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

375 papers77 papers in the last four weeks, up 93% on the four weeks before. 0.5% of all new papers.

Steering the Verifiability of Multimodal AI Hallucinations
Apr 8, 2026Jianhong Pang, Ruoxi Cheng, Ziyi Ye +4Hallucination in Language Models Multimodal Hallucination
Institute of Trustworthy Embodied AI, Fudan University · Shanghai Key Laboratory of Multimodal Embodied AI
Distributed Interpretability and Control for Large Language Models
Apr 7, 2026Dev Arpan Desai, Shaoyi Huang, Zining ZhuLLM Interpretability Activation Steering
The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment
Apr 7, 2026Rishab Balasubramanian, Pin-Jie Lin, Rituraj Sharma +6Transfer Learning Activation Steering
Virginia Tech · Amazon · UNC Chapel Hill
Under Pressure: Emotional Framing Induces Measurable Behavioral Shifts and Structured Internal Geometry in Small Language Models
Apr 6, 2026Rana Muhammad UsmanFraming Effects in Language Models Emotion Representation in Language Models
Independent Researcher
Extracting and Steering Emotion Representations in Small Language Models: A Methodological Comparison
Apr 5, 2026Jihoon JeongEmotion Representation in Language Models Small Language Models
Department of Electrical Engineering and Computer Science Daegu Gyeongbuk Institute of Science and Technology (DGIST) ModuLabs
Where to Steer: Input-Dependent Layer Selection for Steering Improves LLM Alignment
Apr 4, 2026Soham Gadgil, Chris Lin, Su-In LeeLanguage Model Steering Inference-Time Language Model Alignment
University of Washington
Noise Steering for Controlled Text Generation: Improving Diversity and Reading-Level Fidelity in Arabic Educational Story Generation
Apr 3, 2026Haziq Mohammad Khalid, Salsabeel Shapsough, Imran ZualkernanText Generation Diverse Text Generation
American University of Sharjah
Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens
Apr 3, 2026Mohammed Suhail B NadafFunction Vectors Transformer Interpretability
Independent
Reliable Control-Point Selection for Steering Reasoning in Large Language Models
Apr 2, 2026Haomin Zhuang, Hojun Yoo, Xiaonan Luo +2Language Model Steering Activation Steering
University of Notre Dame
Do LLMs Know What Is Private Internally? Probing and Steering Contextual Privacy Norms in Large Language Model Representations
Mar 31, 2026Haoran Wang, Li Xiong, Kai ShuPrivacy Leakage in Language Models Contextual Integrity
Department of Computer Science, Emory University
Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models
Mar 29, 2026Katarzyna Zaleska, Łukasz Popek, Monika Wysoczańska +1T2I Diffusion Models Attention Control in Diffusion Models
Warsaw University of Technology · IDEAS Research Institute
TokenDial: Continuous Attribute Control for Text-to-Video Generation in Visual Dial Space
Mar 29, 2026Zhixuan Liu, Peter Schaldenbrand, Yijun Li +5Controllable Video Generation T2V Generation
Adobe Research · Carnegie Mellon University
Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection
Mar 28, 2026Jinhu Fu, Yihang Lou, Qingyi Si +3VLM Safety Vision-Language Models
Beijing University of Posts and Telecommunications · Huawei Technologies Ltd. · Peking University +1
H-Node Attack and Defense in Large Language Models
Mar 27, 2026Eric Yocam, Varghese Vaidyan, Yong WangLLM Hallucination Mitigation Hallucination in Language Models
Department of Computer Science and Software Engineering, California Polytechnic State University, San Luis Obispo, CA 93407, USA. · Beacom College of Computer and Cyber Sciences, Dakota State University, Madison, SD 57042, USA. · Department of Computer Science, University of Idaho, Moscow, ID 83844, USA.
Sparse Visual Thought Circuits in Vision-Language Models
Mar 26, 2026Yunpeng ZhouVLM Interpretability Sparse Autoencoders
University of Reading, Reading, UK
Closing the Confidence-Faithfulness Gap in Large Language Models
Mar 26, 2026Miranda Muqing Miao, Lyle UngarConfidence Estimation in Language Models Verbalized Confidence Estimation
University of Pennsylvania · Department of Computer and Information Science
Analysing the Safety Pitfalls of Steering Vectors
Mar 25, 2026Yuxiao Li, Alina Fastowski, Efstratios Zaradoukas +2Activation Steering Language Model Safety Evaluation
Technical University of Munich, Munich Center for Machine Learning
Steering Code LLMs with Activation Directions for Language and Library Control
Mar 24, 2026Md Mahbubur Rahman, Arjun Guha, Harshitha MenonLanguage Model Steering Code Language Models
Iowa State University · Northeastern University · Lawrence Livermore National Laboratory
Principled Steering via Null-space Projection for Jailbreak Defense in Vision-Language Models
Mar 23, 2026Xingyu Zhu, Beier Zhu, Shuo Wang +4VLM Robustness Activation Steering
MoE Key Lab of BIPC, University of Science and Technology of China · National University of Singapore · Nanyang Technological University
On the Failure of Topic-Matched Contrast Baselines in Multi-Directional Refusal Abliteration
Mar 23, 2026Valentin PetrovRefusal Behavior in Language Models Causal Interventions in Language Models
INMECHA INC, San Francisco
DSPA: Dynamic SAE Steering for Data-Efficient Preference Alignment
Mar 23, 2026James Wedgwood, Aashiq Muhamed, Mona T. Diab +1Preference Alignment Inference-Time Language Model Alignment
Carnegie Mellon University
Persona Vectors in Games: Measuring and Steering Strategies via Activation Vectors
Mar 22, 2026Johnathan Sun, Andrew ZhangPersonality Modeling in Language Models Language Model Steering
Harvard University
Knowledge Packs: Zero-Token Knowledge Delivery via KV Cache Injection
Mar 22, 2026Andrey PustovitKnowledge Augmentation for Language Models Activation Steering
Independent Researcher
The Anatomy of an Edit: Mechanism-Guided Activation Steering for Knowledge Editing
Mar 21, 2026Yuan Cao, Mingyang Wang, Hinrich SchützeLanguage Model Editing Activation Steering
Technical University of Munich · LMU Munich · Munich Center for Machine Learning (MCML)
Thinking in Different Spaces: Domain-Specific Latent Geometry Survives Cross-Architecture Translation
Mar 20, 2026Marcus Armstrong, Navid Ayoobi, Arjun MukherjeeRepresentation Geometry in Language Models Activation Steering
Department of Computer Science, University of Houston, Houston, TX 77204
Plug-and-Steer: Decoupling Separation and Selection in Audio-Visual Target Speaker Extraction
Mar 20, 2026Doyeop Kwak, Suyeon Lee, Joon Son ChungTarget Speaker Extraction Audio Source Separation
Korea Advanced Institute of Science and Technology, South Korea
Detection Without Correction: A Robust Asymmetry in Activation-Based Hallucination Probing
Mar 20, 2026Dip Roy, Rajiv Misra, Sanjay Kumar Singh +1LLM Hallucination Detection Hidden-State Probing
Department of Computer Science and Engineering, Indian Institute of Technology Patna, India · Department of Computer Science, Rajarshi School of Management and Technology, Varanasi, India · Department of Electronics and Communication Engineering, Jaypee Institute of Information Technology, Noida, India
Steering Video Diffusion Transformers with Massive Activations
Mar 18, 2026Xianhang Cheng, Yujian Zheng, Zhenyu Xie +2Massive Activations Video Diffusion Models
MBZUAI · Pinscreen
Multi-Trait Subspace Steering to Reveal the Dark Side of Human-AI Interaction
Mar 18, 2026Xin Wei Chia, Swee Liang Wong, Jonathan PanHuman-AI Interaction LLM Safety
Home Team Science & Technology Agency (HTX) Singapore
Neuron-Level Emotion Control in Speech-Generative Large Audio-Language Models
Mar 18, 2026Xiutian Zhao, Ismail Rasim Ulgen, Philipp Koehn +2Controllable Speech Generation Emotional Speech Synthesis
Center for Language and Speech Processing (CLSP), Johns Hopkins University, USA · Group on Language, Audio & Music (GLAM), Imperial College London, UK

Apr 8, 2026J/K move · Enter open · S save

Steering the Verifiability of Multimodal AI Hallucinations

Jianhong Pang, Ruoxi Cheng, Ziyi Ye, Xingjun Ma+3

Institute of Trustworthy Embodied AI, Fudan University · Shanghai Key Laboratory of Multimodal Embodied AI

Hallucination in Language Models Multimodal Hallucination Multimodal Large Language Models Activation Steering

Open paper AskPDF arXiv

AI applications driven by multimodal large language models (MLLMs) are prone to hallucinations and pose considerable risks to human users. Crucially, such hallucinations are not equally problematic: some hallucination contents could be detected by human users(i.e., obvious hallucinations), while others are often missed or require more verification effort(i.e., elusive hallucinations). This indicates that multimodal AI hallucinations vary significantly in their verifiability. Yet, little research has explored how to control this property for AI applications with diverse security and usability demands. To address this gap, we construct a dataset from 4,470 human responses to AI-generated hallucinations and categorize these hallucinations into obvious and elusive types based on their verifiability by human users. Further, we propose an activation-space intervention method that learns separate probes for obvious and elusive hallucinations. We reveal that obvious and elusive hallucinations elicit different intervention probes, allowing for fine-grained control over the model's verifiability. Empirical results demonstrate the efficacy of this approach and show that targeted interventions yield superior performance in regulating corresponding verifiability. Moreover, simply mixing these interventions enables flexible control over the verifiability required for different scenarios.