Activation steering has emerged as a promising approach for efficiently adapting large language models (LLMs) to downstream behaviors. However, most existing steering methods rely on a single static direction per task or concept, making them inflexible under task variation and inadequate for complex tasks that require multiple coordinated capabilities. To address this limitation, we propose STEER2ADAPT, a lightweight framework that adapts LLMs by composing steering vectors rather than learning new ones from scratch. In many domains (e.g., reasoning or safety), tasks share a small set of underlying concept dimensions. STEER2ADAPT captures these dimensions as a reusable, low-dimensional semantic prior subspace, and adapts to new tasks by dynamically discovering a linear combination of basis vectors from only a handful of examples. Experiments across 9 tasks and 3 models in both reasoning and safety domains demonstrate the effectiveness of STEER2ADAPT, achieving an average improvement of 8.2%. Extensive analyses further show that STEER2ADAPT is a data-efficient, stable, and transparent inference-time adaptation method for LLMs.
Activation Steering Induces Emergent Misalignment: Steering Vector Construction and Comprehensive Evaluation
Nanyang Technological University(南洋理工大学);Sun Yat-sen University(中山大学);University of Science and Technology of China(中国科学技术大学);National University of Singapore(新加坡国立大学)
BarrierSteer: LLM Safety via Learning Barrier Steering
Department of Computer Science, National University of Singapore(新加坡国立大学计算机科学系);Singapore-MIT Alliance for Research and Technology Centre(新加坡-麻省理工联合研究中心);CSAIL, Massachusetts Institute of Technology(麻省理工学院计算机科学与人工智能实验室);Worcester Polytechnic Institute(沃斯堡理工学院)
Same Values, Different Languages? From Multilingual Probing to Steering LLMs Toward Chinese Social Values
School of Information Science and Technology, Beijing Foreign Studies University(北京外国语大学信息科学与技术学院);Beijing Academy of Artificial Intelligence(北京人工智能研究院);Beihang University(北京航空航天大学)
Northeastern University(东北大学);Washington University in St. Louis(圣路易斯华盛顿大学);Nanjing University of Aeronautics and Astronautics(南京航空航天大学);Tsinghua University(清华大学);Security Capability Center(360安全能力中心);Tencent(腾讯);Beijing University of Posts and Telecommunications(北京邮电大学)
Measuring and Narrowing the Compositionality Gap in Language Models
Paul G. Allen School of Computer Science & Engineering, University of Washington(华盛顿大学保罗·G·艾伦计算机科学与工程学院);MosaicML(MosaicML公司);Meta AI Research(Meta人工智能研究院);Allen Institute for AI(艾伦人工智能研究所)
Department of Statistics and Operations Research, UNC-Chapel Hill(统计与运筹学系,北卡罗来纳大学 Chapel Hill 分校);Imperial College Business School, Imperial College London(帝国理工学院伦敦校区商学院)
LLMs as Cultural Archives: Cultural Commonsense Knowledge Graph Extraction
Mohamed bin Zayed University of Artificial Intelligence(莫扎德·本·扎耶德人工智能大学);Ubiquitous Knowledge Processing Lab (UKP Lab)(无处不在知识处理实验室);Department of Computer Science(计算机科学系);Hessian Center for AI (hessian.AI)(黑森人工智能中心);Technische Universität Darmstadt(德累斯顿技术大学)
AlgoTrace: Algorithmic Primitives and Compositional Geometry of Reasoning in Language Models
Microsoft Research NYC(微软研究院纽约分部);Center for Theoretical Neuroscience(理论神经科学中心);Department of Psychology(心理学系);University of California Los Angeles(加州大学洛杉矶分校);Institute for Pure and Applied Mathematics(纯粹与应用数学研究所);Emory University(埃默里大学);Rice University(里士满大学);Mount Holyoke College(马里兰霍克学院);Technische Universität Berlin(柏林技术大学);BIFOLD-Berlin Institute for the Foundations of Learning and Data(柏林学习与数据基础研究所)
SearchAttack: Red-Teaming LLMs against Knowledge-to-Action Threats under Online Web Search
Institute of Computing Technology, CAS(计算技术研究所,中国科学院);University of Chinese Academy of Sciences(中国科学院大学);People's Public Security University of China(中国人民公安大学);University of Science and Technology of China(中国科学技术大学);Guangdong Laboratory of Artificial Intelligence and Digital Economy(广东人工智能与数字经济实验室);Tsinghua University(清华大学)