mlsanityData to 14 Oct 2026

What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal

arXiv 2604.08524 ↗First posted 9 Apr 2026cs.LG cs.AI cs.CLcited by 7

On the map: LLM alignment / RLHF / preference optimisation → Safety alignment and jailbreak attacks

also in 1 more
Abstract (arXiv)

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works--specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit. Freezing all attention scores during steering drops performance by only 8.83% across three model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

Key facts. Cited by 7 papers in the map (data through 14 Oct 2026). Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 36 of its 54 references that are in the map. First posted on arXiv 9 Apr 2026.

Builds on

36 of 54 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence24 Feb
  2. 2023Universal and Transferable Adversarial Attacks on Aligned Language Models27 Jul

Other references · 34

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 8,065
  3. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  4. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
  5. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  6. 2023Jailbroken: How Does LLM Safety Training Fail?cited by 1,651
  7. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,429
  8. 2023Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversationscited by 1,207
  9. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  10. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  11. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  12. 2023Steering Language Models With Activation Engineeringcited by 922
22 more
  1. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  2. 2023Catastrophic Jailbreak of Open-source LLMs via Exploiting Generationcited by 451
  3. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  4. 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 359
  5. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 345
  6. 2013Direct and Indirect Effectscited by 280
  7. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 218
  8. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 209
  9. 2024Programming Refusal with Conditional Activation Steeringcited by 206
  10. 2024Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanismscited by 149
  11. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 118
  12. 2025Understanding Reasoning in Thinking Language Models via Steering Vectorscited by 72
  13. 2024Robust LLM safeguarding via refusal feature adversarial trainingcited by 64
  14. 2024Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablationcited by 48
  15. 2025Understanding (Un)Reliability of Steering Vectors in Language Modelscited by 47
  16. 2024Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questionscited by 35
  17. 2025Steering off Course: Reliability Challenges in Steering Language Modelscited by 34
  18. 2025HyperSteer: Activation Steering at Scale with Hypernetworkscited by 23
  19. 2025Improved Representation Steering for Language Modelscited by 21
  20. 2025Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectorscited by 8
  21. 2026Fine-Grained Activation Steering: Steering Less, Achieving Morecited by 7
  22. 2026Activation Steering via Generative Causal Mediationcited by 6

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 7.

7 papers.

  1. 2026Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventionscited by 6
  2. 2026Demystifying Variance in Circuit Discovery of LLMscited by 2
  3. 2026Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMscited by 1
  4. 2026Are Arithmetic Heuristic Neurons Form-Invariant? A Mechanistic Analysis of Symbols, Text, and Code in LLMscited by 1
  5. 2026RepBench: Compiling Benchmarks into Capability Representations for Large Language Modelscited by 0
  6. 2026A Unified Mechanistic Analysis of Knowledge- and Safety-Based Refusalscited by 0
  7. 2026From Trait Vectors to Circuits: Tracing Refusal and Sycophancy Through Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026On the Non-Identifiability of Steering Vectors in Large Language Models
  2. 2026On the Steering Dimensionality of Refusal in Language Models
  3. 2026Understanding Unreliability of Steering Vectors in Language Models: Geometric Predictors and the Limits of Linear Approximations
  4. 2026Mechanistic Indicators of Steering Effectiveness in Large Language Models
  5. 2026Does Steering Break Your Model? A Multi-Dimensional Evaluation Suite for LLM Steering Methods
  6. 2026Disentangling Steering Vectors
  7. 2026OPIUM: Mitigating Steering Externalities and Over-Refusal via Dual Objective Latent Optimization
  8. 2026Where Steering Signals Come From: Activation Source Selection in Activation Steering