Datasets:
anonymousNeurIPS2026submission4281
/
reviewarena-eval

Dataset card Data Studio Files
xet
 Community
1
forum_id
stringlengths
10
12
conference
stringclasses
6 values
year
int32
2.02k
2.03k
⌀
track
stringclasses
3 values
venue_id
stringclasses
6 values
paper_number
int32
3
28.9k
title
stringlengths
19
143
abstract
stringlengths
483
2.6k
authors
listlengths
0
31
keywords
listlengths
0
35
tldr
stringlengths
0
250
primary_area
stringclasses
99 values
venue
stringclasses
43 values
decision
stringclasses
14 values
decision_comment
stringlengths
0
7.07k
author_rebuttal
stringclasses
51 values
num_reviews
int32
2
7
reviews_json
stringlengths
3.13k
80.2k
markdown
stringlengths
21.4k
189k
markdown_chars
int64
21.4k
189k
ZlpJLQsr2v
neurips
2,024
main
NeurIPS.cc/2021/Conference
6,178
Generalizable Implicit Motion Modeling for Video Frame Interpolation
Motion modeling is critical in flow-based Video Frame Interpolation (VFI). Existing paradigms either consider linear combinations of bidirectional flows or directly predict bilateral flows for given timestamps without exploring favorable motion priors, thus lacking the capability of effectively modeling spatiotemporal ...
[ "Zujin Guo", "Wei Li", "Chen Change Loy" ]
[ "Motion Modeling; Optical Flow; Video Frame Interpolation" ]
machine_vision
NeurIPS 2024 poster
Accept (poster)
The paper presents a method for continuous motion modeling for video frame interpolation using implicit neural representations. The role of the pretrained flow estimator is questioned by d4Ah, which is sufficiently addressed by the rebuttal. Requested perceptual metrics are evaluated in the rebuttal and should be inclu...
Dear Reviewers, We would like to thank all reviewers for providing constructive feedback that helped improve the paper. Due to the word limit, we provide explanations and experiments for concerns shared by multiple reviewers in the following. 1. **Ablation on Motion Encoder (reviewer d4Ah and ZRiZ)** We conducted ex...
4
[{"review_id": "UMM0w35Bdg", "reviewer": "Reviewer_d4Ah", "summary": "The paper proposed a video frame interpolation model, starting from an optical flow, encoding flows to spatial-temporal motion latent. The motion prediction model GIMM, took the encoded initial motion latent to arbitrary-timestep interpolation motion...
Generalizable Implicit Motion Modeling for Video Frame Interpolation Zujin Guo, Wei Li, Chen Change Loy S S-Lab, Nanyang Technological University (zujin.gue) wei 1,, ccloy)@ntu. edu. edu. https://gseeandaat gi wttthb...o/rojjetss////oooooo Abstract Motion modeling is critical in flow-based Video Frame Interpolation (VF...
73,540
AAYXFyvNbr
emnlp
2,023
main
EMNLP/2023/Conference
4,052
Tokenization Consistency Matters for Generative Models on Extractive NLP Tasks
Generative models have been widely applied to solve extractive tasks, where parts of the input is extracted to form the desired output, and achieved significant success. For example, in extractive question answering (QA), generative models have constantly yielded state-of-the-art results. In this work, we study the iss...
[ "Kaiser Sun", "Peng Qi", "Yuhao Zhang", "Lan Liu", "William Yang Wang", "zhiheng huang" ]
[ "Tokenization", "Question Answering" ]
EMNLP 2023 Findings
Accept-Findings
This paper highlights the issue of inconsistent tokenization between the input and output sequences in extractive QA with BPE tokenization (e.g. as done in the BART pretrained model). The authors then suggest a method to ensure consistency between the input and output which improves performance across several datasets....
4
[{"review_id": "1KO3XqZnpz", "reviewer": "Reviewer_ty6n", "summary": "", "questions": "", "limitations": "", "rating": 4, "confidence": 4, "soundness": 4, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "strength...
Tokenization Consistency Matters for Generative Models on Extractive NLP Tasks Kaiser Sun Peng Qi Yuhao Zhang Lan Liu William Yang Wang Zhiheng Huang Paul G. Allen School of Computer Science & Engineering, University of Washington AWS AI Labs huikas@cs. washington edu {pengqi ( www.haaggaaaazznncccoo (liuall, wy,, www....
36,774
qVc7NWYTRZ6
corl
2,023
main
robot-learning.org/CoRL/2021/Conference
231
An Unbiased Look at Datasets for Visuo-Motor Pre-Training
Visual representation learning hold great promise for robotics, but is severely hampered by the scarcity and homogeneity of robotics datasets. Recent works address this problem by pre-training visual representations on large-scale but out-of-domain data (e.g., videos of egocentric interactions) and then transferring th...
[ "Sudeep Dasari", "Mohan Kumar Srirama", "Unnat Jain", "Abhinav Gupta" ]
[ "Visual Representation Learning", "Datasets", "Manipulation" ]
CoRL 2023 Poster
Accept (Poster)
The authors present a study of a wide variety of visual pre-training choices for subsequently training visuomotor policies. The paper submitted first draft focused on the use of vanilla MAE self-supervision, but swapping out the choice of dataset, then subsequently training a 2-layer policy on top. In the rebuttal th...
4
[{"review_id": "kBLwqMp1nZF", "reviewer": "Reviewer_SPNn", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 4, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "stre...
An Unbiased Look at Datasets for Visuo-Motor Pre-Training Sudeep Dasari" CMU Mohan Kumar Srirama CMU Unnat Jain' FAIR at Meta Abhinav Gupta CMU Visual representation learning hold great promise for robotics, but is severely hampered by the scarcity and homogeneity of robotics datasets. Recent works address this problem...
57,849
TyFrPOKYXw
iclr
2,024
main
ICLR.cc/2020/Conference
7,372
Safe RLHF: Safe Reinforcement Learning from Human Feedback
With the development of large language models (LLMs), striking a balance between the performance and safety of AI systems has never been more critical. However, the inherent tension between the objectives of helpfulness and harmlessness presents a significant challenge during LLM training. To address this issue, we pro...
[ "Josef Dai", "Xuehai Pan", "Ruiyang Sun", "Jiaming Ji", "Xinbo Xu", "Mickel Liu", "Yizhou Wang", "Yaodong Yang" ]
[ "Safe Reinforcement Learning", "Reinforcement Learning from Human Feedback", "Large Language Model", "AI Safety" ]
Safe Reinforcement Learning from Human Feedback
reinforcement learning
ICLR 2024 spotlight
Accept (spotlight)
4
[{"review_id": "1mO6jjKUVl", "reviewer": "Reviewer_NDaV", "summary": "The authors propose safe rlhf, a framework to decouple helpfulness and harmfulness of RLHF model responses. They employ a dynamic λ-trade-off to dual helpfulness and harmlessness objectives. They demonstrate that this approach also results in better ...
Published as a conference paper at ICLR 2024 SAFE RLHF: SAFE REINFORCEMENT LEARNING FROM HUMAN FEEDBACK Juntao Dai Xuehai Pan Ruiyang Sun"! Jiaming Ji Xinbo Xu' Mickel Liu? Yizhou Wang? Yaodong Yang "Center for Al Safety and Governance, Institute for AI, Peking University Schooo of Computer Science, Peking University {...
91,828
g9sWQsqemL
neurips
null
NeurIPS.cc/2021/Conference
1,526
Boosting Resilience of Large Language Models through Causality-Driven Robust Optimization
Large language models (LLMs) have achieved remarkable achievements across diverse applications; however, they remain plagued by spurious correlations and the generation of hallucinated content. Despite extensive efforts to enhance the resilience of LLMs, existing approaches either rely on indiscriminate fine-tuning of ...
[ "Xiaoling Zhou", "Mingjie Zhang", "Zhemg Lee", "YUNCHENG HUA", "chengli xing", "Wei Ye", "Flora D. Salim", "Shikun Zhang" ]
[ "Large language models", "Spurious correlations", "Hallucination", "Knowledge localization", "Logistic regression model", "Policy optimization" ]
This study introduces a causality-driven robust optimization approach that selectively updates model components sensitive to causal reasoning, enhancing model causality while preserving valuable pretrained knowledge to mitigate overfitting.
deep_learning
NeurIPS 2025 poster
Accept (poster)
This paper proposes Causality-driven Robust Optimization (CdRO), a principled method that dynamically identifies and selectively updates causal-sensitive components in large language models, thereby reducing spurious correlations and hallucinations while preserving pretrained knowledge and achieving strong robustness a...
4
[{"review_id": "kXZbOKWdWc", "reviewer": "Reviewer_z2Ph", "summary": "This paper proposes a new robust optimization framework based on causality, called CDRO, aiming to reduce LLMS 'reliance on false correlations and enhance their resilience across various tasks. Specifically, the method proposed in this work first i...
Boosting Resilience of Large Language Models through Carval Robust Optimization Xiaoling Zhou Mingjie Zhang Peking University Peking University xiaol iigolingghouustu. pku. edu. en mjzhang06210stu. pku. edu. cn Zhemg Lee Yuncheng Hua Tianjin University University of New South Wales zhemglee@t.ji edu. devin. hua@unsw ed...
86,709
Coh1A4iSsl
emnlp
2,023
main
EMNLP/2023/Conference
2,926
Revisiting Sparse Retrieval for Few-shot Entity Linking
Entity linking aims to link ambiguous mentions to their corresponding entities in a knowledge base. One of the key challenges comes from insufficient labeled data for specific domains. Although dense retrievers have achieved excellent performance on several benchmarks, their performance decreases significantly when on...
[ "Yulin Chen", "Zhenran Xu", "Baotian Hu", "Min Zhang" ]
[ "entity linking", "sparse retrieval" ]
EMNLP 2023 Main
Accept-Main
The paper presents a method for enhancing the recall of the candidate retrieval phase in the entity linking process. It introduces a keyword extractor based on Pre-trained Language Models (PLM) to create a more refined context and utilizes the BM25 retrieval model for obtaining candidate entities.
3
[{"review_id": "C3JbwC9JUd", "reviewer": "Reviewer_2Pme", "summary": "", "questions": "", "limitations": "", "rating": 4, "confidence": 3, "soundness": 4, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "strength...
Revisiting Sparse Retrieval for Few-shot Entity Linking Yulin Chen", Zhenran Xu", Baotian Hut and Min Zhang Harbin Institute of Technology (Shenzhen), Shenzhen, China {200110528, xuzhenran]@stu. hit. mit.edu.de cn (hubaotian, zhangmin2021 J@hit. edu...duccco cn Abstract 2022). Using bi-encoders and Approximate Near- es...
24,283
us7p0VsOhl
emnlp
2,023
main
EMNLP/2023/Conference
895
GLGR: Question-aware Global-to-Local Graph Reasoning for Multi-party Dialogue Reading Comprehension
Graph reasoning contributes to the integration of discretely-distributed attentive information (clues) for Multi-party Dialogue Reading Comprehension (MDRC). This is attributed primarily to multi-hop reasoning over global conversational structures. However, existing approaches barely apply questions for anti-noise grap...
[ "Yanling Li", "Bowei Zou", "Yifan Fan", "Xibo Li", "AiTi Aw", "Yu Hong" ]
[ "Multi-party dialogue reading comprehension", "Global-to-local graph reasoning" ]
EMNLP 2023 Findings
Accept-Findings
Paper Topic And Main Contributions: * The paper proposes a question-aware global-to-local graph reasoning approach (GLGR) for multi-party dialogue reading comprehension (MDRC). * It introduces two types of graphs to facilitate reasoning. These graphs model global information across utterances and local information with...
4
[{"review_id": "wlhdTKaz8c", "reviewer": "Reviewer_F64H", "summary": "", "questions": "A: Have you conducted an ablation study on Question-aware Reasoning (line 341) to verify its effectiveness?\n\nB: Have you considered reversing the order of graph encoding, i.e., local-to-global? If so, how would it potentially impac...
GLGR: Question-aware Global-to-Local Graph Reasoning for Multi-party Dialogue Reading Comprehension Yanling Li', Bowei Zou', Yifan Fan', Xibo Li', Ai Ti Aw?, Yu Hong 'Scoool of Computer Science and Technology. Soochow University, Suzhou, China Instiiute for Infocomm Research, A SSSAR, Singapore (1i4861988, yifanfannlp,...
41,483
WCxfj3PsWb
emnlp
2,023
main
EMNLP/2023/Conference
1,227
Multi-level Adaptive Contrastive Learning for Knowledge Internalization in Dialogue Generation
Knowledge-grounded dialogue generation aims to mitigate the issue of text degeneration by incorporating external knowledge to supplement the context. However, the model often fails to internalize this information into responses in a human-like manner. Instead, it simply inserts segments of the provided knowledge into g...
[ "Chenxu Yang", "Zheng Lin", "Lanrui Wang", "Chong Tian", "Liang Pang", "Jiangnan Li", "Qirong Ho", "Yanan Cao", "Weiping Wang" ]
[ "knowledge-grounded dialogue generation", "contrastive learning", "text degeneration", "pre-trained language model" ]
EMNLP 2023 Main
Accept-Main
A new degeneration penalizing objective to improve knowledge internalization (or grounding). Some reviewers expressed their uncertainty about the results, however minor, and they think the paper can be benefitted by incorporating them in the final version.
3
[{"review_id": "W9ftkE4mJm", "reviewer": "Reviewer_Q1NT", "summary": "", "questions": "1) Can you provide more details on the implementation of the MACL framework, including the specific architecture, hyperparameters, and training process? \n\n2) How knowledge selection is handled in experiments with MACL and other bas...
Multi-level Adaptive Contrastive Learning for Knowledge Internalization in Dialogue Generation Chenxu Yang! Zheng Lanrul Wang Chong Tian', Liang Pang', Jiangnan Li Qirong Ho², Yanan Cao Weiping Wang 1.2 Lanrui Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China School of Cyber Security, Un...
55,500
dujG4nGClA
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,101
URANIA: Differentially Private Insights into AI Use
We introduce _Urania_, a novel framework for generating insights about LLM chatbot interactions with rigorous differential privacy (DP) guarantees. The framework employs a private clustering mechanism and innovative keyword extraction methods, including frequency-based, TF-IDF-based, and LLM-guided approaches. By lever...
[ "Daogao Liu", "Edith Cohen", "Badih Ghazi", "Peter Kairouz", "Pritish Kamath", "Alexander Knop", "Ravi Kumar", "Pasin Manurangsi", "Adam Sealfon", "Da Yu", "Chiyuan Zhang" ]
[ "Differential Privacy", "Clustering", "Summarization" ]
We introduce a novel framework primarily designed for generating insights about LLM chatbot interactions while providing differential privacy guarantees, while also being applicable to other text corpora.
COLM 2025
Accept
The reviewers are agreed that this paper is ready for publication, addresses an important and timely topic, and makes a useful contribution. I also agree that this paper addresses a very important topic (how to analyze chatbot conversations while preserving privacy) and provides practical methods. The reviewers provide...
3
[{"review_id": "AVGzBAZ45f", "reviewer": "Reviewer_yuSm", "summary": "This paper introduces URANIA, a novel framework for generating insights about LLM chatbot interactions while providing formal differential privacy (DP) guarantees. The authors present a well-structured approach that builds upon previous work (CLIO by...
Published as a conference paper at COLM 2025 URANIA: Differentially Private Insights into AI Use Daogao Liu Edith Cohen edi wwwtthhhewwannncoom com Badih Ghazi Peter Kairouz Liudaogao@gmail.com com badiiggazii@gail.. com kairouzegoogle.co com Pritish Kamath pri gritishkégoogle.co com Alexander Knop Ravi Kumar Pasin Man...
69,999
9aVCUv3nKBg
corl
2,021
main
robot-learning.org/CoRL/2021/Conference
30
Adversarially Robust Imitation Learning
Modern imitation learning (IL) utilizes deep neural networks (DNNs) as function approximators to mimic the policy of the expert demonstrations. However, DNNs can be easily fooled by subtle noise added to the input, which is even non-detectable by humans. This makes the learned agent vulnerable to attacks, especially in...
[ "Jianren Wang", "Ziwen Zhuang", "Yuyang Wang", "Hang Zhao" ]
[ "Imitation Learning", "Adversarial Learning" ]
CoRL2021 Poster
Accept (Poster)
The authors present a system in which policies are trained to be robust to adversarial changes in the observation / state space. Reviewers consistently praised the novelty of the problem formulation as applied to robotic imitation learning (even if such techniques are common in other ML domains), as well as the present...
4
[{"review_id": "zeRhMn5SE1m", "reviewer": "Reviewer_bbzz", "summary": "The paper presents a method for imitation learning that is robust to both sensor noise and physical disturbances, via training adversarial policies with RL. The method is demonstrated on various standard RL benchmarks.", "questions": "", "limitation...
Adversarially Robust Imitation Learning Jianren Wang* j jianrenvandree. cmu edu Ziwen Zhuang ahuaagzw@shanghaitecchhh edu, Yuyang Wang yuyangu@andrew. cmu, cmu.edu Hang Zhao hangzhao01248gmaill com Modern imitation learning (IL) utilizes deep neural networks (DNNs) as function approximators mimic the policy of the expe...
36,794
AmPeAFzU3a4
corl
2,022
main
robot-learning.org/CoRL/2021/Conference
399
MIRA: Mental Imagery for Robotic Affordances
Humans form mental images of 3D scenes to support counterfactual imagination, planning, and motor control. Our abilities to predict the appearance and affordance of the scene from previously unobserved viewpoints aid us in performing manipulation tasks (e.g., 6-DoF kitting) with a level of ease that is currently out of...
[ "Yen-Chen Lin", "Pete Florence", "Andy Zeng", "Jonathan T. Barron", "Yilun Du", "Wei-Chiu Ma", "Anthony Simeonov", "Alberto Rodriguez Garcia", "Phillip Isola" ]
[ "Neural Radiance Fields", "Rearrangement", "Robotic Manipulation" ]
CoRL 2022 Poster
Accept (poster)
3
[{"review_id": "zzuUYwhzQfU", "reviewer": "Reviewer_jzYJ", "summary": "This paper presents MIRA, a supervised pick-and-place manipulation framework that uses a NeRF-based scene representation to efficiently learn 6-DoF policies. The framework first builds an NeRF of a tabletop scene with 30 images captured from various...
MIRA: Mental Imagery for Robotic Affordances Lin Yen-Chen', , Pete Florence", Andy Zeng", Jonathan T. Barronn Yilun Du', Wei-Chiu Ma', Anthony Simeonov', Alberto Rodriguez Garcia', Phillip Isola' MIT Google Abstract: Humans form mental images of 3D scenes support counterfactual imagination, planning, and motor control....
41,938
Jd0bCD12DS
colm
2,024
main
colmweb.org/COLM/2024/Conference
933
Mind the Privacy Unit! User-Level Differential Privacy for Language Model Fine-Tuning
Large language models (LLMs) have emerged as powerful tools for tackling complex tasks across diverse domains, but they also raise privacy concerns when fine-tuned on sensitive data due to potential memorization. While differential privacy (DP) offers a promising solution by ensuring models are “almost indistinguishabl...
[ "Lynn Chua", "Badih Ghazi", "Yangsibo Huang", "Pritish Kamath", "Ravi Kumar", "Daogao Liu", "Pasin Manurangsi", "Amer Sinha", "Chiyuan Zhang" ]
[ "Privacy, User-level privacy, Differential privacy" ]
We present a systematic evaluation of user-level differential privacy for LLM fine-tuning on natural language generation tasks
COLM
Accept
Existing evaluations of privacy-preserving LLMs often treat each example (text sequence) as the privacy unit, leading to uneven privacy guarantees when contributions per user vary. To address this, the authors study user-level differential privacy (DP), driven by the need for uniform privacy protection across users. Th...
3
[{"review_id": "1SrXm0C4UI", "reviewer": "Reviewer_mJeH", "summary": "In this work, the authors study language model fine-tuning with differential privacy guarantees. Since each user generally contributes multiple text sequences to the training corpus, to ensure equal privacy guarantees across users, the privacy unit n...
Published as a conference paper at COLM 2024 Mind the Privacy Unit! User-Level Differential Privacy for Language Model Fine-Tuning Lynn Chua Badih Ghazi Yangsibo Huang Princeton University yangaibo@google.cmm Google Research chualynn@google.com Google Research Mudihghazi@gmail.coo Pritish Kamath Ravi Kumar Daogao Liu U...
62,037
QKICx7eSMJ
neurips
null
NeurIPS.cc/2021/Conference
16,651
Generating Computational Cognitive models using Large Language Models
Computational cognitive models, which formalize theories of cognition, enable researchers to quantify cognitive processes and arbitrate between competing theories by fitting models to behavioral data. Traditionally, these models are handcrafted, which requires significant domain knowledge, coding expertise, and time in...
[ "Milena Rmus", "Akshay Kumar Jagadish", "Marvin Mathony", "Tobias Ludwig", "Eric Schulz" ]
[ "large language models", "cognitive computational models", "cognitive science", "learning", "decision making", "neuroscience" ]
neuroscience_and_cognitive_science
NeurIPS 2025 poster
Accept (poster)
This paper dives into the automatic discovery of computational cognitive models, which are previously hand-crafted by researchers. LLMs with iterative feedback is developed to achieve this. The experiments are conducted on four scenarios: 1. Decision Making, 2. Learning, 3. Planning, 4. Working Memory. The authors fin...
4
[{"review_id": "oXINuDNrbZ", "reviewer": "Reviewer_MXWn", "summary": "This paper introduces the Guided generation of Computational Cognitive Models (GeCCo), a novel pipeline that leverages Large Language Models (LLMs) to automate the discovery of cognitive models. The framework prompts an LLM with task instructions, pa...
Generating Computational Cognitive Models using Large Language Models Milena Rmus f Akshay K. Jagadish Princeton University Helmholtz Munich milena. romna.mmu@@lee www.mmmwssslllleeee.eeeme mmw.ltz-mmmiih..ee munich.de akshay www.aajishppiinettton.oo p Marvin Mathony Tobias Ludwig Helmholtz Munich Tübingen University m...
122,446
mzSwYvwYdC
icml
2,025
main
ICML.cc/2025/Conference
9,129
Independence Tests for Language Models
Motivated by liability and intellectual property concerns over open-weight models we consider the following problem: given the weights of two models, can we test whether they were trained independently---i.e., from independent random initializations? We consider two settings: *constrained* and *unconstrained*. In the c...
[ "Sally Zhu", "Ahmed M Ahmed", "Rohith Kuditipudi", "Percy Liang" ]
[ "language models", "finetuning", "fingerprinting" ]
We propose statistical tests to determine if two open-weight language models are independently trained from each other or not, i.e. one is finetuned.
deep_learning->large_language_models
ICML 2025 spotlightposter
Accept (spotlight poster)
The paper tackles the question of whether two given models were trained independently (precisely, deep networks containing GLU MLPs). A statistical test with an exact p-value is developed under equivariance assumptions for the training algorithms. More precisely, the test leverages permutation invariance and equivarian...
5
[{"review_id": "eGlSeZaZ9M", "reviewer": "Reviewer_5ZNo", "summary": "This paper proposes a statistical test for determining whether the initializations of two language models (really, \"deep networks containing GLU MLPs\", or even really slightly weaker than that) are independent or not, when treating the algorithms t...
Independence Tests for Language Models Sally Zhu Ahmed Ahmed" Rohith Kuditipudi Percy Liang' Abstract 1. Introduction Motivated by liability and intellectual property concerns over open-weight models we consider the following problem: given the weights of two models, can we test whether they were trained independenty--...
84,169
I5hTganf3z
emnlp
2,023
main
EMNLP/2023/Conference
5,478
VECHR: A Dataset for Explainable and Robust Classification of Vulnerability Type in the European Court of Human Rights
Recognizing vulnerability is crucial for understanding and implementing targeted support to empower individuals in need. This is especially important at the European Court of Human Rights (ECtHR), where the court adapts Convention standards to meet actual individual needs and thus to ensure effective human rights prote...
[ "Shanshan Xu", "Leon Staufer", "Santosh T.Y.S.S", "Oana Ichim", "Corina Heri", "Matthias Grabmair" ]
[ "nlp for social good", "in legal domain", "vulnerability classification", "rationale dataset", "robustness" ]
EMNLP 2023 Main
Accept-Main
This paper introduces a dataset for research in legal NLP on assessing the "vulnerability type" (a technical legal term) of cases before the European Court of Human Rights. Pros: - Important task which proves to be challenging for today's NLP systems in preliminary experiments (sufficient for the scope of a short pape...
4
[{"review_id": "N3KSgD4g5A", "reviewer": "Reviewer_S2UF", "summary": "", "questions": "1. In Section 4 \"Dataset Analysis\", you mention that among 1070 documents, 519 documents are considered as “non-vulnerable”. However, in Table 3, total 551 case documents are labelled as “non-vulnerable”. Similarly, in Section 4, ...
VECHR: A Dataset for Explainable and Robust Classification of Vulnerability Type in the European Court of Human Rights Shanshan Xu Leon Staufer Santosh T.Y.S.S', Oana Ichim², Corina Heri', Matthias Grabmair' Technical University of Munich, Germany, 'LMU Munich, Germany 3 Graduate Institute of International and Developm...
53,928
79nO2DPjVX
iclr
2,025
main
ICLR.cc/2020/Conference
2,571
Bad-PFL: Exploiting Backdoor Attacks against Personalized Federated Learning
Data heterogeneity and backdoor attacks rank among the most significant challenges facing federated learning (FL). For data heterogeneity, personalized federated learning (PFL) enables each client to maintain a private personalized model to cater to client-specific knowledge. Meanwhile, vanilla FL has proven vulnerable...
[ "Mingyuan Fan", "Zhanyi Hu", "Fuyi Wang", "Cen Chen" ]
[ "personalized federated learning", "backdoor attacks" ]
alignment, fairness, safety, privacy, and societal considerations
ICLR 2025 Poster
Accept (Poster)
4
[{"review_id": "tJYapIAGVB", "reviewer": "Reviewer_ZRox", "summary": "The authors develop Bad-PFL, a new backdoor attack that leverages natural features from the target label as a trigger, enabling backdoor persistence across both global and personalized models. It uses a dual-component trigger, combining natural targe...
Published as a conference paper at ICLR 2025 Bad-PFL: EXPLORING BACKDOOR ATTACKS AGAINST PERSONALIZED FEDERATED LEARNING Mingyuan Fan', Zhanyi Hu', Fuyi Wang', Cen Chen East China Normal Unversity, 2RMIT University $my2660966@@maii com $12559031108sth www..ecnu....... edu. fuyi wang edu..u ABSTRACT Data heterogeneity a...
92,402
My6Rgv7xXV
emnlp
2,023
main
EMNLP/2023/Conference
3,568
Contextual Interaction for Argument Post Quality Assessment
Recently, there has been an increased emphasis on assessing the quality of natural language arguments. Existing approaches primarily focus on evaluating the quality of individual argument posts. However, they often fall short when it comes to effectively distinguishing arguments that possess a narrow quality margin. To...
[ "Yiran Wang", "Xuanang Chen", "Ben He", "Le Sun" ]
[ "argument", "argument quality", "contrastive learning", "large language models" ]
EMNLP 2023 Main
Accept-Main
The reviewers agreed on the importance of analyzing and understanding the structural similarity between arguments to assess their quality. They found the approach novel and sound. The reviewers also appreciated the author rebuttal which clarified the issues highlighted in the reviews.
3
[{"review_id": "4eZ6kj9kHw", "reviewer": "Reviewer_gESR", "summary": "", "questions": "A. Would be interested in a detailed description of the context provided to the LLMs to score arguments.\nB. Do the topics that the arguments relate to have an impact in terms of scoring and if so, how is that captured in the model p...
Contextual Interaction for Argument Post Quality Assessment Yiran Wang Xuanang Chengl, Ben He Le Sun* UUiieesitty of Chinese Academy of Sciences, Beijing, China 2 Institute of Software, Chinese Academy of Sciences, Beijing, China (wangyiran20, chenxuanang1 (etgg1lmmmmmmsueemmmm uc...ac..ccc.................... cn, benh...
51,716
XfbBiBG46D
iclr
2,026
main
ICLR.cc/2020/Conference
11,444
TAO-Attack: Toward Advanced Optimization-Based Jailbreak Attacks for Large Language Models
Large language models (LLMs) have achieved remarkable success across diverse applications but remain vulnerable to jailbreak attacks, where attackers craft prompts that bypass safety alignment and elicit unsafe responses. Among existing approaches, optimization-based attacks have shown strong effectiveness, yet current...
[ "Zhi Xu", "Jiaqi Li", "Xiaotong Zhang", "Hong Yu", "Han Liu" ]
[ "Jailbreaking Attacks", "Large Language Models" ]
alignment, fairness, safety, privacy, and societal considerations
ICLR 2026 Poster
Accept (Poster)
5
[{"review_id": "Kjbx6JPGOR", "reviewer": "Reviewer_e9qu", "summary": "This work proposes TAO-Attack, a novel optimization-based jailbreak method that enhances both effectiveness and efficiency. It introduces a two-stage loss—first suppressing refusals to maintain harmful prefixes, then penalizing pseudo-harmful outputs...
Published as a conference paper at ICLR 2026 TAO-ATTACK: TOWARD ADVANCED OPTIMIZATION- BASED JAILBREAK ATTACKS FOR LARGE LANGUAGE MODELS Zhi Xu, Jiaqi Li, Xiaotong Zhang, Hong Yu, Han Liu* Dalian University of Technology, Dalian, China xu zhi, dut d @uaai. com, 1i. jiagi dut @ grmail... com, hongyued utt. edu. com, aut...
66,137
SbR9mpTuBn
iclr
2,023
main
ICLR.cc/2020/Conference
1,638
Bag of Tricks for Unsupervised Text-to-Speech
Unsupervised text-to-speech (TTS) aims to train TTS models for a specific language without any paired speech-text training data in that language. Existing methods either use speech and corresponding pseudo text generated by an unsupervised automatic speech recognition (ASR) model as training data, or employ the back-tr...
[ "Yi Ren", "Chen Zhang", "Shuicheng YAN" ]
[ "speech synthesis", "unsupervised learning" ]
ICLR 2023 poster
Accept: poster
3
[{"review_id": "Y8xgy2hSTj9", "reviewer": "Reviewer_G8SV", "summary": "", "questions": "", "limitations": "", "rating": 8, "confidence": 4, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "stre...
Published as a conference paper at ICLR 2023 BAG OF TRICKS FOR UNSUPERVISED TTS Yi Ren', Chen Z Shuicheng Yan' SEA AI Lab, Zhejiang University 3 com, 2c9902 ju. edu. cn, yanscesea.com com ABSTRACT Unsupervised text-to-speech (TTS) aims to train TTS models for specific lan- guage without any paired speech-text training ...
69,400
aRUUFFycNh
icml
2,025
main
ICML.cc/2025/Conference
9,288
On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning
Layer-wise preconditioning methods are a family of memory-efficient optimization algorithms that introduce preconditioners per axis of each layer's weight tensors. These methods have seen a recent resurgence, demonstrating impressive performance relative to entry-wise ("diagonal") preconditioning methods such as Adam(W...
[ "Thomas TCK Zhang", "Behrad Moniri", "Ansh Nagwekar", "Faraz Rahman", "Anton Xue", "Hamed Hassani", "Nikolai Matni" ]
[ "feature learning", "representation learning", "preconditioning", "single-index models", "two-layer networks", "non-convex optimization", "matrix sensing", "quasi-newton methods", "kronecker-factored approximate curvature", "shampoo" ]
Kronecker-Factored preconditioning is gaining popularity as an Adam/SGD alternative. We provide concrete evidence of their usefulness by analyzing how they uniquely enhance feature learning.
theory->optimization
ICML 2025 poster
Accept (poster)
Dear Authors, Thank you for submitting your paper to ICML and for contributing a theoretically grounded and well-motivated analysis of layer-wise preconditioning in the context of feature learning. Your work explores the statistical and algorithmic advantages of Kronecker-factored preconditioning methods, demonstratin...
4
[{"review_id": "Xg7fkeQAuZ", "reviewer": "Reviewer_RnJ3", "summary": "This paper shows that layer-wise preconditioning is statistically necessary for efficient feature learning using two common models: linear representation learning and single-index learning. They prove that SGD struggles in non-isotropic inputs, and d...
On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning Thomas T. Zhang" " Behrad Moniri 1 Ansh Nagwekar' Faraz Rahman Anton Xue Hamed Hassani' Nikolai Matni Abstract Adam-like optimizers theoretically, especially from a statis- tical (e.g. generalization) perspective. In fact, there exis...
123,669
I9DVeu8XKa
emnlp
2,023
main
EMNLP/2023/Conference
4,394
CodeFusion: A Pre-trained Diffusion Model for Code Generation
Imagine a developer who can only change their last line of code—how often would they have to start writing a function from scratch before it is correct? Auto-regressive models for code generation from natural language have a similar limitation: they do not easily allow reconsidering earlier tokens generated. We introdu...
[ "Mukul Singh", "José Cambronero", "Sumit Gulwani", "Vu Le", "Carina Suzana Negreanu", "Gust Verbruggen" ]
[ "Text-to-code generation", "Diffusion models", "Program synthesis", "Language models" ]
A pre-trained diffusion code generation model that employs an attention-based decoding strategy and a mask-denoising objective to generate code.
EMNLP 2023 Main
Accept-Main
This paper introduces a diffusion model for natural language to code generation problems. Code generation is a growing area of research and this paper adds the first diffusion model to improve the diversity of the generated code. Reviewers find that the presentation is clear, the method to be novel and interesting. Rev...
3
[{"review_id": "6SoxVSl1qO", "reviewer": "Reviewer_9CNw", "summary": "", "questions": "* In Table 2, why don't you compare with a stronger baseline such as CodeT5 which is included for comparison in Table 1?\n* It would be better to separate the effects of the diffusion module and CPD pretraing tasks, which are paralle...
CODEFUSION: A Pre-trained Diffusion Model for Code Generation Mukul Singh José Cambronero Sumit Gulwani Carina Negreanu Microsoft Research Cambridge, UK Gust Verbruggen Microsoft Keerbergen, Belgium Microsoft Delhi, India Vu Le Microsoft Redmond, US Abstract approaches then select the vocabulary token with the closest ...
41,582
YeQ8SGjarH
icml
2,025
main
ICML.cc/2025/Conference
1,902
Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts
Graph neural networks (GNNs) have achieved remarkable success, yet most are developed under the in-distribution assumption and fail to generalize to out-of-distribution (OOD) environments. To tackle this problem, some graph invariant learning methods aim to learn invariant subgraph against distribution shifts, which he...
[ "Haoyang Li", "Xin Wang", "Xueling Zhu", "Weigao Wen", "Wenwu Zhu" ]
[ "Disentanglement", "Graph Neural Network", "Distribution Shift" ]
We propose to learn disentangled invariant subgraph via self-supervised contrastive variant subgraph estimation for achieving satisfying OOD generalization.
general_machine_learning
ICML 2025 poster
Accept (poster)
The paper proposes a novel method to learn disentangled invariant subgraph via self-supervised contrastive variant subgraph estimation for achieving satisfying OOD generalization. All reviewers recommended the acceptance. Therefore, I followed the reviewers' recommendations and gave the acceptance.
4
[{"review_id": "zOOrjLQonV", "reviewer": "Reviewer_sHjp", "summary": "This manuscript studies out-of-distribution generalization issue in graph neural networks. The authors propose learning invariant subgraphs via variant subgraph contrastive estimation, which can handle graph distribution shifts with severe bias. The ...
Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts Haoyang Li Xin Wang' Xueling Zhu? Weigao Wen Wenwu Zhu Abstract 1. Introduction Graph ubiquitous in our daily life, which has been widely used to represent the complex relationships between enti- ties in many fields, includin...
63,756
dUo6j3YURS
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
480
MOSAIC: Modular Foundation Models for Assistive and Interactive Cooking
We present MOSAIC, a modular architecture for coordinating multiple robots to (a) interact with users using natural language and (b) manipulate an open vocabulary of everyday objects. At several levels, MOSAIC employs modularity: it leverages multiple large-scale pre-trained models for high-level tasks like language an...
[ "Huaxiaoyue Wang", "Kushal Kedia", "Juntao Ren", "Rahma Abdullah", "Atiksh Bhardwaj", "Angela Chao", "Kelly Y Chen", "Nathaniel Chin", "Prithwish Dan", "Xinyi Fan", "Gonzalo Gonzalez-Pumariega", "Aditya Kompella", "Maximus Adrian Pace", "Yash Sharma", "Xiangwan Sun", "Neha Sunkara", ...
[ "Foundation Models", "Human-Robot Interaction", "Model Learning" ]
MOSAIC is a modular architecture that enables multiple home robots to collaboratively cook with humans.
CoRL 2024
Accept
Strengths: - Interesting problem of multi-robot and human coordination, using LLMs for natural language interaction - Modular architecture, integrating many different SoTA models from literature from object detection, segmentation, human action detection. Impressive overall system, with large number of real-world eval...
3
[{"review_id": "joTsIl8lf5", "reviewer": "Reviewer_MKTe", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 5, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 3, "strength...
MOSAIC: Modular Foundation Models for Assistive and Interactive Cooking Huaxiaoyue Wang", Kushal Kedia", " Juntao Ren", Rahma Abdullah, Atiksh Bhardwaj, Angela Chao, Kelly Y Chen, Nathaniel Chin, Prithwish Dan, Xinyi Fan, Gonzalo Gonzalez-Pumarrie Aditya Kompella, Maximus Adrian Pace, Yash Sharma, Xiangwan Sun, Neha Su...
189,167
lt0Yf8Wh5O
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
72
Differentiable Robot Rendering
Vision foundation models trained on massive amounts of visual data have shown unprecedented reasoning and planning skills in open-world settings. A key challenge in applying them to robotic tasks is the modality gap between visual data and action data. We introduce differentiable robot rendering, a method allowing the ...
[ "Ruoshi Liu", "Alper Canberk", "Shuran Song", "Carl Vondrick" ]
[ "Robot Representation", "Visual Foundation Model" ]
We introduce a differentiable robot rendering method based on deformable Gaussians splattings and show many downstream applications. Abstract:
CoRL 2024
Accept
The reviewers agree that the paper is well-written and presents an innovative approach combining Gaussian Splatting with robot kinematics. They highlight the potential impact on bridging visual foundation models with robotics and the method's versatility across robot types. Quantitative results show significant improve...
3
[{"review_id": "Y0xT8zbluJ", "reviewer": "Reviewer_VuJm", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 3, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 4, "strength...
Differentiable Robot Rendering Ruoshi Liu Alper Canberk Shuran Song² Carl Vondrick CColumbia University Stanford University drrobot co umbiu. edu CLIP ,"a doing IxI") Text-to-Robot Pose with CLIP P1 Robot & Camera Pose Estimation P2 Dr. Robot p.Ox d Motion Rotargeting with Point Tracker PJ_ p Differentiably Rendered Ra...
35,114
6vTv9M9ZAA
colm
2,025
main
colmweb.org/COLM/2024/Conference
359
Building Instruction-Tuning Datasets from Human-Written Instructions with Open-Weight Large Language Models
Instruction tuning is crucial for enabling Large Language Models (LLMs) to solve real-world tasks. Prior work has shown the effectiveness of instruction-tuning data synthesized solely from LLMs, raising a fundamental question: Do we still need human-originated signals for instruction tuning? This work answers the quest...
[ "Youmi Ma", "Sakae Mizuki", "Kazuki Fujii", "Taishi Nakamura", "Masanari Ohi", "Hinari Shimada", "Taihei Shiotani", "Koshiro Saito", "Koki Maeda", "Kakeru Hattori", "Takumi Okamoto", "Shigeki Ishida", "Rio Yokota", "Hiroya Takamura", "Naoaki Okazaki" ]
[ "large language models; instruction tuning; synthetic data generation; cross-lingual datasets" ]
We constructed 4 state-of-the-art datasets in 2 languages for instruction tuning with permissive licenses, by simply appending responses generated by open-weight LLMs to human-written instructions.
COLM 2025
Accept
The paper investigates whether instruction-tuning datasets built from human-written instructions paired with LM-generated responses can outperform fully synthetic datasets like Magpie. Using real instructions from LMSYS-Chat-1M and responses from open-weight LMs (e.g., Llama-3.1, Gemma-2), the authors construct new dat...
4
[{"review_id": "axyUPKrI0p", "reviewer": "Reviewer_7i5D", "summary": "The authors introduce a method to generate instruction-tuning datasets using human\ninstructions from a real-world dataset. The method is tested in several LLMs and performs\nwell. The method is evaluated in English and Japanese.", "questions": "", "...
Published as a conference paper at COLM 2025 Building Instruction-Tuning Datasets from Human-Written Instructions with Open-Weight Large Language Models Youmi Ma² Sakae Mizuki Kazuki Fujiit.2 Taishi Nakamura Masanari Ohil2 Hinari Shimada Taihei Shiotani! Koshiro Saito Koki Maeda! Kakeru Hattori Takumi Okamoto' Shigeki ...
57,589
yubwSWol6K
neurips
2,023
main
NeurIPS.cc/2021/Conference
12,476
Canonical normalizing flows for manifold learning
Manifold learning flows are a class of generative modelling techniques that assume a low-dimensional manifold description of the data. The embedding of such a manifold into the high-dimensional space of the data is achieved via learnable invertible transformations. Therefore, once the manifold is properly aligned via a...
[ "Kyriakos Flouris", "Ender Konukoglu" ]
[ "manifold learning flows", "normalizing flows", "optimization", "orthogonalization", "sparsity", "sparse learning", "generative modeling", "Riemannian manifold", "geometry", "metric tensor", "orthogonal basis" ]
We propose an optimization objective which encourage learning of an orthogonal and sparse manifold basis as applied to canonical manifold learning flows
NeurIPS 2023 poster
Accept (poster)
After the rebuttal and some discussion the reviewers agree that the paper is suitable for acceptance. A number of small issues have been identified during the review process and the authors should revise their manuscript accordingly for the final version.
We express our gratitude to both the reviewer and the chair for their valuable time and insights. We have diligently addressed each of the reviewer's comments individually. Furthermore, we have expanded our testing to encompass additional tabular datasets and incorporated CelebA, 64x64 FID test scores for a more compre...
4
[{"review_id": "h5KRG5rGaZ", "reviewer": "Reviewer_tk9x", "summary": "This paper studies the problem of learning a latent representation for data supported on a low-dimensional manifold. It proposes to promote orthogonality of the tangent vectors arising from a learned chart, on top of existing rectangular flow loss. E...
Canonical normalizing flows for manifold learning Kyriakos Flouris Ender Konukoglu Department of Information Technology and Electrical Engineering Department of Information Technology and Electrical Engineering ETH Zürich kflouris@vision ee. ethz. ch ETH Zürich kender@vision ee ethz. ch Abstract Manifold learning flows...
65,437
SI2CXa5eok
emnlp
2,023
main
EMNLP/2023/Conference
4,994
AMR Parsing with Causal Hierarchical Attention and Pointers
Translation-based AMR parsers have recently gained popularity due to their simplicity and effectiveness. They predict linearized graphs as free texts, avoiding explicit structure modeling. However, this simplicity neglects structural locality in AMR graphs and introduces unnecessary tokens to represent coreferences. In...
[ "Chao Lou", "Kewei Tu" ]
[ "semantic parsing", "AMR parsing", "hierarical attention", "pointer mechanism" ]
EMNLP 2023 Main
Accept-Main
As reviewers indicate, this paper introduces a new method for AMR parsing called CHAP. In contrast to recent work which has focused on using pretrained transformer decoders to output linearized AMR graphs, CHAP attempts to add graph structure to the decoder architecture, while still maintaining compatibility with pretr...
3
[{"review_id": "yyUCwCZAsj", "reviewer": "Reviewer_2oGz", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 3, "soundness": 3, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "strength...
AMR Parsing with Causal Hierarchical Attention and Pointers Chao Lou, Kewei Tu School of Information Science and Technology, ShanghaiTech University Shanghai Engineering Research Center of Intelligent Vision and Imaging {louchao, tukw))@sanghaii tech. edu. cn Abstract even superior performance (Konstas et al., 2017: Xu...
50,488
VHhwhmtx3b
colm
2,024
main
colmweb.org/COLM/2024/Conference
431
Does RoBERTa Perform Better than BERT in Continual Learning: An Attention Sink Perspective
Continual learning (CL) aims to train models that can sequentially learn new tasks without forgetting previous tasks' knowledge. Although previous works observed that pre-training can benefit CL, it remains unclear whether a pre-trained model with higher downstream capacity also performs better in CL. In this paper, we...
[ "Xueying Bai", "Yifan Sun", "Niranjan Balasubramanian" ]
[ "Continual Learning, Attention Sink, Interference, Over-Smoothing, Pre-Scaling" ]
We tackle the LM's interference in CL from an attention sink's perspective, where different models' concentration on common tokens (e.g., special tokens and punctuations) may influence their CL performance besides the models' single-task capacities.
COLM
Accept
The paper identifies the prevalence of attention sinks in Bert and Roberta, and hypothesises that they may lead to interference across tasks in continual learning set ups. To address this, the authors propose an approach to diversifying attention. Reviewers thought the work took an interesting new angle on continual le...
4
[{"review_id": "LxgzWpeiEV", "reviewer": "Reviewer_JZWa", "summary": "The paper studies the attention sink and its effect on continual learning. It also compared BERT and RoBERTa for continual learning with attention sink.", "questions": "", "limitations": "", "rating": 6, "confidence": 4, "soundness": null, "presentat...
Published as a conference paper at COLM 2024 Does RoBERTa Perform Better than BERT in Continual Learn- ing: An Attention Sink Perspective Xueying Bai, Yifan Sun, Niranjan Balasubramanian Department of Computer Science Stony Brook University (xubai, ysun, ni niranjan) wtteeccsttoooooo...... edu Abstract Continual learni...
54,491
A9mHph8GJk
neurips
2,023
main
NeurIPS.cc/2021/Conference
15,111
NAS-X: Neural Adaptive Smoothing via Twisting
Sequential latent variable models (SLVMs) are essential tools in statistics and machine learning, with applications ranging from healthcare to neuroscience. As their flexibility increases, analytic inference and model learning can become challenging, necessitating approximate methods. Here we introduce neural adaptive ...
[ "Dieterich Lawson", "Michael Y. Li", "Scott Linderman" ]
[ "sequence models", "probabilistic inference", "reweighted wake-sleep", "sequential monte carlo", "smoothing", "mechanistic models" ]
We introduce a method for fitting sequential latent variable models that combines the benefits of reweighted wake-sleep and smoothing sequential Monte Carlo.
NeurIPS 2023 poster
Accept (poster)
Based on unanimous approval from all reviewers, this paper is accepted. The authors effectively addressed concerns raised during the review process and provided comprehensive responses during the rebuttal stage. In their rebuttal, the authors present new results showcasing substantial advancements over various existin...
We thank the reviewers for their detailed feedback. We respond to reviewers individually and provide a general response below. We have strengthened our submission with several new experiments (see figures in PDF) and theoretical analyses (see below). If the reviewers feel that the new experimental results, analyses, an...
4
[{"review_id": "huwUxyPsPY", "reviewer": "Reviewer_Dwhv", "summary": "The authors propose to use SIXO particle approximation to calculate the expectations in reweighted wake-sleep algorithm for state space models.", "questions": "What do you think are main limitations of NAS-X?\n\nWhen the algorithm is expected to perf...
NAS-X: Neural Adaptive Smoothing via Twisting Dieterich Lawson * Michael Y. Li" Google Research Meterrich@@googee.com Stanford University mi whwehhallllae ciiceelyl@sstanford.cco Scott W. Linderman Stanford University scott www..iinnermnnttaaooroooeoo edu Abstract Sequential latent variable models (SLVMs) are essential...
88,739
OgWh4J7bkT
colm
2,025
main
colmweb.org/COLM/2024/Conference
146
Enhancing LLM Reasoning with Iterative DPO: A Comprehensive Empirical Investigation
Recent advancements in post-training methodologies for large language models (LLMs) have highlighted reinforcement learning (RL) as a critical component for enhancing reasoning. However, the substantial computational costs associated with RL-based approaches have led to growing interest in alternative paradigms, such a...
[ "Songjun Tu", "Jiahao Lin", "Xiangyu Tian", "Qichao Zhang", "Linjing Li", "Yuqian Fu", "Nan Xu", "Wei He", "Xiangyuan Lan", "Dongmei Jiang", "Dongbin Zhao" ]
[ "LLM Reasoning", "Iterative Optimization" ]
DPO enables iterative self-improvement for LLMs, achieving RL-level reasoning performance with lower computational cost through preference-based learning and verifiable rewards.
COLM 2025
Accept
This paper proposes to enhance LLM reasoning ability with DPO-based algorithms. The central contribution lies in showing that iterative DPO, combined with co-evolutionary training of the policy and verifier, can match the performance of reinforcement learning approaches while being more computationally efficient. The ...
4
[{"review_id": "zdUdpwRMmg", "reviewer": "Reviewer_V4JU", "summary": "This paper makes a compelling case that lightweight, preference-based tuning can match the performance of heavyweight RL approaches in math reasoning tasks—at a fraction of the cost. Starting from a Qwen 2.5-7B baseline, a single application of Direc...
Published as a conference paper at COLM 2025 Enhancing LLM Reasoning with Iterative DPO: A Comprehensive Empirical Investigation Songjun Tu Tu ,Jihhao Lin Xiangyu Tian A Qichao Zhang**, Linjing Li Yuqian Fu Nan Xu Wei He?, Xiangyuan Lan Dongmei Jiang Dongbin * Pengcheng Laboratory School of Artificial Intelligence, Uni...
73,917
cWw5FfVhvl
emnlp
2,023
main
EMNLP/2023/Conference
470
MoT: Memory-of-Thought Enables ChatGPT to Self-Improve
Large Language Models (LLMs) have shown impressive abilities on various tasks. However, fundamentally improving them depends on high-quality datasets or computationally expensive fine-tuning. On the contrary, humans can easily improve themselves by self-thinking and memory, without external resources. In this paper, we...
[ "Xiaonan Li", "Xipeng Qiu" ]
[ "LLM", "ChatGPT", "Self-Improve", "Large Language Model", "Memory" ]
MoT: Memory-of-Thought Enables ChatGPT to Self-Improve
EMNLP 2023 Main
Accept-Main
improvement of large language models (LLMs) without fine-tuning or annotated data. The framework uses a two-stage process: first, it generates and stores high-confidence examples from an unlabeled dataset as external memory; second, it recalls and uses the most relevant memory examples to assist the LLM in answering qu...
3
[{"review_id": "3Ewbju1OWV", "reviewer": "Reviewer_v4Dz", "summary": "", "questions": "", "limitations": "", "rating": 4, "confidence": 4, "soundness": 4, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "strength...
MoT: Memory-of-hhought Enables ChatGPT to Self-Improve Xiaonan Li, Xipeng Qiu School of Computer Science, Fudan University Shanghai Key Laboratory of Intelligent Information Processing, Fudan University {lixn20, xpqiu] @fudan.edu.cm Abstract Dataset labeled Alac Cate LLM Large Language Models (LLMs) have shown impressi...
79,843
ZUtgUA0Fuwd
corl
2,022
main
robot-learning.org/CoRL/2021/Conference
35
Watch and Match: Supercharging Imitation with Regularized Optimal Transport
Imitation learning holds tremendous promise in learning policies efficiently for complex decision making problems. Current state-of-the-art algorithms often use inverse reinforcement learning (IRL), where given a set of expert demonstrations, an agent alternatively infers a reward function and the associated optimal po...
[ "Siddhant Haldar", "Vaibhav Mathur", "Denis Yarats", "Lerrel Pinto" ]
[ "Imitation Learning", "Manipulation", "Robotics" ]
CoRL 2022 Oral
Accept (oral)
4
[{"review_id": "fpeNW2Bt7FG", "reviewer": "Reviewer_Ldfy", "summary": "The paper proposes a method for robustifying a policy obtained via behavior cloning by collecting additional experiences in the environment and minimizing an optimal-transport-based (OT-based) trajectory matching reward between expert demonstrations...
Watch and Match: Supercharging Imitation with Regularized Optimal Transport Siddhant Haldar Vaibhav Mathur Denis Yarats Lerrel Pinto New York University rot-robot.gii citesgiigtuubii Abstract: Imitation learning holds tremendous promise in learning policies efficiently for complex decision making problems. Current stat...
38,464
nBnHXevkjZ
corl
2,022
main
robot-learning.org/CoRL/2021/Conference
522
CC-3DT: Panoramic 3D Object Tracking via Cross-Camera Fusion
To track the 3D locations and trajectories of the other traffic participants at any given time, modern autonomous vehicles are equipped with multiple cameras that cover the vehicle's full surroundings. Yet, camera-based 3D object tracking methods prioritize optimizing the single-camera setup and resort to post-hoc fusi...
[ "Tobias Fischer", "Yung-Hsu Yang", "Suryansh Kumar", "Min Sun", "Fisher Yu" ]
[ "Autonomous driving", "3D multi-object tracking", "cross-camera fusion" ]
CoRL 2022 Poster
Accept (poster)
4
[{"review_id": "mOVFAVPLlZL", "reviewer": "Reviewer_qyz1", "summary": "This paper proposes MV-3DT, a multi-camera 3D object tracking pipeline. The major novelty of MV-3DT is how it handles detections results from multiple cameras, where unlike existing works, it aggerates all cameras' detections together and performs t...
CC-3DT: Panoramic 3D Object Tracking via Cross-Camera Fusion Tobias Fischer Yung-Hsu Yang National Tsing Hua University royyangggapp nthu. edu. tw ETH Zürich tobias fiwcher@viieerini...mmmmmmm. et.eth.ch Suryansh Kumar Min Sun Fisher Yu ETH Zürich National Tsing Hua University ETH Zürich 1@yf 1o sukumardvision. ethz.ch...
40,865
nI6JyFSnyV
colm
2,024
main
colmweb.org/COLM/2024/Conference
860
SKVQ: Sliding-window Key and Value Cache Quantization for Large Language Models
Large language models (LLMs) have demonstrated the capability to process extended token sequences, enabling complex tasks such as book comprehension and long-form text generation. However, as context length increases, the key-value (KV) cache required for LLMs consumes substantial memory, becoming a bottleneck for depl...
[ "Haojie Duanmu", "Zhihang Yuan", "Xiuhong Li", "Jiangfei Duan", "Xingcheng ZHANG", "Dahua Lin" ]
[ "quantization, KVCache compression, model compression" ]
This paper proposes SKVQ, a sliding-window KV cache quantization strategy that achieves high compression ratios while maintaining accuracy, enabling efficient handling of long context lengths in large language models.
COLM
Accept
Quality - Pro - Convincing results, well-motivated - Technically and experimentally rigorous and thorough - Con - Too technical for NLP audiences Clarity - Con - minor grammar issues - explanations of key concepts and ideas could be clearer - notation and results not well explained Originality - Pro -...
4
[{"review_id": "Ixdfz3D4W7", "reviewer": "Reviewer_QPwH", "summary": "The paper introduces a novel method, SKVQ (Sliding-window Key and Value Cache Quantization), that mitigates the memory consumption issue caused by the key-value (KV) cache in large language models (LLMs). The authors propose an extremely low bitwidth...
Published as a conference paper at COLM 2024 SKVQ: Sliding-window Key and Value Cache Quantization for Large Language Models Haojie Duanmu" Zhihang Yuan* Xiuhong Lit Peking University Shanghai AI Laboratory Shanghai Jiao Tong University Houmo AI Jiangfei Duan Xingcheng Zhang Shanghai Al Laboratory Dahua Lin CUHK CUHK S...
49,795
ruGY8v10mK
iclr
2,024
main
ICLR.cc/2020/Conference
1,899
A Data-Driven Measure of Relative Uncertainty for Misclassification Detection
Misclassification detection is an important problem in machine learning, as it allows for the identification of instances where the model's predictions are unreliable. However, conventional uncertainty measures such as Shannon entropy do not provide an effective way to infer the real uncertainty associated with the mod...
[ "Eduardo Dadalto Câmara Gomes", "Marco Romanelli", "Georg Pichler", "Pablo Piantanida" ]
[ "Misclassification detection", "Uncertainty estimation", "Trustworthy AI", "Safety" ]
We introduce a data-driven measure of relative uncertainty as a new method for detecting samples misclassified by machine learning classification models.
societal considerations including fairness, safety, privacy
ICLR 2024 poster
Accept (poster)
4
[{"review_id": "KpOx3Uyvc3", "reviewer": "Reviewer_9edT", "summary": "In this paper, the authors present a technique to learn their proposed uncertainty measure for misclassification detection. The learned uncertainty measure has a closed form solution which makes it easy to implement. The authors present a wide range ...
Published as a conference paper at ICLR 2024 A DATA-DRIVEN MEASURE OF RELATIVE UNCERTAINTY FOR MISCLASSIFICATION DETECTION Eduardo Dadalto' Marco Romanelli New York University New York, NY, USA mr 68652@nyu..ed edu Laboratoire des signaux et systèmes (L2S) Université Paris-Saclay CNRS CentraleSupélec Gif-sur-Yvent Fran...
72,302
SlxH2AbBBC2
neurips
2,021
main
NeurIPS.cc/2021/Conference
3,405
Speech Separation Using an Asynchronous Fully Recurrent Convolutional Neural Network
Recent advances in the design of neural network architectures, in particular those specialized in modeling sequences, have provided significant improvements in speech separation performance. In this work, we propose to use a bio-inspired architecture called Fully Recurrent Convolutional Neural Network (FRCNN) to solve ...
[ "Xiaolin Hu", "Kai Li", "Weiyi Zhang", "Yi Luo", "Jean-Marie Lemercier", "Timo Gerkmann" ]
[ "speech separation", "single-channel", "time-domain", "recurrent convolutional neural network", "deep learning" ]
NeurIPS 2021 Poster
Accept (Poster)
This paper proposes to use a bio-inspired asynchronous fully recurrent convolutional neural network (A-FRCNN) for speech separation. In contrast to the conventional synchronous update, the authors argue that with asynchronous updates via the bottom-up, top-down and lateral connections in the network, the model can fuse...
4
[{"review_id": "yoidmaV4WK9", "reviewer": "Reviewer_EiDq", "summary": "The paper introduce a novel method to separate mixture of sounds. The paper introduce a new neural architecture called Fully Recurrent Convolutional Neural Network (FRCNN) which use neural network with lateral connections. The main variant of the p...
Speech Separation Using an Asynchronous Fully Recurrent Convolutional Neural Network Xiaolin Hu' Kai Li', Weiyi Zhang', Yi Luo Jean-Marie Lemercier Timo Gerkmann' DDpartmeet of Computer Science and Technology. Tsinghua Laboratory of Brain and Intelligence (THBI). IDG/McGovern Institute of Brain Research Tsinghua Univer...
48,979
4HNAwZFDcH
colm
2,024
main
colmweb.org/COLM/2024/Conference
793
WorkBench: a Benchmark Dataset for Agents in a Realistic Workplace Setting
We introduce WorkBench: a benchmark dataset for evaluating agents’ ability to execute tasks in a workplace setting. WorkBench contains a sandbox environment with five databases, 26 tools, and 690 tasks. These tasks represent common business activities, such as sending emails and scheduling meetings. The tasks in WorkBe...
[ "Olly Styles", "Sam Miller", "Patricio Cerda-Mardini", "Tanaya Guha", "Victor Sanchez", "Bertie Vidgen" ]
[ "evaluation, work, LLMs, simulation" ]
A benchmark dataset for evaluating agents in a realistic workplace setting.
COLM
Accept
Reviewers generally agree on the value of the WorkBench challenge set, and the authors have made available additional details about the evaluation in response to reviewer No8N in particular that should be built into the revised paper. The primary weakness of the work raised by reviewers is the lack of situated discussi...
4
[{"review_id": "6sjADFDEtE", "reviewer": "Reviewer_oM5y", "summary": "The paper proposes a new dataset called WorkBench for evaluating the how well agents can perform business-related tasks, such as sending e-mails, given a natural language instruction from the user. The instruction could be simple (such as \"cancel my...
Published as a conference paper at COLM 2024 WorkBench: Benchmark Dataset for Agents in a Realistic Workplace Setting Olly Styles; Sam Miller Patricio Cerda-Mardini Mindsdb ollystylesd@mail. com, sameartanis.ai, patriciommindsdb.com Tanaya Guha University of Glasgow tanaya. augh@ggasggw...omm ac. Victor Sanchez Univers...
78,992
WOt1owGfuN
iclr
2,025
main
ICLR.cc/2020/Conference
12,737
Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing
We introduce Probe Pruning (PP), a novel framework for online, dynamic, structured pruning of Large Language Models (LLMs) applied in a batch-wise manner. PP leverages the insight that not all samples and tokens contribute equally to the model's output, and probing a small portion of each batch effectively identifies c...
[ "Qi Le", "Enmao Diao", "Ziyan Wang", "Xinran Wang", "Jie Ding", "Li Yang", "Ali Anwar" ]
[ "Large Lanuage Model Pruning", "Probe Pruning" ]
We introduce Probe Pruning (PP), a novel framework for online, dynamic, structured pruning of Large Language Models (LLMs) applied in a batch-wise manner.
other topics in machine learning (i.e., none of the above)
ICLR 2025 Poster
Accept (Poster)
4
[{"review_id": "OhQCC1M5gu", "reviewer": "Reviewer_tnLv", "summary": "This paper tackles an important problem and explores saliency-aware structural pruning of neural networks. It utilizes a probing module that analyzes the L2-norm based importance for network slimming and delivers run-time inference speedup. It studie...
Published as a conference paper at ICLR 2025 PROBE PRUNING: ACCELERATING LLMs THROUGH DYNAMIC PRUNING VIA MODEL-PROBINE Qi Le', Enmao Diao, Ziyan Wang", Xinran Wang', Jie Ding', Li Yang', Ali Anwar' University of Minnesota University of North Carolina at Charlotte (1e000288, wang8740, dingj, aanwar) @umneeuu. diao_em@h...
75,298
pQm66IPmeE
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,884
Traceable and Explainable Multimodal Large Language Models: An Information-Theoretic View
Existing multimodal large language models (MLLMs) often lack traceable and explainable mechanisms for visual-textual alignment, making it challenging to understand how textual instructions shape multimodal representations. To address this shortcoming, we propose an information-theoretic framework that clarifies how MLL...
[ "Zihan Huang", "Junda Wu", "Rohan Surana", "Raghav Jain", "Tong Yu", "Raghavendra Addanki", "David Arbour", "Sungchul Kim", "Julian McAuley" ]
[ "multimodal LLM", "information theory" ]
We introduce an information-theoretic framework that uses mutual information, a Concept Bottleneck, and an InfoNCE mechanism to explain how multimodal models align and integrate visual and textual inputs.
COLM 2025
Accept
This paper proposes a new IB-based framework for analyzing multimodal LLMs. After rebuttal, it received scores of 6667. Overall, the authors provided a strong rebuttal, and all reviewers were positive about the paper. They noted that the use of IB theory and concept bottlenecks to analyze MLLMs is novel, and the propos...
4
[{"review_id": "CQHSO74irC", "reviewer": "Reviewer_WuoZ", "summary": "This paper proposes a new IB-based framework for analyzing multimodal LLMs. The key idea is to leverage the concept bottleneck [1] to quantify the layer-wise visual information proportion and obtain the concept vector, which can be utilized for MLLM ...
Published as a conference paper at COLM 2025 Traceable and Explainable Multimodal Large Language Mod- els: An Information-Theorele View Zihan Huang'/ Jaannng/Junde Junda Wul+, Rohan Surana', Raghav Jain', Tong Yu?, Raghavendra Addanki², David Arbour?, Sungchul Kim², Julian McAuley! UC San Diego Adobb Research {zih043, ...
61,020
Nf4MHF1pi5
neurips
2,024
main
NeurIPS.cc/2021/Conference
13,430
Watch Out for Your Agents! Investigating Backdoor Threats to LLM-Based Agents
Driven by the rapid development of Large Language Models (LLMs), LLM-based agents have been developed to handle various real-world applications, including finance, healthcare, and shopping, etc. It is crucial to ensure the reliability and security of LLM-based agents during applications. However, the safety issues of L...
[ "Wenkai Yang", "Xiaohan Bi", "Yankai Lin", "Sishuo Chen", "Jie Zhou", "Xu Sun" ]
[ "LLM-based Agents", "Backdoor Attack" ]
We take the initial step towards investigating backdoor attacks on LLM-based agents, and present a general framework along with 3 concrete forms of agent backdoor attacks.
safety_in_machine_learning
NeurIPS 2024 poster
Accept (poster)
Reviewers unanimously agree that this paper presents an interesting set of attacks on LLM agents. The poisoning rates caused some confusion for many reviewers, so I suggest the authors take care to clarify this point in the final manuscript.
We sincerely thank all the reviewers for their time and efforts on reviewing our paper. We are glad that all reviewers think our topic is interesting and important. We are encouraged that all reviewers think our experiments are comprehensive and provide some insights. Here, we make the general response to a question ab...
4
[{"review_id": "yBHLj1uyHI", "reviewer": "Reviewer_7KQt", "summary": "This paper studies the backdoor vulnerability of LLM-based agents. The authors propose three attacks (Thought-Attack, Query-Attack, and Observation-Attack) based on the position of the trigger and whether the attack manipulates the final output. The ...
Watch Out for Your Agents! Investigating Backdoor Threats to LLM-Based Agents Wenkai Yang", Xiaohan B Yankai Lin'1 1, Sishuo Chen', Jie Zhou", Xu Sun "Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China Center for Data Science, Peking University Pattern Recognition Center, WeChat AI, T...
96,087
meKEKDhdnx
iclr
2,025
main
ICLR.cc/2020/Conference
8,281
Preble: Efficient Distributed Prompt Scheduling for LLM Serving
Prompts to large language models (LLMs) have evolved beyond simple user questions. For LLMs to solve complex problems, today’s practices are to include domain-specific instructions, illustration of tool usages, and/or long context such as textbook chapters in prompts. As such, many parts of prompts are repetitive acros...
[ "Vikranth Srivatsa", "Zijian He", "Reyna Abhyankar", "Dongming Li", "Yiying Zhang" ]
[ "LLM prefix caching", "LLM serving", "Distributed systems for ML" ]
This paper proposes Preble, the first distributed LLM serving platform that targets prompt sharing and improves SOTA serving systems by up to 14.5× on average latency.
infrastructure, software libraries, hardware, systems, etc.
ICLR 2025 Poster
Accept (Poster)
5
[{"review_id": "1P4EZ1H3Rs", "reviewer": "Reviewer_rygC", "summary": "The paper introduces Preble, a distributed LLM serving system specifically designed to handle long-context prompt workloads. The authors propose a tree-based prefix-matching algorithm and an online workload-balancing algorithm that combines explorati...
Published as a conference paper ICLR 2025 PREBLE: EFFICIENT DISTRIBUTED PROMPT SCHEDULING FOR LLM SERVING Vikranth Srivatsa" Zijian He Reyna Abhyankar', Dongming Li', Yiying Zhang University of California, San Diego, "GenseeAl Inc. ABSTRACT Prompts to large language models (LLMs) have evolved beyond simple user questio...
77,518
q5ZwEiLzDft
iclr
2,023
main
ICLR.cc/2020/Conference
1,695
Property Inference Attacks Against t-SNE Plots
With the prevailing of machine learning (ML), researchers have shown that ML models are also vulnerable to various privacy and security attacks. As one of the representative attacks, the property inference attack aims to infer the private/sensitive properties of the training data (e.g., race distribution) given the out...
[ "Jingcan Chen", "Xinlei He", "Boyang Zhang", "Yang Chen", "Yang Zhang" ]
[ "Property Inference Attacks", "t-SNE" ]
Submitted to ICLR 2023
Reject
3
[{"review_id": "f3uzBnzeVhA", "reviewer": "Reviewer_6zD4", "summary": "", "questions": "", "limitations": "", "rating": 5, "confidence": 4, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "stre...
Under review as a conference paper at ICLR 2023 PROPERTY INFERENCE ATTACKS AGAINST PLOTS t-SNE Anonymous authors Paper under double-blind review ABSTRACT With the prevailing of machine learning (ML), researchers have shown that ML models are also vulnerable to various privacy and security attacks. As one of the represe...
51,899
laUd1q5iWW
icml
2,025
main
ICML.cc/2025/Conference
8,812
Directed Graph Grammars for Sequence-based Learning
Directed acyclic graphs (DAGs) are a class of graphs commonly used in practice, with examples that include electronic circuits, Bayesian networks, and neural architectures. While many effective encoders exist for DAGs, it remains challenging to decode them in a principled manner, because the nodes of a DAG can have man...
[ "Michael Sun", "Orion Foo", "Gang Liu", "Wojciech Matusik", "Jie Chen" ]
[ "graph generative model", "graph mining", "grammar", "neurosymbolic", "directed acyclic graphs", "DAG" ]
We establish a principled bijective problem mapping between DAG modeling and sequence modeling.
deep_learning->graph_neural_networks
ICML 2025 poster
Accept (poster)
After considering the rebuttal, all reviewers propose to accept this paper for ICML. They highlight the convincing empirical evidence as well as the theoretical contributions of the paper (really, the appendix).
4
[{"review_id": "tz4QoXZDdO", "reviewer": "Reviewer_ymms", "summary": "This paper proposes representing directed acyclic graphs as sequences of production rules. These sequence-based representations enable generative modeling using language-like models, such as transformers. The authors train and evaluate these models i...
Directed Graph Grammars for Sequence-based Learning Michael Sun Orion Foo? Gang Liu Wojciech Matusik Jie Chen Abstract graphs still lacking. For example, current methods pro- pose decoding graph autoregressively by adding nodes and edges, or fragments and connections, at every time step according to some arbitrary orde...
91,035
tjdXRqKaz5Y
corl
2,021
main
robot-learning.org/CoRL/2021/Conference
178
Aligning an optical interferometer with beam divergence control and continuous action space
Reinforcement learning is finding its way to real-world problem application, transferring from simulated environments to physical setups. In this work, we implement vision-based alignment of an optical Mach-Zehnder interferometer with a confocal telescope in one arm, which controls the diameter and divergence of the co...
[ "Stepan Makarenko", "Dmitry Igorevich Sorokin", "Alexander Ulanov", "Alexander Lvovsky" ]
[ "sim-to-real", "robotics", "optical interferometer" ]
CoRL2021 Poster
Accept (Poster)
The reviewers didn't fully agree on the scores for the paper, but after the discussion with the authors, most of them felt that their points were adequately addressed in the rebuttal. I particularly encourage the authors to further clarify how this work is more than an incremental improvement over a baseline in [12] to...
4
[{"review_id": "cAFLe-VEf1B", "reviewer": "Reviewer_GxfU", "summary": "This paper proposes using reinforcement learning to train a robot for aligning an interferometer. They extend setting to a richer observation and action spaces that are closer to the real-world. To work with continuous action spaces, TD3 algorithm i...
Aligning an optical interferometer with beam divergence control and continuous action space Stepan Makarenko 122,, Dmiiryy Sorokin', Alexander Ulanov', and A. I. Lvovsky 1.3 Russian Quantum Center, Moscow, Russia Moscow Institute of Physics and Technology, Russia 3 University of Oxford, United Kingdom makarenko sd @ @p...
32,821
EiM163eZyg
icml
2,025
main
ICML.cc/2025/Conference
1,823
CFP-Gen: Combinatorial Functional Protein Generation via Diffusion Language Models
Existing PLMs generate protein sequences based on a single-condition constraint from a specific modality, struggling to simultaneously satisfy multiple constraints across different modalities. In this work, we introduce CFP-GEN, a novel diffusion language model for Combinatorial Functional Protein GENeration. CFP-GEN f...
[ "Junbo Yin", "Chao Zha", "Wenjia He", "Chencheng Xu", "Xin Gao" ]
[ "function", "protein design", "diffusion model", "multi-objective" ]
A diffusion model capable of designing functional proteins comparable to natural proteins.
applications->health_medicine
ICML 2025 poster
Accept (poster)
This paper introduces CFP-GEN, a novel protein language model that employs discrete diffusion generation for functional protein design. Its key innovation is the integration of annotated protein labels—such as Gene Ontology (GO) terms, InterPro (IPR) domains, and Enzyme Commission (EC) numbers—during diffusion training...
4
[{"review_id": "fkbb7EpR3v", "reviewer": "Reviewer_1Xgk", "summary": "This paper proposes a novel protein language model, CFP-GEN, which leverages discrete diffusion generation to design functional proteins. The key innovation lies in incorporating annotated protein labels, such as Gene Ontology (GO) terms, InterPro (I...
CFP-GEN: Combinatorial Functional Protein Generation via Diffusion Language Models Junbo Yin 123 Chao Zha 123 Wenjia He 123 Chencheng Xu 123 Xin Gao 1 Abstract Condstooing UC20001227 PLMM OUTOTOEL Function Structure Existing PLMs generate protein sequences based single-condition constraint from spe- cific modality, str...
83,306
A_Aeb-XLozL
neurips
2,021
main
NeurIPS.cc/2021/Conference
1,484
Drop-DTW: Aligning Common Signal Between Sequences While Dropping Outliers
In this work, we consider the problem of sequence-to-sequence alignment for signals containing outliers. Assuming the absence of outliers, the standard Dynamic Time Warping (DTW) algorithm efficiently computes the optimal alignment between two (generally) variable-length sequences. While DTW is robust to temporal shi...
[ "Nikita Dvornik", "Isma Hadji", "Konstantinos G. Derpanis", "Animesh Garg", "Allan Douglas Jepson" ]
[ "Sequence Matching", "video representation learning", "instructional videos", "step localization" ]
NeurIPS 2021 Poster
Accept (Poster)
There has been extensive discussion between the reviewers and authors in the post-rebuttal period to tease apart the novel contributions of this paper and accurately position it wrt prior work. The authors have committed to repositioning the paper, clarifying the contributions, and thoroughly discussing close prior wo...
4
[{"review_id": "nIrHxikN-1r", "reviewer": "Reviewer_n8aa", "summary": "This paper proposed an algorithm named Drop-DTW, which enables DTW to drop outliers from the sequence-to-sequence matching. A differentiable Drop-DTW is trained for temporal step localization tasks, representation learning, audio-visual retrieval an...
Drop-DTW: Aligning Common Signal Between Sequences While Dropping Outliers Nikita Dvornik Isma Hadji Konstantinos G. Derpanis Animesh Garg 2 Allan D. Jepson 'Samsung AI Centre Toronto Universitty of Toronto. Vector Institute (isma. (isma.hadji, allan. allan.jepsont jepson]]sammungg com {n. (n.dvornik, k. derpanis) tner...
45,769
8lcW9ltJx9
neurips
2,024
main
NeurIPS.cc/2021/Conference
16,277
Any2Policy: Learning Visuomotor Policy with Any-Modality
Humans can communicate and observe media with different modalities, such as texts, sounds, and images. For robots to be more generalizable embodied agents, they should be capable of following instructions and perceiving the world with adaptation to diverse modalities. Current robotic learning methodologies often focus ...
[ "Yichen Zhu", "Zhicai Ou", "Feifei Feng", "Jian Tang" ]
[ "multi-modal", "robot learning" ]
robotics
NeurIPS 2024 poster
Accept (poster)
This paper presents a multi-modal robotic manipulation dataset encompassing five input modalities: image, text, audio, point cloud, video. This paper demonstrates that a simple approach integrating all these modalities lead to improved performance in robotic manipulation tasks. Despite its limited technical novelty, ...
4
[{"review_id": "gJBi5TFVVj", "reviewer": "Reviewer_h8z5", "summary": "The paper aims to enhance the generalizability of robotic agents by enabling them to handle tasks using diverse modalities such as text, audio, images, and point clouds. The authors introduce a multi-modal system named Any-to-Policy, which utilizes a...
Any2Policy: Learning Visuomotor Policy with Any-Modality Yichen Zhu, Zhicai Ou, Feifei Feng, Jian Tang" Midea Group Abstract Humans can communicate and observe media with different modalities, such as texts, sounds, and images. For robots to be more generalizable embodied agents, they should be capable of following ins...
79,132
n0dD3d54Wgf
neurips
2,022
main
NeurIPS.cc/2021/Conference
3,736
SparCL: Sparse Continual Learning on the Edge
Existing work in continual learning (CL) focuses on mitigating catastrophic forgetting, i.e., model performance deterioration on past tasks when learning a new task. However, the training efficiency of a CL system is under-investigated, which limits the real-world application of CL systems under resource-limited scenar...
[ "Zifeng Wang", "Zheng Zhan", "Yifan Gong", "Geng Yuan", "Wei Niu", "Tong Jian", "Bin Ren", "Stratis Ioannidis", "Yanzhi Wang", "Jennifer Dy" ]
[ "Continual Learning", "Sparse Training" ]
NeurIPS 2022 Accept
Accept
4
[{"review_id": "hI5fvrY_8cm", "reviewer": "Reviewer_igAw", "summary": "The authors discuss whether continual learning can be investigated under the perspective of training efficiency and propose combining network sparsification techniques in a CL setting. As a result, they notice that sparser network are not necessaril...
SparCL: Sparse Continual Learning on the Edge Zifeng Wang Zheng Zhan" Yifan Gong', Geng Yuan', Wei Niu?, Tong Jian', Bin Ren", Stratis Ioannidis' Yanzhi Wang', Jennifer Dy Northeastern University. College of William and Mary gong-yifa, yifa, geng. yuan, yanz. wang] Pmoorteasteers edu, (zifengwang, jian, ioannidis, @dye...
52,770
xhbIud48JN
neurips
2,023
datasets_and_benchmarks
NeurIPS.cc/2021/Conference
508
SituatedGen: Incorporating Geographical and Temporal Contexts into Generative Commonsense Reasoning
Recently, commonsense reasoning in text generation has attracted much attention. Generative commonsense reasoning is the task that requires machines, given a group of keywords, to compose a single coherent sentence with commonsense plausibility. While existing datasets targeting generative commonsense reasoning focus o...
[ "Yunxiang Zhang", "Xiaojun Wan" ]
[ "generative commonsense reasoning", "generative language model" ]
We introduce a generative commonsense reasoning dataset to evaluate how well machines reason under contrastive geographical and temporal contexts.
NeurIPS 2023 Datasets and Benchmarks Poster
Accept (Poster)
Based on reviewer discussions, the Created dataset has merits.
3
[{"review_id": "8pXPIlFDiL", "reviewer": "Reviewer_63Tn", "summary": "", "questions": "", "limitations": "The authors discussed the limitations in the appendix.", "rating": 6, "confidence": 3, "soundness": null, "presentation": null, "contribution": null, "strengths": "1. The dataset is of high quality and seems meanin...
SITUATEDGEN: Incorporating Geographical and Temporal Contexts into Generative Commonsense Reasoning Yunxiang Zhang University Michigan Ann Arbor, USA yunxiang@umicoa ch edu Xiaojun Wan Peking University Beijing, China vanxiaojun@pku. edu, cn Abstract Recently, commonsense reasoning in text generation has attracted much...
68,429
cvVEkS5yij
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
415
Meta-Control: Automatic Model-based Control Synthesis for Heterogeneous Robot Skills
The requirements for real-world manipulation tasks are diverse and often conflicting; some tasks require precise motion while others require force compliance; some tasks require avoidance of certain regions while others require convergence to certain states. Satisfying these varied requirements with a fixed state-actio...
[ "Tianhao Wei", "Liqian Ma", "Rui Chen", "Weiye Zhao", "Changliu Liu" ]
[ "embodied agent", "model-based control", "LLM", "manipulation" ]
a meta-control system can be built to automate the thought process that human experts use to customize control systems for heterogeneous robot skills.
CoRL 2024
Accept
The paper was initially evaluated with mixed reviews, but after the rebuttal, we the reviewers suggest accepting the paper, although without strong backing. Here is a high level overview of the reviews. (before rebuttal) ### Strengths: - Novel approach using LLMs for model-based controllers with clear practical applic...
3
[{"review_id": "8YqopUxrtb", "reviewer": "Reviewer_9Tq5", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 2, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 3, "strength...
Meta-Control: Automatic Model-based Control Synthesis for Heterogeneous Robot Skills Tianhao Wei Liqian Mat.1.2, Rui Chen', Weiye Zhao', Changliu Liu Equal contribution, Carnegie Mellon University, Tsinghua University meta- caa-ronntolppppppe paa.lo..pppr.ttttteeee github. Abstract: The requirements for real-world mani...
172,649
LmOF7UAOZ7
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
162
A Planar-Symmetric SO(3) Representation for Learning Grasp Detection
Planar-symmetric hands, such as parallel grippers, are widely adopted in both research and industrial fields. Their symmetry, however, introduces ambiguity and discontinuity in the SO(3) representation, which hinders both the training and inference of neural network-based grasp detectors. We propose a novel SO(3) repre...
[ "Tianyi Ko", "Takuya Ikeda", "Hiroya Sato", "Koichi Nishiwaki" ]
[ "Grasp Detection", "Rotation Representation", "Parallel Gripper" ]
We propose a novel SO(3) representation that can parametrize a pair of planar-symmetric poses with a single parameter set by leveraging the 2D Bingham distribution.
CoRL 2024
Accept
Here are the strengths and weaknesses of the paper as identified by the reviewer and area chair: Strength: - the direction of leveraging the symmetry in robotic grippers for better grasp detection is interesting and relevant - The approach of how to incorporate this symmetry is novel - the paper contains experiments...
3
[{"review_id": "IUsdYer2lY", "reviewer": "Reviewer_Ueu5", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 4, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 3, "strength...
A Planar-Symmetric SO(3) Representation for Learning Grasp Detection Tianyi Ko* Takuya Ikeda" Woven by Toyota, Inc. Woven by Toyota, Inc. t cianyi.... kwwwoven. toyota ..kuuaa.keda@@woen.. toyota Hiroya Sato The University of Tokyo h-satol jsk, imi u-tokyo. .c. ip Koichi Nishiwaki Woven by Toyota, Inc. koichi ni waki @...
40,105
jt7oCtYqHE
iclr
2,026
main
ICLR.cc/2020/Conference
14,300
From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones
Does reinforcement learning (RL) teach large language models (LLMs) genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL alone even without preceding supe...
[ "Lifan Yuan", "Weize Chen", "Yuchen Zhang", "Ganqu Cui", "Hanbin Wang", "Ziming You", "Ning Ding", "Zhiyuan Liu", "Maosong Sun", "Hao Peng" ]
[ "Reinforcement Learning", "Large Language Model", "Reasoning", "Exploration" ]
We show that once a model has acquired the necessary atomic skills for a task, RL enables the composition of these skills into more complex capabilities when properly incentivized.
foundation or frontier models, including LLMs
ICLR 2026 Poster
Accept (Poster)
4
[{"review_id": "IliaDpV1QF", "reviewer": "Reviewer_to4c", "summary": "The paper proposes an evaluation of an LLM on compositional coding tasks, mainly to understand what extra value RL actually adds to compositional task performance.\nOverall idea is interesting, but the clarity and the depth of analysis are somewhat l...
Published as a conference paper at ICLR 2026 FROM fx) AND g(x) TO f(g(z)) LLMS LEARN NEW SKILLS IN RL BY COMPOSING OLD ONES Lifan Yuan-*, Weize Chen²*, Yuchen Zhang Ganqu Cuist, Hanbin Wang", Ziming You', Ning Ding2.3r Zhiyuan Liu Maosong Sun?, Hao Peng' 1 University of Illinois Urbana-Champaign Tsinghua University 3 S...
84,329
2jibzAXJzH
emnlp
2,023
main
EMNLP/2023/Conference
3,010
T5Score: Discriminative Fine-tuning of Generative Evaluation Metrics
Modern embedding-based metrics for evaluation of generated text generally fall into one of two paradigms: discriminative metrics that are trained to directly predict which outputs are of higher quality according to supervised human annotations, and generative metrics that are trained to evaluate text based on the proba...
[ "Yiwei Qin", "Weizhe Yuan", "Graham Neubig", "Pengfei Liu" ]
[ "text generation", "evaluation" ]
EMNLP 2023 Findings
Accept-Findings
This paper describes a new text generation metric (T5Score) trained on the signals used by both generative and discriminative metrics. The motivation is clear. The main objective of this framework is to learn evaluation metrics based on the assumption that generative and discriminative objectives can work in concert to...
3
[{"review_id": "pdzIhSjPUE", "reviewer": "Reviewer_zxyM", "summary": "", "questions": "QA: In Finding (1) of 6.2 --> Top-k Analysis, you state that the advantage of the source-based version of T5Score over the reference-based version increases as you evaluate fewer systems. Looking at Figure 6, I understand \"fewer sys...
T5SCORE: Discriminative Fine-tuning of Generative Evaluation Metries Yiwei Qin Weizhe Yuan Graham Neubig Pengfei Liu Carnegie Mellon University, New York University, O Inspired Cognition (yiweiq, gneubig, pliu3 Jecs. cmu.edu wy8858nyu. edu Abstract gen Generative Training Modern embedding- baaaeddd-g- metrics for evalu...
65,394
VUhlMfEekm
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
118
Implicit Grasp Diffusion: Bridging the Gap between Dense Prediction and Sampling-based Grasping
There are two dominant approaches in modern robot grasp planning: dense prediction and sampling-based methods. Dense prediction calculates viable grasps across the robot’s view but is limited to predicting one grasp per voxel. Sampling-based methods, on the other hand, encode multi-modal grasp distributions, allowing f...
[ "Pinhao Song", "Pengteng Li", "Renaud Detry" ]
[ "Grasping", "Implicit Neural Representations", "Diffusion Models" ]
Using diffusion models to generate grasps based on local features
CoRL 2024
Accept
The reviewers generally agree that this is a valuable contribution with some novel ideas and well-demonstrated performance improvements. On the other hand, all three point out weaknesses in terminology and formal description (regarding Eqn. 1, SE(3) invariance, generation/sampling/discretization). There are also questi...
3
[{"review_id": "8ZSG8ki0rz", "reviewer": "Reviewer_cuMq", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 5, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 3, "strength...
Implicit Grasp Diffusion: Bridging the Gap between Dense Prediction and Sampling-based Grasping Pinhao Song', Pengteng Li,,, Renaud Detry KU Leuven, Dept. Mechanical Engineering, Research unit Robotics, Automation and Mechatronics KK Leuven, Dept. Electrical Engineering, Research unit Processing Speech and Images, *HKU...
46,540
muYhNDlxWc
neurips
2,024
main
NeurIPS.cc/2021/Conference
13,494
MGF: Mixed Gaussian Flow for Diverse Trajectory Prediction
To predict future trajectories, the normalizing flow with a standard Gaussian prior suffers from weak diversity. The ineffectiveness comes from the conflict between the fact of asymmetric and multi-modal distribution of likely outcomes and symmetric and single-modal original distribution and supervision losses. Instea...
[ "Jiahe Chen", "Jinkun Cao", "Dahua Lin", "Kris M. Kitani", "Jiangmiao Pang" ]
[ "trajectory prediction", "trajectory forecasting" ]
machine_vision
NeurIPS 2024 poster
Accept (poster)
This paper introduces a normalizing flow (NF)-based approach for diverse trajectory prediction. In contrast to standard Gaussian priors used in NFs, the authors propose using a data-driven Gaussian mixture model (GMM), built from the analysis of trajectory patterns in the training dataset. The resulting Mixed Gaussian ...
# General Response (GR) We thank all the reviewers for their valuable suggestions and comments. We add new experiments in the separate pdf file to assist our responses to the reviewers' questions. We would also resolve the writing issues mentioned by reviewers in the paper revision.
3
[{"review_id": "YepR059hP1", "reviewer": "Reviewer_qRNj", "summary": "The authors proposed a new normalizing flow-based human trajectory prediction method called Mixed Gaussian Flow (MGF), which promotes diversity and controllability of prediction. The model uses a mixture of Gaussian model as the initial distribution ...
MGF: Mixed Gaussian Flow for Diverse Trajectory Prediction Jiahe Chen 1.2+ Jinkun Cao 3x1 Dahua Lin 1,2,5 Kris Kitani Jiangmiao Pang31 1 Zhejiang University Shanghai AI Laboratory Carnegie Mellon University "The Chinese University of Hong Kong CPII under InnoHK *: co-first authors t: co-corresponding authors Abstract T...
87,221
KTF1h2XWKZA
iclr
2,022
main
ICLR.cc/2020/Conference
4,115
Multi-batch Reinforcement Learning via Sample Transfer and Imitation Learning
Reinforcement learning (RL), especially deep reinforcement learning, has achieved impressive performance on different control tasks. Unfortunately, most online reinforcement learning algorithms require a large number of interactions with the environment to learn a reliable control policy. This assumption of the availab...
[ "Di Wu", "Tianyu Li", "David Meger", "Michael Jenkin", "Xue Liu", "Gregory Dudek" ]
[ "Multi-bacth", "batch reinfrocement learning", "sample transfer" ]
ICLR 2022 Submitted
Reject
The main identified issues were the limited contribution and use cases, poor writing and missing baseline comparisons and more needed experiments. These issues were not addressed satisfactorily by the rebuttal and hence, I believe the paper should be revised by the authors and undergo another review process at another ...
4
[{"review_id": "RL4vav7NGn2", "reviewer": "Reviewer_u4cA", "summary": "", "questions": "", "limitations": "", "rating": 5, "confidence": 4, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "stre...
Under review as a conference paper at ICLR 2022 MULTI-BATCH REINFORCEMENT LEARNING VIA SAM- PLE TRANSFER AND IMITATION LEARNING Anonymous authors Paper under double-blind review ABSTRACT Reinforcement learning (RL), especially deep reinforcement learning, has achieved impressive performance on different control tasks. ...
38,126
ppb7gyhc7k
emnlp
2,023
main
EMNLP/2023/Conference
396
Learning Retrieval Augmentation for Personalized Dialogue Generation
Personalized dialogue generation, focusing on generating highly tailored responses by leveraging persona profiles and dialogue context, has gained significant attention in conversational AI applications. However, persona profiles, a prevalent setting in current personalized dialogue datasets, typically composed of mere...
[ "Qiushi Huang", "Shuai Fu", "Xubo Liu", "Wenwu Wang", "Tom Ko", "Yu Zhang", "Lilian Tang" ]
[ "Personalized Dialogue Generation", "Retrieval-augmented Dialogue Generation", "Persona-Based Dialogue Generation" ]
EMNLP 2023 Main
Accept-Main
This paper focuses on personalized dialogue generation by leveraging external story documents. In particular, a retriever and a generator are jointly trained, which eliminates the need for annotated retrieval datasets. The retriever is utilized to extract top-k relevant stories. In general, the reviewers agree that th...
3
[{"review_id": "JrrLY2uGXm", "reviewer": "Reviewer_jXtS", "summary": "", "questions": "The experiments conducted in the study are not solid enough, as most of the baseline models listed in Table 1 demonstrate negative performance. Furthermore, the paper neglects to compare the proposed approach with some traditional me...
Learning Retrieval Augmentation for Personalized Dialogue Generation Qiushi Huang 12, Shuai Fu', Xubo Liu', ,Wewwu Wang', Tom Ko', Yu Zhang", Lilian Tang" Universiity of Surrey, 2Southern University of Science and Technology, 'ByteDance Al Lab fqiushi. huang, xubo.liu, w.wang, h.lang) @surrey.cc.uk akhas,, tomkoese, yu...
57,976
Zlm7F7g9FK
emnlp
2,023
main
EMNLP/2023/Conference
493
NLI4CT: Multi-Evidence Natural Language Inference for Clinical Trial Reports
How can we interpret and retrieve medical evidence to support clinical decisions? Clinical trial reports (CTR) amassed over the years contain indispensable information for the development of personalized medicine. However, it is practically infeasible to manually inspect over 400,000+ clinical trial reports in order to...
[ "Mael Jullien", "Marco Valentino", "Hannah Ruth Frost", "Paul O'Regan", "Dónal Landers", "Andre Freitas" ]
[ "NLI", "Clinical Trial", "Textual entailment", "Evidence retrieval", "NLP" ]
Baselines and description of a novel NLP dataset with two tasks, textual entailment, and evidence retrieval, based on cancer clinical trials
EMNLP 2023 Main
Accept-Main
As a reviewer indicates this paper defined a novel benchmark, called NLI4CT, for clinical trial reasoning and inference tasks. This benchmark includes two tasks. This paper also released a new corpus with 2400 expert annotated entailment relations. This paper tested and compared 7 SOTA NLI models. The main reasons to ...
3
[{"review_id": "ikmKEwr0Q2", "reviewer": "Reviewer_iVg5", "summary": "", "questions": "Some SOTA LLMs are not tested, for example, GPT-4 or LLaMA. ", "limitations": "", "rating": 4, "confidence": 5, "soundness": 4, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity"...
NLI4CT: Multi-Evidence Natural Language Inference for Clinical Trial Reports Maël Jullien', Marco Valentino', Hannah Frost!, Paul O' O'Regan', Donal Landers,, André Freitas1.2,3 Department of Computer Science, University of Manchester, United Kingdom Digital Experimental Cancer Medicine Team, Cancer Research UK Manches...
60,270
um7zVEeyVH1
neurips
2,021
main
NeurIPS.cc/2021/Conference
10,236
Terra: Imperative-Symbolic Co-Execution of Imperative Deep Learning Programs
Imperative programming allows users to implement their deep neural networks (DNNs) easily and has become an essential part of recent deep learning (DL) frameworks. Recently, several systems have been proposed to combine the usability of imperative programming with the optimized performance of symbolic graph execution. ...
[ "Taebum Kim", "Eunji Jeong", "Geon-Woo Kim", "Yunmo Koo", "Sehoon Kim", "Gyeong-In Yu", "Byung-Gon Chun" ]
[ "machine learning system" ]
NeurIPS 2021 Poster
Accept (Poster)
The reviewers appreciated the novel techniques presented in the paper for co-executing the symbolic and imperative representations of a model program. During discussions, one major concern that came up was regarding experimental evaluation and comparisons to other related approaches. The additional experiments of apply...
4
[{"review_id": "w5F3YsnEw94", "reviewer": "Reviewer_JQgH", "summary": "This paper proposes a combined approach based on both imperative and symbolic co-execution of a program that can handle any imperative Deep Learning program while achieving the performance of symbolic graph execution. They evaluated on real-world m...
Terra: Imperative-Symbolic Co-Execution of Imperative Deep Learning Programs Taebum Kim Eunji Jeong" Samsung Research eun-ji, e samsung com Seoul National University, FriendliAl x taebum@ snu ac kr, friendll.iieeni.... iendliidii Geon-Woo Kim Yunmo Koo Scoul National University, FriendliAI Scoul National University, Fr...
49,492
GQ1rtVVIy2
emnlp
2,023
main
EMNLP/2023/Conference
2,380
Identifying {Early Maladaptive Schemas} from Mental Health Question Texts
In Psychotherapy, {maladaptive schemas}-- negative perceptions that {an individual has of the self, others, or the world that endure despite objective reality}-- often lead to resistance to treatments and relapse of mental health issues such as depression, anxiety, panic attacks etc. Identification of early maladapt...
[ "Sujatha Das Gollapalli", "Beng Heng Ang", "See-Kiong Ng" ]
[ "Mental Health", "Schema Therapy", "Early Maladaptive Schema", "Personality Disorders", "Classification" ]
Predicting EMS (Early Maladaptive Schema) labels for mental health QA forum posts
EMNLP 2023 Findings
Accept-Findings
The paper explores the task of identifying Early Maladaptive Schemas (EMS) in mental health texts obtained from community question-answering forums. While the work makes valuable contributions, particularly in dataset curation and initial analysis, it also suffers from a series of shortcomings that affect its overall ...
4
[{"review_id": "OoBRH9iqKM", "reviewer": "Reviewer_CaET", "summary": "", "questions": "Please refer to the weaknesses of the paper above, especially 2, 3, 4. What was the reason for not performing any level of fine-tuning on the dataset.", "limitations": "", "rating": 3, "confidence": 4, "soundness": 4, "presentation":...
Identifying Early Maladaptive Schemas from Mental Health Question Texts $ Sujatha Das Gollapalli, BBeng Heng A SSe--Kiog Ng Institute of Data Science, National University of Singapore Integrative Sciences and Engineering Programme, National University of Singapore idssdgênus. edu. sg, bengheng. angêu. nus .eds, seekion...
43,578
NT8Z5NjwxF
neurips
2,024
main
NeurIPS.cc/2021/Conference
8,486
Dual-Diffusion for Binocular 3D Human Pose Estimation
Binocular 3D human pose estimation (HPE), reconstructing a 3D pose from 2D poses of two views, offers practical advantages by combining multiview geometry with the convenience of a monocular setup. However, compared to a multiview setup, the reduction in the number of cameras increases uncertainty in 3D reconstruction....
[ "Xiaoyue Wan", "Zhuo Chen", "Bingzhi Duan", "Xu Zhao" ]
[ "3D Human Pose Estimation", "Binocular Vision", "Diffusion Model", "Pose Priors" ]
To address the increasing uncertainty of binocular 3D HPE due to the reduction of views compared to multiview setups, we propose a Dual-Diffusion method to simutaneously denoise the 3D and 2D poses.
machine_vision
NeurIPS 2024 poster
Accept (poster)
This paper proposes a dual-diffusion model for binocular 3D human pose estimation where the depth ambiguities are integrated in the binocular configuration. The forward diffusion process is used to simulate the noisy 2D poses from GT 2D poses. The reverse denoising process refines the 3D poses obtained by triangulation...
Dear ACs and Reviewers, We are very grateful for your time and effort in reviewing this manuscript. We value every helpful suggestion and comment. **The two most frequently concerned questions are: 1) comparative evaluation with more 2D pose detectors, and 2) comparative evaluation with Diffpose.** We would like to ma...
4
[{"review_id": "Z4PzT8GWcp", "reviewer": "Reviewer_cBDy", "summary": "This paper presents a method for 3D human body keypoint estimation from binocular images. Different from traditional multi-view settings, such methods only take two views as input, which suffer from larger uncertainty at detph wise. To alleviate the ...
Dual-Diffusion for Binocular 3D Human Pose Estimation Xiaoyue Wan Zhuo Chen Department of Automation Bingzhi Duan Xu Zhao Shanghai Jiao Tong University (sherrywaan, chzh9311, DuanBingzhi, zhaoxu]@sjtu edu. cn Abstract Binocular 3D human pose estimation (HPE), reconstructing 3D pose from 2D poses of two views, offers pr...
87,536
yXLyhKvK4D
neurips
2,023
datasets_and_benchmarks
NeurIPS.cc/2021/Conference
842
OpenGSL: A Comprehensive Benchmark for Graph Structure Learning
Graph Neural Networks (GNNs) have emerged as the *de facto* standard for representation learning on graphs, owing to their ability to effectively integrate graph topology and node attributes. However, the inherent suboptimal nature of node connections, resulting from the complex and contingent formation process of grap...
[ "Zhiyao Zhou", "Sheng Zhou", "Bochao Mao", "Xuanyi Zhou", "Jiawei Chen", "Qiaoyu Tan", "Daochen Zha", "Yan Feng", "Chun Chen", "Can Wang" ]
[ "Graph Representation Learning", "Graph Structure Learning", "Benchmark" ]
The progress in GSL field remains unclear due to inconsistent experimental protocols. We introduce OpenGSL, the first comprehensive benchmark for GSL to enable a fair comparison and provide helpful insights.
NeurIPS 2023 Datasets and Benchmarks Poster
Accept (Poster)
The paper presents a benchmark for graph structure learning for graph neural networks. A total of three training strategies and 12 GSL models have been studied on 10 node classification datasets. Most reviewers generally agree that the evaluation pipeline and protocol look meaningful and well-designed, and the experime...
5
[{"review_id": "nDynhFXcj7", "reviewer": "Reviewer_zkm3", "summary": "", "questions": "", "limitations": "Not applicable. ", "rating": 7, "confidence": 3, "soundness": null, "presentation": null, "contribution": null, "strengths": "- The paper gives a comprehensive overview of the GSL field\n- The authors made a great ...
OpenGSL: A Comprehensive Benchmark for Graph Structure Learning Zhiyao Zhou', Sheng Zochaooochao Bochao Mao', Xuanyi , Jiawei Chen' Qiaoyu Tan", Daochen Zha', Yan Feng'. Chun Chen', Can Wang! *College of Computer Science, Zhejiang University, Hangzhou, China Schooo of Software Technology, Zhejiang University, Ningbo, C...
76,351
12u7diwku0
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,476
ALFA: Aligning LLMs to Ask Good Questions A Case Study in Clinical Reasoning
Large language models (LLMs) often fail to ask effective questions under uncertainty, making them unreliable in domains where proactive information-gathering is essential for decision-making. We present ALignment via Fine-grained Attributes, (ALFA) a framework that improves LLM question-asking by (i) decomposing the no...
[ "Shuyue Stella Li", "Jimin Mun", "Faeze Brahman", "Pedram Hosseini", "Bryceton G. Thomas", "Jessica M. Sin", "Bing Ren", "Jonathan S. Ilgen", "Yulia Tsvetkov", "Maarten Sap" ]
[ "Information Seeking", "Question Asking", "Reliable LLM", "Clinical Reasoning", "Structured Rewards" ]
Novel alignment recipe to teach LLMs perform complex goals by (1) decomposing it into more tangible attributes, (2) creating synthetic data, and (3) integrating the attributes.
COLM 2025
Accept
This paper proposes ALFA, a modular framework for improving question generation in clinical contexts using attribute-specific preference data and alignment techniques. Reviewers agree that the work is well-motivated, clearly written, and presents strong empirical results with rigorous ablations. The use of grounded att...
4
[{"review_id": "GUEzF3E2Kk", "reviewer": "Reviewer_aB4k", "summary": "The main contribution of this paper is the ALFA framework, which aims to improve the question-asking ability of large language models in expert domains, focusing on clinical reasoning.\n\nThe paper also contributes the MediQ-AskDocs dataset, which in...
Published as a conference paper at COLM 2025 ALFA: Aligning LLMs to Ask Good Questions A Case Study in Clinical Reasoning Shuyue Stella Li Jimin Mun Faeze Brahman Pedram Hosseini Bryceton G. Thomas Jessica M. Sin Bing Ren Jonathan S. Ilgen' Yulia Tsvetkov! Maarten Sap² UUniversity of Washington Carmegie Mellon Universi...
102,699
xWYRL1eR74
colm
2,024
main
colmweb.org/COLM/2024/Conference
980
FUSE-ing Language Models: Zero-Shot Adapter Discovery for Prompt Optimization Across Tokenizers
The widespread use of large language models has resulted in a multitude of tokenizers and embedding spaces, making knowledge transfer in prompt discovery tasks difficult. In this work, we propose FUSE (Flexible Unification of Semantic Embeddings), an inexpensive approach to approximating an adapter layer that maps from...
[ "Joshua Nathaniel Williams", "J Zico Kolter" ]
[ "prompt optimization, zero-shot methods, prompt discovery, knowledge transfer, textual embeddings, embedding alignment" ]
We propose a method for computing gradients through the different tokenizers and different embedding spaces of multiple language models.
COLM
Accept
This paper proposes an approach for mapping the semantic embedding of two models trained with different tokenizers in a zero-shot way. More specifically, the authors first map the semantic embedding of models (with different tokenizers), and then derive an approximate gradient of one model's output with respect to the ...
4
[{"review_id": "vvNVBJkUZ8", "reviewer": "Reviewer_1ZFn", "summary": "The paper introduces FUSE (Flexible Unification of Semantic Embeddings), a new method designed to facilitate prompt optimization and knowledge transfer across large language models with diverse tokenizers and embedding spaces. Addressing the challeng...
Published as a conference paper at COLM 2024 FUSE-ing Language Models: Zero-Shot Adapter Discovery for Prompt Optimization Across Tokenizers Joshua Nathaniel Williams J. Zico Kolter Department of Computer Science Department of Machine Learning Camegie Mellon University Pittsburgh, PA 15213, USA zkolter@cs.cmu.edu Carne...
47,529
8XFT1PatHy
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
173
Splat-MOVER: Multi-Stage, Open-Vocabulary Robotic Manipulation via Editable Gaussian Splatting
We present Splat-MOVER, a modular robotics stack for open-vocabulary robotic manipulation, which leverages the editability of Gaussian Splatting (GSplat) scene representations to enable multi-stage manipulation tasks. Splat-MOVER consists of: (i) ASK-Splat, a GSplat representation that distills semantic and grasp affor...
[ "Olaolu Shorinwa", "Johnathan Tucker", "Aliyah Smith", "Aiden Swann", "Timothy Chen", "Roya Firoozi", "Monroe David Kennedy", "Mac Schwager" ]
[ "Gaussian Splatting", "Robotic Grasping", "Robotic Manipulation", "Scene Editing" ]
We present Splat-MOVER, a modular robotics stack for open-vocabulary robotic manipulation, which leverages the editability of Gaussian Splatting scene representations to enable multi-stage manipulation tasks.
CoRL 2024
Accept
# Strengths 1. Utilizing the scene editability of 3DGS in robotics is novel. 1. The paper shows promising results for robotic manipulation tasks. 1. The real-world demonstrations showcase the effectiveness of the approach. 1. The paper is well-written. # Weaknesses 1. The main body sometimes lacks necessary informatio...
3
[{"review_id": "neSVBW5pDL", "reviewer": "Reviewer_wP4F", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 3, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 4, "strength...
Splat-MOVER: Multi-Stage, Open-Vocabulary Robotic Manipulation via Editable Gaussian Splatting Ola Shorinwa" Johnathan Tucker" Aliyah Smith", Aiden Swann Timothy Chen', Roya Firoozi', Monroe Kennedy MI? Mac Schwager Deptt. of Aeronautics and Astronauties, *Dept. of Mechanical Engineering, Stanford University Language I...
77,875
kITPSoTBUo
neurips
null
NeurIPS.cc/2021/Conference
2,210
Exploring Semantic-constrained Adversarial Example with Instruction Uncertainty Reduction
Recently, semantically constrained adversarial examples (SemanticAE), which are directly generated from natural language instructions, have become a promising avenue for future research due to their flexible attacking forms, but have not been thoroughly explored yet. To generate SemanticAEs, current methods fall short ...
[ "Jin Hu", "Jiakai Wang", "linna Jing", "Haolin Li", "Haodong Liu", "Haotong Qin", "Aishan Liu", "Ke Xu", "Xianglong Liu" ]
[ "Adversarial Example", "Generative Models" ]
A multi-dimensional instruction uncertainty reduction (InSUR) framework is introduced to generate highly transferable and robust 2D & 3D adversarial examples directly from natural language instructions.
social_and_economic_aspects_of_machine_learning
NeurIPS 2025 poster
Accept (poster)
The paper proposes the InSUR framework for generating semantically constrained adversarial examples and addresses instruction uncertainty through residual-driven direction stabilization, context-encoded constraints and semantic-abstracted evaluation. The reviewers generally agree that the paper is technically solid, no...
4
[{"review_id": "Ag9gaiHIqI", "reviewer": "Reviewer_t5zq", "summary": "This paper focuses on the topic of semantic-constrained adversarial examples. This paper develops a multi-dimensional instruction uncertainty reduction (InSUR) framework to generate more satisfactory SemanticAE. Their framework proposes residual-driv...
Exploring Semantic-constrained Adversarial Example with Instruction Uncertainty Reduction Jin Hu! 1.2 Jiakai Wang 2 Linna Jing' Haolin Li Haodong Liu Haotong Qin Aishan Liu² Ke Xu Xianglong Liu State Key Laboratory of Complex & Critical Software Environment (CCSE), Beihang University 27hongguancun Laboratory SSchool of...
154,188
VyIe1iVHZ4
emnlp
2,023
main
EMNLP/2023/Conference
5,204
TR-Rules: Rule-based Model for Link Forecasting on Temporal Knowledge Graph Considering Temporal Redundancy
Temporal knowledge graph (TKG) has been proved to be an effective way for modeling dynamic facts in real world. Many efforts have been devoted into predicting future events i.e. extrapolation, on TKGs. Recently, rule-based knowledge graph completion methods which are considered to be more interpretable than embedding-b...
[ "Ningyuan Li", "Haihong E", "Shi Li", "Mingzhi Sun", "Tianyu Yao", "Meina Song", "Yong Wang", "Haoran Luo" ]
[ "Knowledge Graph", "Temporal Knowledge Graph", "Link forecasting", "Temporal Rules" ]
EMNLP 2023 Findings
Accept-Findings
This paper proposes a new method which is TR-Rules, a rule-based model for extrapolating temporal knowledge graphs. This innovative approach effectively addresses temporal redundancy through a straightforward algorithm. However, the reviewers also state the technical contribution is incremental, more experimental studi...
3
[{"review_id": "Wh5mHGU8DX", "reviewer": "Reviewer_uQVS", "summary": "", "questions": "A. Could the authors provide the rationale behind the selection of these baselines?\nB. Have the authors considered conducting multiple experiments and averaging the results?\nC. Have the authors considered releasing the paper’s code...
TR-Rules: Rule-based Model for Link Forecasting on Temporal Knowledge Graph Considering Temporal Redundancy Ningyuan Li', Haihong Li Shi?, Mingzhi Sun', Tianyu Yao!, Meina Song', Yong Wang', Haoran Luo! Beijing University of Posts and Telecommunications: National Computer Network Emergency Response Technical Team/Coord...
39,081
y7JnjDcIQa
colm
2,024
main
colmweb.org/COLM/2024/Conference
945
How Susceptible are LLMs to Influence in Prompts?
Large Language Models (LLMs) are highly sensitive to prompts, including additional context provided therein. As LLMs grow in capability, understanding their prompt-sensitivity becomes increasingly crucial for ensuring reliable and robust performance, particularly since evaluating these models becomes more challenging. ...
[ "Sotiris Anagnostidis", "Jannis Bulian" ]
[ "Assistance, scalable oversight, sycophancy" ]
We investigate bias in AI assistance in current LLMs and characterize properties of successful assistance.
COLM
Accept
This paper investigates the susceptibility of LLMs to influence from prompts, focusing on how additional context (especially explanations and confidence) from another model affects LLM responses in QA tasks. Strengths: 1. The paper studies a timely topic, and the impact of external knowledge sources on responses is ve...
3
[{"review_id": "A5iUMUIoHo", "reviewer": "Reviewer_YHyH", "summary": "This paper studies the sensitivity of LLMs’ responses to the injected context in the prompt. It examines how the output response of an LLM will be swayed by the prediction and explanations provided by an external source. It shows that models are gene...
Published as a conference paper at COLM 2024 How Susceptible are LLMs to Influence in Prompts? Sotiris Anagnostidis ETH Zürich Jannis Bulian Google DeepMind Abstract Large Language Models (LLMs) are highly sensitive to prompts, including Additionalcccmn additional context provided therein. As L.Ms grow in capability, u...
102,247
RjS0j6tsSrf
neurips
2,022
main
NeurIPS.cc/2021/Conference
4,642
Diagonal State Spaces are as Effective as Structured State Spaces
Modeling long range dependencies in sequential data is a fundamental step towards attaining human-level performance in many modalities such as text, vision, audio and video. While attention-based models are a popular and effective choice in modeling short-range interactions, their performance on tasks requiring long ra...
[ "Ankit Gupta", "Albert Gu", "Jonathan Berant" ]
[ "state spaces", "long range models", "efficient", "Transformer", "speech recognition", "language modeling", "time series model" ]
NeurIPS 2022 Accept
Accept
3
[{"review_id": "dyBxYbIq243", "reviewer": "Reviewer_NmPw", "summary": "The paper builds upon the recently proposed S4 architecture, which was shown to be effective at modeling long-range dependencies. The authors propose simplifications to the S4 model by modifying the kernel such that, instead of the originally propos...
Diagonal State Spaces are as Effective as Structured State Spaces Ankit Gupta* Albert Gu Jonathan Berant Tel Aviv University joberant&cs tau. tau.ac.il IBM Research ikitgupta.iittkanpun igmaa com Stanford University albertguustanford. Abstract Modeling long range dependencies in sequential data is fundamental step towa...
42,807
gwd3MQufGP
neurips
2,024
main
NeurIPS.cc/2021/Conference
9,290
KptLLM: Unveiling the Power of Large Language Model for Keypoint Comprehension
Recent advancements in Multimodal Large Language Models (MLLMs) have greatly improved their abilities in image understanding. However, these models often struggle with grasping pixel-level semantic details, e.g., the keypoints of an object. To bridge this gap, we introduce the novel challenge of Semantic Keypoint Compr...
[ "Jie Yang", "Wang ZENG", "Sheng Jin", "Lumin Xu", "Wentao Liu", "Chen Qian", "Ruimao Zhang" ]
[ "Keypoint Detection", "Pose Estimation", "Multimodal Large Language Model" ]
machine_vision
NeurIPS 2024 poster
Accept (poster)
**Summary:** This work introduces KptLLM, an MLLM that addresses the challenges of semantic keypoint comprehension, where a model must reason about fine-grained semantic visual information through keypoint. KptLLM shows that using a support set and LoRA on a frozen LLM leads to improvements over other baseline models o...
We appreciate the efforts of all reviewers in reviewing our paper and providing insightful comments and valuable suggestions. The supplementary visualization results (To \# Reviewer Hwow) have been included in the rebuttal PDF.
4
[{"review_id": "FVzhhM0mP2", "reviewer": "Reviewer_j2Jb", "summary": "The paper proposes to build a MLLM for keypoint detection. They propose three main points to achieve this. First, the paper introduces a new benchmark that measures various types of keypoint detection including both text and visual-prompt-based, as w...
KptLLM: Unveiling the Power of Large Language Model for Keypoint Comprehension Jie Yang 125 Wang Zeng45 125 Sheng Jin Lumin Xu* Wentao Liu Chen Qian Ruimao Zhang Sun Yat-sen University "The Chinese University of Hong Kong, Shenzhen 'The University of Hong Kong *The Chinese University of Hong Kong Sense'Time Research an...
74,215
cCYWeCzAv0
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,131
MS-SSM: A Multi-Scale State Space Model for Efficient Sequence Modeling
State-space models (SSMs) have recently attention as an efficient alternative to computationally expensive attention-based models for sequence modeling. They rely on linear recurrences to integrate information over time, enabling fast inference, parallelizable training, and control over recurrence stability. However, ...
[ "Mahdi Karami", "Ali Behrouz", "Peilin Zhong", "Razvan Pascanu", "Vahab Mirrokni" ]
[ "MS-SSM: Sequence Models", "Language models", "State Space Model", "Multi-Scale", "Multi-Resolution" ]
We introduce MS-SSM which enhances traditional SSMs by modeling sequence dynamics at multiple resolutions using independent SSMs, scale-dependent initialization, and an input-dependent scale-mixer.
COLM 2025
Accept
This paper introduces MS-SSM, a multi-scale state-space model framework that addresses limitations in traditional SSMs by processing sequences at multiple resolutions with specialized dynamics and an input-dependent scale mixer. The work presents a well-motivated and technically sound approach to enhance SSMs' ability ...
4
[{"review_id": "qw91NXa2cl", "reviewer": "Reviewer_xBk6", "summary": "This work introduces an extension to the State Space Model (SSM) by adding multi-scale convolution layer. The motivation for this additional layer is to capture multi-scale dependencies in order to better model complex structures. \nFirst, the input ...
Published as a conference paper at COLM 2025 MS-SSM: A Multi-Scale State Space Model for Efficient Se- quence Modeling Mahdi Karami 1 Ali Behrouz ! Peilin Zhong' 1 Razvan Pascanu 2, Vahab Mirrokni 'Google Research, "Google DeepMind (mahdika, al ibehrouz, pei weilinz, razp, mirrokni @google.coom com Abstract State-space...
63,309
QIv5aXEAcc
neurips
null
NeurIPS.cc/2021/Conference
12,434
Theoretical Investigation of Adafactor for Non-Convex Smooth Optimization
Adafactor is an early memory-efficient optimization algorithm proposed as an alternative to Adam. By eliminating first-order momentum and employing a rank-$1$ matrix factorization to approximate the second-moment matrix, Adafactor achieves near-zero memory overhead compared to traditional gradient descent methods. Desp...
[ "Yusu Hong", "Junhong Lin" ]
[ "Adafactor", "stochastic optimization", "non-convex smooth optimization", "convergence rate" ]
optimization
NeurIPS 2025 poster
Accept (poster)
This paper presents the first convergence analysis of Adafactor in the setting of nonconvex smooth optimization. To the best of my knowledge, such a result is novel and a timely contribution. The analysis is clearly written and technically sound, with the authors providing satisfactory responses to several technical co...
4
[{"review_id": "TgHPmtrID7", "reviewer": "Reviewer_qPLs", "summary": "The paper analyzes the convergence rate to a stationary point of Adafactor.", "questions": "# Suggestions\n1. Some symbols look too close to each other, I suggest changing some of them.\n 1. $c$ and $c_0$ have different purposes, but almost identi...
Theoretical Investigation of Adafactor for Non-Convex Smooth Optimization Yusu Hong Center for Data Science Junhong Lin" Center for Data Science Zhejiang University junhongāzju. edu. edu. cn and School of Mathematical Sciences Zhejiang University yusuhongōzju. edu, Abstract Adafactor is early memory-efficient optimizat...
93,601
6FEpFCqH7o
neurips
null
NeurIPS.cc/2021/Conference
23,161
DictPFL: Efficient and Private Federated Learning on Encrypted Gradients
Federated Learning (FL) enables collaborative model training across institutions without sharing raw data. However, gradient sharing still risks privacy leakage, such as gradient inversion attacks. Homomorphic Encryption (HE) can secure aggregation but often incurs prohibitive computational and communication overhead. ...
[ "Jiaqi Xue", "Mayank Kumar", "Yuzhang Shang", "Shangqian Gao", "Rui Ning", "Mengxin Zheng", "Xiaoqian Jiang", "Qian Lou" ]
[ "Federated Learning", "Homomorphic Encryption", "Privacy" ]
We propose DictPFL, a framework that ensures efficient and private federated learning (FL) by encrypting shared gradients and keeping most gradients local, while still preserving the performance of global gradient aggregation.
social_and_economic_aspects_of_machine_learning
NeurIPS 2025 poster
Accept (poster)
This paper proposes **DictPFL**, a framework for efficient and private federated learning through selective encryption of model weights; with experiments across vision and NLP tasks showing clear improvements over prior art. The **final scores** of the paper read as 1 × Accept and 2 × Borderline Accept, an increase ove...
3
[{"review_id": "8B81vVQydx", "reviewer": "Reviewer_oqcz", "summary": "This paper introduces a new Homomorphic encryption (HE) method defending against gradient inversion attacks in Federated Learning (FL).\nPrior HE methods encrypt every gradient update, which is time-consuming, or encrypt most sensitive gradients, whi...
DictPFL: Efficient and Private Federated Learning on Encrypted Gradients Jiaqi Xue', Mayank Kumar', Yuzhang Shang', Shangqian Gao', Rui Ning² Mengxin Zheng', Xiaoqian Jiang', Qian Lou' Uniiversity of Central Florida FFlorida State University 'Old Dominion University *University of Texas Health Science Center at Houston...
80,226
2Kl8Ztw6wk
colm
2,025
main
colmweb.org/COLM/2024/Conference
16
PredGen: Accelerated Inference of Large Language Models through Input-Time Speculation for Real-Time Speech Interaction
Large Language Models (LLMs) are widely used in real-time voice chat applications, typically in combination with text-to-speech (TTS) systems to generate audio responses. However, their large size often leads to noticeable latency between the end of user input and the start of audio output, resulting in suboptimal user...
[ "Shufan Li", "Aditya Grover" ]
[ "Large Language Models", "Inference", "Speculative Decoding" ]
We leverage speculative decoding at user input time to reduce the latency of speech interactions.
COLM 2025
Accept
The paper focuses on latency in real-time voice applications from waiting for the LLM to generate a sequence to be passed into the text-to-speech system. The authors propose an (iterative) speculative decoding approach which proposes multiple candidate responses while the user is still speaking---allowing the TTS to st...
4
[{"review_id": "r3TYUWp0UE", "reviewer": "Reviewer_iE4i", "summary": "This paper proposes PredGen, a framework for accelerating real-time LLM-powered voice assistants by performing speculative response generation during user input. It reduces time-to-first-sentence (TTFS) latency by generating and verifying candidate r...
Published as a conference paper at COLM 2025 PredGen: Accelerated Inference of Large Language Models through Input-Time Speculation for Real-Time Speech Inter- action Shufan Li, Aditya Grover University of California, Los Angeles Abstract Large Language Models (LLMs) are widely used in real-time voice chat applications...
49,054
7DtgxVZGj-y
iclr
2,023
main
ICLR.cc/2020/Conference
2,924
Contrastive Unsupervised Learning of World Model with Invariant Causal Features
In this paper we present a world model, which learns the causal features using invariance principle. We use contrastive unsupervised learning to learn the invariant causal features, which enforces invariance across augmentations of irrelevant parts or styles of the observation. Since the world model based reinforcement...
[ "Rudra P. K. Poudel", "Harit Pandya", "Roberto Cipolla" ]
[ "world models", "causality", "contrastive learning", "model-based reinforcement learning", "reinforcement learning", "out-of-distribution generalisation", "sim-to-real transfer", "robot navigation" ]
Submitted to ICLR 2023
Reject
4
[{"review_id": "uHXcO5RuRQ1", "reviewer": "Reviewer_tiQV", "summary": "", "questions": "", "limitations": "", "rating": 6, "confidence": 4, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "stre...
Under review as a conference paper at ICLR 2023 CONTRASTIVE UNSUPERVISED LEARNING OF WORLD MODEL WITH INVARIANT CAUSAL FEATURES Anonymous authors Paper under double-blind review ABSTRACT In this paper we present world model, which learns causal features using the in- variance principle. In particular, we use contrastiv...
43,292
DKabt9MFnT
neurips
2,021
main
NeurIPS.cc/2021/Conference
748
How Gradient Descent Separates Data with Neural Collapse: A Layer-Peeled Perspective
In this paper, we derive a landscape analysis to the surrogate model to study the inductive bias of the neural features and parameters from neural networks with cross-entropy. We show that once the training cross-entropy loss decreases below a certain threshold, the features and classifiers in the last layer of the neu...
[ "Wenlong Ji", "Yiping Lu", "Yiliang Zhang", "Zhun Deng", "Weijie J Su" ]
[]
NeurIPS 2021 Submitted
Reject
This paper studies an asymptotic implicit bias phenomenon termed "neural collapse". It constitutes a very interesting line of work, however reviewers overall had many concerns, and especially during internal discussion it was highlighted that the paper is very hard to follow and feels unpolished in many places. There...
4
[{"review_id": "qeOfbP1iWO", "reviewer": "Reviewer_6dN9", "summary": "This paper studies representations of neural networks trained with cross entropy, when trained beyond zero training error, utilizing the unconstrained layer peeled model (ULPM). ULPM is motivated by the fact that many neural networks are highly overp...
How Gradient Descent Separates Data with Neural Collapse: A Layer-Peeled Perspective Anonymous Author(s) Affiliation Address email Abstract In this paper, study the inductive bias of the neural features and parameters from neural networks with cross-entropy loss. We study a surrogate model named unconstrained layer pee...
41,560
lEQnUI5lEA
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,016
EvidenceBench: A Benchmark for Extracting Evidence from Biomedical Papers
We study the task of automatically finding evidence relevant to hypotheses in biomedical papers. Finding relevant evidence is an important step when researchers investigate scientific hypotheses. We introduce EvidenceBench to measure models performance on this task, which is created by a novel pipeline that consists of...
[ "Jianyou Wang", "Weili Cao", "Kaicheng Wang", "Xiaoyue Wang", "Ashish Dalvi", "Gino Prasad", "Qishan Liang", "Hsuan-lin Her", "Mingwang", "Qin Yang", "Gene W. Yeo", "David E Neal", "Maxim Khan", "Christopher D. Rosin", "Ramamohan Paturi", "Leon Bergen" ]
[ "Biomedical Benchmark", "Scientific Information Extraction", "Large Language Models", "BioNLP" ]
A large biomedical benchmark to evaluate LLMs' evidence-finding ability for scientific hypotheses and an automatic pipeline to create such benchmarks.
COLM 2025
Accept
This paper introduces EvidenceBench, a corpus instantiating an "evidence extraction" task in which the goal is identify the sentences in an article that are most relevant to a given hypothesis. There was consensus amongst reviewers that this is an important task, and that the authors have done a laudable job of validat...
4
[{"review_id": "MEolDJTBty", "reviewer": "Reviewer_F3p6", "summary": "This paper introduces EvidenceBench to measure LLM performance on finding evidence relevant to hypotheses in biomedical papers. The paper proposes a LLM driven pipeline for hypothesis generation and sentence-by-sentence annotation of biomedical paper...
Published as a conference paper at COLM 2025 CvidenceBenche A Benchmark for Extracting Evidence from Biomedical Papers Jianyou Wang"; Weili Cao Kaicheng Wang', Xiaoyue Wang', Ashish Dalvi', Gino Prasad,, Qishan Liang', Hsuan-lin Her?, Ming Wang*, Qin Yang", Gene W. Yeo?, David Neal?, Maxim Khan?, Christopher D. Rosin",...
101,127
RgJwDQwW82y
corl
2,022
main
robot-learning.org/CoRL/2021/Conference
64
Last-Mile Embodied Visual Navigation
Realistic long-horizon tasks like image-goal navigation involve exploratory and exploitative phases. Assigned with an image of the goal, an embodied agent must explore to discover the goal, i.e., search efficiently using learned priors. Once the goal is discovered, the agent must accurately calibrate the last-mile of n...
[ "Justin Wasserman", "Karmesh Yadav", "Girish Chowdhary", "Abhinav Gupta", "Unnat Jain" ]
[ "Visual Navigation", "Embodied AI", "Image-Goal Navigation", "Perspective-n-Point", "AI Habitat", "Sim-to-Real" ]
CoRL 2022 Poster
Accept (poster)
3
[{"review_id": "ld5xwWiD2Vz", "reviewer": "Reviewer_FqFC", "summary": "This paper tackles the problem of last-mile embodied visual navigation, in which the agent has to navigate to the goal after the goal image is visible in the current view. Once the image is visible, the correspondences between the current and the go...
Last-Mile Embodied Visual Navigation Justin Wasserman "1, KKrmmshh Yadav?, Girish Chowdhary"!, Abhinav Gupta', Unnat Jain "University of Illinois Urbana- Champaign, Meta AI Research, Carnegie Mellon University Abstract: Realistic long-horizon tasks like image-goal navigation involve ex- ploratory and exploitative phase...
45,689
DktAODDdbt
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,903
Evaluating Large Language Models as Expert Annotators
Textual data annotation, the process of labeling or tagging text with relevant information, is typically costly, time-consuming, and labor-intensive. While large language models (LLMs) have demonstrated their potential as direct alternatives to human annotators for general domains natural language processing (NLP) task...
[ "Yu-Min Tseng", "Wei-Lin Chen", "Chung-Chi Chen", "Hsin-Hsi Chen" ]
[ "LLMs-as-expert-annotators", "reasoning models", "multi-agent framework" ]
We investigate: whether top-performing LLMs, which might be perceived as having expert-level proficiency in academic and professional benchmarks, can be an direct alternative to human-expert annotators?
COLM 2025
Accept
I recommend accepting this paper. It provides a thorough evaluation of LLMs as expert annotators across specialized domains, with novel insights about their collaborative behaviors in a multi-agent discussion framework. Three of four reviewers supported acceptance, and the authors addressed key concerns in their rebutt...
4
[{"review_id": "9sYJspHLMy", "reviewer": "Reviewer_jNni", "summary": "This paper explores the use of large language models (LLMs) as annotators in specialized domains such as finance, law, and biomedicine. The authors evaluate both instruction-tuned and reasoning-focused models on five datasets, using only annotation g...
Published as a conference paper at COLM 2025 Evaluating Large Language Models as Expert Annotators Yu-Min Tseng Wei-Lin Chen Chung-Chi Chen Hsin-Hsi Chen R7 *National Taiwan University /Virginia Tech 'University of Virginia 6 AISTT Japan " "AINTU, Taiwan ymtseng@vt. edu, wlchen@virginia.edu c. chencaccccmorrro hhchen@n...
66,073
oEsuNpkA8d
emnlp
2,023
main
EMNLP/2023/Conference
2,464
Gold: A Global and Local-aware Denoising Framework for Commonsense Knowledge Graph Noise Detection
Commonsense Knowledge Graphs (CSKGs) are crucial for commonsense reasoning, yet constructing them through human annotations can be costly. As a result, various automatic methods have been proposed to construct CSKG with larger semantic coverage. However, these unsupervised approaches introduce spurious noise that can l...
[ "Zheye Deng", "Weiqi Wang", "Zhaowei Wang", "Xin Liu", "Yangqiu Song" ]
[ "Knowledege Graph Denoising", "Commonsense Reasoning", "Question Answering" ]
EMNLP 2023 Findings
Accept-Findings
The authors present a denoising framework for Commonsense Knowledge Graphs (CSKGs), which often contain spurious data. Leveraging semantic information, global rules, and local structures, this framework differentiates between authentic and noisy triples. Emphasizing link prediction across two datasets using Recall and ...
3
[{"review_id": "vuT7PQIFnH", "reviewer": "Reviewer_68ji", "summary": "", "questions": "I am only concerned about the novelty of the implementation of global and local information modeling. Have they been separately proposed in existing studies, and you are simply combining them, or is there a new approach to combining ...
GOLD: A Global and Local-aware Denoising Framework for Commonsense Knowledge Graph Noise Detection Zheye Deng, Weiqi Wang, Zhaowei Wang, Xin Liu, Yangqiu Song Department of Computer Science and Engineering, HKUST, Hong Kong SAR, China {zdengah, wwangbw, zwanggy, xliucr, yasong) ust.hk Abstract Rene Magritte run out of ...
75,703
cNI0ZkK1yC
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
238
Flow as the Cross-domain Manipulation Interface
We present Im2Flow2Act, a scalable learning framework that enables robots to acquire real-world manipulation skills without the need of real-world robot training data. The key idea behind Im2Flow2Act is to use object flow as the manipulation interface, bridging domain gaps between different embodiments (i.e., human and...
[ "Mengda Xu", "Zhenjia Xu", "Yinghao Xu", "Cheng Chi", "Gordon Wetzstein", "Manuela Veloso", "Shuran Song" ]
[ "Robots", "Learning", "cross-domain", "cross-embodiment" ]
We present Im2Flow2Act, a scalable learning framework that enables robots to acquire manipulation skills from diverse data sources.
CoRL 2024
Accept
The paper introduces a framework for imitation learning across different domains and embodiments using image inputs. Reviewers generally agree that the paper addresses a significant problem and presents an interesting framework. However, concerns are raised about the originality of using flow as an interface, as it has...
3
[{"review_id": "zw4qqMG0rH", "reviewer": "Reviewer_WWt7", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 3, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 3, "strength...
Flow as the Cross-Domain Manipulation Interface Mengda Xu 1,2,3 Zhenjia Xu 1.2 Yinghao Xu Cheng Chi 1.2 Gordon Wetzstein Manuela Veloso Shuran Song 1.2 Stanford University, Columbia University, 3 J.P. Morgan AI Research, 4 Camegie Mellon University https:// flow- aelowwattt..... .ett.githbb, Abstract: We present Im2Flo...
87,450
W6ijeWfHFU
emnlp
2,023
main
EMNLP/2023/Conference
2,971
Improving Factual Consistency for Knowledge-Grounded Dialogue Systems via Knowledge Enhancement and Alignment
Pretrained language models (PLMs) based knowledge-grounded dialogue systems are prone to generate responses that are factually inconsistent with the provided knowledge source. In such inconsistent responses, the dialogue models fail to accurately express the external factual knowledge they rely upon. Inspired by previo...
[ "Boyang XUE", "Weichao Wang", "Hongru WANG", "Fei Mi", "Rui Wang", "Yasheng Wang", "Lifeng Shang", "Xin Jiang", "Qun Liu", "Kam-Fai Wong" ]
[ "knowledge-grounded dialogue system", "factual consistency", "knowledge enhancement" ]
EMNLP 2023 Findings
Accept-Findings
This paper presents K-DIAL which introduces extensions to the FFN in Transformer neural architecture, and apply a reinforcement learning for factual consistency method to align the responses with ground-truth knowledge. The presentation clarity in the paper leaves room for further improvement. While the authors did cla...
4
[{"review_id": "mk9C3V6JYz", "reviewer": "Reviewer_waXu", "summary": "", "questions": "* Could you provide the number of parameters used for each approach in Table 4? \n* Is the result for CMU_DoG still consistent, similar to that of Table 1 (RLFC+K-DIAL)? Could you also report that result?\n* Additional clarification ...
Improving Factual Consistency for Showledge-Groundee Dialogue Systems via Knowledge Enhancement and Alignment Boyang Xue Weichao Wang",, Hongru Wang', Fei Mi², Rui Wang', Yasheng Wang', Lifeng Shang', Xin Jiang', Qun Liu', and Kam-Fai Wong". The Chinese University of Hong Kong HHuawei Noah's s Ark Lab Harrbin Institute...
65,228
8uZ5UdIul2
iclr
2,026
main
ICLR.cc/2020/Conference
13,749
Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models
Log-likelihood evaluation enables important capabilities in generative models, including model comparison, certain fine-tuning objectives, and many downstream applications. Yet paradoxically, some of today's best generative models -- diffusion and flow-based models -- still require hundreds to thousands of neural funct...
[ "Xinyue Ai", "Yutong He", "Albert Gu", "Ruslan Salakhutdinov", "J Zico Kolter", "Nicholas Matthew Boffi", "Max Simchowitz" ]
[ "flow matching", "distillation models", "fast likelihood evaluation", "fast sampling", "generative models" ]
We propose a modular distillation framework for simultaneous fast likelihood evaluation and fast sampling for flow matching models
generative models
ICLR 2026 Poster
Accept (Poster)
4
[{"review_id": "FqoTTv958K", "reviewer": "Reviewer_tED2", "summary": "This paper introduces a technique to add log probability estimates to popular shortcut distillation techniques. They achieve this by considering the combined sample and log-density differential equation and writing down additional loss terms for the ...
Published as a conference paper at ICLR 2026 JOINT DISTILLATION FOR FAST LIKELIHOOD EVALU- ATION AND SAMPLING IN FLOW-BASED MODELS Xinyue - Yutong Gu', Ruslan Salakhutdinov', J. Zico Kolter', Nicholas M. Boffi', Max Simchowitz! 'Carnegie Mellon University, "Peking University yutonghe@cs. cmu. edu, axy253c7estu. edu. cn...
74,063
YwrNePfb3E
colm
2,024
main
colmweb.org/COLM/2024/Conference
239
Prompt Exploration with Prompt Regression
In the advent of democratized usage of large language models (LLMs), there is a growing desire to systematize LLM prompt creation and selection processes beyond iterative trial-and-error. Prior works majorly focus on searching the space of prompts without accounting for relations between prompt variations. Here we prop...
[ "Michael Feffer", "Ronald Xu", "Yuekai Sun", "Mikhail Yurochkin" ]
[ "prompt engineering, prompt selection, large language models, regression" ]
We introduce a regression method to predict effects of prompt combinations as well as select effective prompts based on the regression model accordingly.
COLM
Accept
This paper explores the problem of automatically identifying suitable prompts. This work takes advantage of similarities between prompt variations and their performance, and proposes a method called Prompt Exploration with Prompt Regression that predicts performance of prompt combinations given performance of component...
4
[{"review_id": "0BLyzJ7djC", "reviewer": "Reviewer_t5Qb", "summary": "This work studies the prompt library search problem where the model needs to predict prompt combinations in a library that can lead to optimal performance. This work further proposes a method including three steps: (1) building a prompt library; (2) ...
Published as a conference paper at COLM 2024 Prompt Exploration with Prompt Regression Michael Feffer *+ Ronald Xu Software and Societal Systems Department Carnegie Mellon University EECS Department Massachusetts Institute of Technology MIT-IBM Watson Al Lab mfeffer@andres www.ccmu.cmmmm edu ronaldxu@mit edu Yuekai Sun...
62,708
XeeTWJvAQl
neurips
2,021
main
NeurIPS.cc/2021/Conference
6,970
High-probability Bounds for Non-Convex Stochastic Optimization with Heavy Tails
We consider non-convex stochastic optimization using first-order algorithms for which the gradient estimates may have heavy tails. We show that a combination of gradient clipping, momentum, and normalized gradient descent yields convergence to critical points in high-probability with best-known rates for smooth losses ...
[ "Ashok Cutkosky", "Harsh Mehta" ]
[ "stochastic optimization", "heavy tails", "non-convex optimization", "optimization for deep learning" ]
NeurIPS 2021 Oral
Accept (Oral)
The paper considers non-convex SO using first-order algorithms for which the gradient estimates may have heavy tail distributions. Using gradient clipping normalized GD with momentum is shown to convergence to near critical points in high-probability with best-known rates for smooth losses when the gradients only have ...
3
[{"review_id": "j76n3qQkW7", "reviewer": "Reviewer_WuQp", "summary": "This paper focuses on high probability convergence behavior of nnormalized SGD with momentum and gradient clipping under heavy-tailed noise but finite moments. Specifically, without the light-tail assumption, e.g. sub-Gaussian noise, and with bounded...
High-probability bounds for Non-Convex Stochastic Optimization with Heavy Tails Ashok Cutkosky Boston University ashokOcutkosky.o com Harsh Mehta Google Research harshn@goggle.com com Abstract We consider non-convex stochastic optimization using first-order algorithms for which the gradient estimates may have heavy tai...
36,046
OPOBV0zXu7
neurips
null
NeurIPS.cc/2021/Conference
14,500
Multi-Scale Finetuning for Encoder-based Time Series Foundation Models
Time series foundation models (TSFMs) demonstrate impressive zero-shot performance for time series forecasting. However, an important yet underexplored challenge is how to effectively finetune TSFMs on specific downstream tasks. While naive finetuning can yield performance gains, we argue that it falls short of fully l...
[ "Zhongzheng Qiao", "Chenghao Liu", "Yiming Zhang", "Ming Jin", "Quang Pham", "Qingsong Wen", "Ponnuthurai Nagaratnam Suganthan", "Xudong Jiang", "Savitha Ramasamy" ]
[ "Time Series Foundation Model", "Time Series Forecasting", "Multiscale modeling", "Finetuning" ]
deep_learning
NeurIPS 2025 poster
Accept (poster)
The paper argues that current fine-tuning techniques for time series (TS) foundation models are suboptimal because they fail to exploit the inherent multi-scale nature of the data and the ability of foundation models to capture dynamics at multiple temporal resolutions. The authors propose a causal analysis of fine-tun...
5
[{"review_id": "GgYmq6lfgY", "reviewer": "Reviewer_7Ko1", "summary": "This paper focuses on the task of fine-tuning time series foundation model. To address the issues of overfitting and suboptimal performance of naive fine-tuning methods, the paper develops a diversified fine-tuning approach based on different samplin...
Multi-Scale Finetuning for Encoder-based Time Series Foundation Models Zhongzheng Qiao!:2 1,2,3 Chenghao Liu' Yiming Zhang' Ming Jin Quang Pham 4 Qingsong Wen P.N.Suantthan 7 Xudong Jiang' Savitha Ramasamy 2,3 Abstract Time series foundation models (TSFMs) demonstrate impressive zero-shot per- formance for time series ...
114,731
B41hNBoWLo
colm
2,024
main
colmweb.org/COLM/2024/Conference
424
TOFU: A Task of Fictitious Unlearning for LLMs
Large language models trained on massive corpora of data from the web can memorize and reproduce sensitive or private data raising both legal and ethical concerns. Unlearning, or tuning models to forget information present in their training data, provides us with a way to protect private data after training. Although s...
[ "Pratyush Maini", "Zhili Feng", "Avi Schwarzschild", "Zachary Chase Lipton", "J Zico Kolter" ]
[ "Machine unlearning" ]
Synthetic benchmark for machine unlearning for LLMs
COLM
Accept
This paper presents a benchmark, metrics, and implementation of baselines for unlearning LLM, which is an urgent and important topic. The paper is well-written and the benchmark is well-thought. There are some concerns about whether the dataset can represent real-world scenarios as the dataset is artificial and smal...
4
[{"review_id": "DsdNAP8vI8", "reviewer": "Reviewer_Q175", "summary": "This paper proposes a new dataset of 200 diverse synthetic author profiles used to control the experiments for unlearning. It also compiles a diverse set of metrics to holistically evaluate unlearning and observes that current unlearning baselines fa...
Published as a conference paper at COLM 2024 TOFU: A Task of Fictitious Unlearning for LLMs Pratyush Maini' Zhili Feng" pratyushmaini@cm edu zhilif@ zhilif@andrew. cmu, edu Carnegie Mellon University Carnegie Mellon University Avi Schwarzschild" schwarzschil Idemmu edu Carnegie Mellon University Zachary C. Lipton Carne...
76,879
19uudhc1s8
emnlp
2,023
main
EMNLP/2023/Conference
801
Analyzing Film Adaptation through Narrative Alignment
Novels are often adapted into feature films, but the differences between the two media usually require dropping sections of the source text from the movie script. Here we study this screen adaptation process by constructing narrative alignments using the Smith-Waterman local alignment algorithm coupled with SBERT embed...
[ "Tanzir Pial", "Shahreen Salim Aunti", "Charuta Pethe", "Allen Kim", "Steven Skiena" ]
[ "Text Alignment", "Book Movie Alignment" ]
We analyze the book to film adaptation process through aligning source book text with the the film adaptation's script.
EMNLP 2023 Main
Accept-Main
There is a perfect agreement among reviewers about the soundness and excitement of this paper. Some reasons to reject do not have to do with soundness or excitement about the paper (research question is too niche, neglect of cinematic elements, literature on alignment, …), but may be explored in future work. Paper Top...
3
[{"review_id": "Oq7ffjg07P", "reviewer": "Reviewer_MAg3", "summary": "", "questions": "A)\tGiven the limitations of a text-centric approach, have you explored how incorporating visual or auditory analysis to capture essential cinematic elements could enhance the quality of interpreting book-film adaptations?", "limitat...
Analyzing Film Adaptation through Narrative Alignment Tanzir Pial, Shahreen Salim, Charuta Pethe, Allen Kim, Steven Skiena Department of Computer Science, Stony Brook University, NY, USA (tpial, ssalimaunti,,peehe,, allekim, skiieaa@esssttonbbrrdd edu Abstract tation process. Novels are often adapted into feature films...
66,396
0VQImEvjPJ
emnlp
2,023
main
EMNLP/2023/Conference
3,912
NormDial: A Comparable Bilingual Synthetic Dialog Dataset for Modeling Social Norm Adherence and Violation
Social norms fundamentally shape interpersonal communication. We present NormDial, a high-quality dyadic dialogue dataset with turn-by-turn annotations of social norm adherences and violations for Chinese and American cultures. Introducing the task of social norm observance detection, our dataset is synthetically gener...
[ "Oliver Li", "Mallika Subramanian", "Arkadiy Saakyan", "Sky CH-Wang", "Smaranda Muresan" ]
[ "social norms", "resources and evaluation", "large language models" ]
EMNLP 2023 Main
Accept-Main
This short paper introduces a dataset of dialogues which are annotated as to whether they adhere to social norms. The dataset contains English and Chinese dialogues. The dialogues are synthetic - they were generated using ChatGPT and then annotated. The authors also evaluate whether ChatGPT can recognize norm violation...
3
[{"review_id": "FdF1nDPq1y", "reviewer": "Reviewer_nowG", "summary": "", "questions": "", "limitations": "", "rating": 3, "confidence": 4, "soundness": 4, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "strength...
NORMDIAL: A Comparable Bilingual Synthetic Dialogue Dataset for Modeling Social Norm Adherence and Violation Oliver Li Mallika Subramanian Arkadiy Saakyan Sky CH-Wang' Smaranda Muresan Department of Computer Science, Columbia University Data Science Institute, Columbia University (a14143, ms6544, smara]ecolumbil edu, {...
45,530
1dd7q3Ktkz
icml
2,025
main
ICML.cc/2025/Conference
9,292
Towards a Unified Framework of Clustering-based Anomaly Detection
Unsupervised Anomaly Detection (UAD) plays a crucial role in identifying abnormal patterns within data without labeled examples, holding significant practical implications across various domains. Although the individual contributions of representation learning and clustering to anomaly detection are well-established, t...
[ "Zeyu Fang", "Ming Gu", "Sheng Zhou", "Jiawei Chen", "Qiaoyu Tan", "Haishuai Wang", "Jiajun Bu" ]
[ "Anomaly Detection", "Clustering" ]
A unified theoretical framework models the intrinsic connections among representation learning, clustering, and anomaly detection.
deep_learning->generative_models_and_autoencoders
ICML 2025 poster
Accept (poster)
This paper received one accept, two weak accepts, and one weak reject. All reviewers agree that the combined use of clustering and representation learning for anomaly detection is a valid and meaningful approach. Initial concerns were raised regarding the use of a t-distribution in place of a Gaussian, the adoption of ...
4
[{"review_id": "1pr7mk8Rco", "reviewer": "Reviewer_GPqn", "summary": "The paper proposes UniCAD, a novel model for Unsupervised Anomaly Detection (UAD) that unifies representation learning, clustering, and anomaly detection within a single theoretical framework. By leveraging a mixture model with the Student-t distribu...
Towards a Unified Framework of Clustering- based Anomaly Detection Zeyu Fang' Ming Gu Sheng Zhou' Jiawei Chen? Qiaoyu Tan Haishuai Wang* Jiajun Bu Abstract Representation Learning Clustering Unsupervised Anomaly Detection (UAD) plays a crucial role in identifying abnormal patterns within data without labeled examples, ...
65,256
oaCUsn391F
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,280
SlimMoE: Structured Compression of Large MoE Models via Expert Slimming and Distillation
The Mixture of Experts (MoE) architecture has emerged as a powerful paradigm for scaling large language models (LLMs) while maintaining inference efficiency. However, their substantial memory requirements make them prohibitively expensive to fine-tune or deploy in resource-constrained environments. To address this chal...
[ "Zichong Li", "Chen Liang", "Zixuan Zhang", "Ilgee Hong", "Young Jin Kim", "Weizhu Chen", "Tuo Zhao" ]
[ "Large Language Model", "Mixture of Experts", "Structured Pruning", "Knowledge Distillation" ]
This paper introduces a novel multi-stage prune-and-distill framework that efficiently compresses Phi-MoE 3.5 into compact 7.6B and 3.8B parameter models that significantly outperform similarly-sized alternatives.
COLM 2025
Accept
In this paper the authors propose a methodology for compressing MoE language models that combines structured pruning (within experts) with knowledge distillation (general-purpose, on hundreds of billions of tokens of pretraining data). They demonstrate the approach on Phi-3 MoE models (1-7B range) across a variety of t...
3
[{"review_id": "IL2hxUDqPO", "reviewer": "Reviewer_6UVt", "summary": "This paper proposes SlimMoE, a multi-stage compression framework for large Mixture-of-Experts (MoE) language models. By combining structured pruning (specifically, intra-expert neuron slimming) with staged knowledge distillation, the method compresse...
Published as a conference paper at COLM 2025 SlimMoE: Structured Compression of Large MoE Models via Expert Slimming and Distillation Zichong Li Chen Liang"!Ziiuaa Zhang', Ilgee Hong', Young Jin Kim², Weizhu Chen?, Tuo Zhao GGeoriia Institute of Technology Microsoff Abstract The Mixture of Experts (MoF) architecture ha...
66,616
QsQatTzATT
colm
2,025
main
colmweb.org/COLM/2024/Conference
1,118
Humans overrely on overconfident language models, across languages
As large language models (LLMs) are deployed globally, it is crucial that their responses are calibrated across languages to accurately convey uncertainty and limitations. Prior work shows that LLMs are linguistically overconfident in English, leading users to overrely on confident generations. However, the usage and i...
[ "Neil Rathi", "Dan Jurafsky", "Kaitlyn Zhou" ]
[ "multilingual language models", "uncertainty" ]
Multilingual LLMs are overconfident across languages, and that users overrely on confident responses
COLM 2025
Accept
This paper evaluates three language models across five languages, analyzing the differences between English and non-English outputs through human evaluations and model overconfidence, finding that model over-confidence differs across linguistic settings. The paper is clear and well-explained, and the authors engaged ex...
4
[{"review_id": "hb1JftjGfW", "reviewer": "Reviewer_EaEa", "summary": "This paper addresses the important and timely issue of model overconfidence in multilingual settings. The authors evaluate three language models from two model families across five languages, analyzing disparities between English and non-English outp...
Published as a conference paper at COLM 2025 Humans overrely on overconfident language models, across languages Neil Rathi, Dan Jurafsky, Kaitlyn Zhou Stanford University rathi@stanford.cm Abstract As large language models (LLMs) are deployed globally, it is crucial that their responses are calibrated across languages ...
57,740
kjNJYWvfPA
colm
2,025
main
colmweb.org/COLM/2024/Conference
510
How Multimodal LLMs Solve Image Tasks: A Lens on Visual Grounding, Task Reasoning, and Answer Decoding
Multimodal Large Language Models (MLLMs) have demonstrated strong performance across a wide range of vision-language tasks, yet their internal processing dynamics remain underexplored. In this work, we introduce a probing framework to systematically analyze how MLLMs process visual and textual inputs across layers. We ...
[ "Zhuoran Yu", "Yong Jae Lee" ]
[ "Multimodal LLM", "Interpretability" ]
We propose a probing framework to analyze how multimodal LLMs process visual and textual inputs across layers
COLM 2025
Accept
This paper presents a thoughtful study on the internal workings of Multimodal LLMs using a novel probing framework. I find the research question timely and the approach clever. The reviewers raised several valid points during the initial review phase. They questioned the limited scope and specific design choices, which...
4
[{"review_id": "05sp8Gs2KI", "reviewer": "Reviewer_oi97", "summary": "This paper investigates how recent multimodal large language models (MLLMs) process information across their layers, aiming to determine whether there is a hierarchical structure in how these models handle different types of information. The main obj...
Published as a conference paper at COLM 2025 How Multimodal LLMs Solve Image Tasks: A Lens on Visual Grounding, Task Reasoning, and Answer Decoding Zhuoran Yu University of Wisconsin-Madison zhuoran. Mnuoran.y@@vicc.ee yu@wisc. Yong Jae Lee University of yongjaelee@css. wisc.edu Abstract Multimodal Large Language Model...
43,994
E22I6z7qFy
iclr
2,026
main
ICLR.cc/2020/Conference
11,024
Learning Adaptive Distribution Alignment with Neural Characteristic Function for Graph Domain Adaptation
Graph Domain Adaptation (GDA) transfers knowledge from labeled source graphs to unlabeled target graphs but is challenged by complex, multi-faceted distributional shifts. Existing methods attempt to reduce distributional shifts by aligning manually selected graph elements (e.g., node attributes or structural statistics...
[ "Wei Chen", "Xingyu Guo", "Shuang Li", "Zhao Zhang", "Yan Zhong", "Fuzhen Zhuang", "deqing wang" ]
[ "Graph Domain Adaptation", "Graph Neural Networks", "Characteristic Function" ]
transfer learning, meta learning, and lifelong learning
ICLR 2026 Poster
Accept (Poster)
4
[{"review_id": "JkTEYZOSbN", "reviewer": "Reviewer_A6vr", "summary": "This paper introduced the ADAlign framework, a novel approach to handling composite\ndistribution shifts in GDA. By leveraging characteristic function in the complex Fourier domain, ADAlign dynamically identifies and aligns discriminative spectral co...
Published as a conference paper at ICLR 2026 LEARNING ADAPTIVE DISTRIBUTION ALIGNMENT WITH NEURAL CHARACTERISTIC FUNCTION FOR GRAPH DOMAIN ADAPTATION Wei Chen', Xingyu Guo', Shuang Li Zhao Zhang?, Yan Zhong', Fuzhen Zhuang Deqing Wang? 'School of Artificial Intelligence, Beihang University, Beijing. China Schooo of Com...
76,611
WXdSp8k0TMn
neurips
2,022
main
NeurIPS.cc/2021/Conference
6,669
Revisiting Non-Parametric Matching Cost Volumes for Robust and Generalizable Stereo Matching
Stereo matching is a classic challenging problem in computer vision, which has recently witnessed remarkable progress by Deep Neural Networks (DNNs). This paradigm shift leads to two interesting and entangled questions that have not been addressed well. First, it is unclear whether stereo matching DNNs that are trained...
[ "Kelvin Cheng", "Tianfu Wu", "Christopher G. Healey" ]
[ "Stereo Matching", "Contextualized Non-Parametric Cost Volume", "Adversarial Robustness", "Simulation-to-Real Generalizability" ]
NeurIPS 2022 Accept
Accept
4
[{"review_id": "lylcQm-JXUK", "reviewer": "Reviewer_WPxi", "summary": "The paper considers the problem of adversarial on binocular stereo matching systems. It shows that an adversarial attack based on PGD that is photometrically consistent and consistent with stereo constraints significantly affects the performance of ...
Revisiting Non-Parametric Matching Cost Volumes for Robust and Generalizable Stereo Matching Kelvin Cheng', Tianfu Wu and Christopher Healey? CSt and ECE at NC State University, Raleigh NC 27695 (kbcheng, twu19, healey)(ncsu.een Abstract Stereo matching is a classic challenging problem in computer vision, which has rec...
56,447
z6GEZ2ogct
icml
2,025
main
ICML.cc/2025/Conference
9,875
DRAG: Data Reconstruction Attack using Guided Diffusion
With the rise of large foundation models, split inference (SI) has emerged as a popular computational paradigm for deploying models across lightweight edge devices and cloud servers, addressing data privacy and computational cost concerns. However, most existing data reconstruction attacks have focused on smaller CNN c...
[ "Wa-Kin Lei", "Jun-Cheng Chen", "Shang-Tse Chen" ]
[ "Data Reconstruction Attack", "Privacy", "Diffusion Model" ]
We use diffusion model as the image prior to improve data reconstruction attack in the context of split inference.
social_aspects->privacy
ICML 2025 poster
Accept (poster)
This paper proposes DRAG, a new data reconstruction attack against split inference. The key idea is to use a pre-trained diffusion model as the image prior, which significantly reduces the dimensionality of the feature inversion optimization problem. The authors show that DRAG is able to invert deeper layer representat...
4
[{"review_id": "Tk5jnvfCH3", "reviewer": "Reviewer_u9u9", "summary": "The paper introduces a new reconstruction attack method, DRAG (Data Reconstruction Attack using Guided Diffusion), that reconstructs private data from intermediate representations in split Inference settings. Unlike previous attacks on small CNNs, DR...
DRAG: Data Reconstruction Attack using Guided Diffusion Wa-Kin Lei Jun-Cheng Chen 2 Shang-Tse Chen Abstract promising solutions, as it balances computational and pri- vacy concerns. This approach enables efficient utilization of cloud resources, reduces the computational burden on local devices, and facilitates the int...
60,207
tl3FlgWScA
icml
2,025
main
ICML.cc/2025/Conference
3,677
Ad Hoc Teamwork via Offline Goal-Based Decision Transformers
The ability of agents to collaborate with previously unknown teammates on the fly, known as ad hoc teamwork (AHT), is crucial in many real-world applications. Existing approaches to AHT require online interactions with the environment and some carefully designed teammates. However, these prerequisites can be infeasible...
[ "Xinzhi Zhang", "Hohei Chan", "Deheng Ye", "Yi Cai", "Mengchen Zhao" ]
[ "Ad Hoc Teamwork", "Offline Reinforcement Learning" ]
This paper frames offline ad hoc teamwork as a sequence modeling problem and proposes goal-based Decision Transformers to train the ego agent for effective collaboration with unknown teammates.
reinforcement_learning->multiagent
ICML 2025 poster
Accept (poster)
This paper proposes an approach for learning an ad hoc teamwork policy offline. It does so by making modifications to the decision transformer approach as well as additional pre-processing of the offline data to enable easier identification of teammate goals and capabilities. The proposed approach seems to be the first...
4
[{"review_id": "rjjSKyoGl9", "reviewer": "Reviewer_cnPV", "summary": "This paper addresses ad hoc teamwork in the offline setting. It proposes a method called TAGET, which is based off the decision-transformer architecture and learns from a dataset of offline cooperative multi-agent interactions. It has a couple of mai...
Ad Hoc Teamwork via Offline Goal-Based Decision Transformers Xinzhi Zhang Hohei Chan 1 Deheng Ye Yi Cai Mengehen Zhao Abstract Previous research on AHT primarily focuses on online rein- forcement learning (RL) methods (Barrett & Stone, 2015; Durugkar et al., 2020; Mirsky et al., 2020; Ye et al., 2020), which typically ...
57,170
zHdSCtNmM4
colm
2,025
main
colmweb.org/COLM/2024/Conference
320
Deep Binding of Language Model Virtual Personas: a Study on Approximating Political Partisan Misperceptions
Large language models (LLMs) are increasingly capable of simulating human behavior, offering cost-effective ways to estimate user responses during the early phases of survey design. While previous studies have examined whether models can reflect individual opinions or attitudes, we argue that a higher-order binding of ...
[ "Minwoo Kang", "Suhong Moon", "Seung Hyeong Lee", "Ayush Raj", "Joseph Suh", "David Chan" ]
[ "user approximation", "metaperception", "social psycholog", "democratic backsliding", "outgroup hostility" ]
We propose a method to build virtual personas for deeper user binding and demonstrate its superiority in approximating metaperception in political science.
COLM 2025
Accept
The reviewers were unanimous in their recommendation to accept this paper (all 7s). Additionally, the authors provided very thorough responses to reviewer concerns, including new experiments on non-political domains (ATP Waves 34 & 99), detailed demographic breakdowns showing consistent improvements across subgroups, a...
4
[{"review_id": "YVZRHVPBuk", "reviewer": "Reviewer_3XgY", "summary": "Most work on \"silicon samples\" or \"virtual personas\" check whether the LLM is mimicking a person's own opinions and beliefs. This paper goes beyond this and looks at whether LLMs are able to reproduce higher-order bindings: how people think about...
Published as a conference paper at COLM 2025 Deep Binding of Language Model Virtual Personas: Study on Approximating Political Partisan Misperceptions Minwoo Kang" Suhong Moon Seung Hyeong Lee: Ayush Raj* Joseph Suh! David M. Chan' 'University of California, Berkeley, WNorthwestern University, (minwoo kaggsuuongg.oon/B...
161,317
fIj88Tn3fc
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
634
ReMix: Optimizing Data Mixtures for Large Scale Imitation Learning
Increasingly large robotics datasets are being collected to train larger foundation models in robotics. However, despite the fact that data selection has been of utmost importance to scaling in vision and natural language processing (NLP), little work in robotics has questioned what data such models should actually be ...
[ "Joey Hejna", "Chethan Anand Bhateja", "Yichen Jiang", "Karl Pertsch", "Dorsa Sadigh" ]
[ "Data Curation", "Data Quality", "Robot Imitation Learning" ]
We use techniques from robust optimization to learn data mixture weights for Bridge and RT-X datasets, and show they improve downstream performance.
CoRL 2024
Accept
**Paper summary** This paper adapts techniques from NLP to improve the generalizability of robot foundation models by re-weighting datapoints in large-scale datasets. The paper presents extensive experiments in real-world and simulated tasks, demonstrating how the method improves the data-efficiency and generalizabili...
3
[{"review_id": "9mY7sCwHYq", "reviewer": "Reviewer_aygU", "summary": "", "questions": "", "limitations": "", "rating": 4, "confidence": 5, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 4, "strength...
ReMix: Optimizing Data Mixtures for Large Scale Imitation Learning Joey Hejna Stanford Chethan Bhateja Stanford Yichen Jiang Stanford Karl Pertsch Stanford, UC Berkeley Dorsa Sadigh Stanford Abstract: Increasingly large imitation learning datasets are being collected with the goal of training foundation models for robo...
68,686
rThtgkXuvZ
corl
2,024
main
robot-learning.org/CoRL/2021/Conference
434
NOD-TAMP: Generalizable Long-Horizon Planning with Neural Object Descriptors
Solving complex manipulation tasks in household and factory settings remains challenging due to long-horizon reasoning, fine-grained interactions, and broad object and scene diversity. Learning skills from demonstrations can be an effective strategy, but such methods often have limited generalizability beyond training ...
[ "Shuo Cheng", "Caelan Reed Garrett", "Ajay Mandlekar", "Danfei Xu" ]
[ "Robot Learning", "Robot Planning", "Manipulation" ]
We introduce NOD-TAMP, a TAMP-based framework that can solve broad long-horizon manipulation tasks by adapting and composing short manipulation trajectories from a handful of human demonstration.
CoRL 2024
Accept
The reviewers raised concerns regarding the fact that a plan skeleton is provided, so the task planning is essentially given, lack of implementation details, unstated assumptions. The rebuttal submitted by the authors addressed these concerns to a large extent, and two out of three reviewers recommend acceptance.
3
[{"review_id": "RRyNIWMUbv", "reviewer": "Reviewer_3zeS", "summary": "", "questions": "", "limitations": "", "rating": 2, "confidence": 5, "soundness": null, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": 2, "strength...
NOD-TAMP: Generalizable Long-Horizon Planning with Neural Object Descriptors Shuo Cheng , Caelan Garrett" 2, Ajay Mandlekar" Danfei Xu' GGoorgia Institute of Technology "NVIDIA Corporation Abstract: Solving complex manipulation tasks in household and factory settings remains challenging due to long-horizon reasoning, f...
64,923