Research map / Large language models / Language, safety and interpretability
RepIt: Steering Language Models with Concept-Specific Refusal Vectors
Vincent Siu, Nathan W. Henry, Nicholas Crispino, Yang Liu, Dawn Song, Chenguang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| keep only heavy-tail (z>2) coordinates of the removed projection (ablation) | RepIt · Bioweaponry / Chemical / Cyber, five models | change in ASR: deviations remain within +-0.05 absolute ASR | No clear effect |
| build target vector from only 12 or 24 prompts instead of full set (ablation) | RepIt · WMD target vs non-target | ASR: averages closely match full dataset | No clear effect |
| remove overlap with non-target vectors from target vector (v_t to v_RepIt) (ablation) | difference-in-means target vector · WMD target and non-target | ASR: non-target ASR falls to roughly 0.1; target ASR 0.4-0.7 | Mixed |
| use non-target DIM vector alone (ablation) | refusal steering · WMD target prompts | ASR | Improved |
| use projection component alpha P alone (ablation) | refusal steering · WMD target prompts | ASR | Improved |
Closest papers by meaning
- Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation (ICLR 2025 · similarity 0.90)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.89)
- The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence (ICML 2025 · similarity 0.89)
- Refusal Direction is Universal Across Safety-Aligned Languages (NeurIPS 2025 · similarity 0.89)
- Latent-space Attacks for Refusal Evasion in Language Models (NeurIPS 2026 · similarity 0.89)
- SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
