Overview

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

CardsList

417 papers59 papers in the last four weeks, down 9% on the four weeks before. 0.4% of all new papers.

  1. Restoring Diversity in Distilled Image and Video Generators by Probing Activations

    Nov 13, 2026Furkan Horoz, Ozgur Kara, Kaan Oktay +4Diffusion Model DistillationDiffusion-Based Image Generation

  2. Right Token, Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering

    Nov 13, 2026Nima Eshraghi, Lovedeep Gondara, Sagarika Suresh +3Language Model SteeringActivation Steering

  3. AmbiSteer: Selectively Eliciting Ambiguity-Aware Responses in Vision-Language Models

    Nov 13, 2026Monica Riedler, Samuele Antonelli, Barbara Plank +1Activation SteeringVisual Question Answering

  4. Activation Steering the Nan-Shan in Typhoon

    Nov 13, 2026Megan Wei, Damon Pham, John Cayley +1Controllable Text GenerationLanguage Model Steering

  5. Safety Side Effects of Preventative Persona Steering: An Empirical Audit

    Nov 13, 2026Sabrina Tang, Nikolai Vo, Carson Murillo +2Language Model SteeringActivation Steering

  6. What Roles Do System Prompts Really Assign? Validating and Steering Occupational Role Vectors in LLMs

    Nov 11, 2026Jose L. Sampedro Mazon, Ilmi Yoon, Nadim SaadActivation SteeringLLM Interpretability

  7. MAST: Medoid-Anchored Subspace Steering for Visual Grounding in Multimodal LLMs

    Nov 11, 2026Zhitao Chen, Zhenxin Cai, Jingyuan Li +1Activation SteeringMultimodal Large Language Models

  8. Breaking Bad: "Going Badly" Activations Drive Reward Hacking

    Nov 11, 2026Natalia Fischl-Lanzoni, Punya Syon Pandey, Merlin Stein +1Reward HackingCoding Agents

  9. Do Language Models Feel the Sentiment? Investigating Emotion Vectors

    Nov 10, 2026Ali Hassan Khan, Saad Bin Ather, Muhammad Saif +2Emotion Representation in Language ModelsLarge Language Model-Based Text Classification

  10. MicroLens: Read, Trace, and Rewire the Residual Stream with Two Lenses

    Nov 9, 2026Nika Zahedi, Fangyi Zhu, Anna WuLLM InterpretabilityLanguage Model Introspection

  11. Represented but Not Used: Value-Policy Dissociation in LLM Agents

    Nov 7, 2026Yuzhang Luo, Mohamed Aghzal, Liangming Pan +1LLM AgentsMechanistic Interpretability

  12. When Should an LLM Steering Edit Depend on the Input?

    Nov 3, 2026Rachael Harkness, Konrad RawlikLanguage Model SteeringActivation Steering

  13. SKILL2STEER: One-Pass Hypernetwork for Mapping Natural-Language Skills to Low-Rank Post-Block Activation Shifts

    Nov 3, 2026Ran Li, Lei ChenLLM Agent Skill LearningActivation Steering

  14. Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI

    Nov 3, 2026Vishal Pramanik, Maisha Maliha, Susmit Jha +1Multi-Agent LLM SystemsLong-Horizon Agent Tasks

  15. Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs

    Nov 3, 2026Zihao Feng, Qingzhao Zhang, Chunyu Xia +3Adversarial Attacks on VLA ModelsReasoning for VLA Models

  16. VIAS: Value-Isolating Activation Steering for Multi-Value Alignment and Compliance for Organisations

    Nov 2, 2026Avni Mittal, Avinash Anand, Dikshant Kukreja +3LLM AlignmentLanguage Model Steering

  17. Steering with Multilingual Lexicons for Sentiment Control in Low-Resource Languages

    Nov 2, 2026Lucinda Laurent, Muhammad Dehan Al Kautsar, Alfan Farizki Wicaksono +1Language Model SteeringActivation Steering

  18. SYNC: A Training-Free Inference Controller for Dynamic Layer Selection in Multi-Persona Alignment

    Nov 2, 2026Yosef Nuraga Wicaksana, Muhammad Reza Qorib, Fadhil Muhammad +1Inference-Time Language Model AlignmentActivation Steering

  19. Towards Multi-Value Alignment via Value Localization in LLMs

    Nov 2, 2026Xueqi Ma, Yanbei Jiang, Xingjun Ma +2Multi-Objective Language Model AlignmentInference-Time Language Model Alignment

  20. Sweep Less, Steer More: Using Activation Steering Geometry for Reduced Hyperparameter Search

    Nov 1, 2026Nishkal Hundia, Swastik Agrawal, Navita Goyal +1Language Model SteeringActivation Steering

  21. TopoSteer: Topological Layer Selection for Activation Steering in LLMs

    Nov 1, 2026Tuc Nguyen, Marcos Gutiérrez-del-Pozo, Eduardo Paluzo-Hidalgo +3Language Model SteeringActivation Steering

  22. Latent Harness: Compiling Agent Behaviors into Post-Block Activation Shifts

    Nov 1, 2026Ran Li, Shimin Di, Zeyu Ren +3LLM Agent HarnessesInference-Time Adaptation of Language Models

  23. J-Anchor: Defending Against Prefill Attacks via J-Space Selective Refusal Restoration

    Oct 31, 2026Running Yang, Beidi Zhao, Xiaoxiao LiJailbreak DefenseLLM Jailbreak Attacks