through conformity score averaging. It introduces weighted averaging of multiple score functions and explores various data-splitting strategies to optimize the weight selection process. Theoretical guarantees for covera... | Rebuttal 1:
Rebuttal: Thank you very much for your positive review. Most of your questions are high-level and insightful. We will address them in order, starting with simpler questions and moving to more complex, open-ended ones.
---
### **1. About additional experiments**
We have conducted additional experiments on... | Summary: In this paper, the authors presented an approach that enhances conformal prediction for multi-class classification by optimally averaging multiple conformity score functions, and a set of evaluation experiments showed that the weighted averaging approach consistently outperforms single-score methods by produci... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and for appreciating the organization and motivation of our paper. Below, we address your questions and comments.
---
### **1. About the complexity of the proposed approach**
In our algorithm, the main source of complexity lies in finding the optimal weight ... | Summary: Existing conformal prediction methods typically rely on a single conformity score function, limiting both their efficiency and informativeness. In this paper, they propose a new approach that enhances conformal prediction by averaging multiple conformity score functions for the same classification task. They a... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and helpful suggestions. We also appreciate your careful review of the supplemental materials. Below, we will address your comments.
---
### **1. About the intuition of VC dimension in our proof**
We agree that providing this intuition would benefit readers.... | null | null | null | null | null | null |
Improving Multi-Class Calibration through Normalization-Aware Isotonic Techniques | Accept (poster) | Summary: This paper addresses extending isotonic regression from binary calibration to multi-class calibration. It proposes isotonic normalization-aware techniques for multi-class calibration. In particular, it introduces two techniques to account for probability normalization: (1) NA-FIR that incorporates normalizati... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review.
As you rightfully pointed out, validating our methods on larger datasets such as ImageNet-1K is important for demonstrating scalability and generality. In the last few days we have made a concentrated effort and conducted additional experiments on ImageNet-1... | Summary: The paper addresses the critical challenge of multi-class calibration in supervised learning.
While isotonic regression has proven effective for binary calibration problems, its extension to multi-class settings through one-vs-rest (OvR) calibration has historically underperformed compared to parametric meth... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and positive review.
As far as theoretical claims we would like to also refer the reviewer to Appendix 4 where we provide a theoretically valid motivation for our isotonic approach. | Summary: he paper discusses a tweak on a post-hoc recalibration method for the multi-class classification algorithm. One of the widely used methods for post-hoc recalibration in the binary setting is the isotonic regression: given already a classification algorithm, we can consider a new regression problem where the co... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and positive review.
The reviewer rightfully mentions that only final results for the best heuristic are being presented, but we refer to our Appendix C and comments to Reviewer 4 where we provide more details on computational considerations for adopting our suggest... | Summary: This paper proposes two isotonic regression based approaches that incorporate normalization into problem formulation of multiclass calibration:
* Normalized Aware Flattened Isotonic Regression, which finds a mapping g such that g(p(x)) is normalized before computing the NLL objective.
* Sorted Cumulative Isoto... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review.
The major concern is regarding statistical properties of the algorithms we propose, in particular NA-FIR. This is expressed in the review sections on Claims and Evidence, Theoretical Claims and the last point of Questions for Authors.
Regarding the statisti... | null | null | null | null | null | null |
Human Cognition-Inspired Hierarchical Fuzzy Learning Machine | Accept (poster) | Summary: The paper extracts the similarities between concepts from the human knowledge system and uses these similarities to guide the learning process. As a result, the similarities between concepts are integrated into the sample similarity, thereby improving model performance. Meanwhile, the paper guarantees the effe... | Rebuttal 1:
Rebuttal: # 0. General Response
We sincerely appreciate the reviewer’s positive feedback and valuable comments. Below, we provide a point-by-point response to each comment.
# 1. Response to “Weaknesses”
(1) CLIP (Contrastive Language-Image Pre-training) is a self-supervised learning method that learns repr... | Summary: In general terms, the paper presents a new and innovative method called Human Cognition-Inspired Hierarchical Fuzzy Learning Machine (HC-HFLM), which is aimed at improving interpretability and performance in classification tasks. This method is based on human cognition and takes into account the ambiguity pres... | Rebuttal 1:
Rebuttal: # 0. General Response
We thank reviewer for the appreciation of our work and valuable comments. Below, we provide a point-by-point response to each comment.
# 1. Response to “Methods And Evaluation Criteria”
(1) We add the standard deviation of accuracy for all methods based on 5-fold cross-vali... | Summary: The authors propose a human cognition-inspired classifier. The method first mines the fuzzy similarity relation between concepts from human knowledge system. And then the authors design the hierarchical alignment loss based on the principles of concept cognition. Using this loss, the fuzzy similarity relation ... | Rebuttal 1:
Rebuttal: # 0. General Response
We thank reviewer for the appreciation of our work and valuable comments. Below, we provide a point-by-point response to each comment.
# 1. Response to “Weakness”
- The fuzzy equivalence relation (FER) is a more restrictive form of the fuzzy similarity relation (FSR).
In p... | Summary: This paper advocates solving classification problems from the perspective of concept cognition. Inspired by human cognition, this paper utilizes the relationships between concepts embedded in the human knowledge system to guide the learning process. This deepens the model's understanding of concepts. In additi... | Rebuttal 1:
Rebuttal: # 0. General Response
We sincerely appreciate the reviewer’s positive feedback and valuable comments. Below, we provide a point-by-point response to each comment.
# 1. Response to “Weakness”
- In the experiments, to directly highlight the performance improvement achieved by incorporating class ... | null | null | null | null | null | null |
Scaling Off-Policy Reinforcement Learning with Batch and Weight Normalization | Reject | Summary: The paper discusses the failure of a previously proposed DRL algorithm, CrossQ, to reliably scale up to more complex environments than those considered in the original paper. To achieve this, the authors propose to use weight normalization.
### Update after rebuttal period
With additional context, I suppor... | Rebuttal 1:
Rebuttal: We thank the reviewer for their review and positive feedback, as well as the additional questions. We will extend the paper with discussions, additional experiments for each question, and mentioned baselines and feedback. To answer them here:
- **Could Unit Sphere Normalization work (instead of t... | Summary: The paper studies the scaling property of a previously proposed RL method, CrossQ, with high update-to-data ratio. CrossQ does not use target network updates and is known to be brittle to tune as also shown by the authors. The paper proposes to stabilize the training dynamics of CrossQ using weight normalizati... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough and extensive review. We were happy to read that they found our experimental results convincing. We hope that our rebuttal manages to address their remaining concerns and open questions. We ran many additional experiments and ablations for the rebuttal, fo... | Summary: This paper enhances the sample efficiency of reinforcement learning (RL) by improving CrossQ, a model-free algorithm that leverages Batch Normalization (BN). While CrossQ excelled at low update-to-data (UTD) ratios, it struggled to scale reliably. The authors found that scaling the UTD ratio in CrossQ leads to... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough and extensive review. We were happy to read that they appreciate the idea and direction of our work and are open to increasing the score. We hope that our rebuttal manages to address their concerns and open questions. We ran many additional experiments and... | Summary: The paper proposes an enhancement to the CrossQ reinforcement learning framework by integrating weight normalization (WN) with the existing batch normalization (BN) approach. The primary goal is to stabilize training when using higher update-to-data (UTD) ratios, which are typically associated with improved sa... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thorough review and questions. We were especially pleased that they acknowledged the appropriateness of our methods and evaluation criteria and that they found the paper easy to read.
In the following, we want to address the reviewer’s three main concerns individua... | null | null | null | null | null | null |
Channel Normalization for Time Series Channel Identification | Accept (poster) | Summary: The paper talk about the importance of Channel Identifiability (CID) when modeling multivariate time series data. The paper talk about how existing methods failed to provide CID capability. To solve this problem, the model proposed various Channel Normalization (CN) method. CN is a type of normalization method... | Rebuttal 1:
Rebuttal: ## Weakness 1. PCN’s CID capability
> Reviewer: *I am not fully convinced that PCN provides CID capability. ~ I believe PCN is more of an extension of CN (i.e., an effective channel normalization method that does not provide CID) and should be discussed in the appendix.*
Thank you for pointing th... | Summary: - The Channel Normalization (CN) strategy is proposed to enhance the Channel Identifiability (CID) of Time Series (TS) models by assigning specific parameters to each channel.
- Two variants of CN, Adaptive CN (ACN) and Prototypical CN (PCN), are introduced to dynamically adjust parameters and handle datasets ... | Rebuttal 1:
Rebuttal: ## Weakness 1. Miscolored Figure 1
> Reviewer: *"Figure 1: ~ producing different outputs (green) even with same inputs (yellow)." Maybe the colors are mispositioned.*
Thank you for pointing that out. We will fix it in the revised version.
## Question 1. Details about the performa... | Summary: The authors propose a new method to adaptively normalize each time series channel distinctly through learned channel specific adaptive parameters. These adaptive parameters for each channel are data dependent are computed through a dynamic weighted summation of a similarity matrix computed between channel toke... | Rebuttal 1:
Rebuttal: ## Suggestions 1. Explanation of Legends in Figures
> Reviewer: *"It would be helpful for readers to provide what different legend in different figures represent. For example, LN in Figure 5? Is that layer normalization? This needs to be clarified*
Thank you for your feedback. Due to **space limi... | null | null | null | null | null | null | null | null |
One-Shot Heterogeneous Federated Learning with Local Model-Guided Diffusion Models | Accept (poster) | Summary: This paper addresses the important and interesting problem of one-shot federated learning (OSFL), aiming to reduce the communication round of FL to 1. With the help of pretrained Classifier-Guided Diffusion Models, this paper proposes to generate local clients' data distribution in the server side with the gui... | Rebuttal 1:
Rebuttal: We sincerely appreciate your recognition of our work and your valuable feedback. Below, we provide detailed responses to the key concerns you raised:
>**(Essential References Not Discussed)** *"BN loss has been widely used, for example, [Yin'2020], while the related citations are missing."*
Than... | Summary: This paper introduces FedLMG, a novel One-Shot Federated Learning (OSFL) method addressing limitations of diffusion model-based OSFL. FedLMG leverages locally trained client models to guide a server-side diffusion model in generating synthetic datasets tailored to individual client distributions. This approach... | Rebuttal 1:
Rebuttal: Thank you for recognizing our work. Below, we provide detailed responses to your concerns:
>**(Server Cost)** *Questions #3"... server computing costs."*
In FL, the server is generally designed to have sufficient resources to handle the model aggregation, but clients often exhibit significant he... | Summary: This paper introduces FedLMG, a novel approach for One-shot Federated Learning (OSFL) designed to establish an aggregated model within a single communication round. Specifically, FedLMG leverages fully-trained client models as classifier guidance to facilitate diffusion generation at the server. The generated ... | Rebuttal 1:
Rebuttal: We appreciate your positive comments of our work and address each of your concerns as follows:
>**(Server Cost)** *Claims #1 & Questions #3: "… DMs on the server, which can be an additional computation cost"*
In FL, the server is generally designed to have sufficient resources to handle the aggr... | Summary: In response to the increasing demand for efficient One-Shot Federated Learning (OSFL) solutions, this paper introduces FedLMG, a novel OSFL method leveraging Local Model-Guided diffusion models. Unlike existing OSFL methods that rely on foundation models deployed on client devices—causing significant computati... | Rebuttal 1:
Rebuttal: We sincerely appreciate your review and valuable comments and provide detailed responses to the key concerns:
>**(Privacy Concerns)** *Claims And Evidence #1 & Weaknesses "general quantitative metrics for privacy of FL, such as Gradient Leakage (GL) or Differential privacy (DP)."*
Regarding GL, ... | null | null | null | null | null | null |
A Generalization Result for Convergence in Learning-to-Optimize | Accept (oral) | Summary: This paper proposes a new method for analyzing the generalization ability of learning to optimize (L2O). The authors aim to formulate the convergence of L2O to unseen data as a random event measured by a posterior distribution of neural network's (NN) parameters. By assuming training ensures that the L2O gener... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to provide this feedback, even it was unfortunately rather negative.
Regarding the claim that the paper is hard to follow and that most technical details are not clearly introduced:
1) Could you please be more specific here? Otherwise, we c... | Summary: Learn-to-optimize has been a popular research topic in recent years.
However, many theoretical guarantees are still lacking.
This paper develops a probabilistic framework that resembles
classical optimization and allows for transferring geometric
arguments into learn-to-optimize. The paper establishes
a genera... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for giving this feedback, and we would like to take the opportunity to shortly comment on the weaknesses and to answer the questions.
Regarding your first posed weakness: It is true that, on an abstract level, in the end the result follows by combining Theorems... | Summary: While learning-to-optimize has shown to be a powerful paradigm to enhance the efficiency of the optimisation phase for problems similar to the one encountered during training, it is unsure how such a trained algorithm will behave on unseen problems with different internal structure. This work tackles this issu... | Rebuttal 1:
Rebuttal: Also here, we would like to thank the reviewer for taking the time to provide this feedback. We are glad that you considered the proof-strategy as “well-explained”, because it is one of our main contributions.
Regarding your question whether we could expand a bit more on the comparison between ou... | Summary: This paper presents a probabilistic framework to establish convergence guarantees for L2O algorithms, addressing the challenge that conventional geometric arguments for convergence do not readily apply to learned optimizers. The key contribution is a generalization result that combines PAC-Bayesian learning th... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for taking the time to provide this detailed feedback.
We shortly want to comment on the stated weaknesses:
We agree that such convergence guarantees are highly desirable, but we also think that
this is asking for too much: We are considering an abstract algor... | null | null | null | null | null | null |
New Bounds for Sparse Variational Gaussian Processes | Accept (spotlight poster) | Summary: The authors introduce a tighter ELBO bound for inducing-point-based Gaussian process regression à la SVGP (Titsias 2009) and its mini-batchable extension (Hensman et al. 2013). The main idea is to use a more flexible ansatz for $q({\bf f} | {\bf u})$ than the conditional prior $p({\bf f} | {\bf u})$, in partic... | Rebuttal 1:
Rebuttal: Thank you for the insightful comments.
> I would say that the evidence for 2 and 3 is convincing, while the evidence for 1 is somewhat weak. If this is to be a central claim of the submission, it should be supported with more empirical evidence, in particular in the simulated data setting where... | Summary: The paper revisits the widely-used variational approximation for sparse Gaussian processes (GPs). It proposes a refined variational formulation, introducing a more flexible conditional posterior distribution in place of the traditional assumption (where the conditional posterior matches the prior). This adjust... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer. We respond to the main comments below.
> Yes, mainly. The proposed methods and evaluation criteria are appropriate for the problem at hand, and the experiments effectively demonstrate the benefits of the tighter variational bound. However, the datasets used in... | Summary: The paper introduces new evidence lower bounds (ELBOs) for sparse variational Gaussian processes (SVGP) by relaxing the traditional assumption that the variational distribution must factorize with the conditional GP prior p(f|u). Instead, the authors propose a more flexible variational distribution q(f|u), whi... | Rebuttal 1:
Rebuttal: Thank you for your comments. Below we provide some responses.
> The key insight, replacing $p(f|u)$ with a diagonal-covariance $q(f|u)$, is sound. However, the gap between diagonal V and spherical V is not thoroughly assessed.
Please note that in the medium-size regression experiments reported... | Summary: The authors present an improvement on the standard SVGP approximation by departing from the standard conditional GP prior distribution. The approach introduces an additional $N$ variational parameters which modify the covariance matrix of the conditional distribution. This leads to an improvement on the result... | Rebuttal 1:
Rebuttal: Thank you for very accurately describing the contribution of the paper and for pointing to concurrent work. As also mentioned in the response to Reviewer Pyzb below, we plan to discuss the concurrent work in the next version of our paper. | null | null | null | null | null | null |
Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation | Accept (spotlight poster) | Summary: The paper proposes utilizing an adapter for cross-domain few-shot semantic segmentation. The authors first demonstrate that the adapter naturally serves as a decoupler, and then design a DFN network to decouple source domain information into domain-agnostic and domain-specific components. They also propose SAM... | Rebuttal 1:
Rebuttal: ## 1. The computational efficiency of SAM-SVN：
Our SAM-SVN is used only during source domain training and not during fine-tuning or inference, so it does not affect the computational efficiency during inference. Regarding efficiency during training in the source domain, although it requires doubl... | Summary: This paper find an interesting phenomenon that a sort of adapters naturally serve as domain-information decoupler for the CDFSS task. By comprehensive experiments, the authors validate the condition that makes adapters to be decouplers. Then, they extend such a natural decoupler by sharpness-aware minimization... | Rebuttal 1:
Rebuttal: ## 1. Our method can fit different structures：
Our structure and the APM[1] structure are both based on HSNet, while our ViT structure is based on FPTrans[2]. Additionally, in the appendix, we used the SSP-based architecture for comparison with IFA. Thus, our method can be applied to networks with... | Summary: The paper proposes that adapters naturally serve as domain information decouplers in Cross-Domain Few-Shot Segmentation (CD-FSS) by separating domain-specific and domain-agnostic features. Based on this insight, the authors introduce Domain Feature Navigator (DFN), a structure-based decoupler that captures dom... | Rebuttal 1:
Rebuttal: ## 1. Deeper theoretical analysis for “natural decoupling”：
Due to space limitation, please refer to reviewer ueDb's reply 1 for theoretical analysis.
## 2. Analysis of trade-off between decoupling and domain knowledge retention:
For any $\rho>0$ and any distribution $\mathscr{D}$, with probab... | Summary: This paper introduces a novel perspective on using adapters as structural domain decouplers for cross-domain few-shot semantic segmentation (CD-FSS). They introduce the Domain Feature Navigator (DFN), a structure-based decoupler inserted into deeper network layers with residual connections, and SAM-SVN, a shar... | Rebuttal 1:
Rebuttal: ## 1. Deeper theoretical analysis for “natural decoupling”：
The behavior of adapters as decouplers can be analyzed through the Information Bottleneck (IB) theory. The IB objective is:
$\mathcal{L}_{IB} = I(X;Z) - \beta I(Z;Y)$
where $I(\cdot;\cdot)$ is mutual information, X is the input, Y is t... | null | null | null | null | null | null |
PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering | Accept (poster) | Summary: The paper addresses the class imbalance issue in multi-view clustering by combining UOT and POT to perceive class imbalance, and uses POT-enhanced class rebalance to mitigate the representation degradation of minority samples in contrastive learning. Through comparisons across multiple datasets and multi-view ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your recognition of our work's motivation and method.We are also deeply grateful for your thorough review and valuable suggestions.
**Q1:** A intuitive explanation of how Optimal Transport addresses class imbalance.
**A1:** Thanks for your suggestion. We would like to dra... | Summary: In this paper, a novel Partial Optimal Transmission (POT) enhanced contrast learning framework, PROTOCOL, is proposed to address the class imbalance challenge in multi-view clustering. A two-level rebalancing strategy achieves balanced feature learning as well as consistency in view-specific and view-sharing a... | Rebuttal 1:
Rebuttal: We sincerely appreciate your recognition of both the novelty of our method and the practical value of our motivation, as well as your positive feedback on our paper's representation and experiments. We are also deeply grateful for your thorough review and valuable suggestions.
**Q1:** The empiric... | Summary: The paper introduces PROTOCOL, a new method for imbalanced multi-view clustering. It combines partial optimal transport (POT) with contrastive learning. The approach solves two main problems: perceiving class imbalance distributions through POT-based label assignment and reducing the representation degradation... | Rebuttal 1:
Rebuttal: We sincerely appreciate your recognition of our work's novelty and its potential impact in enhancing multi-view clustering for real-world imbalanced scenarios, as well as your positive feedback on our experimental results. We are also deeply grateful for your thorough review and helpful suggestion... | Summary: This paper proposes the first systematic study on the common class imbalance problem in multi-view clustering and develops a new framework called PROTOCOL. This method reformulates the imbalanced clustering problem as a partial optimal transfer problem by mapping multi-view features to a consensus space, and i... | Rebuttal 1:
Rebuttal: We sincerely appreciate your recognition of our work's novelty as the first to identify and systematically study the class imbalance problem in multi-view clustering, as well as your positive feedback on our method's effectiveness and robustness. Furthermore, we are deeply grateful for your thorou... | null | null | null | null | null | null |
Aguvis: Unified Pure Vision Agents for Autonomous GUI Interaction | Accept (poster) | Summary: This paper focuses on the autonomous GUI interaction task from the pure vision agent perspective. A large-scale cross-platform dataset of GUI agent trajectories is constructed. A two-stage training pipeline is proposed to separate GUI grounding from planning and reasoning. The experiments demonstrate the effec... | Rebuttal 1:
Rebuttal: Thank you for taking the time to review our work and providing constructive feedback! We greatly appreciate your recognition of our contribution in the realm of pure-vision based autonomous GUI agents, including our grounding-then-planning training pipeline and open-sourcing the large-scale, curat... | Summary: The paper introduces AGUVIS, a unified vision-based framework for autonomous GUI agents designed to overcome limitations of existing approaches, which rely on textual representations, platform-specific actions, and closed-source models for reasoning. AGUVIS enables direct operation on screen images, standardiz... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful review and the opportunity to further clarify our contributions.
> **W1: Technical Innovation in Training Method**
Thank you for your insightful comments regarding the training methodology! We recognize that recent approaches share similar high-level compo... | Summary: This paper introduces Aguvis, a vision-based framework that operates directly on screen images, providing a standardized cross-platform interaction method enhanced by structured reasoning through inner monologue. The researchers developed a comprehensive dataset with multimodal annotations and implemented a tw... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review of our AGUVIS paper. We appreciate your recognition of our comprehensive roadmap for developing a pure-vision GUI agent, particularly our data curation approach and training strategies. Your positive assessment of our experimental validations across multiple be... | Summary: This paper introduces AGUVIS, a vision-based UI agent designed to operate across diverse digital platforms. The authors collect data from existing resources and do some essential augmentation. They then leverage a vision-language model to train AGUVIS in two stages, grounding and planning, to improve interacti... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thorough review and positive assessment of our work. We're particularly encouraged by your recognition of our comprehensive evaluation benchmarks, the engineering effort involved in dataset aggregation and standardization, and the value of our open-sourced models and d... | null | null | null | null | null | null |
Towards Trustworthy Federated Learning with Untrusted Participants | Accept (poster) | Summary: Achieving robustness against Byzantine workers and preserving data privacy are two important objectives in distributed learning. Existing work primarily studies each problem separately, and achieving both simultaneously is a challenging task. In this paper, the authors propose CAFCOR, an algorithm designed to ... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful feedback. Below, we address the key points raised:
### **Experimental Scope**
> evaluating on MNIST and Fashion-MNIST is insufficient [...]
The reviewer’s suggestion to evaluate additional datasets is highly valid. We stress, however, that our current ... | Summary: The paper proposes a technique to perform distributed mean estimation with differential privacy guarantees and robustness to byzantine participants. To achieve privacy, if first adopts the anti-correlated noise method of [1,2]. To achieve robustness, it uses the empirical covariance matrix of the contributions... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed and thoughtful feedback. Below, we address the points raised:
### **Empirical Comparison with SMEA**
> it is not clear why [...] SMEA [3] is not included in the comparison. Theoretically, it seems that SMEA has the same robustness as CafCor.
We thank th... | Summary: This paper proposes a novel methodology to achieve resilience in the face of malicious parties colluding with an untrusted server in the distributed learning framework as well as privacy guarantees, based on a weak assumption that each pair of communicating workers secretly share a seed of randomness, used to ... | Rebuttal 1:
Rebuttal: We clarify our intent and recall our contributions below following the reviewer's comments, and we welcome further discussion to refine our presentation.
### **Use of "Byzantine” Terminology**
This is a standard term in distributed computing literature (Lamport et al., 1982), referring strictly... | Summary: The paper introduces an algorithm (CAFCOR) to achieve privacy and robustness in distributed learning without relying on a trusted central server. In particular, it employs correlated noise injection inspired by secret sharing and combines it with a robust aggregation technique to mitigate Byzantine workers' im... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful feedback. Below, we address the key points raised:
### **Experimental Scope**
> Limited evaluation datasets.
The reviewer’s suggestion to evaluate additional datasets is highly valid. We stress, however, that our current evaluation on MNIST and Fashion... | null | null | null | null | null | null |
Tuning Sequential Monte Carlo Samplers via Greedy Incremental Divergence Minimization | Accept (poster) | Summary: The authors propose a method for tuning kernels in SMC by minimising a KL-divergence between proposal paths and target paths. This optimisation is done using a gradient free method. They show that by adapting the step size with their method they get better normalising constant estimates than using fixed step s... | Rebuttal 1:
Rebuttal: Thank you for your review.
> The datasets used seem quite standard for evaluating SMC methods. One thing I am not sure of is whether just reporting normalising constant is sufficient to show the superiority of this method, it would have been nice to see something like ESS or some assessment of po... | Summary: The paper provides a gradient-free, hyperparameter-free tuning algorithm for proposal step sizes and particle refresh/resampling rates in Sequential Monte Carlo pipelines. Experiments on example graphical models from PosteriorDB demonstrate that the new algorithm not only provides a great advantage in log-evi... | Rebuttal 1:
Rebuttal: Thank you for your review.
> Do the authors want to minimize incremental divergences in order to target the path measure of the annealed importance sampling procedure, or fine-tune an annealed importance sampling procedure in order to later perform SMC on the target density?
We kindly request mo... | Summary: This paper proposes a novel method for tuning sequential Monte Carlo (SMC) samplers by greedily minimizing the incremental KL divergence between the target and proposal path measures. The authors develop efficient, gradient-free algorithms for tuning key parameters—such as step sizes in unadjusted Langevin Mon... | Rebuttal 1:
Rebuttal: Thank you for your review.
> One potential caveat is that some theoretical guarantees rely on assumptions (e.g., unimodality of the tuning objective) that may limit generalizability,
We agree that those theoretical assumptions are restrictive. As such, we are happy to mention that we were able t... | Summary: Main problem and approach: The performance of sequential Monte Carlo (SMC) samplers heavily depends on the tuning of the Markov kernels used in the path proposal. The paper proposes a framework for tuning the Markov kernels in SMC samplers by minimizing the incremental Kullback-Leibler (KL) divergence between ... | Rebuttal 1:
Rebuttal: Thank you for your review.
> The experiment set up is too simple on a collection of toy benchmark datasets. It is not clear how it is applicable to the application domains of SMC, such as steering large language models and conditional generation from diffusion models.
We agree that more realisti... | null | null | null | null | null | null |
Representative Ranking for Deliberation in the Public Sphere | Accept (poster) | Summary: The paper studies a setting of algorithmic comment ranking/selection incorporating fairness. There is a given set of comments, together with "likes" and based on these likes, a representative set of comments needs to be selected. The paper studies the impact of a group fairness concept called "justified repres... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comprehensive and thoughtful comments.
The reviewer seems to be evaluating our paper primarily as a contribution to social choice. However, the main goal of our work was not to contribute to social choice, but rather to facilitate deliberation online (particularly ... | Summary: The authors propose a comment ranking approach for public deliberation that incorporates justified representation, a concept from the social choice literature. The goal is to rank high quality comments, without losing the representation of groups that are present in the discussion. The approach relies on user ... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s careful reading of our paper and thank them for their constructive comments. We will correct all the typos that they identified.
With regards to the Perspective API bridging score $f_C$, it is the average of the scores for the seven available bridging attributes:
* ... | Summary: The authors take the problem of content ranking in online social deliberation, and adds justified representation (JR) constraints to the quality optimization problem to ensure diversity and representation. Theoretically, they show that under assumptions of user clusterization, the extra constraint leads to low... | Rebuttal 1:
Rebuttal: Thank you to the reviewer for their comprehensive review. We address their questions below.
**Connection Between Theoretical and Empirical Results**
The reviewer asks about how our theoretical results connect to the empirical findings. Even without considering Theorem 5.4, the bound from Theorem... | Summary: This work applies the principle of "justified representation" as a means to algorithmically surface public comment for the end-goal of public deliberation. This is, in part, motivated by the ideals of deliberative democracy and normative reasons for selected public comments to satisfy some notion of "represent... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s thoughtful comments and expertise. The reviewer’s main question is about their interpretation of Sections 5 and 6, which we address in the following. In Section 5, we first introduce novel theoretical results for the price of JR for arbitrary score functions, and then ... | null | null | null | null | null | null |
Explanatory Instructions: Towards Unified Vision Tasks Understanding and Zero-shot Generalization | Accept (poster) | Summary: This work proposed a large-scale explanatory instruction dataset to unify multiple CV tasks for AR VLM understanding and then generation. It uses VQ-VAE style tokens for vision and then an AR model to merge the 2 modality. It provides qualitative results on various tasks, showing the zero-shot capabilities.
C... | Rebuttal 1:
Rebuttal: Thank you for your feedback. We realize there may be some misunderstandings regarding both the dataset and idea presented in our paper. We hope the following responses will help clarify these points.
**Experimental Designs**: *The model is trained for 2 epcohs.*
**Response**: Thank you for your ... | Summary: This paper proposes a concept called “Explanatory Instructions” to move beyond the conventional limitations of computer vision (CV) tasks. The authors argue that the currently common terminological definitions (e.g., “semantic segmentation”) oversimplify the expression of CV objectives, limiting the model’s ab... | Rebuttal 1:
Rebuttal: Thank you for your positive evaluation of our paper. Below are our responses to the concerns you raised.
**Essential References Not Discussed**: *While the paper covers recent VLM-related literature well, it would benefit from discussing more explicitly related previous work on task-level general... | Summary: This paper proposes Explanatory Instructions to address the challenge of task-level zero-shot generalization in computer vision, inspired by the success of instruction-driven models in NLP. The authors hypothesize that conventional terminological task definitions (e.g., "semantic segmentation") limit models' a... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback on our paper. In the following responses, we have addressed the concerns you raised during the review process, and we hope these answers will resolve your questions.
**Question 1 / Weakness 1**: *The authors used GPT-4o for training data generation, especially... | Summary: This work proposes a new method for fine-tuning visual tasks, explanatory Instructions. Inspired by the work on text instruction fine-tuning, the authors aim to explore whether there is a generalization phenomenon in instruction fine-tuning for visual tasks. Therefore, they construct pure-text instructions for... | Rebuttal 1:
Rebuttal: We greatly appreciate your recognition of the idea presented in our paper. Below are our responses to the concerns you raised.
**Question 1 & 4 / Weakness 3**: *I am curious about the cross-task generalization ability of the initial VLM model, Lumina-mGPT-7B-768-Omni, without any training. Does t... | null | null | null | null | null | null |
De-coupled NeuroGF for Shortest Path Distance Approximations on Large Terrain Graphs | Accept (poster) | Summary: This paper proposes a new learning-based approach for answering shortest path distances on large-scale terrain DEMs. Overall, the proposed method extends the prior work of NeuroGF while providing a comprehensive and in-depth analyses on the training mechanisms and design choices of neural components. Extensive... | Rebuttal 1:
Rebuttal: Thank you for your review! We are glad that you appreciate our proposed advancement in the problem of neural data structures for SP queries on terrains.
**Regarding your question about $xy$-monotone surfaces**, a continuous surface in $\mathbb{R}^3$ is called $xy$-monotone if every line parallel... | Summary: This paper proposes a De-coupled NeuroGF for efficiently approximating SPD queries on large-scale DEMs. The key contribution authors decouples the Siamese embedding module and the distance calculation module in NeuroGF. By combining an efficient two-stage hybrid training strategy, the method significantly redu... | Rebuttal 1:
Rebuttal: We thank you for your time and constructive feedback. We are glad the reviewer appreciates our de-coupled and mixed coarse-to-refined training strategy for efficiently processing large terrain graphs at scales previously which were not able to be considered. As you point out, our de-coupled traini... | Summary: This paper presents decoupled-NeuroGF framework for efficient approximate SPD queries on large terrain DEMs, based on the NeuroGF framework. This paper appropriately abstracts high-resolution terrain datasets as weighted graphs. The proposed decoupled-NeuroGF with a two-stage mixed-training strategy significan... | Rebuttal 1:
Rebuttal: We thank the reviewer for your time and constructive feedback. We are happy that the reviewer appreciated our innovative training strategy: with our de-coupled and mixed coarse-to-refined training strategy, we introduce a lightweight neural data structures that can efficiently answer many shortest... | null | null | null | null | null | null | null | null |
How Much Can We Forget about Data Contamination? | Accept (poster) | Summary: - the paper studies the effect of data contamination during the pre-training of language models, through a series of *controlled* contamination experiments
- the paper studies the effect by considering: (1) scaling the amount of contamination (i.e. repetitions), (2) scaling the model size, (3) scaling the (unr... | Rebuttal 1:
Rebuttal: Thank you for the detailed review of our paper. We are happy to hear that our paper is “a pleasure to read”! Below, we respond to your questions/comments.
*“deduplicating the contamination benchmark data (e.g. HellaSwag) from the pre-training tokens (e.g. FineWeb) is also recommended for a truly ... | Summary: The paper investigates the impact of data contamination in LLMs, specifically addressing whether small-scale contamination significantly affects benchmark evaluations. The authors analyze contamination effects along three dimensions: model size, number of training tokens, and repetition of contaminated example... | Rebuttal 1:
Rebuttal: Thank you for the detailed review of our paper and insightful questions. Below, we give detailed answers to your questions/comments.
*“The main limitation is the relatively small scale of experiments, though this is understandable given computational constraints.”*
Running experiments with a 7B ... | Summary: This paper investigates the impact of data contamination in large language models (LLMs), challenging the assumption that minor contamination invalidates benchmark evaluations. Through controlled experiments, the authors study how contamination effects scale with model size (up to 1.6B parameters), training to... | Rebuttal 1:
Rebuttal: Thank you for the detailed review of our paper. We are happy to hear that you appreciate our experimental design. Below, we give answers to your questions/comments.
*“The proof assumes constant learning rates”*
To clarify, the proof does not assume constant learning rates (in the proof, the lear... | Summary: This paper provides a very important perspective in data contamination of LLM and show that not all data leakage will lead to false evaluation in benchmarks.
Claims And Evidence: Strengths:
1. This paper question the severity mentioned in the previous paper. The assumption or the settings of data contaminatio... | Rebuttal 1:
Rebuttal: Thank you for reviewing our paper and for your positive assessment.
It seems that you ask under what conditions we can be confident that a benchmark evaluation is not contaminated. This is an interesting question that lies somewhat beyond the scope of our paper. In our paper, we demonstrate that ... | null | null | null | null | null | null |
e-GAI: e-value-based Generalized $\alpha$-Investing for Online False Discovery Rate Control | Accept (poster) | Summary: The paper proposed the e-GAI framework which can control the FDR under arbitrary dependence structures by defining a conservative e-value-based FDP estimator and adopting a risk-averse strategy.
Claims And Evidence: yes
Methods And Evaluation Criteria: yes
Theoretical Claims: no
Experimental Designs Or Ana... | Rebuttal 1:
Rebuttal: Thank you for your helpful comments and suggestions. We would like to part-wisely respond to your comments.
> In financial bubble detection, ω1 is set to 0.0001 without justification.
This application can be seen as a long-term time series (i.e., $T=10000$). As response to your W1&W2, we sugges... | Summary: The paper proposes a framework for online multiple testing with false discovery rate (FDR) control that utilizes e-values with generalized alpha-investing methods to improve power for e-values that satisfy conditional validity. This paper then establishes connections between these methods and existing generali... | Rebuttal 1:
Rebuttal: We appreciate your careful reading and constructive suggestions. Per your comments, we would like to clarify key **contributions** of our work as follows:
1. We **propose the e-GAI framework** for online testing with using e-values, which **achieves FDR control under arbitrary dependence among va... | Summary: This paper proposes a framework for generalized $\alpha$-investing (GAI) with e-values, an approach for online multiple hypothesis testing. While prior work had considered GAI with p-values, this paper contributes two things: First (Section 3.1) they derive bounds on the false discovery proportion (FDP) for ... | Rebuttal 1:
Rebuttal: Thanks for your comments on our paper. We would like to part-wisely respond to your comments.
> First, the empirical argument would have been stronger in Section 5.1 with a simulated scenario that caused LORD++ to exceed the desired FDR bounds.
Thank you for your valuable suggestion! We have to... | Summary: The paper extends the generalized $\alpha$-investing (GAI) framework for online testing by allowing it to be based on e-values as well (hence the name e-GAI). This allows for online false discovery rate control under arbitrary dependencies of the hypotheses and can lead to improved power under good dynamic all... | Rebuttal 1:
Rebuttal: Thanks for your comments on our paper. We would like to part-wisely respond to your comments.
> One small note is that maybe Figure 2 could benefit from some improvements in its display as currently the squares and dots are quite small and subtle (the latter could be fixed by different coloring ... | null | null | null | null | null | null |
Adaptive Median Smoothing: Adversarial Defense for Unlearned Text-to-Image Diffusion Models at Inference Time | Accept (poster) | Summary: This paper seeks to enhance the adversarial robustness of unlearned t2i diffusion models, specifically aiming to balance the adversarial robustness and generative capabilities of the original t2i models. The proposed method, Adaptive Median Smoothing, starts by formulating the target task as a regression probl... | Rebuttal 1:
Rebuttal: # Generalizability Across T2I Model Variants
We evaluated our method on other widely used T2I models, SD 1.5 and SD 2.1, to assess the generalizability of our method. Due to time constraints, we used the UCE unlearned model with "*nudity*" removed, keeping $\sigma_0$ and $k$ consistent with versio... | Summary: This paper proposes an inference-time defense method, named Adaptive Median Smoothing, to protect unlearned text-to-image diffusion models against adversarial prompt attacks. Specially, the promposed method reformulates robustness as a regression problem and extends median smoothing by using anisotropic noise.... | Rebuttal 1:
Rebuttal: Thank you for your insightful comment regarding the relationship between our method and AdvUnlearn [1]. Our response is elaborated on the following three aspects:
# Methodological Differences
AdvUnlearn [1] is a pre-inference method that fine-tunes the text encoder. It falls into the category of... | Summary: This paper proposes "Adaptive Median Smoothing" as an inference-time defense for adversarial attacks on unlearned diffusion models. The defense goal can be formulated as minimizing MSE predicted noise before and after adversarial perturbation. Based on this formulation, the paper then introduces the naive medi... | Rebuttal 1:
Rebuttal: # Evaluation Across Object and Style Concepts
We would like to clarify that our original manuscript includes experiments not only on the concept of nudity but also on **violence**. Following your suggestion, we expanded to include **object** and **style** concepts.
- For the **object** concept, we... | Summary: Even after unlearning, models are still vulnerable to adversarial inputs that can expose users to inappropriate contents. Existing adversarial defense methods still have difficulty with balancing the adversarial robustness and the generation quality. To address these issues, the paper proposes an inference-tim... | Rebuttal 1:
Rebuttal: # Analysis of Vague Adversarial Prompts
Thanks for your insightful comment. Our method’s robustness against vague adversarial prompts is validated through evaluations on the **MMA attack** [1], which constructs adversarial prompts avoiding sensitive words while inducing unsafe generations. As sho... | null | null | null | null | null | null |
AI for Global Climate Cooperation: Modeling Global Climate Negotiations, Agreements, and Long-Term Cooperation in RICE-N | Accept (poster) | Summary: ## update after rebuttal
The paper proposed a multi-agent reinforcement learning (MARL) approach to model global climate negotiations, agreements, and long-term cooperation. Two novel negotiation protocols were proposed: Bilateral Negotiation and Basic Club. The proposed approach and protocols take into accou... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback and for pointing out the positive aspects of our submission, such as the novelty of the protocols, the realism of the components, the empirical support for our claims and the overall quality of the evaluation criteria, experimental design and res... | Summary: The paper adds reinforcement learning to a multi-agent Nordhaus RICE model.
It allows two communication protocols between agents, whereby agents can make binding commitments to curtail their greenhouse gas emissions.
Simulation allows these two protocols to be compared to the 'no negotiation' (Nordhaus RICE)... | Rebuttal 1:
Rebuttal: We thank the reviewer for the incisive review of our submission. Below, we respond to the concerns raised.
### > does the 'no negotiation' baseline match the Nordhaus-Yang results?
We can compare our results to Nordhaus’ RICE from 2010 [1]. The aforementioned has two results: baseline and optima... | Summary: The paper introduces and analyses a climate policy modelling framework for assessing the effect of different international agreemtns on future climate. It introduces RICE-N, a multi-region integrated assessment model that simulates global climate negotiations and agreements using multi-agent reinforcement lear... | Rebuttal 1:
Rebuttal: Thank you for the thoughtful review and for highlighting the positive aspects, such as RICE-N’s suitability for studying strategic climate negotiations, the evidence-based claims, and the originality of our MARL application. We are pleased that you consider both the paper and the code accessible, ... | Summary: The paper introduces RICE-N, a multi-region integrated assessment model designed to simulate global climate negotiations, agreements, and long-term cooperation using multi-agent reinforcement learning (MARL). The model extends the Regional Integrated Model of Climate and Economy (RICE) by incorporating negotia... | Rebuttal 1:
Rebuttal: We thank the reviewer for their favorable assessment regarding the innovativeness and suitability of our chosen approach, as well as the constructive feedback. Below, we carefully address any outstanding concerns:
### > detailed sensitivity analysis
Thank you for this suggestion. We select param... | null | null | null | null | null | null |
Efficient First-Order Optimization on the Pareto Set for Multi-Objective Learning under Preference Guidance | Accept (spotlight poster) | Summary: This paper considers the problem of preference-guided multi-objective optimization. The authors first formulate it as a semivectorial bilevel optimization problem, which optimizes the upper-level preference objective, subject to the constraint that the model parameters are weakly Pareto optimal or Pareto stati... | Rebuttal 1:
Rebuttal: Thanks for acknowledging that **we propose a novel formulation and an easy-to-understand novel method for preference-guided multi-objective optimization, our proof is sound and interesting, and the experiments are sufficient with sound analysis**.
Below we address your concerns point by point. Th... | Summary: This paper studies multi-objective optimization with user-specified preferences. The authors formulate the problem as a bilevel optimization problem, where the upper-level is a preference function, and the lower-level problem is the minimization of a smoothed version of merit function. Merit function usually s... | Rebuttal 1:
Rebuttal: Thanks for acknowledging that the problem is important and the bilevel perspective is new. We would like to emphasize that the bilevel problem in this paper with *non-convex vector-valued lower-level objective* is much more challenging and nontrivial, as pointed out by Reviewer XkmK.
Below we add... | Summary: In this work, the authors frame preference-guided multi-objective learning as an optimization problem on the Pareto set and propose a first-order penalty method to address it, where the penalty function is a polynomial of a smoothed merit function. They begin by establishing key properties of the merit functio... | Rebuttal 1:
Rebuttal: Thanks for acknowledging that we consider BLO with *non-convex vector-valued lower-level objective* with stronger guarantees, which is different from prior works, and the experiment results are promising.
We would like to emphasize that the *non-convex vector-valued lower-level objective* is much... | Summary: This paper proposes a new method for solving the semivector bilevel optimization problem. The authors first reformulate the multi-objective subproblem as a single objective constraint and then use a penalty-based method to solve the reformulated optimization problem. The results demonstrate the effectiveness o... | Rebuttal 1:
Rebuttal: Thanks for supporting our work, acknowledging that **we propose a new efficient method to solve a very challenging semivectorial bilevel optimization problem, with a strong theoretical guarantee, and with intuitive experimental result demonstrating the authors' claims**.
We address your other co... | null | null | null | null | null | null |
QEM-Bench: Benchmarking Learning-based Quantum Error Mitigation and QEMFormer as a Multi-ranged Context Learning Baseline | Accept (poster) | Summary: This paper presents QEM-Bench, a benchmarking suite for machine learning-based quantum error mitigation (ML-QEM), addressing the lack of standardized datasets in the field. The benchmark includes twenty datasets spanning different circuit types and noise models to enable consistent evaluation of ML-QEM techniq... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the insightful comments and inquiries, and the positive evaluation of our work. **We have summarized the newly added Figs and Tabs at [this link](https://anonymous.4open.science/r/Rebuttal-iomM-B8AB/rebuttal_iomM.pdf).**
Below are our responses.
> **1. Rea... | Summary: This paper introduces a dataset for benchmarking quantum error mitigation techniques, as well as a graph-transformer model to serve as baseline. The dataset consists of three evaluation settings - each with different levels of added noise; Standard (general purpose testing with Trotterized TFIM Circuits, Rando... | Rebuttal 1:
Rebuttal: We appreciate your constructive feedback and positive evaluation of our work. We acknowledge that quantum error correction is aimed to be solved on the hardware side ultimately. However, the significant qubit overhead associated with QEC renders it less feasible in the near term, especially for la... | Summary: The paper introduces QEM-Bench, a benchmarking suite designed to evaluate machine learning-based Quantum Error Mitigation (QEM) techniques.
The benchmark includes 20 datasets covering various circuit types and noise models to standardize QEM evaluation.
Furthermore, the paper proposes QEMFormer, a novel learni... | Rebuttal 1:
Rebuttal: We would like to thank the reviewer for the insightful questions and positive evaluation of our work. **We summarize the Tabs. and Figs. for newly added experiments at [this link](https://anonymous.4open.science/r/Rebuttal-6Drn-C2D8/rebuttal_%206Drn.pdf).**
Below are our responses to each questio... | Summary: The authors make two primary contributions in their manuscript. First, they compile QEM-Bench, a set of twenty datasets that the community can use to benchmark ML-based approaches to quantum error mitigation (QEM). Second, they introduce a new ML-based approach to QEM called QEMFormer, which combines multi-lay... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful comments and inquiries. **We summarized all Tabs. and Figs. of the newly added experiments at [this link](https://anonymous.4open.science/r/Rebuttal-fTJA-433C/rebuttal_%20fTJA.pdf).**
Below are our responses.
> **1: RF [1] and GTraQEM [2] on QEM-Bench.**
We apologi... | null | null | null | null | null | null |
FlexControl: Computation-Aware Conditional Control with Differentiable Router for Text-to-Image Generation | Accept (poster) | Summary: This paper proposes FlexControl, a framework that introduces a novel gating mechanism for dynamically selecting blocks to activate in the control network, reducing computational overhead while preserving or improving image quality. The authors have conducted experiments on both UNet-based (SD1.5) and DiT-based... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for acknowledging the novelty and performance of our paper. We hope the following answers reflect your questions.
> Quantitative comparison and ablation on ControlNeXt..
We appreciate the reviewer’s concern regarding the need for additional comparisons with effic... | Summary: This paper proposes FlexControl, a novel method aimed at improving the computational efficiency of ControlNet, an important model for adding controllability in text-to-image generation tasks. Unlike the original ControlNet, which utilizes half of the diffusion architecture as its encoder, FlexControl introduce... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for the detailed and constructive feedback.
> ControlNext.. LoRA-based…
We appreciate the reviewer’s feedback regarding the comparison with other methods. While a direct comparison is not applicable (as our work focuses on control block integration rather than pa... | Summary: This paper studies the Computation-Aware ControlNet by proposing a dynamic routing strategy which dynamically selects blocks to activate at each denoising step. It aims at adjusting control blocks based on timestep and
conditional information while maintaining (or even improving) generation quality. The experi... | Rebuttal 1:
Rebuttal: We hope the answer below solves all the clarity issues.
> The proposed method requires more parameters than the typical controller.
Yes, our method requires more parameters than standard ControlNet. However, compared to ControlNet-Large, **it achieves better generation quality and controllabili... | Summary: The paper addresses the limitations of existing ControlNet implementations in diffusion-based generative models, which often rely on ad-hoc heuristics for selecting control blocks. The authors employs a trainable gating mechanism to dynamically select which blocks to activate at each denosing step.
Claims And... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive feedback and support of our work. We hope to have answered all of your questions satisfactorily below. Please let us know if you see any further issues in the paper that must be clarified or addressed.
> The ablation study … with simpler alternatives, such ... | null | null | null | null | null | null |
SToFM: a Multi-scale Foundation Model for Spatial Transcriptomics | Accept (poster) | Summary: The authors introduce SToFM, a single-cell foundation model that incorporates not only the single-cell expression data but also their spatial locations. They propose SToCorpus-88M, one of the largest single-cell pretraining datasets curated to date, and also pretrain their model on the large pretraining datase... | Rebuttal 1:
Rebuttal: Dear Reviewer tnYZ:
Thanks for your appreciation and detailed review! We try our best to response your questions. Due to space limitations, we include **tables and references in the anonymous link** https://anonymous.4open.science/api/repo/stofm-rebuttal/file/rebuttal-tnYZ.pdf?v=af96a9f5, and ref... | Summary: This paper introduces a foundation model for cell spot representation of spatial transcriptomics. It fine-tunes the pretrained cell embedding (from existing scRNA foundation models) by incorporating the location information via masked feature prediction and noised distance information recovery, as well as invo... | Rebuttal 1:
Rebuttal: Dear Reviewer LBxH:
Thanks for your appreciation and detailed review. We try our best to response the questions below:
>Q1: The ablation study is not sufficient to support the effectiveness of the fine-tuning with the incorporation of distance information.
Thank you for your suggestions! We con... | Summary: The paper proposes a multi-scale foundation model to integrate macro-scale tissue morphology, micro-scale cellular microenvironment and gene-scale gene expression profile of spatial transcriptomics. The author constructs a large-scale spatial transcriptomics corpus containing approximately 2,000 tissue slices... | Rebuttal 1:
Rebuttal: Dear Reviewer Rgcj:
Thanks for your appreciation and detailed review. We try our best to response the questions below:
>Q1: Ablation experiment on micro-scale components. **&** Q2: The effect of the spatial distance matrix. **&** Q4: The effect of incorporating spatial information.
Thank you for... | Summary: The paper proposes SToFM, a multi-scale Spatial Transcriptomics foundation model, to effectively integrate macro-, micro-, and gene-scale information from Spatial Transcriptomics (ST) data. SToFM uses a combination of gene expression profiles, cell coordinates, and spatial relationships to learn representation... | Rebuttal 1:
Rebuttal: Dear Reviewer TZDp:
Thanks for your appreciation and detailed review. We try our best to response the questions below:
>Q1: Bridging the gap between ST and scRNA-seq data through transfer learning.
ST data consists of two parts: **spatial location** and **gene expression values**. As SToFM is a ... | null | null | null | null | null | null |
MiraGe: Editable 2D Images using Gaussian Splatting | Accept (poster) | Summary: This paper presents approaches to editing 2D images represented by Gaussian Splatting. The authors propose to use 3D flat Gaussians optimized with mirrored cameras from two opposite sides to represent 2D images, with quality better than other prior works. With the GS-represented images, this paper demonstrates... | Rebuttal 1:
Rebuttal: We appreciate the Reviewer’s feedback. We also thank Reviewer for appreciating the concept and work: “utilizing two opposite cameras to fit 3D Gaussians to represent 2D images is an interesting idea, and evaluation also shows the effectiveness of this idea”. Since the section for questions direct... | Summary: The paper introduces MiraGe, a method for representing and editing 2D images using parameterized 3D Gaussian components. By embedding 2D images in 3D space with flat Gaussians and leveraging mirror cameras for training, MiraGe achieves high-fidelity reconstruction and enables intuitive 3D-like editing (e.g., ... | Rebuttal 1:
Rebuttal: We appreciate the Reviewer's thoughtful feedback. We are pleased for the recognition of the distinct advantage of our proposed method in seamlessly integrating with physics engines, allowing for more dynamic and realistic modifications compared to traditional tools.
W1/2
In our paper, we address... | Summary: The paper proposes MiraGe, a novel approach for representing and editing 2D images using Gaussian Splatting. MiraGe uses flat-controlled Gaussian components positioned in 3D space, providing intuitive editing capabilities with a 3D perception. Key contributions include high-quality reconstruction results that ... | Rebuttal 1:
Rebuttal: We thank the Reviewer for the feedback and constructive remarks regarding our paper that we believe will improve our paper. In particular, we are grateful for the Reviewer’s recognition that "the methods and evaluation criteria employed are appropriate and well-chosen".
W1
We acknowledge that s... | Summary: The paper introduces MiraGe, a novel method for representing and editing 2D images using flat 3D Gaussian components. The approach leverages Gaussian splatting in 3D space to enable high-quality image reconstruction and realistic editing capabilities. MiraGe allows for both 2D and 3D manipulations of images, c... | Rebuttal 1:
Rebuttal: We sincerely thank the Reviewer for their valuable feedback and are pleased with the appreciation of our work. In particular, we are especially grateful for the recognition of the breadth and depth of our experiments "covering multiple datasets and providing both quantitative and qualitative compa... | null | null | null | null | null | null |
SpeCache: Speculative Key-Value Caching for Efficient Generation of LLMs | Accept (poster) | Summary: The paper introduces SpeCache, a speculative KV caching mechanism designed to enhance the efficiency of LLM inference. SpeCache mitigates these drawbacks by offloading the complete KV cache to CPU memory and dynamically fetching KV pairs back into GPU memory during decoding. To minimize CPU-GPU transfer latenc... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback! We address specific concerns and questions below.
> Q1. Unclear Overhead of Parallel Prefetching and Computation
Since our method is built upon KIVI, we can consider KIVI as the baseline when parallel pre-fetching and computation are not used... | Summary: This paper proposes to offload the KV cache to CPU memory and retrieve KV pairs based on the similarity between the query of a speculative token with quantized KV pairs.
Claims And Evidence: This paper has two claims: 1) Attention is sparse while each token requires different KV pairs. It emphasizes the impor... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback! We address specific concerns and questions below.
> Q1. Discuss the difference between SpeCache and QUEST and the contribution of SpeCache.
Although both QUEST and SpeCache focus on how to accurately select sparse KV pairs, their emphasis diff... | Summary: This paper presents SPECACHE, a novel method to address the memory bottleneck caused by key-value (KV) caches in large language models when processing long sequences. The authors propose a training-free approach that offloads the complete KV cache to CPU memory while maintaining a low-bit copy in GPU VRAM. The... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback.
> Q1. How sensitive is SpeCache to changes in the distribution of attention patterns?
We add the phrase "How many paragraph are there in the article? Translate the first sentence into German." to the end of a 15k-token text composed of several... | Summary: This paper proposes storing low-bit KV on the GPU while offloading full-precision KV. Attention is performed between the overall full-precision KV of the top-k keys selected from the low-bit keys.
Additionally, the paper introduces the use of speculative tokens to speculatively prefetch the KV cache needed fo... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback! We address specific concerns and questions below.
> Q1. Why the 1-bit improvement is so large? Performance of KVQuant + SpeCache.
+ The improvement of SpeCache on 2-bit KIVI is relatively small because the performance of 2-bit KIVI on longben... | null | null | null | null | null | null |
Improving Rationality in the Reasoning Process of Language Models through Self-playing Game | Accept (poster) | Summary: The paper presents the Critic-Discernment Game (CDG), a self-play approach that enhances the reasoning of large language models (LLMs). In CDG, a "prover" generates solutions, while two critics—helpful and misleading—offer feedback. The prover learns to correct mistakes from the helpful critic and resist misle... | Rebuttal 1:
Rebuttal: > **Experimental Designs Or Analyses:**
> While the paper shows improvements on GSM8K and MATH500, one limitation is the lack of explicit testing for generalization to other problem domains or tasks outside of mathematical reasoning.
> **Weakness:**
> While the paper reports significant improvemen... | Summary: This paper introduces the Critic-Discernment Game (CDG), a self-play approach to improve reasoning in language models without human supervision. In CDG, three roles interact: a prover that solves problems, a helpful critic that identifies errors in incorrect solutions, and a misleading critic that fabricates e... | Rebuttal 1:
Rebuttal: Thanks for your constructive feedback on our paper. Our response to your questions is as follows:
> **Claims And Evidence:**
> One limitation is that the mechanism by which CDG improves understanding of reasoning processes is somewhat indirect. It's unclear that if LLMs are truly doing the reason... | Summary: This paper introduces a self-play reinforcement learning approach called the Critic-Discernment Game (CDG) to improve language models' reasoning capabilities. In CDG, three roles interact: a prover provides solutions to problems, a helpful critic identifies genuine errors in incorrect solutions, and a misleadi... | Rebuttal 1:
Rebuttal: Thanks for your constructive feedback on our paper. Our responses to your questions are as follows:
> **Claims And Evidence:**
> The paper claims to improve "rationality in reasoning" but doesn't directly measure this construct beyond task performance.
As mentioned in the introduction, recent st... | Summary: The paper introduces a framework that involves training three models (prover, helpful critic, misleading critic) via reinforcement learning with the goal of improving the reasoning capability of the prover model. Through the proposed training process, the prover learns to rely only on helpful feedback, and to ... | Rebuttal 1:
Rebuttal: We greatly appreciate your recognition of our idea and your constructive feedback. Here are our responses to your concerns:
> **Claims And Evidence:**
> The goal of the technique is unclear and the evidence provided to support this is not super convincing.
> **Experimental Designs Or Analyses:**
... | null | null | null | null | null | null |
SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models | Accept (poster) | Summary: This work proposes a mixed-precision quantization approach with coarse-level and fine-level partitioning via proposed Salience-Determined Bit Allocation and Salience-Weighted Quantizer Calibration, respectively. The former leverages the double-pointer search algorithm, optimizing KL-divergence between the orig... | Rebuttal 1:
Rebuttal: Dear Reviewer H7Hr,
Thank you for your feedback. We will address your questions and recommendations one by one.
> Q1: The key aspect of the proposed method ... for uneven weight sensitivity. The proposed saliency measure ... However, no reference is provided.
A: We would like to clarify that in... | Summary: The paper introduces SliM-LLM, a novel PTQ framework for LLMs. The proposed method leverages the authors’ observation that important weights follow a structured distribution to preserve the model performance at extremely low-bit precision. Their two key contributions are:
- **Salience-Determined Bit Allocatio... | Rebuttal 1:
Rebuttal: Dear Reviewer iFeq,
Thank you for your valuable feedback and suggestions. We will address your questions and recommendations one by one.
> Q1: (1)The analysis of local salience is relatively underdeveloped, and it has also been explored in recent previous works. (2)The authors may add references... | Summary: This paper introduces a group-wise mixed-precision quantization method for LLMs, addressing challenges in accuracy and efficiency. The key contributions are two strategies: SBA, which optimally allocates bit-widths by minimizing entropy divergence through Hessian and weight salience analysis, and SQC, which en... | Rebuttal 1:
Rebuttal: Dear Reviewer Ut3n,
We sincerely appreciate your insightful feedback and suggestions. Below, we will respond to your questions and recommendations individually.
> Q1: The authors are advised to consider testing on more challenging LLM benchmarks, such as GSM8K in the mathematics domain. Compared... | Summary: This paper proposes SliM-LLM, a post-training quantization (PTQ) framework for large language models (LLMs). Its core idea is to allocate bit-widths to weight groups adaptively and locally preserve important (salient) weights. The approach combines two techniques:
1. Salience-Determined Bit Allocation (SBA): ... | Rebuttal 1:
Rebuttal: Dear Reviewer BegC,
Thank you for your valuable feedback. We have summarized your questions and concerns. Due to the character limit for reply, if there are any question that are not detailed, we will further reply in next stage. Thanks you so much.
> Q1:(1) There is...diverse deployment conditi... | null | null | null | null | null | null |
Inducing, Detecting and Characterising Neural Modules: A Pipeline for Functional Interpretability in Reinforcement Learning | Accept (poster) | Summary: This paper addresses the challenge of interpretability in reinforcement learning (RL) models by proposing a method based on functional modules, aiming to overcome the scalability limitations of traditional neuron-level interpretability approaches. The authors introduce spatially aware regularization and neuron... | Rebuttal 1:
Rebuttal: Thank you for your detailed consideration of our work. We appreciate the acknowledgment of our robust experimentation and the insightful questions raised.
**Figures and Captions.** We appreciate the advice on improving clarity. We have updated the network plots to include legends and have extend... | Summary: The authors identify that most (post hoc) interpretability methods focus on explaining models' units (e.g. neurons), which does not scale. They propose to have interpretability at the level of *functional modularity*. They just try to identify neural modules, which are groups of neurons that are functionally r... | Rebuttal 1:
Rebuttal: Thank you for your detailed and thoughtful review. We are grateful for your recognition of the potential of our work and the constructive feedback which we have applied as follows. We have also updated the figures and draft PDF on the project page to reflect the changes: https://sites.google.com/v... | Summary: This paper presents a pipeline for inducing, detecting, and characterizing neural modules within reinforcement learning (RL) policy networks to enhance interpretability. By penalizing non-local connectivity and encouraging sparsity and locality in network weights, the fully connected networks used in the study... | Rebuttal 1:
Rebuttal: Thank you for your detailed response and consideration of our work. We appreciate your recognition of the compelling nature of our proposed approach, and respond to the points raised as follows:
**Scalability.** We agree that application to more complex tasks will be valuable for further evidenci... | Summary: This paper proposes a method to learn a functionally modular and interpretable model in an RL policy network. It combines a few ideas:
- Spatially embedded neurons with a distance-weighted loss to encourage locality
- Neuron relocalization (Algorithm 1)
- Partitioning the model into different modules using va... | Rebuttal 1:
Rebuttal: Thank you for the thorough review. We appreciate the positive comments on the quality of our execution and are grateful for the constructive feedback, which we respond to below.
**Contributions.** We recognise that our training approach builds on Liu et al (2023). Specifically, we do this by adap... | null | null | null | null | null | null |
Conformal Prediction as Bayesian Quadrature | Accept (oral) | Summary: This paper proposes a new Bayesian interpretation of conformal prediction, which recovers standard conformal prediction as its mean and provides additional finite sample uncertainty estimate. Interestingly, as a Bayesian algorithm, the uncertainty comes from both 'lack of precise input locations' and finite (i... | Rebuttal 1:
Rebuttal: We thank the reviewer for writing a thoughtful and detailed review. We are happy that the review has recognized that the choice of prior can be a headache in practical problems and how our method circumvents this. We also thank the reviewer for reading our proofs. We are glad that the reviewer fou... | Summary: The paper proposes a Bayesian quadrature approach as a Bayesian alternative to conformal prediction, encompassing two widely used methods: split conformal prediction and conformal risk control. The equivalence between these approaches and Bayesian quadrature is clearly established through theoretical proofs. E... | Rebuttal 1:
Rebuttal: We thank the reviewer for writing a detailed review that recognizes the bridging nature of our work and its benefits for various decision-making domains.
**Generalizing Conformal Prediction.** One of our main contributions is to show how both split conformal prediction and Conformal Risk Control ... | Summary: The authors propose a Bayesian version of conformal prediction, which guarantees conditional coverage, rather than marginal coverage. The technique is distribution free, since it considers the worst case risk by maximizing over all possible priors. This is made tractable by leveraging some prior results on di... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to write a thoughtful and detailed review. We are pleased that the review recognized the novelty of creatively combining Bayesian quadrature and distribution-free analysis of random quantile spacings.
We will update the paper to clarify that our guarantee... | Summary: This paper proposes a Bayesian reinterpretation of conformal prediction, framing it within a Bayesian quadrature framework. The authors shows that split conformal prediction and conformal risk control can be derived as special cases of Bayesian quadrature. By modeling uncertainty over quantile functions and le... | Rebuttal 1:
Rebuttal: We thank the reviewer for taking the time to write a thoughtful and detailed review of our work. We appreciate that the review has recognized the "very interesting and novel insights about UQ & decision making" provided by our paper.
We appreciate the desire for additional experiments to provide ... | null | null | null | null | null | null |
Learning to Plan & Reason for Evaluation with Thinking-LLM-as-a-Judge | Accept (poster) | Summary: The paper proposes a new training pipeline for LLM-as-a-judge models, using online preference optimization techniques as well as an agentic workflow that lets the model first output a plan then the detailed execution followed by the verdict. Experiments indicate the superiority of their new approach
Claims An... | Rebuttal 1:
Rebuttal: Thank you for your review!
> it is unclear this manual decomposition of plan/execute is most optimal for models
We added additional experiments to showcase the effectiveness of EvalPlanner (planning+execution) for smaller models. In particular, we experimented with Llama-3.1-8B-Instruct and obta... | Summary: This paper introduces EvalPlanner, which is a preference optimization algorithm for thinking-llm-as-a-judge. It first generates an unconstrained evaluation plan, followed by its execution, and then the final judgement. It uses a selftraining loop to iteratively optimizes data and evaluation predictions. The pa... | Rebuttal 1:
Rebuttal: Thank you for your positive comments about the novelty and the extensive experiments of our paper.
> it remains unclear whether the proposed method would also be effective on LLMs from other families or smaller-sized models
EvalPlanner, does in fact, work well even with smaller sized models. To ... | Summary: The paper introduces EvalPlanner, a novel preference optimization algorithm designed to enhance the Thinking-LLM-as-a-Judge framework for evaluating LLM responses. The approach employs a self-training loop that iteratively optimizes synthetic evaluation plans and executions using Direct Preference Optimization... | Rebuttal 1:
Rebuttal: Thank you for your review and appreciating our work! We are also glad to hear that you’re willing to adjust your scores. We respond to your comments below.
> Evalplanner’s applicability to Best-of-N settings
Upon your suggestion, we conducted some experiments and obtained promising results. Plea... | Summary: This paper proposes EvalPlanner, a method that separates planning from reasoning to enhance LLM-as-a-Judge evaluation. EvalPlanner iteratively improves itself using synthetic preference pairs, achieving state-of-the-art performance (93.9%) on RewardBench and strong results on RM-Bench, JudgeBench, and FollowBe... | Rebuttal 1:
Rebuttal: Thank you for your review!
> Assess whether the proposed judge can be applied to rejection sampling …
First, note that performing extensive RLHF experiments with Evalplanner is beyond the scope of this work and requires separate studies. That said, upon your suggestion, we conducted additional ... | null | null | null | null | null | null |
SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images | Accept (poster) | Summary: This paper proposes an image steganography method based on the diffusion model. By introducing a reference image and adaptive keys, it solves the problems of "limited control of text prompts" and "insufficient key security" in current methods, improves the quality of the revealed secret images, and enhances th... | Rebuttal 1:
Rebuttal: We greatly appreciate your thorough feedback and the time you've dedicated to reviewing our work. We sincerely hope that our clarifications will address your concerns and strengthen your confidence in our work.
A1: (1) Conceal and Reveal Process. In the proposed model, the secret image $x_{sec}$... | Summary: The paper presents a novel generative steganography method, SSHR, which incorporates the diffusion model to address challenges in image steganography. It replaces the traditionally used text prompts with reference images and adaptive symmetric keys to generate stego images, providing greater control over the i... | Rebuttal 1:
Rebuttal: We greatly appreciate the very detailed feedback and your recognition of our contributions! We sincerely hope our response below will further enhance your confidence in our work.
A1: (1) Reference Image Selection and Model Performance. The selection of the reference image does not significantly ... | Summary: This paper proposed a targeted solution to some drawbacks in diffusion model-based generative steganography with text prompts. Although various experiments indicate the proposed model can outperform existing methods in terms of recovery quality and secret image security, there still exist some issues：
1) In t... | Rebuttal 1:
Rebuttal: We greatly appreciate your valuable feedback and sincerely hope our response adequately addresses your points and restores your confidence in our work.
A1: Fundamental Disparities from Cover-based Steganography. We clarify that the proposed method effectively bridges the gap between cover-based... | null | null | null | null | null | null | null | null |
Integer Programming for Generalized Causal Bootstrap Designs | Accept (poster) | Summary: The authors proposed to numerically estimate the joint distribution with the highest variance thus leading to the ATE estimate with the lowest variance for an RCT.
Their optimization problem is modeled by all the posible choices of assignment rules but instead of optimizing over all possible assignments they... | Rebuttal 1:
Rebuttal: We thank the reviewer for their question about theoretical guarantees on the ATE. Our results all bound the variance of the ATE, so we believe the reviewer is asking about the bias of the ATE. Our results hold for a general class of quadratic-in-treatment estimators, which each have their own bias... | Summary: This paper proposed a new integer program to jointly address two sources of uncertainty in causal inference, the design uncertainty due to the treatment assignment mechanism, and sampling uncertainty. Traditional methods tend to address one of the two uncertainties, but do not handle them at the same time. Mot... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful consideration of our work and overall positive review. The main concern seems to be with the number of datasets and baselines. We were unfortunately limited by space considerations, and relegated additional experimental results to Appendix A.1. We are excite... | Summary: This paper proposes a novel method for quantifying **design uncertainty** in causal inference settings, particularly when experiments involve small samples, heterogeneous treatment effects, or non-standard assignment mechanisms. The standard bootstrap only captures sampling uncertainty, while existing causal b... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful engagement with our work.
## Responses to Questions
1. Thank you for this question; please see the discussion of scalability in our response to Reviewer bwtT.
2. We currently discretize the continuous outcomes using the grid of observed outcomes, so th... | Summary: The paper presents a method that employs integer programming to maximize the variance of proposed estimators in randomized experimental design, addressing the issue of design uncertainty. It extends linear-in-treatment and quadratic-in-treatment estimators and generalizes assignment mechanisms using integer pr... | Rebuttal 1:
Rebuttal: We thank the reviewer for their careful and positive review of our paper. Below, we address the two questions raised.
Q1: Unconfoundedness implies that the treatment assignment is independent of any potential outcome, usually achieved through randomization. We distinguish between conditional an... | null | null | null | null | null | null |
Model Uncertainty Quantification by Conformal Prediction in Continual Learning | Accept (poster) | Summary: The paper addresses the problem of continual learning with calibration guarantees. More precisely, the purpose is to train a model to address a series of tasks, in a sequential way (i.e., one task after the other). The datasets used to train the model on the successive tasks are not exchangeable, and may even ... | Rebuttal 1:
Rebuttal: **R to OCOS1.** The end of the "Nonconformity score function" paragraph in Section 4 mainly discusses the score set calculated by the nonconformity score function. We are sorry that our notations is hard to follow. We will revised the confusing notations.
**R to OCOS2.** We will improve our writi... | Summary: The authors propose a Conformal prediction-based methodology to address the calibration problem, which is reliably quantifying model prediction uncertainty in continual learning settings. The authors first enumerate reasons why a standard conformal prediction method cannot be extended to continual learning set... | Rebuttal 1:
Rebuttal: **R to EDOA1.** We conduct experiments on simulated data by split conformal prediction (SCP) . Please refer to **R to Q5.** of Reviewer KEqT.
**R to EDOA2.** Here we conduct experiments on real-world data by creating the scenarios where tasks are ordered from easy to hard and hard to easy. In Sec... | Summary: The paper introduces a conformal prediction-based continual learning (CPCL) method to quantify model uncertainty in continual learning models. CPCL constructs a calibration set using replay techniques and applies a nonconformity score function to measure prediction errors. Theoretical analysis and experiments ... | Rebuttal 1:
Rebuttal: **R to W1.** Here we conduct experiments in domain-incremental scenarios with gradual distribution shifts. We use the dataset of CORe50 [1], which contains 50 objects (classes). Each object has been collected in 8 distinct indoor sessions characterized by different backgrounds and lighting. Due to... | Summary: This paper explores **calibration in continual learning**, specifically focusing on **model uncertainty quantification** using **Conformal Prediction (CP)**. CP provides **theoretical coverage guarantees** under the assumption that data are **exchangeable**, but this assumption is violated in **continual learn... | Rebuttal 1:
Rebuttal: **C1. Lack of ...**
***R to C1.*** Here we discuss the essential difference of QRF against quantile regression (QR) approaches [1] in continual learning. QR approaches estimate the conditional quantiles of the response variable over varying predictor variables. At training time, QR approaches nee... | null | null | null | null | null | null |
EncryptedLLM: Privacy-Preserving Large Language Model Inference via GPU-Accelerated Fully Homomorphic Encryption | Accept (poster) | Summary: This privacy preservation for cloud-deployed LLMs is considered. This work proposes a GPU-accelerated Fully Homomorphic Encryption(FHE) for LLMs. Evaluations are made on a GPT-2 LLM.
Claims And Evidence: See Strengths And Weaknesses below
Methods And Evaluation Criteria: See Strengths And Weaknesses below
T... | Rebuttal 1:
Rebuttal: Thank you for your review.
We refer to the rebuttal of reviewer gAoN for additional benchmarks with larger models. We will include these in the next version. | Summary: This paper addresses the practical challenge of performing LLM evaluation, where clients preserve the privacy of their inputs and model owners retain privacy of the model. They propose using FHE to achieve the goal: encrypt the client inputs, perform evaluation homomorphically over encrypted data, and the resu... | Rebuttal 1:
Rebuttal: Thank you for your review.
Q1: We set the parameters of our approximation so that accuracy is essentially unchanged. To see the evidence of how well our approximations scale, please see the rebuttal for reviewer gAoN.
Q2: We added roughly 10k lines of code to the core OpenFHE library as well a... | Summary: The paper presents a novel approach to privacy-preserving inference for large language models (LLMs) using GPU-accelerated fully homomorphic encryption (FHE), specifically targeting the GPT-2 architecture. It addresses significant privacy concerns associated with LLMs, particularly when deployed on third-party... | Rebuttal 1:
Rebuttal: Thank you for your review.
We assure the reviewer that implementing GPU-accelerated FHE is a highly non-trivial task, requiring the synthesis of dozens of algorithms & optimizations from prior works. We additionally implemented several optimizations derived from the design of custom ASIC & FPGA ... | Summary: This paper presents a GPU-accelerated implementation of CKKS-based fully homomorphic encryption (FHE) for non-interactive private LLM inference. Specifically, it focuses on enabling privacy-preserving (for users' sensitive data) access to proprietary LLMs (e.g., ChatGPT) for latency-tolerant tasks such as docu... | Rebuttal 1:
Rebuttal: Thank you for your review.
The threat model here is semi-honest.
We wrote this paper with a broad audience in mind, including cryptographers who may not be familiar with the LLM circuit. The paper describes the details of the LLM circuit as it is necessary to completely implement the LLM as a h... | null | null | null | null | null | null |
Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector | Accept (poster) | Summary: The paper focuses on catastrophic forgetting in the two-stage object detector. The authors first analyse the forgetting in component-level and reveal that RoI Head classifier is the primary cause of catastrophic forgetting. Then the authors propose Regional Prototype Replay (RePRE) to mitigate forgetting via r... | Rebuttal 1:
Rebuttal: Thanks for the insightful comments.
**Q1:** Dataset concern.
**R1:** To show the generalizability of our key findings, we conducted experiments on a wildly used remote sensing detection datasets DIOR. **The three key findings still hold with different two-stage IODs.** As shown in this link [DIO... | Summary: The paper addresses the critical challenge of catastrophic forgetting in incremental object detection (IOD). The authors focus on the Faster R-CNN architecture and identify that catastrophic forgetting predominantly occurs in the RoI Head classifier, while the regressor remains robust across incremental stages... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's positvie feedback and insightful comment. The Questions and Responses are as follows.
**Q1:** More theorical analysis, such as complexity analysis can be provided.
**R1:** Thanks for the suggestions. We provide a comprehensive analysis of the parameter, com... | Summary: The paper investigates catastrophic forgetting in incremental object detection using standard Faster-RCNN architecture. The authors show that catastrophic forgetting mainly happens in the RoI part of the model, and the regressor behaves more robustly while learning subsequent tasks. Based on their observations... | Rebuttal 1:
Rebuttal: Thanks for the insightful comments.
**Q0: On the Avg metric.**
**R0:** We show the Avg performance at every step.
|10-10|Step1|\||Step 2|Base|New|Avg|All|
|-|-|-|-|-|-|-|-|
|Baseline|77.8|\||Baseline|69.3|73.3|71.3|71.3|
|BPF*|77.8|\||BPF*|71.8|73.4|72.6|72.6|
|NSGP-RePRE|77.8|\||NSGP-RePRE|75.... | Summary: This paper addresses the challenge of catastrophic forgetting in incremental object detection, particularly in two-stage detectors like Faster R-CNN. The authors identify that catastrophic forgetting predominantly occurs in the RoI Head classifier, while the RPN and regression branches remain robust across inc... | Rebuttal 1:
Rebuttal: We sincerely appreciate the reviewer's positvie feedback and insightful comment. The Questions and Responses are as follows.
**Q1**: Using NSGP has been explored in the literature for continual learning.
**R1**:Although NSGP has been explored in incremental classification, it is non-trivial to a... | null | null | null | null | null | null |
Meta-Reinforcement Learning with Adaptation from Human Feedback via Preference-Order-Preserving Task Embedding | Accept (poster) | Summary: This paper focuses on meta-reinforcement learning with human in the loop adaptation scenario and proposes the Preference-Order-preserving EMbedding framework. The core idea of this framework is that if the optimal strategy of the environment achieves better performance in the other task, then the two tasks are... | Rebuttal 1:
Rebuttal: We are grateful and indebted for the time and effort invested in evaluating our manuscript and for all the suggestions to make our manuscript better.
>**Weakness 1 and Q1**
**Answer:** Thanks for pointing out the important observation. In both the existing context-based meta-RL methods, such as ... | Summary: This paper presents a novel meta-reinforcement learning (meta-RL) framework called Preference-Order-Preserving EMbedding (POEM), which enables test-time preference-based human-in-the-loop adaptation of the meta-RL policy. The main research problem is how to meta-train a policy when there exists a discrepancy b... | Rebuttal 1:
Rebuttal: We are grateful and indebted for the time and effort invested in evaluating our manuscript. Thanks for the typo reminders and the suggestions to make our manuscript a better and stronger contribution.
>**Methods And Evaluation Criteria: One minor weakness in the evaluation is the experiments are... | Summary: The paper presents a framework for meta-reinforcement learning (meta-RL) called Preference-Order-preserving Embedding (POEM), which aims to facilitate few-shot adaptation to new tasks using human preference queries instead of traditional reward signals. The framework comprises a task encoder that maps tasks in... | Rebuttal 1:
Rebuttal: >**Claims and Evidence: The approach is quite similar to the work proposed by ANOLE.**
**Answer:** The partitioning of the embedding into reward embedding and policy embedding spaces is only an initial and minor design of the paper. The main contribution of the proposed method is that we train a ... | Summary: The authors present the adaption via Preference-Order-preserving EMbedding (POEM) framework. Their key insights that are if the trajectory of a task is distilled into an embedding, the similarities between tasks should be evident in these embeddings and that the optimal policy on one task should do sufficientl... | Rebuttal 1:
Rebuttal: We are grateful and indebted for the time and effort invested in evaluating our manuscript and for all the suggestions to make our manuscript a better and stronger contribution.
>**Methods And Evaluation Criteria 1: I could see issues with Property 1 being called into question in scenarios where ... | null | null | null | null | null | null |
From Complex to Atomic: Enhancing Augmented Generation via Knowledge-Aware Dual Rewriting and Reasoning | Accept (poster) | Summary: The authors propose a knowledge aware rewriting and reasoning framework, a variant of retrieval augmented reneration (RAG), which is suitable for multi hop question answering tasks since it can aggregate knowledge from different documents. It consists of 4 steps: knowledge atomizer, query proposer, atomic retr... | Rebuttal 1:
Rebuttal: We appreciate the valuable feedback and insightful comments from the reviewer. Below, we address each concern point by point. All tables are accessible via [hyperlinks](https://tinyurl.com/R4-tables).
### *Q1.* Embedding model used for retrieval.
The text-embedding-ada-002 model is used across ... | Summary: This paper addresses the challenge of solving complex, multi-hop queries in domain-specific contexts by introducing a method called KAR^3-RAG. Traditional Retrieval-Augmented Generation (RAG) techniques often rely on straightforward text retrieval methods, which can struggle when queries require multiple steps... | Rebuttal 1:
Rebuttal: We appreciate reviewer's feedback. All tables are accessible via hyperlinks.
### *Q1-1.* How large was the knowledge base in the experiments?
[Table 1](https://tinyurl.com/R3-tables) provides detailed statistics. All chunks are derived from the context paragraphs of the sampled QA, with chunk co... | Summary: This paper proposes a framework for handling multi-hop questions which require complex reasoning which has 4 main components; an atomizer which generates atomic questions from document chunks, a query processor which iteratively generates atomic queries using the input question and current context, an atomic r... | Rebuttal 1:
Rebuttal: We appreciate the valuable feedback and insightful comments from the reviewer. Below, we address the raised concerns point by point.
### *Q1.* Discussion on the important limitation of the method that it is a resource heavy method and can not be used practically when retrieving from dynamic data ... | Summary: The authors present a new RAG framework suitable for addressing complex questions with a focus on multi-hop. The main idea is based on an iterative process of collecting evidence and generating follow-up questions as required. To devise this iterative process, the authors describe four main components: (1) Kno... | Rebuttal 1:
Rebuttal: We appreciate the valuable feedback from the reviewer. Below, we address each concern point by point.
### *Q1.* How KAR-RAG system would fare on questions that require parallel subtasks.
KAR³ is specifically designed to handle complex questions by decomposing them into multiple subqueries, enab... | null | null | null | null | null | null |
PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning | Accept (poster) | Summary: The paper tackles the problem of learning controllable object-centric video prediction from actionless videos. The proposed method, PlaySlot, combines slot-attention-based object-centric representations with a previous slot dynamics prediction module (OCVP) conditioned on learned latent actions (inverse module... | Rebuttal 1:
Rebuttal: Thanks for the constructive review. We are happy that you find our claims clear and our experimental design sound.
For the final version, we will try to improve the description of the latent action modules and we will rename the subsection to “Unsupervised Slot-based Object-centric Learning”.
Belo... | Summary: This paper introduces PlaySlot, an object-centric video prediction model that learns inverse latent dynamics for controllable future frame forecasting and can be used in downstream tasks. PlaySlot infers object representations and latent actions from unlabeled video sequences instead of action annotations, lev... | Rebuttal 1:
Rebuttal: Thanks for the constructive comments. We are delighted that you consider our experimental section and model analysis as a strength.
Below we address the highlighted weaknesses and questions, and clarify some misunderstandings.
**The action decoder still requires ground-truth actions for trainin... | Summary: This work introduces a novel approach for the video prediction task using object-centric representations. It proposes an InvDyn module to learn latent action embeddings and a conditional object-centric predictor to forecast future object slots. Extensive experimental results demonstrate superior performance co... | Rebuttal 1:
Rebuttal: Thanks for the positive review, acknowledging our experimental design, highlighting that our method is novel and interesting, and that our paper is well-written.
Below we address some of your questions and add clarifications about certain weaknesses:
**Clarification about the lower performance o... | Summary: The papers proposes PlaySlot - a novel approach to controllable video prediction that builds on previous work on object-centric learning and latent action learning. Unlike previous approaches to video prediction based on object-centric learning, PlaySlot incorporates the InvDyn module for inferring latent acti... | Rebuttal 1:
Rebuttal: Thanks for the constructive review and for finding our proposed approach novel, exciting and potentially useful.
Below we address your questions and main comments. We will address other comments (e.g. remove exaggerated claims and add missing references) in the final version of the paper. Further... | null | null | null | null | null | null |
Synthesizing Software Engineering Data in a Test-Driven Manner | Accept (poster) | Summary: This paper proposes a TDD-driven data synthesis framework `UnitFlow`, which can
generate data samples for incremental development tasks based on real-world
GitHub project and unit tests. Based on this framework, this paper constructs a
promising benchmark `UnitFlow-Eval`, which contains data samples from 74
re... | Rebuttal 1:
Rebuttal: Dear reviewer,
We are grateful for the reviewer’s valuable suggestions. Our detailed responses to the concerns are provided below.
---
> **Concern D1:** The paper would benefit from additional empirical evidence to substantiate the claimed improvements in LLM performance on incremental developme... | Summary: This paper introduces UnitFlow, a novel data synthesis framework based on Test-Driven Development (TDD), which automatically generates software engineering data by inferring incremental development steps directly from unit tests. The framework constructs a Runtime Dependency Graph (RDG) to capture function int... | Rebuttal 1:
Rebuttal: Dear reviewer,
We sincerely appreciate your thoughtful and constructive feedback. Please find our responses to the concerns below.
---
> **Concern C1:** Is the synthesized data valid and of high quality?
We fine-tuned the **Qwen2.5-Coder-32B-Instruct** model using data synthesized by UnitFlow,... | Summary: This paper proposes a data synthesis framework UnitFlow that leverages Test-Driven Development to automatically generate high-quality, structured, and verifiable training data for LLMs in software engineering. It constructs a Runtime Dependency Graph (RDG) from unit tests to capture function interactions and g... | Rebuttal 1:
Rebuttal: Dear reviewer,
We are grateful for your valuable suggestions. Our detailed responses to the concerns are provided below.
---
> **Concern B1:** The evaluation of the proposed method is limited to the UnitFlow-Eval benchmark, which shares the same distribution as the training data. The authors di... | Summary: The paper introduces UnitFlow, a novel framework for synthesizing test-driven software engineering data. Unlike prior datasets that rely on human-submitted issues, UnitFlow generates incremental development steps directly from unit tests. The framework constructs a Runtime Dependency Graph (RDG) to capture fun... | Rebuttal 1:
Rebuttal: Dear reviewer,
We deeply appreciate your thoughtful review and your recognition of our contributions. Below, we provide point-by-point responses to your concerns and suggestions.
---
> **Concern A1:** The evaluation does not compare against human-written commit sequences, which could serve as a... | null | null | null | null | null | null |
A Closer Look at Generalized BH Algorithm for Out-of-Distribution Detection | Accept (poster) | Summary: The paper investigates methods for setting a decision threshold for out-of-distribution (OOD) detection using a calibration set, ensuring that the resulting detector achieves a desired performance on unseen data. The authors examine the g-BH algorithm, a recently proposed method for threshold selection [Ma 202... | Rebuttal 1:
Rebuttal: **Q1**: about the claim and goal of our paper.
**Ans1**: We first emphasize that FDR is closely related to TPR and FPR. Based on [1], we have
$$ FDR = E(\frac{1}{1 + \frac{P}{N}\cdot \frac{1-FPR}{1-TPR} }) $$
where P is the number of ID data in test set and N is the number of OOD data in test set... | Summary: The paper explores the role of the calibrated set in the performance of the g-BH algorithm for OOD detection. Theoretical results indicates the large calibrated set will improve the performance of the g-BH algorithm but small calibrated set tends to degrade the performance of
It. Then, the authors propose a no... | Rebuttal 1:
Rebuttal: __Weakness 1__: about the interpretations of some concepts, including type-1 error, significant level and FDR.
__Ans1__: we interpret these concepts as follows:
Significance level: If the probability of obtaining a result as extreme as the one obtained, supposing that the null hypothesis were tr... | Summary: Based on the recent work [1], this paper aims to study the influence of the calibrated set on the generalized BH (g-BH) algorithm for out-of-distribution (OOD) detection task. By theoretical analysis and experimental results on the real data, the authors show that the small calibrated set tends to degrade the ... | Rebuttal 1:
Rebuttal: __Weakness 1__: the proposed method requires a hold-out set. However, since the hold-out set consists of ID examples, this is not a significant restriction.
__Ans1__: We emphasize that the calibrated set consists of ID data, without the need of OOD data. We directly extract some examples from tr... | Summary: This paper investigates the impact of the calibrated set on the generalized BH (g-BH) algorithm[1] for Out-of-Distribution (OOD) detection. The authors provide a theoretical analysis showing that the conditional expectation of the true positive rate (TPR) follows a beta distribution, demonstrating that a small... | Rebuttal 1:
Rebuttal: __W1__: The proposed eg-BH algorithm depends on multiple p-values. the advantages of integrating multiple p-values should be discussed
__Ans1__. A small calibrated set leads to under-representative empirical p-values, which fail to capture the distributional
characteristics of the ID data. To a... | null | null | null | null | null | null |
The Meta-Representation Hypothesis | Reject | Summary: This paper proposes "the meta-representation hypothesis", i.e., learning a representation that reflects the abstract, high-level understanding of inputs can lead to better generalization of RL agents. The paper models the generalization problem in RL using an MDP generator and assumes that all different MDPs t... | Rebuttal 1:
Rebuttal: Dear Reviewer twjt,
Thank you for your positive assessment of our work. Below, we will address your concerns.
>I think prior work in two related areas should be discussed:
>- **Invariant representation learning** [1, 2].
>- **Representation learning in RL** [3].
Thank you for your valuable sug... | Summary: The paper deals with RL environments where the observations presented to an agent are noisy transformations of the true state via a rendering function that is drawn from an environment dependent distribution.
This paper first rigorously demonstrates how generalization performance of RL agents can suffer if th... | Rebuttal 1:
Rebuttal: Dear Reviewer Rz5E,
Thank you for your positive feedback on our paper. Below, we will address your concerns.
>The only concern I have here is that the evaluation is limited to only specific kinds of distributions of rendering functions.
>I, however, found that the evaluation in the paper is lim... | Summary: The paper proposes to combine Deep Mutual Learning with RL. In Deep Mutual Learning, several learners learn independently but at the same try to minimize the KL between their predictive distributions. The paper hypothesizes that two RL policies can learn from different MDPs — where each MDP has its own randoml... | Rebuttal 1:
Rebuttal: Dear Reviewer ByPb,
Thank you for your careful evaluation on our paper. Below, we will address your concerns.
>The idea of RL + DML appears to be not very novel. Therefore I believe the key novelty is mainly about how one can get perturbed MDPs. However, I tend to think that the randomized CNN a... | Summary: The paper tackles generalisation in reinfocement learning introducing the concept of "meta-representation", which is an abstract representation of a state shared by all instances with shared semantics, and separated from the details of a particular high-dimensional observation.
The paper brings two (almost se... | Rebuttal 1:
Rebuttal: Dear Reviewer bqXb,
Thank you for your constructive feedback on our paper! Below, we will address your concerns.
>Did you correct for the "actual" number of environment interactions? Does the DML combo with two policies see double the states compared to the PPO baseline?
Thanks. First, the two ... | null | null | null | null | null | null |
Taming Diffusion for Dataset Distillation with High Representativeness | Accept (poster) | Summary: This paper utilizes DDIM inversion to map the VAE latent space into a high-normality Gaussian space. Numerous subsets are then sampled from this Gaussian distribution. The final selection is determined by identifying the subset whose distribution has the smallest loss compared to the Gaussian distribution.
The... | Rebuttal 1:
Rebuttal: **Q1: The concern about the multi-component Gaussian mixture distribution assumption.**
A1: Thank you for highlighting this. To clarify the assumption, we offer a more comprehensive explanation here.
For current diffusion-based methods and ours, the pre-trained VAE are not conditional VAE, so bo... | Summary: This paper addresses limitations in diffusion-based dataset distillation methods and introduces D3HR, a novel framework that enhances the representativeness of distilled datasets. This paper reveal that current methods suffer from issues like inaccurate distribution matching, distribution deviation due to rand... | Rebuttal 1:
Rebuttal: **Q1: Figure 1 and 3 less informative**
A1: Thanks for the suggestions. The “blue lines” are the “blue contour lines” which indicate the probability density of the distribution. We will revise the caption to clarify this.
Figure 1 is intended to convey two key messages: (1) The VAE latent space e... | Summary: This article introduces the D³HR framework (Taming Diffusion for Dataset Distillation with High Representativeness) to tackle the issue of inaccurate distribution matching in existing diffusion-based dataset distillation methods. D³HR enhances distribution matching accuracy by utilizing DDIM inversion to trans... | Rebuttal 1:
Rebuttal: **Q1: primarily focuses on optimal subset selection**
In line 027-044 right column, we identify three key limitations in diffusion-based methods due to their reliance on optimization in the VAE latent space. This motivates us to identify a more effective space by DDIM inversion, aiming to provide... | Summary: This work proposes a novel diffusion-based dataset distillation solution. Based on the fact that previous methods suffer from inaccurate distribution matching, the authors propose to convert the images to latents with DDIM inversion and model it as Gaussian. Then, sample multiple subsets from the Gaussian and ... | Rebuttal 1:
Rebuttal: **Q1: Comparison to MiniMax with hard labels**
As noted in Lines 297–300 right column, we did not include Minimax in the main table as it focuses on small subsets of ImageNet-1K. For large datasets, it requires extra training of multiple diffusion models with high computational cost. They only re... | Summary: This paper proposes a diffusion-based dataset distillation method. The paper claims that the VAE space of the diffusion model is more difficult for distribution matching. To tackle this challenge, the core idea of the proposed method is to apply DDIM inversion to each sample in the original dataset and then mo... | Rebuttal 1:
Rebuttal: **Q1: no clear definition of normality**
Thanks for pointing this out. In our context, normality refers to the degree of the latent space data conforms to a normal distribution. A higher level of normality indicates that the latent distribution more closely resembles a normal distribution. This u... | null | null | null | null |
Diversified Flow Matching with Translation Identifiability | Accept (poster) | Summary: This paper proposed a flow matching model for diversified distribution matching, called DFM. The proposed method formulates a bilevel optimization problem to learn an interpolant and train a flow model on this interpolant. This paper demonstrates that the standard flow matching models fail in the DDM task. To ... | Rebuttal 1:
Rebuttal: **[Novelty of Proposition 3.4]**
The novelty lies in how to use FM-based losses to attain the same conclusion of Thm 2.2. Note that Thm 2.2 assumes that distribution matching is already attained. But Prop 3.4 specifically needs the distribution matching part to be realized by FM. How to achieve t... | Summary: The paper introduces Diversified Flow Matching (DFM), a novel unsupervised domain translation (UDT) framework that extends ODE-based Flow Matching (FM) from linear to nonlinear interpolants, addressing the critical limitation of translation identifiability.
Key Contributions:
Overcoming Linear Flow Matching... | Rebuttal 1:
Rebuttal:
**Anonymized URL**: (https://drive.google.com/file/d/1U4gdB5qy1d98AJ1YWxR2v3fzDQp167nt/view?usp=sharing)
**[Limitations of Claim 2: Bilevel Loss implies Identifiability]**
To clarify, the proposed DFM and the baselines FM, FM-OT, and FM-cond do use the same architecture, and are... | Summary: The paper introduce Diversified Flow Matching (DFM), an FM-based framework for DDM. They design a custom loss function and nonlinear interpolant to ensure translation identifiability, addressing the limitations of conventional FM methods that use linear interpolants. By leveraging the non-overlapping property ... | Rebuttal 1:
Rebuttal:
**[Quantitative Metrics for Image to Image Translation]**
Image similarity metrics were not presented because we observed that metrics such as LPIPS do not make sense when the images have large domain gaps (i.e., when the geometric representations of the feature spaces are largely differe... | Summary: This paper aims to address the unpaired domain translation problem with conditional information. The previous method is based on GAN. However, GAN training may not be stable; this paper proposes a Flow Matching-based method. However, the naive FM method may not work because conditional distributions may be mis... | Rebuttal 1:
Rebuttal: **[Intersecting trajectories in figures]**
Note that Fig. 3(b) does **not** show the velocity field but the interpolant $I^{\rm linear}(x, y, t)$ where $x, y \sim \rho(x, y | u^{(q)})$. The wording was meant to imply that interpolants guide the learning of the vector field. Our term “intended tra... | null | null | null | null | null | null |
Gradient-based Explanations for Deep Learning Survival Models | Accept (poster) | Summary: The paper benchmarks previously proposed gradient-based explanation methods across three previously proposed deep survival analysis methods. Experimental results on synthetic and real-world datasets highlight differences in performance across scenarios.
Claims And Evidence: - The paper claims an extension of ... | Rebuttal 1:
Rebuttal: Thank you for your valuable and insightful feedback. Before addressing your concerns in detail, we want to clarify a few crucial aspects and potential misunderstandings:
* This is **not** a benchmark paper.
* We do provide the definition and explanation of local accuracy in the appendix.
* We do... | Summary: This paper shows a comparative study on various explanation methods for survival analysis. While there are several model-agnostic methods to interpret models for survival analysis, this paper considers gradient-based methods. The applicability of the gradient-based methods is limited to models that can compu... | Rebuttal 1:
Rebuttal: Thank you for your careful evaluation and suggestions. We acknowledge that our contributions may not have been communicated clearly enough in the original submission. To address this, we will revise the manuscript to better clarify these key contributions:
* Our work follows adaptations common in... | Summary: The authors introduce GradSHAP(t), an extension of SurvSHAP(t) that analyzes the gradients to better explain the model’s predictions. The authors also propose extensions of other gradient-focused XAI methods to align with the survival task.
Claims And Evidence: Yes.
Methods And Evaluation Criteria: GradSHAP(... | Rebuttal 1:
Rebuttal: We appreciate the suggestion to include semiparametric AFT-based survival deep learning models, such as Deep AFT Rank-regression for Time-to-event prediction model (DART). It is an interesting approach, which estimates the survival function in similar fashion to a non-Cox-based version of the Deep... | Summary: This paper addresses the challenge of interpreting "black box" deep learning models used for survival analysis, which predict time-to-event outcomes. The authors introduce a framework for gradient-based explanation methods to capture the time-dependent influence of various features, including those from multi-... | Rebuttal 1:
Rebuttal: We are grateful for your constructive feedback. In response, we conducted additional experiments on the feasibility and computational efficiency of GradSHAP(t) and SurvSHAP(t) on the multi-modal real data example, which we will include in the final paper. Here, GradSHAP(t) took 5 minutes to comput... | null | null | null | null | null | null |
ShapeEmbed: a self-supervised learning framework for shape quantification | Reject | Summary: This paper presents a self-supervised method for learning object shape, given a binary segmentation mask, that by construction is invariant to translation/scaling/rotation/reflection/outline-point-indexing. This method, ShapeEmbed, consists of extracting a normalized distance matrix from points sampled along ... | Rebuttal 1:
Rebuttal: # Envisioned use of the method in biological imaging
Morphological features extracted from 2D images serve as phenotypic fingerprints to reveal cell identity, cell states, and response to chemical treatments (see 10.1038/s41592-024-02241-6 for a recent example). Shape, as captured in 2D contours, ... | Summary: Authors introduce a network for 2D shape analysis (silhouettes of objects in images). It works as follows: shape outline is interpolated via spline curve with fixed number of points N across all data samples; pairwise distance matrix is constructed based on those points and normalized to unit Frobenius norm; r... | Rebuttal 1:
Rebuttal: # ShapeEmbed assumptions on 2D shapes connectedness
We thank you for your insightful questions and hereafter answer them one by one.
**Does the method assume that 2D shapes are simply-connected?** Yes, ShapeEmbed operates with contours that are simply connected and described by a sequence of ord... | Summary: The paper proposes a novel self-supervised framework for shape embedding, which is invariant to translation, scale, and outline pixel-indexing. The learned shape representation is used in a classification task and outperforms all previous works.
## update after rebuttal
My final rating is accept.
Claims And E... | Rebuttal 1:
Rebuttal: Thank you for your positive comments on the strengths of ShapeEmbed and for your appreciation of the way we describe and evaluate the method. | Summary: This work presents ShapeEmbed, a new approach for representation learning of 2D shapes (represented as contours/outlines). The main desiderata for such a representation are invariance to translation, rotation, scaling, reflection, and indexing.
The key ideas behind the design of this approach build upon using... | Rebuttal 1:
Rebuttal: # Significance of investigating self-supervised learning of 2D silhouettes
We refer to our reply to Reviewer MfnJ, where we clarify how we envision the method to be used in biological imaging.
# Comparison with prevailing computer vision self-supervised learning algorithms
We hereafter provide ... | null | null | null | null | null | null |
Subgroups Matter for Robust Bias Mitigation | Accept (poster) | Summary: This paper studies how the definition of subgroups affects the efficacy of bias mitigation techniques for spurious correlations. A causal graph approach is introduced to formalize correlations between classes, attributes, and subgroups, which are then manipulated to study AUC with respect to an ERM baseline on... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful and comprehensive review. We agree with the comments and feel they have helped us substantially improve the paper. We have made clarifications to the manuscript and run >200 additional experiments. We include details of our changes below and share all res... | Summary: The authors investigate the impact of group definitions on the performance of bias mitigation methods using semi-synthetic experiments on binary classification of images. Specifically, the authors introduce a spurious correlation into the training datasets by selecting examples based on two attributes and a la... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful comments, and we are glad to hear they appreciate the paper and the originality of our findings. We respond to the comments raised in the review below, and refer to additional results, which are all presented in this [folder](https://drive.google.com/driv... | Summary: The paper investigates how subgroup definition impacts the effectiveness of bias mitigation methods in machine learning, hypothesizing that inconsistent success stems from this often-overlooked factor. Through experiments on semi-synthetic image classification tasks with varied subgroup definitions (coarse, fi... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed review and helpful suggestions. We summarise the >200 supplementary experiments and edits we have made to address them below. All extra results are shared in this [folder](https://rb.gy/xebqmp). We hope this will give the reviewer confidence to raise their ... | null | null | null | null | null | null | null | null |
Correlation Clustering Beyond the Pivot Algorithm | Accept (poster) | Summary: The paper studied a variant of the pivot algorithm for correlation clustering and gave a dynamic implementation with polylog(n) update time and 2.997 approximation. Correlation clustering is a classical problem in TCS and machine learning. Here, we are given a labeled complete graph $G=(V, E^+ \cup E^-)$, and ... | Rebuttal 1:
Rebuttal: We appreciate your recommendation to run multiple experiments using different random seeds, especially since the Pivot algorithm is known to have instability. To clarify, Figures 1 and 2 already show both average and worst-case outcomes, based on multiple random seeds per dataset. As illustrated, ... | Summary: This paper presents a modification of the standard 3-approximation Pivot algorithm for correlation clustering (select a node in a graph, cluster it with its neighbors, and remove it) called ModifiedPivot, which avoids the worst-case errors of Pivot by clustering some its neighbors as singletons (if their neigh... | Rebuttal 1:
Rebuttal: Thanks for pointing out the confusion regarding Algorithm 1, we will make sure to clarify this in the next version of the paper. To clarify this, the vertices in $A_v$ will be allowed to be picked as pivots later on, but even if they start clusters they won't themselves be added to those clusters ... | Summary: This paper studies the classic correlation clustering problem, where the objective is to partition objects into clusters while minimizing disagreements with given similarity and dissimilarity labels. The PIVOT algorithm by Ailon et al. (STOC’05) provides a 3-approximation for this problem, but its analysis is ... | Rebuttal 1:
Rebuttal: We thank the reviewer for thoughtful comments. We note that while indeed the improvement from 3 to 2.997 in the approximation ratio is rather small *quantitatively*, it breaks a longstanding barrier of 3-approximation for combinatorial algorithms and has an important *qualitative* value. Additiona... | null | null | null | null | null | null | null | null |
BiAssemble: Learning Collaborative Affordance for Bimanual Geometric Assembly | Accept (poster) | Summary: This paper proposes a framework for learning collaborative affordance in bimanual geometric assembly. The task is assembling fractured parts into complete objects, which is a long-horizon task requiring pick-up, alignment, and assembly. The paper tackles this task through predicting collaborative affordance an... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and questions. We have carefully addressed them below.
> W1. Low reported success rates
The relatively low scores across all models and baselines stem from the highly diverse and complex nature of geometric shape assembly task. As detailed in Appendix G, our ... | Summary: In this paper, the authors present a framework for bimanual geometri assembly. They formulate the task into 3 steps: pick-up, alignment and assembly. For pick-up, a point-level affordance prediction module is trained and used; For alignment, a SE(3) transformation is predicted; For assembly, a collision-free d... | Rebuttal 1:
Rebuttal: We sincerely appreciate your thoughtful questions. We've addressed them in detail below.
**For detailed paper references, please refer to our response to Reviewer1 (z3UB).**
> W1. Thresholds of distance and rotation. There is a gap between two parts after assembly.
Thank you for the suggestion.... | Summary: This work focuses the shape assembly task aimed at reconstructing broken objects. A multi-stage BiAssembly framework is put forward to carry out this task. Initially, the BiAssembly framework utilizes SOTA techniques to obtain an imagined assembled shape. Subsequently, it forecasts the disassembly direction, a... | Rebuttal 1:
Rebuttal: Thank you for your valuable questions. We've addressed each of your concerns below.
**For detailed paper references, please refer to our reply to Reviewer1 (Reviewer z3UB).**
> W1. The multi-stage framework has some assumptions. For example, it assumes the object has two broken parts, the imagi... | Summary: This paper addresses the challenges in the observation space and action space by proposing the BiAssemble framework to solve the collaborative problem of bimanual robots in geometric assembly tasks. Specifically, the task is decomposed into three steps: pick-up, alignment, and assembly, which are addressed by ... | Rebuttal 1:
Rebuttal: We sincerely appreciate your valuable questions, and we have provided detailed responses below.
> W1. The algorithm relies on predefined assembly shapes...
The assumption of "imaginary assembled shape" is justified based on two well-established research areas that together ensure both adaptabili... | null | null | null | null | null | null |
Differential Coding for Training-Free ANN-to-SNN Conversion | Accept (poster) | Summary: In this work, the authors proposed differentiable neural coding. Based on the proposed coding, the authors provided differential graded units, differential spiking neurons, and differential coding for linear layer. According to the authors’ experiments, they could achieve state-of-the-art accuracy on image cla... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful and very detailed feedback. We are delighted that you find our paper theoretically supported well and the results state-of-the-art. We would like to address your concerns and answer your questions in the following:
### 1. Answer to "MT neurons are more complex to im... | Summary: This paper introduces a novel differential coding scheme for training-free ANN-to-SNN conversion. The authors propose using time-weighted spikes as incremental updates rather than direct rate representations, significantly reducing energy consumption and spike counts. They detail an algorithmic framework integ... | Rebuttal 1:
Rebuttal: Thank you for your positive and thoughtful comments. We are encouraged that you find our method novel and well-explained, and our empirical results thorough. We would like to address your concerns and answer your questions in the following.
### 1. Hardware implementation of MT Neuron and discussi... | Summary: ANN-to-SNN conversion has been known to produce so-called ‘conversion’ errors. Recent studies proposed methods that can reduce conversion errors, and in this study, the authors propose to improve the earlier studies with a novel algorithm named ‘differential coding’. Specifically, they focus on preventing the ... | Rebuttal 1:
Rebuttal: Thank you for your positive and constructive comments. We are delighted that you find our idea novel, interesting and well explained. We would like to address your concerns and answer your questions in the following.
## 1. It would be great to see evaluations on a few more datasets to strengthen ... | null | null | null | null | null | null | null | null |
Staged and Physics-Grounded Learning Framework with Hyperintensity Prior for Pre-Contrast MRI Synthesis | Accept (poster) | Summary: This paper is about using post-contrast MRI to create pre-contrast MRI via deep learning. Physics principles are built into the model. To tackle the complexity in setting up the model and its training, a two-stage approach is presented, which alleviates the challenge in handling the complexity. The approach is... | Rebuttal 1:
Rebuttal: Thank you for taking time and efforts to review our paper, your opinon is really appreciated. Thank you for acknowledging our work’s theoretical contribution, motivation and usage of the physcis principles. The goal of this project is to leverage the power of AI4Science to tackle unsolved challeng... | Summary: This work discusses a deep learning method for recovering pre-contrasted MRI from post-contrasted MRI. The authors propose to first estimate a thresholded map as a mask indicating contrast agent update. This mask is then fed as an additional conditioning signal for recovering the pre-contrast image. The author... | Rebuttal 1:
Rebuttal: Thank you for your in-depth review of our manuscript. Your comments are constructive and will help improve the quality of our work. In this manuscript, we propose a novel MRI theory-driven method to address a challenging problem in MRI imaging. Please find our detailed responses to your comments b... | Summary: This paper proposes SPHERE, a staged and physics-grounded learning framework for synthesizing Pre-Contrast MRI images from Post-Contrast MRI scans. The key innovation lies in incorporating MRI physics principles and a hyperintensity prior into a two-stage deep learning model. The framework consists of segmenta... | Rebuttal 1:
Rebuttal: Thank you for your meticulous and comprehensive review. We appreciate your recognition of our work on quantitative performance, mathematical support, prior knowledge incorporation, and model generalizability. Regarding the identified weaknesses, we provide the following responses:
1.Yes, the h... | Summary: This paper proposes a novel staged, physics-grounded learning framework with a hyperintensity prior to synthesize Pre-Contrast images directly from Post-Contrast MRIs. The proposed method can generate high-quality Pre-Contrast images, thus, enabling comprehensive diagnostics while reducing the need for additio... | Rebuttal 1:
Rebuttal: Thank you for reviewing our manuscript from the clinical application perspective, your comments are truly constructive for us. We appreciate your recognition of our novelty, integration of domain knowledge, and potential clinical applicability. The motivation of our study is to provide a theoretic... | null | null | null | null | null | null |
Enhancing Adversarial Robustness with Conformal Prediction: A Framework for Guaranteed Model Reliability | Accept (poster) | Summary: This paper studies adversarial robustness of Conformal Prediction. Specifically, this paper develops an attack method that does not require coverage guarantees and integrates it with a conformal training-based defense strategy by minimizing the size of the prediction sets under adversarial perturbations while ... | Rebuttal 1:
Rebuttal: We appreciate the recognition of our paper’s clarity and thank the reviewer for their careful attention to detail. While we share many points of agreement, the main misunderstanding lies in the interpretation of outcomes, which we would like to clarify.
**Tabular Data Interpretation**:
W... | Summary: The paper proposed a conformal prediction based adversarial attack and training method. To enable computationally tractable implementations, the authors propose a smoothed surrogate loss. The attack and defense methods are tested on CIFAR10 and CIFAR100.
## update after rebuttal
I believe the authors have s... | Rebuttal 1:
Rebuttal: We appreciate the recognition of the novelty and clarity of our work. Thanks for your valuable feedback on addressing the effectiveness problem. We would like to address your concern.
**Time Consumption**:
Table 4 reports the time (in seconds) per epoch for each adversarial training model on 100 ... | Summary: This paper introduces a novel approach that integrates Conformal Prediction (CP) with Adversarial Training (AT) to enhance the adversarial robustness of deep learning models. The authors frame adversarial robustness as a bi-level optimization problem, where an attacker maximizes the uncertainty by enlarging th... | Rebuttal 1:
Rebuttal: Thank you for recognizing the novelty, strong theoretical foundation, comprehensive experiments, and practical implications of our work. We appreciate your valuable suggestions and feedback. Below, we address your concerns.
**Time Consumption**: See response to Reviewer 7J63.
**Experimental Dive... | Summary: The paper proposes a framework that integrates adversarial training with conformal prediction (CP) to enhance model robustness against adversarial attacks while maintaining reliable uncertainty estimates. It formulates adversarial training within the CP framework as a bi-level optimization problem, where an at... | Rebuttal 1:
Rebuttal: We truly appreciate your recognition of the novelty and potential impact of our work. Your comments on clarifying the missing experimental details and time consumption are greatly appreciated. Below, we address your points and provide further explanations.
**Experimental Setup**: In principle, an... | null | null | null | null | null | null |
ExPLoRA: Parameter-Efficient Extended Pre-Training to Adapt Vision Transformers under Domain Shifts | Accept (poster) | Summary: The authors show that continued pre-training (with PEFT) on the target domains before supervised fine tuning is an effective methodolofy for adapting natural-image foundation models to non-natural image target tasks. They show particular gains on RGB tasks, but also show some benefit on multi-spectral imagery.... | Rebuttal 1:
Rebuttal: Thank you for your feedback and recognition of the empirical performance and novelty of using parameter-efficency for pre-training in ExPLoRA. Here are our responses:
**Q: Is PEFT needed in either pre-training or fine-tuning beyond memory savings?**
Our primary motivation for parameter-efficien... | Summary: This paper proposes a continual pre-training method with a parameter-efficient fine-tuning (PEFT) module such as LoRA (Hu et al. 2021) to improve the adaptability of visual foundation models on specific domains. By inserting and training the PEFT module inside the general-domain pre-trained backbone model with... | Rebuttal 1:
Rebuttal: Thank you for your valuable feedback. We appreciate your recognition of our extensive, well-designed experimental validation and for ExPLoRA’s improved performance over baselines. Here are our responses:
**Q: Can you provide qualitative analysis that illustrates the difference between pre-trained... | Summary: The paper proposes ExPLoRA, a parameter-efficient way to extend pre-training of a large vision transformer from its original domain (e.g. natural images) to a new domain (e.g satellite imagery). ExPLoRA accomplishes this by unfreezing 1-2 transformer blocks for full training and applying low-rank (LoRA) update... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments and for recognizing our thorough experimental validation across baselines and datasets and for providing useful insights into in-domain parameter-efficient pre-training. You may find responses to your questions below:
**Q: Further justification for which block... | Summary: The authors aim to transfer knowledge from large pre-trained vision models to new domains efficiently, and address different downstream tasks. So, given a set of downstream tasks on a new domain, the straightforward approach is to either pre-train from scratch a large model on this new domain and then fine-tun... | Rebuttal 1:
Rebuttal: Thank you for your detailed and insightful feedback. We appreciate your recognition of ExPLoRA’s novelty in combining fundamental ideas and its strong empirical performance across datasets.
**Q: Discussion of required compute of ExPLoRA vs fine-tuning baselines**
We agree that more details on c... | null | null | null | null | null | null |
Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching | Accept (spotlight poster) | Summary: This paper proposes a method to predict gene expression from histology whole-slide images using generative flow matching. Spatial transcriptomics (ST) datasets are used to train and evaluate the model. A foundation-model encoder is used to extract visual features. Spatial attention is used to model spatial dep... | Rebuttal 1:
Rebuttal: We sincerely thank you for the insightful comments and will revise the manuscript accordingly. Please see our detailed responses below:
```
A complex system with several modules and hyperparameters makes it challenging to evaluate.
```
Ans: Thank you so much for the feedback. STFlow primarily co... | Summary: This paper proposes a flow matching (FM) approach (called STFlow) for predicting the spatial transcriptomics (ST) from pathological Whole-Slide Images (WSIs). The core designs of STFlow contain i) learning the joint distribution $p(Y_0,\cdots,Y_N|I_0,\cdots,I_N)$ using the FM approach and ii) the E(2)-invarian... | Rebuttal 1:
Rebuttal: We sincerely thank you for the insightful comments and will revise the manuscript accordingly. Due to character limits, all additional results are available in our anonymous codebase: https://anonymous.4open.science/r/Anonymous_STFlow-3616/.
```
The Introduction analyzes the issues of existing sc... | Summary: The paper introduces a scalable and efficient framework for predicting spatial transcriptomics from histology images. By integrating flow matching for progressive gene refinement, E(2)-invariant spatial attention for robust spatial modeling, and whole-slide scalability, STFlow formulates gene expression predic... | Rebuttal 1:
Rebuttal: Thanks for your feedback on our work! We will explain your concerns point by point.
```
The authors should perform a hyperparameter study on the number of refinement steps in the flow matching process.
```
Ans: We report STFlow’s performance across different numbers of refinement steps below. Th... | Summary: The authors propose STFlow, a model for spatially resolved gene-expression prediction from WSIs.
STFlow is based on flow matching, modelling the joint distribution of the full spatial gene-expression data across each WSI, through an iterative refinement process. This enables explicit modelling of spot-to-spot... | Rebuttal 1:
Rebuttal: We greatly appreciate your valuable suggestions and will revise the manuscript accordingly. Due to character limits, all additional results are available in our anonymous codebase: https://anonymous.4open.science/r/Anonymous_STFlow-3616/.
```
I found it quite difficult to follow and understand th... | null | null | null | null | null | null |
Heterogeneous Treatment Effect in Time-to-Event Outcomes: Harnessing Censored Data with Recursively Imputed Trees | Accept (poster) | Summary: The authors consider the problem of estimating an individual treatment effect from survival data, where some of the observations may be right-censored. They propose a two-stage approach, MISTR, for this problem. In the first stage, they use recursively imputed survival trees (RIST) to impute the censored survi... | Rebuttal 1:
Rebuttal: Thank you very much for the detailed review, constractive comments, and your support for the acceptance of our paper. The following is our point-by-point response.
* The max time for RIST and the RMST horizon are very close, so it seems that the imputation procedure should have a very minimal ... | Summary: The authors propose a tree-based method for estimating Heterogeneous Treatment Effects (HTE) in survival analysis, and further extend it by incorporating instrumental variables to account for unobserved confounders. The authors conduct thorough and detailed experiments on both synthetic and real-world datasets... | Rebuttal 1:
Rebuttal: Thank you very much for your thorough review and insightful feedback.
* Why does replacing censoring rate estimation in methods such as CSF with the conditional
survival distribution proposed by RIST mitigate the impact of extreme cases?
Thank you for raising this important issue. It is well... | Summary: The paper proposes MISTR—a novel non‐parametric approach for estimating heterogeneous treatment effects (HTE) in time-to-event (survival) data, where right censoring is prevalent. MISTR tackles censoring by employing multiple imputations through Recursively Imputed Survival Trees (RIST) to generate several “co... | Rebuttal 1:
Rebuttal: Thank you very much for the careful review, thoughtful comments, and your support for the acceptance of our paper. The following is our point-by-point response.
* Explain the equations presented in the main body of the paper.
We apologize for this omission. The revised version of the paper will ... | null | null | null | null | null | null | null | null |
Learn Singularly Perturbed Solutions via Homotopy Dynamics | Accept (poster) | Summary: This paper introduces homotopy dynamics as a strategy to solve PDEs with sharp interfaces using PINNs. The key idea is to start training with a larger interface width parameter $\epsilon$ (corresponding to a smoother solution), then gradually decrease $\epsilon$ to the desired sharp-interface regime. This appr... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful review and constructive suggestions. Below, we address the concerns you raised.
- **Why do we use PINN for this type of problem instead of well-established numerical methods?**
Neural networks offer strong approximation capabilities [1] and help mitigate the curse... | Summary: In this paper, the authors present a training based method based on homotopy dynamics for handling sharp interface problems. The authors provide a proof of the convergence of the Homotopy dynamics for stable training. The experiments results demonstrate that the proposed method can help capture the sharp inter... | Rebuttal 1:
Rebuttal: Thank you for your valuable suggestions on our paper.
First, We would like to emphasize the main focus and contribution of our work. This paper addresses the core challenge of training neural networks to solve PDEs, particularly those involving sharp interfaces, where specific parameters in the ... | Summary: The authors look at the physics-informed neural network (PINNs) setting of solving a PDE via minimizing the PDE residual. They look at cases where there are “sharp” interfaces (introducing near singularities). They propose a method based on homotopy dynamics, which involves starting with an easier to learn pro... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and valuable suggestions.
First, we would like to emphasize the main motivation and contribution of our work. **From a theoretical perspective, we analyze how such parameters affect training convergence speed**. To overcome these difficulties, we propose a **ho... | Summary: This paper proposes Homotopy Dynamics to train neural network for solving sharp interface problems. For sharp interface problems, the parameter $\epsilon$ in the PDE affects the singularity of the solution. As $\epsilon \to 0$, the PDE becomes increasingly singular and thus the solution is difficult to compute... | Rebuttal 1:
Rebuttal: Thank you for your careful reading and valuable suggestions. Our work addresses the core challenge of training neural networks to solve PDEs with sharp interfaces, where small parameters introduce near-singularities that hinder optimization. **From a theoretical perspective, we analyze how such pa... | null | null | null | null | null | null |
A Theory for Conditional Generative Modeling on Multiple Data Sources | Accept (poster) | Summary: This work analyzes the effect of training with multiple data sources on conditional generative models. The authors establish a bound on the total variation distance between true and model distributions in terms of the bracketing number.
## update after rebuttal
The other reviews and the rebuttal have increas... | Rebuttal 1:
Rebuttal: # Experimental suggestion: Error estimate for FID scores
We thank the reviewer for the valuable suggestion regarding error estimation.
We would like to clarify that the real-world experiments in Section 5.2 were run only once due to the long training time. Based on these trained models, we additi... | Summary: This paper takes the first step toward a rigorous analysis of multi-source training in conditional generative modeling, where each condition represents a distinct data source. Specifically, the article establishes a general distribution estimation error bound in average total variation distance for conditional... | Rebuttal 1:
Rebuttal: # Q1: Intuitive explanation for upper bracketing number
The $\epsilon$-upper bracketing number is a notion to quantify the complexity of an infinite set of functions. The key idea is to construct a finite collection of "brackets" that enclose every function in the set within a small margin.
To i... | Summary: This paper investigates conditional generative models with multiple data sources. It establishes a general upper bound on the MLE error. The theoretical result is then specialized to conditional Gaussian distributions, autoregressive models, and energy-based models. Finally, the theoretical findings are valida... | Rebuttal 1:
Rebuttal: # Q1: Characterization of multi-source advantage
We thank the reviewer for the insightful comment.
We would like to clarify that the advantage of multi-source training is indeed measured by the model parameter sharing, while the degree of the model parameter sharing reflects the source distribut... | Summary: This paper provides a theoretical framework proving that training conditional generative models on multiple data sources outperforms single-source training when sources share similarities. The authors instantiate their theory across Gaussian distributions, autoregressive models, and energy-based models, demons... | Rebuttal 1:
Rebuttal: # Q1: Close alignment in Figure 1
We appreciate the reviewer’s careful examination of Figure 1.
As detailed in lines 339-340 of our submission (the caption of Figure 1), the empirical and theoretical values are plotted on **separate vertical axes**: empirical values correspond to the left axis, ... | Summary: The paper establishes a distribution estimation error bound in average total variation distance for conditional maximum likelihood estimation. The main result is based on the bracketing number; it shows that when source distributions share certain similarities and the model is expressive enough, multi-source t... | Rebuttal 1:
Rebuttal: # Q1: Experiments for ARMs or EBMs
We thank the reviewer for the valuable comment.
Following the reviewer's suggestion, we have conducted supplementary simulations for ARMs according to the formulation in Section 4.2. Experimental settings and results are presented below. Generally speaking, th... | null | null | null | null |
Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning | Accept (spotlight poster) | Summary: This paper investigates the role of image token continuity in Vision Transformers (ViTs) for Cross-Domain Few-Shot Learning (CDFSL). The authors observe that disrupting token continuity (e.g., shuffling patches or perturbing frequency components) significantly degrades source-domain performance but only margin... | Rebuttal 1:
Rebuttal: Thank you for your thoughtful feedback and constructive suggestions. Below are our responses to your concerns:
## **1. More Proofs for Continuity**
Our method is consistent with the proofs of the previous works [1-3] in that the small patterns are easier to transfer than larger ones. But we diff... | Summary: This paper investigates the role of image tokens' continuity in Vision Transformers for Cross-Domain Few-Shot Learning. The authors identify an interesting phenomenon: disrupting the continuity of image tokens significantly affects performance in source domains but has only a marginal impact on target domains ... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and constructive feedback. Below are our detailed responses to your concerns:
## **1. Justification of Clustering Hyperparameters**
The clustering threshold in Eq. 16 controls the **granularity of patch grouping** in the frequency domain. Higher threshold value... | Summary: This article explores the impact of image token continuity on model performance in the context of cross-domain few-shot learning (CDFSL). The study demonstrates that disrupting image token continuity can reduce the gap between the source and target domains to some extent, thereby improving the model's generali... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and constructive feedback. Below are our detailed responses to your concerns:
## **1. Classification Values in Figure 1**
The numerical values in Fig.1 are summarized below:
| Disruption Method | Source Acc. | Target Acc. |
| :---------------------- | :... | Summary: This paper provides a novel perspective to improve the performance cross-domain few-shot learning (CDFSL). The key insight is that disrupting the continuity of image tokens in ViT will force the model to learn smaller patterns which are more easily transferred under extreme domain gaps. The observation is inte... | Rebuttal 1:
Rebuttal: Thank you for your constructive feedback and valuable suggestions. Below are our responses to your comments:
## **1. Source domain performance**
As suggested, we have provided the performance on the source domain (miniImageNet) for reference.
| Model | Source Domain | Target Domain |
| ----... | null | null | null | null | null | null |
Towards Understanding Gradient Dynamics of the Sliced-Wasserstein Distance via Critical Point Analysis | Accept (poster) | Summary: The paper studies the existence and stability of critical points for semi-discrete sliced Wasserstein loss functions. In particular, they prove that there exist critical points which do not coincide with the global minimum, but also that any of those critical points is unstable under small perturbations. The a... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback and relevant questions.
- Concerning the limitations of the analysis in Section 5 : it is indeed true that most of the examples we discuss are in 2D, with the exception of our Proposition 5.1(b) which gives some examples of critical points in arbitrary dimensi... | Summary: This paper proposes a systematic study of the Sliced-Wasserstein functional in an optimization context where the target measure is continuous and the measure to optimize is discrete. The authors use the notion of ‘Lagrangian critical point’ of a functional defined over probability measures, since it aligns wel... | Rebuttal 1:
Rebuttal: Thank you for your relevant remarks and questions.
- Regarding the possible generalizations of our results to other types of optimized or target measures (notably discrete ones), we refer to our answer to reviewer BmZL which addresses these points extensively. Note in particular that even though ... | Summary: The paper under consideration presents theoretical results regarding (Section 3) properties of discrete gradient descent w.r.t. SW functional (with absolute continuous target measure) and (Section 4) properties of (lagrangian) critical points of SW functional; (Section 5) provides some examples of lower-then-d... | Rebuttal 1:
Rebuttal: Thank you for your positive remarks and comments. Regarding the questions you raised :
- We meant by "descent lemma" a result that guarantees that, provided some conditions on the initial point and the step size are satisfied, one step of the gradient descent will decrease the loss. In particular... | Summary: This paper investigates the properties of gradient flows for the Sliced-Wasserstein (SW) distance when used as an objective functional. It rigorously develops different notions of critical points—Eulerian, Wasserstein, and Lagrangian (including a barycentric variant)—and studies the convergence and stability p... | Rebuttal 1:
Rebuttal: Thank you for your positive remarks and relevant questions. Even though we restricted our theoretical analysis of the $SW$ distance to absolutely continuous target measures (and in Section 3 to discrete approximating candidates), this was mostly for the sake of simplicity, and it is indeed possibl... | null | null | null | null | null | null |
An Efficient Private GPT Never Autoregressively Decodes | Accept (poster) | Summary: The authors mainly aim to improve the efficiency of private inference for autoregressive language models. First, they observe that the decoding time is relatively insensitive to the input length. Next, they adapt speculative decoding to the private inference setting. The authors employ a small public model as ... | Rebuttal 1:
Rebuttal: Thank you for the positive feedback and appreciation of our work. We appreciate your insights and would like to provide more clarification.
# Question 1: Figure 1&2 claim that the decoding time is relatively insensitive to the input length. However, for softmax, the cost increases drastically as t... | Summary: The paper proposes an efficient method for secure inference in generative pre-trained transformer (GPT) models by replacing the traditional autoregressive secure decoding process with a Public decOding and Secure verificaTion (POST) approach. The POST method leverages publicly available GPT models to generate ... | Rebuttal 1:
Rebuttal: Thank you for the reviewer's appreciation of our efforts and valuable feedback on our paper. We address the main concerns as follows.
# Question 1: The approach relies on the acceptance ratio achieved by the public model; if the alignment is suboptimal, the speedup may diminish.
Our experiments de... | Summary: This paper focus on secure inference on GPT, and presents POST, which contains (1) a private sampling protocol optimized for cryptographic primitives and (2) model alignment using knowledge distillation to speedup the secure inference. Experiments demonstrate speedup compared to standard decoding across three ... | Rebuttal 1:
Rebuttal: Thank you for your thorough examination and thoughtful feedback on our paper. We address the main concerns as follows.
# Question 1: In Appendix D, the authors try to prove the division can be refactored as multiplication. However, it is not stated what is the range of p(x)/q(x). Therefore, I am n... | Summary: To accelerate privacy-preserving inference, the authors propose a Public Decoding and Secure verificaTion (POST) approach that utilizes public GPT models, based on the observation that securely decoding one token vs. multiple tokens takes a similar latency.
Since the efficiency of secure decoding depends on th... | Rebuttal 1:
Rebuttal: We are grateful for the reviewer's appreciation of our efforts. Below, we respond to your constructive comments in detail.
# Question 1: Could the author clarify the ambiguity in Line 90 in Introduction and the Line 27 in abstract?
Thank you for pointing out the potential ambiguity. We clarify it ... | null | null | null | null | null | null |
Locality-Sensitive Hashing for Efficient Hard Negative Sampling in Contrastive Learning | Reject | Summary: This paper addresses the computational challenge of efficiently finding high-quality hard negative (HN) examples in large, high-dimensional datasets for contrastive learning. The authors propose a novel GPU-friendly Locality-Sensitive Hashing (LSH) technique which projects the input vectors to binary buckets, ... | Rebuttal 1:
Rebuttal: 1. Thank you for your suggestion. We agree that including a recall-QPS trade-off chart for our LSH approaches would provide additional clarity. We will replace Figure 1 with this graph. While Figure 1 was originally intended to illustrate processing time from a dataset size perspective by comparin... | Summary: This paper explores hard negative (HN) sampling in contrastive learning and proposes a Locality-Sensitive Hashing (LSH)-based Approximate Nearest Neighbor (ANN) approach to improve computational efficiency while maintaining competitive performance. The proposed method enables fast and efficient pre-epoch HN se... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and will address them accordingly:
1. Thank you for the suggestion, we will address this and structure the paper more clearly by, for example, removing redundancies and improve citation clarity as suggested by udPs.
2. See 1.
3. See 1.
4. LSH has the advanta... | Summary: This paper introduces a Locality-Sensitive Hashing-based method for efficient Hard Negative sampling in contrastive learning. This method converts feature embeddings into binary representations, which enables fast approximate nearest neighbor searches.
Claims And Evidence: Most of the claims are supported by ... | Rebuttal 1:
Rebuttal: We thank the reviewer for the comments and will address them accordingly:
1. Yes we agree with Reviewer XybM, that investigating dynamic bit sizes can be interesting. Therefore we already explored this idea on both the VIGOR dataset and MS MARCO. We started with 8 bits and gradually increased to ... | Summary: The paper addresses the efficient sampling of hard negatives in contrastive learning by introducing an approximate nearest neighbor method based on Locality Sensitive Hashing (LSH). This method quantizes real-valued feature vectors into binary representations for approximate nearest neighbor search, thereby re... | Rebuttal 1:
Rebuttal: We thank the reviewer for their comments, and begin by addressing them below:
1. We agree with the reviewer that LSH has been extensively studied in the context of approximate nearest neighbor search. However, its application in contrastive learning, especially for hard negative sampling during tr... | Summary: This paper proposes to use locality-sensitive hashing to extract nearest neighbors as hard negatives when performing contrastive learning to train sentence embedding. The authors perform experiments to verify that the proposed method can achieve almost the same embedding quality and that the search time runs a... | Rebuttal 1:
Rebuttal: We thank the reviewer for his insightful comments. To address the concern about our motivation, we measured the training times per epoch (averaged over 3 epochs) using a DGX-2 on different datasets and compared these to the pre-epoch HN search times shown in Figure 1 and the search time with 128 b... | null | null | null | null |
Modification-Considering Value Learning for Reward Hacking Mitigation in RL | Reject | Summary: The paper proposes a novel value learning RL algorithm intended to reduce the probability of having agents developing reward hacking (i.e., unsafe and unintended behavior due to non-optimal definition of the reward function). The paper proposes two variations of their algorithm, which in high-level basically u... | Rebuttal 1:
Rebuttal: Thank you for your detailed review and insightful questions. We appreciate the chance to clarify MCVL's mechanism, assumptions, and evaluation. We will revise the paper to make these points clearer.
> For the method, I couldn't really get why the proposal would avoid reward hacking. [...] how wou... | Summary: Traditional Reinforcement Learning (RL) agents often demonstrate reward hacking, which is defined as the ability to maximize rewards without providing the desired outcomes. The paper studies reward hacking in RL by using General Utility in order to learn and update utility functions at the trajectory level. In... | Rebuttal 1:
Rebuttal: Thank you for your review and valuable suggestions. We address your points below and will incorporate the responses into the paper.
> Learned Modifications: [...] But wouldn't the new updated policy with fresh T trajectory samples always be better? [...] comparing policy discrepancies is [not] a ... | Summary: This paper studies how to mitigate reward hacking by considering the change of trajectory utilities. The agent is initially trained in a Safe environment in which exploiting the reward leads to the intended behavior, and then continued in a Full environment with different dynamics/rewards. The paper claims tha... | Rebuttal 1:
Rebuttal: Thank you for your thorough review and valuable suggestions. We would like to clarify several points and will incorporate these clarifications in the paper.
> The paper claims that MCVL iteratively refines the initial coarse utility function, but it seems that the experiment results cannot reflec... | Summary: This paper addresses the problem of reward hacking by framing it within the General Utility Reinforcement Learning (GU-RL) framework. The authors introduce trajectory value functions and a mechanism for explicit utility inconsistency detection. Their proposed utility update technique can be integrated into sta... | Rebuttal 1:
Rebuttal: Thank you very much for your thorough review and valuable suggestions. We are happy to answer your questions and will incorporate all answers in the final manuscript.
> How robust is MCVL to inaccuracies in the learned transition model? For example, if the learned model deviates slightly from the... | null | null | null | null | null | null |
The Harder Path: Last Iterate Convergence for Uncoupled Learning in Zero-Sum Games with Bandit Feedback | Accept (poster) | Summary: This paper improves the lower bound for learning matrix games with bandit feedback and last-iterate convergence in the uncoupled setting from $O(T^{-1/2})$ to $O(T^{-1/4})$. The authors then propose a black-box reduction from an algorithm with the so-called “output convergence” to the last-iterate convergence,... | Rebuttal 1:
Rebuttal: We thank reviewer TkNJ for taking the time to review our article and for the clarity suggestion. We address the concerns below.
>Nevertheless, I think there is no fundamental difference between the definitions of “output convergence” and the average-iterate convergence.
We preferred to state the... | Summary: This paper studies the last-iterate convergence rates of uncoupled learning dynamics in two-player zero-sum games with bandit feedback. One of the main contributions of the paper are lower bounds for uncoupled learning dynamics: (1) $\Omega(t^{-1/(2+p)})$ lower bound for any-time $\ell_p$ last-iterate converge... | Rebuttal 1:
Rebuttal: We thank Reviewer eajt for the review and for pointing out a result that we missed in one of our references.
>I think related works have been substantially discussed. The only comment I have is regarding the results in [1]. For uncoupled learning in zero-sum games, there are two results in [1] fo... | Summary: The paper aims to improve the upper and lower bounds in Last Iterate Convergence for Uncoupled Learning in Zero-Sum Games with Bandit Feedback
Claims And Evidence: The evidence is fairly clear, but I will provide a detailed explanation of my questions in the following sections.
Methods And Evaluation Criter... | Rebuttal 1:
Rebuttal: We would like to express our sincere gratitude to Reviewer U4sa for their very comprehensive and constructive feedback. We appreciate the encouragement, particularly the acknowledgement of our "honest and transparent presentation of results".
We prepared a response for all of the questions, but ... | Summary: This paper studies a zero-sum matrix game in which two players repeatedly select stochastic policies, sample actions, and receive stochastic losses—without access to the underlying payoff matrix—that depend on their joint actions. The authors aim to develop an uncoupled algorithm that independently controls ea... | Rebuttal 1:
Rebuttal: We thank Reviewer bHVB for taking the time to review our submission and especially for pointing out the issues in the notations. We address the concerns below.
>It is concerning that the abstract claims both proposed algorithms achieve the optimal convergence rate. Could the authors kindly clarif... | null | null | null | null | null | null |
Linear Bandits with Partially Observable Features | Accept (poster) | Summary: The paper proposes a method to solve the linear bandit problem with only a subset of features per arm visible to the learner without assuming any structural property beforehand. The authors do this by transforming each context vector onto an augmented space with dimension $K$ and learn the respective augmented... | Rebuttal 1:
Rebuttal: We are thankful for your careful review and for acknowledging the impact of our contributions. We will address the following questions one by one, and we believe that the answers will collectively provide a comprehensive response.
* **On Use of Compatibility Condition**
* It is well known tha... | Summary: ### Problem Setting
The paper studies the **linear bandit problem with partially observable features**, where rewards depend on a full set of features, but the learner only observes a subset of them. This setting models real-world scenarios (e.g., recommendation systems) where unobserved latent features (e.g... | Rebuttal 1:
Rebuttal: We appreciate your feedback on page usage. We will make full use of the page limit and avoid leaving unused space in the revision.
* **On Explanation and Mislocation of Table 1**
* We appreciate you for giving us an opportunity to clarify Table 1. We will relocate the table to a more suitabl... | Summary: This paper studies linear bandits with partially observable features. The authors suggest an epsilon greedy type algorithm based on a doubly robust estimator. A regret bound of the algorithm is provided with supporting numerical experiments.
Claims And Evidence: The authors claim that the basis of orthogonal ... | Rebuttal 1:
Rebuttal: We are grateful for your valuable feedback and acknowledgment of our contributions. We are happy to address each of your comments.
* **On Descriptions of Notation and Methods**
- Thank you for the opportunity to clarify this point. Definitions of $p$ and $\delta$ are given after Eq. (9) (line... | null | null | null | null | null | null | null | null |
From Local Details to Global Context: Advancing Vision-Language Models with Attention-Based Selection | Accept (poster) | Summary: The paper introduces **Attention-Based Selection (ABS)**, a method to enhance vision-language models (VLMs) like CLIP by addressing limitations of random cropping, which often introduces background noise and compromises global semantic understanding. ABS leverages **DINO's attention maps** to guide cropping in... | Rebuttal 1:
Rebuttal: **Q1**: Experimental details in Table 3.
**A1:** Thanks. The results of baselines in Table 3 follow WCA protocol: Tuning-based methods are 16-shot source-trained and target-evaluated for OOD generalization. Notably, our method requires no fine-tuning, operating in a zero-shot manner on both sourc... | Summary: Recent studies have explored the use of multiple image crops obtained through random cropping, utilizing text descriptions generated by LLMs to assess the similarity between image and text embeddings for zero-shot classification tasks. This paper builds on that concept by addressing the noise caused by random ... | Rebuttal 1:
Rebuttal: **Q1:** Claims And Evidence.
**A1:** Thanks. The purpose of employing cropping is to enable the model to focus on local features of objects, thereby achieving better alignment with certain LLM descriptions. However, random crop exhibits two inherent limitations: When using smaller crop sizes, it ... | Summary: The paper proposes an Attention-Based Selection (ABS) method to improve zero-shot classification and out-of-distribution generalization capabilities of vision-language models (VLMs). ABS leverages DINO’s attention maps to guide the cropping of images, thus preventing random crops from focusing on irrelevant ba... | Rebuttal 1:
Rebuttal: **Q1:** About contribution.
**A1:** Thanks. Our core contribution resides not in proposing incremental adjustments to established frameworks, but rather in advancing systematic methodologies that demystify stochastic factors while enhancing holistic semantic comprehension for this research domain... | Summary: This paper introduces ABS, a training-free Attention-Based Selection method that uses Vision-Language Pretraining (VLP) model’s attention maps (e.g., DINO and CLIP) to guide cropping in both raw image and feature space, effectively integrating local details with global semantic context via soft matching to ach... | Rebuttal 1:
Rebuttal: **Q1:** The concerns in Claims And Evidence.
**A1:** Thanks.
1. The table below compares applying soft matching alone vs. combined two selections to the baseline. Combined with the ablations in our paper, it shows that individual components yield improvement when used independently, but integrati... | null | null | null | null | null | null |
Hypothesis Testing for Generalized Thurstone Models | Accept (poster) | Summary: In this paper, the authors look at the problem of hypothesis testing of generalized thurstone models. The later models are used to model ranking among several entities based on pairwise comparisons. Extensive research has been done to learn the parameters of such a model. However, an important questions is giv... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback and for dedicating time to review our paper! We respond to the specific questions as follows:
**Testing problem with respect to other norms**: We believe the Frobenius norm is a natural choice for our problem as it allows tractable analysis of maximum-likeli... | Summary: Covers “Generalized Thurstone models,” in which each player has a utility, and the probability of winning is a function of the difference in utilities – but this choice function is an arbitrary CDF. Nice motivation to observe that GTMs do not capture certain types of choice dynamics, so asking whether data are... | Rebuttal 1:
Rebuttal: Thank you for detailed feedback. Due to 5000 character limit, our responses are concise. We will incorporate the addressed points and new experiments in the final version
**Experimental Concerns**
* Experiments not explored in detail: Our focus was on theoretical aspects of the testing problem. T... | Summary: This work develops a hypothesis testing framework to determine whether pairwise comparison data follows a generalized Thurstone model for a given choice function, introducing a minimax separation distance to quantify deviations from such models. The study establishes theoretical bounds on the critical threshol... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback and for dedicating time to review our paper! We respond to the specific questions as follows:
**Intuitive explanation for definition of test statistic**: In addition to our existing discussion after Eq 15 in the paper, we provide here with an additional intu... | Summary: The paper addresses the problem of hypothesis testing for whether a given pairwise comparison dataset follows a Generalized Thurstone Model (GTM), which is formally stated in equation (12). It proposes a test statistic along with a corresponding testing threshold that matches the lower bound on the critical th... | Rebuttal 1:
Rebuttal: Thank you for your insightful feedback and for dedicating time to review our paper! We respond to the specific questions as follows:
**Comparisons to other methods for testing of Thurstone models**: Minimax testing for generalized Thurstone models (for a fixed choice function $F$) has not been st... | null | null | null | null | null | null |
Enhancing Spectral GNNs: From Topology and Perturbation Perspectives | Accept (poster) | Summary: This paper proposes a higher-dimensional sheaf Laplacian matrix based on perturbation theory and the theory of cellular sheaves. The perturbation is controlled from the block form of the normalized graph Laplacian matrix, and can contain more distinct eigenvalues. The paper provides theoretical analyses on the... | Rebuttal 1:
Rebuttal: Thank you for your valuable suggestion.
**Response to Q1**
We apologize for any confusion caused. To clarify: Table 1 demonstrates the performance gains achieved by integrating PSL into various models, while Table 2 highlights the comparison between PSL, GSL, and the conventional normalized grap... | Summary: This paper claims that the presence of repeated eigenvalues limits the expressive power of spectral GNNs. To address this issue, this paper proposes perturbed sheaf Laplacian, which achieves optimal model performance due to its more distinct eigenvalues.
Claims And Evidence: The occurrence of repeated eigenva... | Rebuttal 1:
Rebuttal: Thank you for your valuable suggestion.
**Response to Comment 1 in Experimental Designs or Analyses**
We reconducted the experiments by adding four new datasets (10 in total). All baseline algorithms use the best parameter settings in their original paper. Additionally, we also tuned the parame... | Summary: This paper aims to solve the repeated eigenvalues of graph Laplacian by proposing a novel perturbed sheaf Laplacian (PSL). The authors claim that PSL can increase the number of distinct eigenvalues and improve the expressive power of spectral GNNs. Experiments on the node classification task validate the effec... | Rebuttal 1:
Rebuttal: Thank you for your valuable suggestion.
**Response to C1**
We apologize for the confusion.
Briefly, the Laplacian spectrum measures graph connectivity. Cheeger's inequality $2h_G > \lambda_1 > \frac{h_G^2}{2}$ shows that a larger spectral gap $\lambda_1$ implies better connectivity (More details... | Summary: This paper presents a novel solution to the repeated eigenvalue problem in Spectral GNNs.
Through the formal definition of cellular sheaf on graphs,
the paper formally introduces the definition of cellular sheaf,
which essentially specifies that when a signal with dimension $d$ propagates from node $i$ along... | Rebuttal 1:
Rebuttal: Thank you for your valuable suggestion.
**Response to the raised issues in Experimental Designs or Analyses**
We reconducted the experiments by adding four new datasets (10 total) and using a full-supervised split (60%/20%/20%), which follows Wang & Zhang’s (*How Powerful are Spectral Graph Neu... | null | null | null | null | null | null |
Token Coordinated Prompt Attention is Needed for Visual Prompting | Accept (poster) | Summary: This paper proposes a Token Coordinated Prompt Attention (TCPA) module to enhance the effectiveness of visual prompting in Vision Transformers (ViT). Existing methods use shared prompts for all tokens, overlooking the distinct roles of CLS and image tokens, leading to limited representational capacity. TCPA ad... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our **novelty**, **effectiveness** and **comprehensive experiments**.
(The images mentioned below are available at the anonymous link: https://anonymous.4open.science/r/ICML-2025-Paper35-Rebuttal-7E9E.)
### Q1: More Visualizations
1. Thank you for your valuable ... | Summary: The paper introduces Token Coordinated Prompt Attention (TCPA), a novel module for visual prompting in Vision Transformers (ViTs). TCPA assigns specific prompts to CLS and image tokens, enhancing their discriminative abilities through targeted attention interactions. It uses a matching function to dynamically ... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our **novelty**, **effectiveness** and **comprehensive experiments**.
(The images mentioned below are available at the anonymous link: https://anonymous.4open.science/r/ICML-2025-Paper35-Rebuttal-7E9E.)
### Q1: Cosine Distance-based Matching Function
1. We condu... | Summary: This paper proposes Token Coordinated Prompt Attention (TCPA) to enhance visual prompting for Vision Transformers. By disentangling and adaptively assigning prompts to different CLS and image tokens based on their distinct roles, this method effectively mitigates the limitations of conventional visual promptin... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our **motivation**, **effectiveness** and **comprehensive experiments**.
### Q1: Theorem
1. Theorem 4.1 and Theorem 4.2 mentioned in our paper are established theories from existing work, which we have *appropriately cited*.
2. In their original paper, these th... | Summary: This paper introduces a token-wise prompt termed as TCPA to enrich discriminative information of tokens by assigning specific prompts to each different tokens. As a plug-and-play strategy，TCPA can be seamlessly integrate with existing prompt based methods. Experiments show that TCPA can achieve consistent perf... | Rebuttal 1:
Rebuttal: Thank you for your appreciation of our **clearity**, **motivation** and **sound experiments**.
(The images mentioned below are available at the anonymous link: https://anonymous.4open.science/r/ICML-2025-Paper35-Rebuttal-7E9E.)
### Q1: Load Balance of Different Prompts
1. Since the number of CLS ... | null | null | null | null | null | null |
Lower Bounds for Chain-of-Thought Reasoning in Hard-Attention Transformers | Accept (poster) | Summary: This paper explores the use of CoT reasoning and scratchpads in enhancing the computational capabilities of transformers. The authors propose new lower bounds for the number of CoT steps required for various algorithmic problems, challenging optimistic bounds from circuit complexity.
Claims And Evidence: The ... | Rebuttal 1:
Rebuttal: We thank the reviewer for acknowledging that our claims are supported by clear and convincing evidence. We now address all the concerns mentioned in the review:
1. On the restrictiveness of our approach due to the use of hard attention and binary input/output format:
We agree that hard atten... | Summary: This paper establishes lower bounds on the required chain-of-thought (CoT) length that unique hard-attention (UHAT) transformers need for solving certain classes of problems.
In particular, lower bounds are established for PARITY ($\Omega(N)$), MULTIPLICATION ($\Omega(N)$), MEDIAN ($\Omega(N)$), and REACHABILI... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive assessment of our paper. We particularly appreciate the reviewer's point that complexity lower bounds, as shown in this paper, tend to be harder to prove than upper bounds.
If possible, we'd like to kindly ask the reviewer to provide more justification suppo... | Summary: This paper establishes that hard-attention transformers require chain-of-thought (CoT) sequences of length linear in the input size to solve high-sensitivity algorithmic tasks like MEDIAN, and REACHABILITY in layered DAGs, with bounds tight up to logarithmic factors. By leveraging sensitivity analysis and a no... | Rebuttal 1:
Rebuttal: We thank the reviewer for the positive assessment of our experiments and theorems and pointing out the strengths of our paper. We now address the concerns raised in the review.
1. In Appendix A, we explain in detail why our results for hard attention are relevant for real-world transformers, espe... | null | null | null | null | null | null | null | null |
CALM: Consensus-Aware Localized Merging for Multi-Task Learning | Accept (poster) | Summary: The authors introduce a novel model merging approach called CALM to address multi-task learning integration. The core idea involves identifying localized parameters aligned with global task consensus through three key components:
1. Class-Balanced Entropy Minimization Sampling (CB-EMS): A method to extract rel... | Rebuttal 1:
Rebuttal: ***Question 1: Does the Efficient-Aware Framework apply to existing model merging methods?***
***Answer:*** Thanks for the inspiring question.
The Efficient-Aware Framework (EAF) introduces a novel serialized merging approach for model merging, which impacts existing methods as follows:
- **For... | Summary: The paper introduces a test-time adaption method named CALM, which optimizes an equally-sized mask on pre-fine-tuned task-specific models through a reliable unsupervised dataset. The mask aims to extract locally shared parameters with global task consensus, offering a new perspective for model merging by explo... | Rebuttal 1:
Rebuttal: ***Question 1: The sampling rates of CB-EMS in experiments appear inconsistent; how would this be resolved in practical applications?***
***Answer:*** Thanks for the practical comments. We would like to explain it as follows:
- **CB-EMS Sampling Rates Show Cross-Task Generalizability.**
Exper... | Summary: This paper focuses on model merging in multi-task learning, aiming to identify locally shared information with global task consensus while addressing existing limitations of parameter conflict in global information and diluted local features during merging.
The method comprises three components: 1) Class-Bala... | Rebuttal 1:
Rebuttal: ***Question 1: The authors need to provide a clearer definition and explanation of global task consensus to establish the core idea of this work.***
***Answer:*** Thanks for your valuable question. We clarify the concept of global task consensus as follows:
- **Definition:** Global task consens... | Summary: Localized Information with Global Consensus: CALM proposes a method to extract localized parameters that align with global task consensus, ensuring that the merged model maintains effectiveness across all tasks.
* A new sampling technique that leverages unsupervised data more effectively by balancing class re... | Rebuttal 1:
Rebuttal: ***Question 1: A more formal theoretical analysis of the effectiveness of localized information with global consensus***
***Answer 1***: Thanks for the inspiring question. We perform a theoretical analysis of the following three aspects based on error.
- **Localized information reduces interfer... | null | null | null | null | null | null |
Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors | Accept (poster) | Summary: Understanding the behavior of complex black-box models has always been a challenge since the inception of deep neural networks (DNNs). This problem has worsened after the introduction of large language models (LLMs). In this paper, the authors investigate whether understanding the internal causal mechanisms of... | Rebuttal 1:
Rebuttal: We are grateful that the reviewer found our perspective on correctness prediction novel, and we appreciate the opportunity to discuss the practical value of our work!
---
### Q1. Generalizability to complex real-world tasks
We believe it's helpful to take a step back to examine the source of gen... | Summary: This paper presents a method for estimating whether a language model’s output is correct by examining “causal” internal representations. It studies both symbolic tasks and open-ended tasks, finding that features which directly mediate model behavior are more reliable than simpler methods. The authors claim tha... | Rebuttal 1:
Rebuttal: We very much appreciate the reviewer’s suggestion to clarify what is distinctively *causal* about our methods, and also how what we are doing relates to important neighboring areas of search, like **causal discovery** and **causal representation learning (CRL)**.
We see this as an opportunity to ... | Summary: This paper focuses on correctness prediction of large language models. It separates internal features into causal features and background features and suggests two approaches for predicting the correctness of model outputs. In one, permutations are used to determine whether predictions are robust against chang... | Rebuttal 1:
Rebuttal: We thank the reviewer for the thoughtful feedback! We are especially glad that the reviewer recognizes our work as the first to address correctness prediction in LLMs from a causal perspective.
---
---
### Q1. Comparison with strong baselines in correctness estimation
A major concern is that
> t... | Summary: This paper investigates the use of internal causal mechanisms within language models (LMs) to predict the correctness of their outputs. Rather than relying on traditional confidence scores or heuristic probing of internal activations, the authors propose two methods grounded in causal interpretability, Counter... | Rebuttal 1:
Rebuttal: We thank the reviewer for their insightful comments! We are encouraged that they recognized the novelty of our methods, the strength of our empirical results, and the significance of bridging the explanatory and predictive aspects of interpretability analysis.
---
---
### Q1. Clarification on the... | Summary: For this paper the authors try to identify internal features of LLM that mediate causal effects on the final prediction output. In settings where LLM predictions align with the actual real-world causal process, causal features are assumed to resemble the ground-truth mechanism and, therefore, allow for predict... | Rebuttal 1:
Rebuttal: We thank the reviewer for their detailed feedback! We are glad they found our proposed methods “well formalized” and “soundly incorporate the notion of causality” and appreciated our task design, metrics, and supporting results.
---
### Q1. Experimental Details
We provide detailed experimental se... | null | null | null | null |
EnIGMA: Interactive Tools Substantially Assist LM Agents in Finding Security Vulnerabilities | Accept (poster) | Summary: This paper presents EnIGMA, an LM agent enhanced with Interactive Agent Tools (IATs) to solve CTF challenges, achieving state-of-the-art results. Their experiments use 390 challenges from diverse benchmarks to evaluate EnIGMA with different LLMs. They also provided several ablation studies and analyses to demo... | Rebuttal 1:
Rebuttal: Thank you so much for your time and consideration. You’ve brought up excellent points in your feedback that we address below.
**Q1: Previous best methods comparison - does the superiority come from the better LLMs or the proposed framework?**
To address your concern, we present agent performance... | Summary: The paper proposes EnIGMA, an LM agent designed for CTF challenges. EnIGMA is built based on SWE-agent for code generation, which is based on the ReAct framework. On top of SWE-agent, EnIGMA incorporates actions and tools specially designed for the CTF challenges, including a debugger and a remote connection s... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their feedback on our work and for finding our results convincing. We address each of your concerns below.
**Q1: Novelty of the method and whether the performance gain due to the agent workflow design or harnessing the power of the tools.**
In EnIGMA, we are t... | Summary: The paper describes a new and improved agent for solving computer security Capture the Flag challenges.
Claims And Evidence: Mostly. There is one claim about interactive tools that I think is overstated (see Other comments below). The evaluation of leakage has some limitations.
Methods And Evaluation Crite... | Rebuttal 1:
Rebuttal: Thanks for finding our work helpful. Your thorough review of the paper and suggestions has been helpful to clarify several details.
**Q1: What "main shell" means. Does this mean a connection to a Unix shell (e.g., bash)? Or does this mean an agent running a REACT loop?**
“Main shell” is a connec... | Summary: This paper presents EnIGMA, an LM agent designed for autonomously solving Capture The Flag (CTF) challenges.
- The authors introduce Interactive Agent Tools (IATs), which enable the LM agent to execute interactive cybersecurity tools such as debuggers and remote server connection utilities. These tools addres... | Rebuttal 1:
Rebuttal: Thank you for your interest in our research and the acknowledgment of IATs as a meaningful addition to LM agents. We greatly appreciate your feedback and insights which will help us improve our work. We’ve addressed your concerns below:
**Q1: Data leakage analysis is incomplete + It is unclear wh... | null | null | null | null | null | null |
Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers | Accept (poster) | Summary: This paper proposed a quantization method named Q-VDiT tailored specifically for video Diffusion Transformers. The proposed Q-VDiT aims to address severe model quantization information loss in video models. Specifically, the authors proposed Token-aware Quantization Estimator to compensate for quantization err... | Rebuttal 1:
Rebuttal: We thank the reviewer for the constructive comments on our paper. Regarding the concerns, we provide the following responses.
> Q1: Optimization details.
Sorry for the misunderstanding, **we have reported optimization details including training data in Appendix Sec. B. and show the training cost ... | Summary: This paper addresses the issue of information loss and misalignment of optimization objectives that arise when applying existing quantization methods to video generation models. Current quantization techniques, which are primarily designed for image generation models, may not be directly suitable for video gen... | Rebuttal 1:
Rebuttal: Thank you very much for your high recognition of our work and the valuable suggestions you provided. Our response is as follows:
> Q1: Quantitative analysis on quantization error.
We add quantitative experiments on W3A6 model last layer weight's quantization error and information entropy mention... | Summary: The paper introduces Q-VDiT, a quantization framework for video DiT to reduce computational complexity while preserving video quality. It addresses two key challenges: quantization error compensation through a Token-aware Quantization Estimator (TQE) and spatiotemporal consistency via Temporal Maintenance Dist... | Rebuttal 1:
Rebuttal: Thank you for your detailed review of our work. Here are our responses to your concerns:
> Q1: Statement of our Q-VDiT.
We apologize for the misunderstanding caused by our statement. **We absolutely do not deny ViDiT-Q's contribution. ViDiT-Q is the first method to explore quantization for video ... | Summary: Diffusion transformers (DiT) are powerful for video generation but face deployment challenges due to large parameter sizes and high computational complexity. To tackle the issues of information loss and mismatched objectives during quantization, the authors propose Q-VDiT, introducing the Token aware Quantizat... | Rebuttal 1:
Rebuttal: Thank you for reviewing our manuscript and providing valuable suggestions. Here are our responses to some of the concerns you raised:
> Q1: Original model’s qualitative results.
We apologize for our negligence and we have released all original videos in [[https://anonymous.4open.science/r/Generat... | null | null | null | null | null | null |
Reinforced Learning Explicit Circuit Representations for Quantum State Characterization from Local Measurements | Accept (poster) | Summary: The paper introduces a novel approach termed "explicit circuit representations" for quantum state characterization. Unlike traditional implicit representations, this method allows for direct experimental reconstruction of quantum states. The representations are designed to predict quantum properties accurately... | Rebuttal 1:
Rebuttal: **1. Suggestion 1. Refinement of tables and plots.**
To improve readability, we will refine the tables and figures by moving the legends to the top or embedding them within the figures and abbreviating some metrics in tables. The figures in current paper is in PDF format thus can be scaled larger... | Summary: The paper uses a deep reinforcement learning algorithm to construct quantum circuits in a manner that avoids barren plateaus by not requiring gradients with respect to the circuit parameters. The approach appears to outperform alternative approaches, including VQE and QAQA.
A nice addition to the work is the ... | Rebuttal 1:
Rebuttal: **1. Suggestion 1. Refinement of figures.**
We will adjust the legend position and font size. Current figures are in PDF format and can be enlarged without loss of resolution.
**2. Q1. Benefits of transformer architecture. Capture of entanglement.**
Compared to simpler architecture like MLP, th... | Summary: This work develops explicit quantum state representations by generating surrogate preparation circuits through reinforcement learning. The approach uses a local fidelity reward function and a quantum measurement feature aggregation block that extracts global features from local measurement data. The paper atte... | Rebuttal 1:
Rebuttal: **1. W1 & Q1. The accuracy-sample cost tradeoff, estimation of density matrix fidelity.**
The impact of finite sampling (inaccurate expectation value estimation) to the accuracy has been discussed in Appendix G1.
We would like to clarify that while estimating density matrix fidelity is difficult... | Summary: This paper introduces QCrep, a novel reinforcement learning framework for quantum state characterization that generates explicit circuit representations rather than implicit neural encodings. The innovation lies in using local measurements from neighboring qubits to learn circuit descriptions that can faithful... | Rebuttal 1:
Rebuttal: **1. W1. Exploration in highly entangled states.**
Although fully reconstructing the target states for highly entangled states using local fidelity is generally difficult, our framework would work if the state characterization task is to estimate some properties of interests like correlations, wh... | null | null | null | null | null | null |
M+: Extending MemoryLLM with Scalable Long-Term Memory | Accept (poster) | Summary: The paper proposes an LLM memory-augmentation method called M+. By building on MemoryLLM, it improves the long-context understanding and information retention of a base LLM. They introduce what they call long-term memory vectors that are extracted by a trained retriever. M+ outperforms the base model and other... | Rebuttal 1:
Rebuttal: **Claims And Evidence:**
We would like to clarify that by “CPU offloading,” we specifically mean **offloading the memory vectors** present in each layer of the model. In our setup, each layer contains 12,800 memory vectors, and it is unnecessary to keep all of them simultaneously in GPU memory. I... | Summary: This paper presents M+, an enhanced memory-augmented language model that extends long-term memory retention beyond conventional limits. Building on MemoryLLM, M+ integrates long-term memory with a co-trained retriever. Extensive experiments across long-context understanding, question answering, and knowledge ... | Rebuttal 1:
Rebuttal: We are sincerely grateful for Reviewer v8Qd’s recognition of our work. Below, we address the reviewer’s questions in detail:
**[Q1] Direct Comparison between M+ and MemoryLLM: ** Thank you for raising this important point. In developing M+, we incorporated a number of significant improvements ov... | Summary: Memory model: memory pool (based on MemoryLLM) and a long-term memory with additional temporal information. Every time, the memory pool is updated, and a subset of tokens is dropped to the Long-term memory. For recall from memory, a small subset of the long-term memory vectors is retrieved according to the dot... | Rebuttal 1:
Rebuttal: We sincerely thank reviewer f5Uz for their recognition of the value of our work. We address the reviewer's concerns below:
**Relation To Broader Scientific Literature:**
To the best of our knowledge, following MemoryLLM, the most recent works on parametric memory include Titans [3] and Memory at... | Summary: **Main Findings:**
Equipping large language models (LLMs) with latent-space memory has gained significant interest, as it extends the effective context window of existing models. However, preserving and retrieving information from distant past contexts remains challenging. To address this, this paper proposes... | Rebuttal 1:
Rebuttal: **Essential References Not Discussed:**
Thank you for highlighting these important works. We will incorporate [1], [2], and [3] into our related work section. Specifically, [1] aligns with the core motivation behind incorporating memory into language models. Both [2] and [3] explore architectural... | null | null | null | null | null | null |
Task-Gated Multi-Expert Collaboration Network for Degraded Multi-Modal Image Fusion | Accept (poster) | Summary: In this paper, authors proposes a framework, TG-ECNet, for degraded multi-modal fusion by unifying restoration and fusion. The key design involves task-gated routing and expert collaboration. The paper conduct a series of experiments to demonstrate the effectiveness of TG-ECNet.
Claims And Evidence: Yes, the ... | Rebuttal 1:
Rebuttal: Thanks for recognizing our contributions in this work. We will reply to your questions in order.
---
## Response to the essential references
#### Concerning the references you suggested, we will incorporate them into our bibliography. Furthermore, our experimental configuration utilizes state-of... | Summary: The paper introduces Task-Gated Multi-Expert Collaboration Network (TG-ECNet), a novel framework designed to address the challenges of degraded multimodal image fusion. The key innovation lies in its task-gated router, which integrates degradation-aware gating in the encoder and fusion-aware gating in the deco... | Rebuttal 1:
Rebuttal: Thanks for recognizing our contributions in this work. We will reply to your questions in order.
---
## Response to the real-world experiments
#### We utilize real-world data from AWMM dataset to validate the robustness of our method in [Newfig2](https://anonymous.4open.science/r/TG-ECNet/NewFig... | Summary: This paper presents TG-ECNet, a unified framework that concurrently addresses restoration and fusion of degraded visible images (affected by noise, blur, and haze) and infrared images (with stripe noise) through a task-gated router and multi-expert collaboration mechanism. The proposed integration of restorati... | Rebuttal 1:
Rebuttal: Thanks for recognizing our contributions in this work. We will reply to your questions in order.
---
## Response to the selection of experts
#### In our work, we select the top 6 experts from 11 experts to cope with different degradations. In our paper, there is a typo in Line 218. We utilize de... | Summary: This paper introduces TG-ECNe, a novel framework designed to address the challenges of degraded multimodal image fusion. Multimodal images, such as visible and infrared images, often suffer from degradations like noise, blur, haze, and stripe noise, which negatively impact fusion quality. TG-ECNet tackles thes... | Rebuttal 1:
Rebuttal: Thanks for recognizing our contributions in this work. We will reply to your questions in order.
---
## Response to visualization results in combined degradation scenarios
#### In Fig.1, we’ve shown the performance of some methods like Text-if and DRMF in combined degradation scenarios. We use A... | null | null | null | null | null | null |
Understanding the learned look-ahead behavior of chess neural networks | Reject | Summary: The paper analyzes the behavior of a transformer model trained on chess games using activation patching, probing, and attention ablation, as originally proposed in [1]. The paper finds that the model considers up to the 7th future move when selecting the best next move, and its lookahead behavior is highly con... | Rebuttal 1:
Rebuttal: Thank you for your review highlighting concerns about the paper's distinctiveness from Jenner et al. and its self-containedness. We would make substantial revisions in an updated version:
1. **Novelty vs. Jenner et al.**: We would revise the introduction to more clearly articulate how our work di... | Summary: This paper builds on Jenner et al.'s work investigating the look-ahead capabilities of chess-playing neural networks, specifically the Leela Chess Zero policy network. The authors employ patching, probing, and ablation techniques to demonstrate that: 1) Chess models can consider moves up to 7 steps ahead, 2) M... | Rebuttal 1:
Rebuttal: Thank you for your positive review and helpful suggestions for improving clarity. We would implement all your recommended changes in a revised version of the paper:
1. **Peculiarities of model**: We would clarify in Section 2.1 that Leela originally takes in past board states in addition to the c... | Summary: This paper extends findings by Jenner et al (2024), which is a mechanistic interpretability paper examining how a chess network--specifically the Leela model, which has transformer architecture--"looks ahead" of game play by several moves. Specifically, the authors examine longer move sequences and possible br... | Rebuttal 1:
Rebuttal: Thank you for your detailed review. We address each point below:
1. **Regarding indirect effects in patching**: We agree that coordinated interventions could strengthen our conclusions, though in this work we restricted our focus to patching at the layer or head level. During our initial prelimin... | Summary: The authors use an existing technique for examining chess model internal states to expand the analysis of chess games to more complex positions.
Claims And Evidence: This paper has a common issue for interp papers, the authors don't make strong claims.
Of the three key contributions two (first and last) are ... | Rebuttal 1:
Rebuttal: Thank you for your review. We appreciate your feedback on our paper's contributions and their broader relevance.
We wish to clarify that our work makes several substantive contributions beyond Jenner et al. (2024):
1. While Jenner et al. showed evidence of look-ahead to the 3rd move, we demonstra... | null | null | null | null | null | null |
Algorithms with Calibrated Machine Learning Predictions | Accept (spotlight poster) | Summary: This paper introduces calibration as a tool to enhance learning-augmented algorithms, focusing on two problems: ski rental and online job scheduling. The authors propose algorithms that leverage calibrated predictions to achieve instance-specific competitive ratios, theoretically and empirically outperforming ... | Rebuttal 1:
Rebuttal: We’re glad that the reviewer shares our excitement about calibration as a novel and practical prediction concept for algorithms with predictions with lots of theoretical potential!
**“The expected performance bound for ski rental scales with max calibration error, which can be large, so the algor... | Summary: The paper studies how calibrated machine learning predictions can improve online algorithms. The paper focuses on two settings: the ski rental problem and online job scheduling.
In the ski rental problem, the predictions is about whether the skier will ski for more than a given threshold $b$ days using a cal... | Rebuttal 1:
Rebuttal: We appreciate the reviewer’s feedback and are glad that they value our high-level goal of using calibration as a tool to ensure the trustworthiness of predictions provided to decision makers.
**“Do the job scheduling results generalize to predictors with non-zero calibration error?”**
Yes, our re... | Summary: The paper introduces a novel idea of leveraging calibrated predictors to design learning-augmented algorithms. Instead of proving consistency, robustness, and smoothness guarantees, which are worst-case guarantees, the authors derive bounds that depend on the predictor's maximum calibration error. They apply t... | Rebuttal 1:
Rebuttal: We sincerely thank the reviewer for their detailed feedback. However, we believe there may have been a significant misunderstanding regarding key aspects of our paper. Below, we address each concern in detail and respectfully ask that the reviewer reconsider their evaluation based on these clarifi... | Summary: The paper initiates the study of the effect of calibration in algorithms with predictions through two case studies:
1. Ski Rental: The authors design an algorithm that achieves optimal prediction-dependent performance, bounding the expected performance using both the squared error and the calibration error. Th... | Rebuttal 1:
Rebuttal: We thank the reviewer for their thoughtful feedback! We’re encouraged that they find our application of calibration to algorithms with predictions innovative, and our algorithms and analysis novel and interesting. In addition to the modifications detailed below, we will correct the typo on line 77... | null | null | null | null | null | null |
One Image is Worth a Thousand Words: A Usability Preservable Text-Image Collaborative Erasing Framework | Accept (poster) | Summary: This paper aims to solve the problem of concept erasing for images with visually undesirable or even harmful content. The authors first analyze the issues presenting in the prior works, which are actually caused by the sole use of text prompts. To overcome this, a new framework, called Co-Erasing, is proposed,... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments! Due to space constraints, we include `additional_tables.pdf` and figures at the following link: https://anonymous.4open.science/r/icml25_rebuttal-8608. References to $\textcolor{blue}{\text{Table}}$ and $\textcolor{blue}{\text{Figure}}$ in our responses corres... | Summary: This paper introduces Co-Erasing, a text-image collaborative framework designed to address the challenge of generating undesirable content (e.g., NSFW, inappropriate styles) in text-to-image diffusion models. By leveraging both text prompts and self-generated images of the target concept, Co-Erasing aims to im... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments! Due to space constraints, we include `additional_tables.pdf` and figures at: https://anonymous.4open.science/r/icml25_rebuttal-8608. References to $\textcolor{blue}{\text{Table}}$ and $\textcolor{blue}{\text{Figure}}$ correspond to those provided **in this lin... | Summary: This paper proposes Co-Erasing, a framework for concept erasure in text-to-image diffusion models. Existing methods that rely solely on text-based erasure often struggle to balance efficacy (removing unwanted content) and usability (preserving benign generation quality) due to the inherent gap between text and... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments! Due to space constraints, we include `additional_tables.pdf` and figures at: https://anonymous.4open.science/r/icml25_rebuttal-8608. References to $\textcolor{blue}{\text{Table}}$ and $\textcolor{blue}{\text{Figure}}$ correspond to those provided **in this lin... | Summary: This work is proposing a concept erasing method for diffusion models by exploiting images to aid text prompts during training. Image features related to text prompts that we wish to erase in the diffusion models are provided and then combined together after cross-attention layers so that the cross-attention la... | Rebuttal 1:
Rebuttal: Thank you for your valuable comments! Due to space constraints, we include `additional_tables.pdf` and figures at: https://anonymous.4open.science/r/icml25_rebuttal-8608. References to $\textcolor{blue}{\text{Table}}$ and $\textcolor{blue}{\text{Figure}}$ correspond to those provided **in this lin... | null | null | null | null | null | null |
MATH-Perturb: Benchmarking LLMs' Math Reasoning Abilities against Hard Perturbations | Accept (poster) | Summary: This paper constructs a new dataset by applying simple an hard perturbations to the hard problems in the original MATH dataset. Experimental results show drop in performance for almost all models.
## Update after rebuttal
I remain positive about the paper after reading author rebuttal.
Claims And Evidence: ... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback on our work! **We evaluated 12 new long-CoT models that appeared near or after the ICML submission deadline.** The results [here](https://anonymous.4open.science/r/icml2025_13579_math_perturb_rebuttal-C0F8/) show no sign of saturation on MATH-P-Hard. We would l... | Summary: This paper proposes a new benchmark by modifying 279 MATH hard problems and evaluates the popular model on these questions. They also provide various analyses of the performance on these questions.
Claims And Evidence: Yes.
Methods And Evaluation Criteria: Yes.
Theoretical Claims: Not applicable.
Experimen... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback on our work! **We evaluated 12 new long-CoT models that appeared near or after the ICML submission deadline.** The results [here](https://anonymous.4open.science/r/icml2025_13579_math_perturb_rebuttal-C0F8/) show no sign of saturation on MATH-P-Hard. We would l... | Summary: This paper investigates the robustness of mathematical reasoning models when faced with out-of-distribution problem modifications. The authors introduce MATH-P-Simple and MATH-P-Hard, two benchmark datasets that test models under simple and hard perturbations, respectively. Their evaluation reveals significant... | Rebuttal 1:
Rebuttal: > **Q1** The authors evaluate instruction-tuned MLLMs …
**A1**. We would like to first clarify that our dataset only contains textual input, and we evaluated on text-only LLMs, not Multimodal LLMs.
---
> **Q2**: A key limitation of this study is that the authors do not evaluate RL-based models.... | Summary: The paper constructs MATH-Perturb to evaluate the math reasoning generalization of LLMs under simple and especially hard perturbations. The authors create MATH-P-Simple (279 problems) and MATH-P-Hard (279 problems) datasets from level-5 problems in the MATH dataset. Experiment results on 18 LLMs show significa... | Rebuttal 1:
Rebuttal: Thank you for your positive feedback on our work! **We evaluated 12 new long-CoT models that appeared near or after the ICML submission deadline.** The results [here](https://anonymous.4open.science/r/icml2025_13579_math_perturb_rebuttal-C0F8/) show no sign of saturation on MATH-P-Hard. We would l... | 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.