Get Started
Home
Topics
Search
Library
Research questionWhich internal circuits do steering vectors alter to induce refusal in large language models?Steering vectors can change whether an LLM refuses, but it is unclear which internal pathways mediate that change. Without that understanding, it is difficult to determine which vector dimensions matter or how much of a vector can be removed.
AI
Alignment & Safety
Inference Optimization
Mechanistic Interpretability
Latest papersRecent research connected to this question, newest first.What Drives Representation Steering? A Mechanistic Case Study on Steering RefusalThis concerns a mechanistic case study of refusal steering across three model families and multiple steering methodologies applied at the same layer. The reported evidence analyzes attention circuits, especially OV versus QK involvement, and finds that vectors can be sparsified by 85–96% while retaining most performance; it does not establish that these findings generalize to other behaviors or models.research paper · Sep 1, 2026
Related questions
How does refusal-prefix diversity during training affect language-model refusal vulnerability to activation-vector ablation?How can post-training make language-model refusals robust without sacrificing general capability?How can inference-time hidden-state steering of LLMs become efficient and extensible for deployment?Do refusal representations transfer across language-model architectures, and where should safety interventions read them?