Research map / Large language models / Language, safety and interpretability
SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives
Vincent Siu, Nicholas Crispino, David Park, Nathan W. Henry, Zhun Wang, 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 |
|---|---|---|---|
| remove the KL filter (NoKL) from direction selection (ablation) | steering methods · refusal and hallucination perspectives | effectiveness / entanglement: often more than doubles entanglement | Mixed |
| add CAST-style conditional gating instead of the KL filter (ablation) | steering methods · refusal perspective | effectiveness / entanglement: matching NoKL effectiveness at entanglement near Standard | Improved |
| add CAST-style conditional gating instead of the KL filter (ablation) | steering methods · bias perspective | effectiveness | Hurt |
| add conditional steering to DIM | DIM on Gemma-2-2B · FaithEval Inconsistent | hallucination reduction: 54.5% reduction | Improved |
| steer refusal with ACE or DIM instead of CAA, PCA or LAT | steering methods · refusal perspective | effectiveness: exceed 50% on every model except Gemma-2-2B | Improved |
| use PCA for extrinsic hallucination steering in Llama-3.1-8B | PCA · hallucination perspective | effectiveness/entanglement ratio: 1.76 | Improved |
| use LAT for bias steering | LAT · bias perspective | effectiveness/entanglement ratio: 7.05 and 8.37 | Improved |
| steer refusal (suppress it) and measure social behaviors | refusal steering · social behaviors | entanglement: degradation up to 76% | Hurt |
| steer refusal (jailbreak) and measure commonsense morality | refusal steering · commonsense morality | accuracy: up to 26% | Hurt |
| steer intrinsic hallucination and measure political views | hallucination steering · TwinViews political views | left-leaning agreement shift: 25% shift to the right to a 28% shift to the left | Mixed |
| steer any perspective and measure reasoning | steering methods · reasoning (GPQA, ARC-C) | entanglement: below 2% in every case | No clear effect |
| scale Qwen-2.5 from 1.5B and 3B to 7B (ablation) | steering methods · effectiveness-entanglement ratio ranking | ranking stability | No clear effect |
Closest papers by meaning
- Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models (NeurIPS 2026 · similarity 0.88)
- The Hawthorne Effect in Reasoning Models: Evaluating and Steering Test Awareness (NeurIPS 2025 · similarity 0.88)
- SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering (AAAI 2025 · similarity 0.87)
- LLMs Encode Harmfulness and Refusal Separately (NeurIPS 2025 · similarity 0.87)
- The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions (ICML 2025 · similarity 0.87)
- SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal (ICLR 2025 · similarity 0.86)
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.
