ICML

When Meaning Stays the Same, but Models Drift: Evaluating Quality of Service under Token-Level Behavioral Instability in LLMs

Xiao Li, Joel Kreuzwieser, Alan Peters

机构 * Electrical and Computer Engineering, Vanderbilt University（维珍尼亚大学电气与计算机工程系） ； Radiology and Radiological Sciences, Vanderbilt University（维珍尼亚大学放射学与放射科学系） 纠错

Comments This paper was developed for presentation at ICML 2025 Tokshop Workshop, but is now submitted as a standalone contribution

scSSL-Bench: Benchmarking Self-Supervised Learning for Single-Cell Data

Olga Ovcharenko, Florian Barkmann, Philip Toma, Imant Daunhawer, Julia Vogt, Sebastian Schelter, Valentina Boeva

Comments Accepted at ICML 2025 (Spotlight)

FedRAG: A Framework for Fine-Tuning Retrieval-Augmented Generation Systems

Val Andrei Fajardo, David B. Emerson, Amandeep Singh, Veronica Chatrath, Marcelo Lotif, Ravi Theja, Alex Cheung, Izuki Matsuba

机构 * Vector Institute（向量研究所） 纠错

Comments 9 pages, 4 figures, 2 tables. Accepted for the CODEML Workshop at ICML 2025. Framework code available at https://github.com/VectorInstitute/fed-rag

ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Preference Optimization

Hee Suk Yoon, Eunseop Yoon, Mark Hasegawa-Johnson, Sungwoong Kim, Chang D. Yoo

机构 * Korea Advanced Institute of Science and Technology (KAIST)（韩国科学技术院） ； University of Illinois Urbana-Champaign (UIUC)（伊利诺伊大学厄巴纳-香槟分校） ； Korea University（韩国大学） 纠错

Comments ICML 2025

When Maximum Entropy Misleads Policy Optimization

Ruipeng Zhang, Ya-Chien Chang, Sicun Gao

Journal ref ICML 2025

Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Models

Patrick Leask, Neel Nanda, Noura Al Moubayed

机构 * Department of Computer Science, Durham University（杜伦大学计算机科学系） 纠错

Comments ICML 2025

Elucidating the Design Space of Multimodal Protein Language Models

Cheng-Yen Hsieh, Xinyou Wang, Daiheng Zhang, Dongyu Xue, Fei Ye, Shujian Huang, Zaixiang Zheng, Quanquan Gu

Comments ICML 2025 Spotlight; Project Page: https://bytedance.github.io/dplm/dplm-2.1/

An energy-efficient learning solution for the Agile Earth Observation Satellite Scheduling Problem

Antonio M. Mercado-Martínez, Beatriz Soret, Antonio Jurado-Navas

机构 * University of Malaga（马拉加大学） 纠错

Comments This paper has been accepted for presentation at the IEEE International Conference on Machine Learning for Communication and Networking (ICMLCN) Special Sessions 2025

Heavy-Tailed Linear Bandits: Huber Regression with One-Pass Update

Jing Wang, Yu-Jie Zhang, Peng Zhao, Zhi-Hua Zhou

机构 * National Key Laboratory for Novel Software Technology, Nanjing University, China（新型软件技术国家实验室，南京大学，中国） ； School of Artificial Intelligence, Nanjing University, China（人工智能学院，南京大学，中国） ； Center for Advanced Intelligence Project, RIKEN, Japan（先进人工智能项目中心，RIKEN，日本） 纠错

Comments ICML 2025

Density Ratio Estimation with Conditional Probability Paths

Hanlin Yu, Arto Klami, Aapo Hyvärinen, Anna Korba, Omar Chehab

机构 * University of Helsinki, Finland（赫尔辛基大学） 纠错

Comments To appear in ICML 2025

Few-Shot Learner Generalizes Across AI-Generated Image Detection

Shiyu Wu, Jing Liu, Jing Li, Yequan Wang

机构 * Institute of Automation, Chinese Academy of Sciences, Beijing, China（中国科学院自动化研究所） ； University of Chinese Academy of Sciences, Beijing, China（中国科学院大学） ； Harbin Institute of Technology, Shenzhen, China（哈尔滨工业大学（深圳）） ； Beijing Academy of Artificial Intelligence, Beijing, China（北京人工智能研究院） 纠错

Comments 12 pages, 6 figures, Accepted at ICML 2025

Position: Theory of Mind Benchmarks are Broken for Large Language Models

Matthew Riemer, Zahra Ashktorab, Djallel Bouneffouf, Payel Das, Miao Liu, Justin D. Weisz, Murray Campbell

机构 * IBM Research（IBM研究院） 纠错

Comments ICML 2025

Decision Making under the Exponential Family: Distributionally Robust Optimisation with Bayesian Ambiguity Sets

Charita Dellaporta, Patrick O'Hara, Theodoros Damoulas

机构 * Department of Statistics, University of Warwick, Coventry, UK（威斯敏斯特大学统计学系） ； Department of Computer Science, University of Warwick, Coventry, UK（威斯敏斯特大学计算机科学系） ； Department of Statistics, University College London, UK（伦敦大学学院统计学系） ； The Alan Turing Institute, London, UK（艾伦·图灵研究所） 纠错

Comments Accepted for publication (spotlight) at ICML 2025

Simplicity bias and optimization threshold in two-layer ReLU networks

Etienne Boursier, Nicolas Flammarion

机构 * INRIA, LMO, Université Paris-Saclay, Orsay, France（INRIA、LMO、巴黎-萨克雷大学、欧萨斯分校、法国） ； TML Lab, EPFL, Switzerland（TML实验室、瑞士联邦理工学院） 纠错

Comments ICML camera ready version

Efficient Length-Generalizable Attention via Causal Retrieval for Long-Context Language Modeling

Xiang Hu, Zhihao Teng, Jun Zhao, Wei Wu, Kewei Tu

机构 * ShanghaiTech University（上海科技大学） ； Fudan University（复旦大学） 纠错

Comments accepted to ICML 2025

Flexible Tails for Normalizing Flows

Tennessee Hickling, Dennis Prangle

机构 * School of Mathematics University of Bristol（数学学院 英国布里斯托尔大学） 纠错

Comments Accepted at ICML 2025

Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance

Linxi Zhao, Yihe Deng, Weitong Zhang, Quanquan Gu

Comments 25 pages, 13 figures, 25 tables

Journal ref In Proceedings of the 42nd International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025

The Sample Complexity of Online Strategic Decision Making with Information Asymmetry and Knowledge Transportability

Jiachen Hu, Rui Ai, Han Zhong, Xiaoyu Chen, Liwei Wang, Zhaoran Wang, Zhuoran Yang

机构 * School of Computer Science, Peking University（北京大学计算机科学学院） ； Center for Data Science, Peking University（北京大学数据科学中心） ； National Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University（北京大学通用人工智能国家重点实验室） ； Northwestern University（西北大学） ； Yale University（耶鲁大学） 纠错

Comments Accepted at ICML 2025

The Four Color Theorem for Cell Instance Segmentation

Ye Zhang, Yu Zhou, Yifeng Wang, Jun Xiao, Ziyue Wang, Yongbing Zhang, Jianxu Chen

机构 * School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), China（计算机科学与技术学院，哈尔滨工业大学（深圳）） ； School of Science, Harbin Institute of Technology (Shenzhen), China（科学学院，哈尔滨工业大学（深圳）） ； Department of Electrical and Computer Engineering, National University of Singapore, Singapore（电气与计算机工程系，新加坡国立大学） 纠错

Comments Accepted at ICML 2025

Do Multiple Instance Learning Models Transfer?

Daniel Shao, Richard J. Chen, Andrew H. Song, Joel Runevic, Ming Y. Lu, Tong Ding, Faisal Mahmood

机构 * Massachusetts Institute of Technology（麻省理工学院） ； Harvard University（哈佛大学） 纠错

Comments ICML 2025 (Spotlight). 20 pages, 8 figures

NTPP: Generative Speech Language Modeling for Dual-Channel Spoken Dialogue via Next-Token-Pair Prediction

Qichao Wang, Ziqiao Meng, Wenqian Cui, Yifei Zhang, Pengcheng Wu, Bingzhe Wu, Irwin King, Liang Chen, Peilin Zhao

机构 * Tencent（腾讯） ； National University of Singapore（新加坡国立大学） ； The Chinese University of Hong Kong（香港中文大学） ； Nanyang Technological University（南洋理工大学） ； Shenzhen University（深圳大学） 纠错

Comments Accepted by ICML 2025

Product of Experts with LLMs: Boosting Performance on ARC Is a Matter of Perspective

Daniel Franzen, Jan Disselhoff, David Hartmann

机构 * Johannes Gutenberg University Mainz（莱茵河畔明斯特大学） ； Lambda, Inc.（Lambda公司） 纠错

Comments ICML 2025 camera-ready; 15 pages, 6 figures, 5 tables

Griffin: Towards a Graph-Centric Relational Database Foundation Model

Yanbo Wang, Xiyuan Wang, Quan Gan, Minjie Wang, Qibin Yang, David Wipf, Muhan Zhang

机构 * Institute for Artificial Intelligence, Peking University（北京大学人工智能研究院） ； Amazon Web Services（亚马逊网络服务） 纠错

Comments Published at ICML 2025

Conformal Prediction as Bayesian Quadrature

Jake C. Snell, Thomas L. Griffiths

机构 * Department of Psychology, Princeton University（心理学系，普林斯顿大学） ； Department of Computer Science, Princeton University（计算机科学系，普林斯顿大学） 纠错

Comments ICML 2025 camera-ready version (accepted as an oral presentation). 16 pages, 4 figures. Code available at https://github.com/jakesnell/conformal-as-bayes-quad

Scalable First-order Method for Certifying Optimal k-Sparse GLMs

Jiachang Liu, Soroosh Shafiee, Andrea Lodi

机构 * School of Operations Research and Information Engineering, Cornell University（运营研究与信息工程学院，康奈尔大学） ； Jacobs Technion-Cornell Institute, Cornell Tech（雅各布斯技术ion-康奈尔研究所，康奈尔科技） 纠错

Comments ICML 2025 camera ready, typo fixed

No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces

Daniel Marczak, Simone Magistri, Sebastian Cygert, Bartłomiej Twardowski, Andrew D. Bagdanov, Joost van de Weijer

机构 * Warsaw University of Technology, Poland ； IDEAS Research Center, Warsaw, Poland ； Department of Information Engineering, Media Integration ； Communication Center (MICC), University of Florence, Italy ； NASK - PIB, National Research Institute, Warsaw, Poland ； Gdańsk University of Technology, Poland ； Computer Vision Center, Barcelona, Spain ； Department of Computer Science, Universitat Autonoma de Barcelona, Spain 纠错

Comments Accepted at ICML 2025

CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization

Nay Myat Min, Long H. Pham, Yige Li, Jun Sun

机构 * School of Computing and Information Systems, Singapore Management University, Singapore（计算与信息系统学院，新加坡国立管理大学） 纠错

Comments Accepted at ICML 2025, 20 pages

Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing

Kento Nishi, Rahul Ramesh, Maya Okawa, Mikail Khona, Hidenori Tanaka, Ekdeep Singh Lubana

机构 * CBS-NTT Program in Physics of Intelligence, Harvard University ； Harvard College ； Informatics Lab, NTT Research Inc. ； Information Science, University of Pennsylvania ； Department of Physics, Massachusetts Institute of Technology 纠错

Comments Accepted to ICML 2025

Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment

Yifan Zhang, Ge Zhang, Yue Wu, Kangping Xu, Quanquan Gu

机构 * IIIS, Tsinghua University, Beijing, China（清华大学信息科学技术学院，北京，中国） ； Shanghai Qi Zhi Institute, Shanghai, China（上海启智研究院，上海，中国） ； Department of Computer Science, University of California, Los Angeles, California, USA（计算机科学系，加州大学洛杉矶分校，美国，加州） 纠错

Comments Accepted to the 42nd International Conference on Machine Learning (ICML 2025)

Gradient Aligned Regression via Pairwise Losses

Dixian Zhu, Tianbao Yang, Livnat Jerby

机构 * Department of Genetics, Stanford University, CA, USA（遗传学系，斯坦福大学，加州，美国） ； Department of Computer Science, Texas A\&M University, TX, USA（计算机科学系，德克萨斯A&M大学，德克萨斯州，美国） 纠错

Comments ICML 2025; 23 pages, 12 figures, 7 tables