Large Language Models (LLMs) are trained with next-token prediction, implemented in autoregressive Transformers via causal masking for parallelism. This creates a subtle misalignment: residual connections tie activations to the current token, while supervision targets the next token, potentially propagating mismatched information if the current token is not the most informative for prediction. In this work, we empirically localize this input-output alignment shift in pretrained LLMs, using decoding trajectories over tied embedding spaces and similarity-based metrics. Our experiments reveal that the hidden token representations switch from input alignment to output alignment deep within the network. Motivated by this observation, we propose a lightweight residual-path mitigation based on residual attenuation, implemented either as a fixed-layer intervention or as a learnable gating mechanism. Experiments on multiple benchmarks show that these strategies alleviate the representation misalignment and yield improvements, providing an efficient and general architectural enhancement for autoregressive Transformers.
More Than Meets the Eye? Uncovering the Reasoning-Planning Disconnect in Training Vision-Language Driving Models
S-Lab, Nanyang Technological University, Singapore（南洋理工大学S实验室）；College of Computing and Data Science, Nanyang Technological University, Singapore（南洋理工大学计算与数据科学学院）
MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
School of Computer Science and Technology, East China Normal University（东华大学计算机科学与技术学院）；Shanghai AI Lab, Shanghai Innovation Institute（上海人工智能实验室，上海创新研究院）
What Gets Activated: Uncovering Domain and Driver Experts in MoE Language Models
Guangdong University of Technology（广东工业大学）；Soochow University（苏州大学）；University of Copenhagen（哥本哈根大学）；Mohamed bin Zayed University of Artificial Intelligence（穆罕默德·本·扎耶德人工智能大学）
Intrinsic Guardrails: How Semantic Geometry of Personality Interacts with Emergent Misalignment in LLMs
International Institute of Information Technology, Hyderabad（国际信息研究所，海得拉巴）；Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi（Mohamed bin Zayed人工智能大学，阿布扎比）
CoSToM:Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in Large Language Models
National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab, Huazhong University of Science and Technology（华中科技大学国家大数据技术与系统工程技术研究中心、服务计算技术与系统实验室、集群与网格计算实验室）；Singapore Management University（新加坡管理大学）
CausalGame: Benchmarking Causal Thinking of LLM Agents in Games
MBZUAI（穆罕默德·本·扎耶德人工智能大学）；Carnegie Mellon University（卡内基梅隆大学）；Hong Kong Baptist University（香港浸会大学）；University of Oxford（牛津大学）；New York University, Abu Dhabi（纽约大学阿布扎比分校）
Uncovering Main Causalities for Long-tailed Information Extraction
Singapore University of Technology and Design（新加坡科技设计大学）；Carnegie Mellon University（卡内基梅隆大学）；National University of Singapore（新加坡国立大学）；University of Cambridge（剑桥大学）
Revealing Multimodal Causality with Large Language Models
University of Technology Sydney（技术大学悉尼大学）；Tongji University（同济大学）；University of Melbourne（墨尔本大学）；University of Sydney（悉尼大学）；Macquarie University（麦考瑞大学）
School of Computer and Communication Sciences, EPFL, Lausanne, Switzerland（苏黎世联邦理工学院计算机与通信科学系）；Leiden Institute of Advanced Computer Science (LIACS), Leiden University, Leiden, The Netherlands（莱顿大学高级计算机科学研究所）；College of Management of Technology, EPFL, Lausanne, Switzerland（苏黎世联邦理工学院技术管理学院）
Hidden Heroes and Gradient Bloats: Layer-Wise Redundancy Inverts Attribution in Transformers
Department of Computer Science, Fu Foundation School of Engineering and Applied Science, Columbia University, New York, NY, USA（计算机科学系，工程与应用科学基金会学院，哥伦比亚大学，纽约，纽约州，美国）
