Mechanistically Interpreting Compression in Vision-Language Models
Indian Institute of Technology Jodhpur（印度理工学院焦特布尔分校）；AIM Intelligence；National Institute of Technology Agartala（国立阿加尔塔拉理工学院）；Fordham University（福特汉姆大学）；University of Fukui（福井大学）；Independent Researcher（独立研究员）
Compressed vision-language models (VLMs) are widely used to reduce memory and compute costs, making them a suitable choice for real-world deployment. However, compressing these models raises concerns about whether internal computations and safety behaviors are preserved. In this work, we use causal circuit analysis and crosscoder-based feature comparisons to examine how pruning and quantization fundamentally change the internals across representative VLMs. We observe that pruning generally keeps circuit structure intact but rotates and attenuates internal features, while quantization modifies the circuits at a higher level yet leaves the surviving features better aligned. Leveraging this insight, we also introduce VLMSafe-420, a novel benchmark that pairs harmful inputs with matched benign counterfactuals across various safety categories. Our findings show that pruning causes a sharp drop in genuine refusal behavior, suggesting that the choice of compression has safety implications.
Mechanistically Interpreting the Role of Sample Difficulty in RLVR for LLMs
Beijing Jiaotong University（北京交通大学）；AntGroup（蚂蚁集团）；Northwestern Polytechnical University（西北工业大学）；University of Leeds（利兹大学）；University of Southampton（南安普顿大学）
Beyond Dense States: Sparse Transcoders as Causally Testable Operators for LLM Latent Reasoning
College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics（南京航空航天大学计算机科学与技术学院）；MIIT Key Laboratory of Pattern Analysis and Machine Intelligence（信息产业部模式分析与机器智能重点实验室）；Institute for AI, Tsinghua University, Beijing, China（清华大学人工智能研究院）
Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws
Institute for Interdisciplinary Information Sciences, Tsinghua University（清华大学交叉信息研究院）；School of Computer Science and Engineering, Beihang University（北航计算机科学与工程学院）
In-context superposition: human-like working memory interference in large language models
School of Psychological and Brain Sciences, Georgia Tech（佐治亚理工学院心理与脑科学学院）；Department of Psychology, New York University（纽约大学心理学系）；Department of Cognitive Science, Indiana University Bloomington（印第安纳大学布卢明顿分校认知科学系）；Honda Research Institute（本田研究所）；Center of Excellence for Computational Cognition, Georgia Tech（佐治亚理工学院计算认知卓越中心）；Departments of Neuroscience and Psychology, The University of Texas at Austin（德克萨斯大学奥斯汀分校神经科学和心理学系）
MedPruner: Training-Free Hierarchical Token Pruning for Efficient 3D Medical Image Understanding in Vision-Language Models
Chinese University of Hong Kong（香港中文大学）；Westlake University（西湖大学）；Southern Medical University（南方医科大学）；Jiangnan University（江南大学）；Dalian University of Technology（大连理工大学）；Shanghai Jiao Tong University（上海交通大学）；University of Copenhagen（哥本哈根大学）
ILLC: Iterative Layer-by-Layer Compression for Enhancing Structural Faithfulness in SpArX
Institute for Clarity in Documentation（文献 clarity 研究所）；Inria Paris-Rocquencourt（法国国家信息与自动化研究所巴黎-罗康校区）；Rajiv Gandhi University（拉吉夫·甘地大学）；Tsinghua University（清华大学）；Palmer Research Laboratories（帕尔默研究实验室）；The Kumquat Consortium（金桔联盟）
