When an LLM agent misbehaves, a human steps in, and that intervention is usually discarded or replayed as training data. We treat it instead as a measurement: a label pointing at the internal computation that failed. We study it in a controlled agent environment, on one failure (asking a clarifying question when the request was already unambiguous), with labels from a rule-based overseer validated against blind human judgment. Claims cover eligible models, which both over-ask and sometimes act on matched unambiguous requests: five of thirteen qualified (Qwen2.5-3B, gemma-2-2b, OLMo-3-7B, Mistral-Nemo-12B, Phi-4). Four findings. First, the ask-vs-act decision is causally localizable in every eligible family: a direction learned from intervention labels flips 72-100% of held-out over-asks at a definite residual-stream depth, with site and dose chosen on training data, and random directions under the same procedure flip at most one episode. Second, over-asking has two mechanistic signatures: in four families the model demonstrably knows the request is unambiguous (the belief is decodable, causally inert and orthogonal to the decision) and asks anyway; in the fifth, the rare over-asks track different world-representations. Third, steering gated on a probe readout removes most of the simulated oversight burden while legitimate questions survive (bootstrap 95% intervals exclude zero); a prompt instruction makes it worse. Fourth, consolidating the fix into the weights yields a map, not a rule: on one model a LoRA confined to the causal band passes every criterion fixed in advance while behavior cloning on identical data breaks retention; on three families the reverse holds; on the fifth the outcome crosses the boundary across seeds. A geometric rule written down in advance to predict the regime failed on two out-of-family tests; we report the failure and stop rather than fit a replacement.
SLIM: Sparse Latent Steering for Interpretable and Property-Directed LLM-Based Molecular Editing
The Hong Kong University of Science and Technology (Guangzhou)（香港科技大学（广州））；Nanjing University of Aeronautics and Astronautics（南京航空航天大学）；The 63rd Research Institute, National University of Defense Technology, Nanjing（国防科技大学第六三研究所，南京）
DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing
School of Computer Science, Peking University（北京大学计算机学院）；Center on Frontiers of Computing Studies, Peking University（北京大学计算前沿研究中心）；National Research and Engineering Center of Software Engineering, Peking University（北京大学国家软件工程研究与工程中心）
EVA: Editing for Versatile Alignment against Jailbreaks
ShanghaiTech University（上海科技大学）；Sun Yat-sen University（中山大学）；State Key Laboratory of Blockchain and Data Security（区块链与数据安全国家重点实验室）；Tsinghua University（清华大学）
Is Model Editing Built on Sand? Revealing Its Illusory Success and Fragile Foundation
National University of Singapore（新加坡国立大学）；Huazhong University of Science and Technology（华中科技大学）；Central China Normal University（中国地质大学）；iWudao Tech（iWudao科技）；Oxford（牛津大学）；Google Deepmind（谷歌DeepMind）
DPN-LE: Dual Personality Neuron Localization and Editing for Large Language Models
Southeast University（东南大学）；Shanghai Jiao Tong University（上海交通大学）；East China Normal University（华东师范大学）；Zhongguancun Academy（中关村学院）；Zhejiang University of Technology（浙江工业大学）；Kuaishou Technology（快手科技）；Northeastern University（东北大学）；Tsinghua University（清华大学）；Nanjing University of Aeronautics and Astronautics（南京航空航天大学）
Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing
SKLCCSE, Beihang University（北航智能计算与复杂系统研究院）；National University of Singapore（新加坡国立大学）；Zhongguancun Laboratory, Beijing（中关村实验室）；Institute of Dataspace, Hefei（合肥数据空间研究院）；Nanyang Technological University（南洋理工大学）
Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors
Zhejiang University（浙江大学）；Palo Alto Networks（帕洛阿尔托网络）；Ningbo Global Innovation Center, Zhejiang University（宁波全球创新中心，浙江大学）；OPPO Research Institute（OPPO研究院）；Qilu University of Technology（齐鲁大学）
Controllable Value Alignment in Large Language Models through Neuron-Level Editing
National University of Singapore（新加坡国立大学）；Hefei University of Technology（合肥工业大学）；University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）；ST Engineering Ltd.（ST工程有限公司）
On Fairness of Task Arithmetic: The Role of Task Vectors
Stanford University（斯坦福大学）；Mila（Mila研究所）；Université de Montréal（蒙特利尔大学）；Institute of Science Tokyo（东京科学研究所）；Kyoto University（京都大学）；ZOZO Research（ZOZO研究所）
Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models
The University of Hong Kong（香港大学）；Fudan University（复旦大学）；LMU Munich（慕尼黑大学）；Tsinghua University（清华大学）；Technische Universität Darmstadt（达姆施塔特技术大学）；Technische Universität Berlin（柏林技术大学）；Technische Universität Dresden（德累斯顿技术大学）；The Chinese University of Hong Kong（香港中文大学）；University of Pennsylvania（宾夕法尼亚大学）；Nanjing University（南京大学）；University of Manchester（曼彻斯特大学）；Dartmouth College（达特茅斯学院）；University of California Los Angeles（加州大学洛杉矶分校）；University of Michigan（密歇根大学）；Microsoft（微软）；Tencent（腾讯）
Beyond Max Tokens: Stealthy Resource Amplification via Tool Calling Chains in LLM Agents
Nanyang Technological University, Singapore（南洋理工大学，新加坡）；University of Illinois Urbana-Champaign, United States（伊利诺伊大学厄巴纳-香槟分校，美国）；The Hong Kong University of Science and Technology, Hong Kong（香港科学与技术大学，香港）；Shanghai Jiao Tong University, China（上海交通大学，中国）
