Activation Steering

Follow this topic Ask about this topic

Activation steering studies how edits to a model’s hidden activations can control its behavior at inference time, enabling changes without retraining the full model. Common approaches add or remove contrastive steering vectors, target sparse autoencoder features, neurons or attention heads, and use conditional or feedback-based controllers to adapt interventions to the input or generation state. Researchers apply these methods to language tasks such as reasoning, refusal, factuality and hallucination mitigation, as well as image and video generation, speech, and embodied agents; evaluations include benchmarks such as GSM8K, TruthfulQA, POPE and LIBERO. A central challenge is that readable or decodable features are not necessarily controllable, and effective edits can vary by context and architecture or cause collateral changes to fluency, utility and safety.

Latest papers

377 papers78 papers in the last four weeks, up 95% on the four weeks before. 0.5% of all new papers.

The Answer-Basin Representation Hypothesis: We Are Not Probing or Steering Concepts
Sep 21, 2026Manjiang Yu, Hongji Li, Zihan Wang +5LLM Interpretability Linear Representation Hypothesis
The University of Queensland · Mohamed bin Zayed University of Artificial Intelligence · Institute of Science Tokyo
What do VLM-Based Vision-Language Navigation Models Rely on: Interpreting and Steering Policy Behavior
Sep 21, 2026Débora Oliveira Makowski, Samiran Gode, Abhijeet Nayak +4Vision-Language Models VLM Interpretability
University of Technology Nuremberg. · ETH Zürich. · Inria, Ecole Normale Sup´erieure, CNRS, PSL Research University.
Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation-Level Enforcement
Sep 21, 2026Xutao Mao, Jianing Zhu, Jinman Zhao +4RL for VLMs Faithfulness in VLMs
City University of Hong Kong · Hong Kong Baptist University · University of Toronto +2
Look Before You Steer: Geometry Predicts SAE Feature Steerability
Sep 19, 2026Muhammad Khan, Shlok Channawar, Akshaj Gurugubelli +2Sparse Autoencoders Activation Steering
Algoverse AI Research.
When Steering Fails in Latent Reasoning: A Latent-to-Language Transition Gap
Sep 18, 2026Gaoxiang Huang, Lei QiLanguage Model Steering Activation Steering
HKUSTGZ · SEU
Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30
Sep 18, 2026Hans Andersen, David DichasLanguage Model Steering Moral Reasoning in Language Models
Interpreting and Steering LLM Agents for Social Simulations
Sep 14, 2026Jiayue Gaveal Fan, Arul Murugan, Shreyas Krishnan +1LLM Interpretability Human Behavior Simulation
1Data Innovation and AI Lab (DIAL), University of California, Berkeley · 2National Bureau of Economic Research (NBER)
Metacognitive Steering: Learning the Structure of Scientific Judgment
Sep 14, 2026Vincent Karpf, Joseph Reth, Eike Gerhardt +5Inference-Time Guidance Scientific Reasoning in Language Models
Autopoiesis Sciences
Signatures of Steerability in Activation Space of Language Models
Sep 12, 2026Prajjwal Bhattarai, Tuka AlhanaiLanguage Model Steering Activation Steering
New York University Abu Dhabi, Abu Dhabi, UAE · New York University, Tandon School of Engineering, Brooklyn, NY, USA
MAxBench: A Multinomial Concept Recovery Benchmark
Sep 11, 2026Divya Appapogu, Freya Behrens, Yonatan Belinkov +1LLM Interpretability Representation Geometry in Language Models
Department of Computer Science, Boston University, Boston, MA 02215, USA · Department of Computer Science, Technion – Israel Institute of Technology, Israel · Department of Computer Science, Kempner Institute, Harvard University, Boston, USA
AcFlow: Controlling Text-to-Image Diffusion Transformers via Learned Conditional Activation Flow
Sep 9, 2026Junran Wang, Zehao Jin, Tianyu Luan +3Diffusion Transformer T2I Generation
Georgia Institute of technology · South China University of Technology
GEOSTEER: Geodesic Optimization for Activation Steering in Large Language Models
Sep 9, 2026Xuan Cuong Ngo, Hao Vo, Ngan LeLanguage Model Steering Representation Geometry in Language Models
University of Arkansas, Fayetteville, Arkansas, USA
Same Values, Different Languages? From Multilingual Probing to Steering LLMs Toward Chinese Social Values
Sep 8, 2026Yuemei Xu, Kexin Xu, Jian Zhou +3Multilingual Language Model Evaluation Cultural Alignment
School of Information Science and Technology, Beijing Foreign Studies University · Beijing Academy of Artificial Intelligence · The State Key Lab of Software Development Environment, Beihang University
Key Path Identification for Resolving Knowledge Conflicts via SAE-based Steering
Sep 8, 2026Wenbo Zhang, Zhongxiang Sun, Zhiguang Han +1Language Model Steering Knowledge Conflicts in Language Models
Gaoling School of Artificial Intelligence, Renmin University of China · Nanyang Technological University
Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
Sep 7, 2026Srikanth Malla, Chiho Choi, Joon Hee ChoiLanguage Model Steering Causal Interventions in Language Models
Samsung Semiconductor US
AutoLexSteer: Automatic Contrast Construction for Lexical Activation Steering
Sep 6, 2026Shuhe Wang, Lachlan Cowley, Eduard Hovy +1Language Model Steering Activation Steering
School of Computing and Information Systems, The University of Melbourne, Australia
Steering Under Compression: Dose-Response, Capability Cost, and Failure Asymmetry in Quantized LLMs
Sep 6, 2026Saurav Bhandari, Benjamin WadeActivation Steering LLM Quantization
Department of Electronics and Computer Engineering Institute of Engineering, Thapathali Campus Tribhuvan University, Kathmandu, Nepal · Independent Researcher
Spillover-Aware Multi-Value Steering for Pluralistic LLM Alignment
Sep 5, 2026Weici Pan, Xander Barron, Jiawei Zhou +1Language Model Steering Inference-Time Language Model Alignment
Stony Brook University
Can Activation Steering Capture Multidimensional Authorship Style?
Sep 4, 2026Hieu Tran, Calvin Bao, Marine CarpuatLanguage Model Steering Activation Steering
University of Maryland, College Park
Locating and Steering Refusal Beyond Attention
Sep 4, 2026Preethi Carmel Bosco, Gopalakrishnan SrinivasanRefusal Mechanisms in Language Models Refusal Behavior in Language Models
Department of Computer Science and Engineering, Indian Institute of Technology Madras
Reducing Hallucinated Transcripts in Whisper via Hallucination Space Projection
Sep 3, 2026Maryam Abbasihafshejani, Murtuza JadliwalaSpeech Foundation Models Automatic Speech Recognition
The university of Texas at San Antonio
From Trait Vectors to Circuits: Tracing Refusal and Sycophancy Through Language Models
Sep 2, 2026Oscar Miró López-Feliu, Maya OzbayogluCircuits in Language Models Activation Steering
University of Amsterdam · Safe AI Netherlands
IDEEA: training-free Input-Dependent stEEring via Activation cluster matching
Sep 2, 2026Zheng Wang, Muchen Li, Renjie Liao +1Activation Steering
University of British Columbia · Vector Institute for AI · Canada CIFAR AI Chair +1
GAPS: Dimension-Level Gates for Conditional Activation Steering
Sep 1, 2026Moghis Fereidouni, Muhammad Umair Haider, Hassan Sajjad +1Language Model Steering LLM Safety
University of Kentucky · Dalhousie University
Polished but Unresolved: Identifying Late-Stage Pressure States in Long-Horizon Tool-Use Agents
Sep 1, 2026Haoyang Chen, Yi Liu, Jianzhi Shao +3LLM Agents Long-Horizon LLM Agents
State Key Laboratory for Novel Software Technology, Nanjing University, China · Alibaba Group, China
Investigating Assistant Bias in LLM User Simulators Using a Role Vector
Sep 1, 2026Daeheon Jeong, Yoonjoo Lee, Eugene Choi +2Large Language Model-Based Role-Play Simulation User Simulation
1KAIST · University of Michigan · 3Seoul National University +2
Controlling Refusal Behavior of LLMs via Stiefel-Constrained Rotation Steering
Aug 31, 2026Kirill Bunin, Dmitry Bylinkin, Vladimir Aletov +3Language Model Steering Refusal Behavior in Language Models
1Basic Research of Artificial Intelligence Laboratory (BRAIn Lab) · 2Kandinsky Lab · 3Innopolis University
Enhancing Low-Resource Language Reasoning via High-Resource Language Feature Transfer
Aug 31, 2026Minju Song, Hyeon Hwang, Junhyun Lee +1Cross-Lingual Reasoning in Language Models Low-Resource Language Processing
Korea University · Hankuk University of Foreign Studies · Noah’s Farm +1
Beyond Token-Level Guidance: Inference-Time Alignment of Specialized LLMs via Cross-Family Representation Steering
Aug 31, 2026Jin Gan, Xin Li, Jun LuoLLM Safety Alignment Language Modeling
College of Computing and Data Science, Nanyang Technological University, Singapore · College of Cryptology and Cyber Science, Nankai University, Tianjin, China
Interpreting and Steering for Safe and Correct Code Generation
Aug 30, 2026Hao Yan, Ziyu YaoMechanistic Interpretability LLM Security
Department of Computer Science George Mason University, Fairfax, VA

Sep 21, 2026J/K move · Enter open · S save

The Answer-Basin Representation Hypothesis: We Are Not Probing or Steering Concepts

Manjiang Yu, Hongji Li, Zihan Wang, Junwei Chen+4

The University of Queensland · Mohamed bin Zayed University of Artificial Intelligence · Institute of Science Tokyo

LLM Interpretability Linear Representation Hypothesis Representation Geometry in Language Models Activation Steering Language Model Probing

Open paper AskPDF arXiv

Linear probing and activation steering use linear directions to predict and control behavior-level concepts such as correctness, safety, and social bias in question answering. We call these behavior-level concept directions. Drawing on the Linear Representation Hypothesis (LRH) and circuit studies, these directions are often interpreted as concept representations. Yet controlled evidence for linear concept structure and local mechanisms does not establish that these directions represent the intended behavioral concepts, leaving probing and steering without a unified theoretical account. We propose the Answer-Basin Representation Hypothesis (ABRH): the model's own answer measure organizes the linear structure of these directions. For each question, all continuations yielding the same answer form an answer basin, whose mass is their total probability; these masses define the model's answer distribution. ABRH posits that its concentration before generation and the relative mass of each answer after generation are represented along linear directions shared across questions. Experiments span four Qwen2.5 and Gemma-3 models on correctness, social bias, and safety tasks. Decoupling concept labels from basin mass shows that probing and steering exhibit concept-consistent effects when labels align with mass orderings, weaken as mass gaps shrink, and reverse under conflict. Probes trained solely on mass orderings among wrong answers still select correct answers; a fixed steering vector can instead favor wrong answers when labels conflict with mass orderings. These results support an answer-measure account of behavior-level concept directions and of when probing and steering succeed, fail, or reverse.