Research map / Large language models / Language, safety and interpretability
AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint
Leheng Sheng, Changshuo Shen, Weixiang Zhao, Junfeng Fang, Xiaohao Liu, Zhenkai Liang, Xiang Wang, An Zhang, Tat-Seng Chua
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 |
|---|---|---|---|
| learn a null-space-constrained steering transform (AlphaSteer) instead of a fixed refusal vector | Llama-3.1-8B-Instruct · Jailbreak attacks (7 attacks, avg) | DSR %: 48.00 -> 91.93 | Improved |
| add AlphaSteer | Qwen2.5-7B-Instruct · Jailbreak attacks (7 attacks, avg) | DSR %: 20.57 -> 97.29 | Improved |
| add AlphaSteer | Gemma-2-9b-IT · Jailbreak attacks (7 attacks, avg) | DSR %: 24.69 -> 98.20 | Improved |
| add AlphaSteer | Llama-3.1-8B-Instruct · Utility Score (XSTest, AlpacaEval, MATH, GSM8K) | Utility Score %: 67.1 -> 67.3 | No clear effect |
| add AlphaSteer | Qwen2.5-7B-Instruct · Utility Score (XSTest, AlpacaEval, MATH, GSM8K) | Utility Score %: 77.6 -> 75.9 | Hurt |
| add AlphaSteer | Gemma-2-9b-IT · Utility Score (XSTest, AlpacaEval, MATH, GSM8K) | Utility Score %: 63.8 -> 62.4 | Hurt |
| replace the learned null-space transform with directly adding the refusal vector to all prompts (RV) (ablation) | AlphaSteer · Utility Score (XSTest, AlpacaEval, MATH, GSM8K) | Utility Score %: 67.3 -> 29.1 | Hurt |
| replace the learned null-space transform with RV on Qwen2.5-7B-Instruct (ablation) | AlphaSteer · Utility Score (XSTest, AlpacaEval, MATH, GSM8K) | Utility Score %: 75.9 -> 19.7 | Hurt |
| replace the learned null-space transform with RV on Gemma-2-9b-IT (ablation) | AlphaSteer · Utility Score (XSTest, AlpacaEval, MATH, GSM8K) | Utility Score %: 62.4 -> 2.4 | Hurt |
| replace the learned null-space transform with RV on Llama-3.1-8B-Instruct (ablation) | AlphaSteer · Jailbreak attacks (7 attacks, avg) | DSR %: 91.93 -> 100.00 | Improved |
| use conditional steering (CAST) instead of null-space steering | Llama-3.1-8B-Instruct · MATH and GSM8K | accuracy %: MATH 0.0, GSM8K 0.0 | Hurt |
| increase steering strength lambda in AlphaSteer (ablation) | AlphaSteer · Jailbreak DSR vs utility | DSR and utility score | Improved |
Later papers that compared with this method
1 later accepted paper reports AlphaSteer as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Mitigating Memorization Where It Happens (NeurIPS 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Jailbreak Antidote — Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering (AAAI 2025 · similarity 0.89)
- Programming Refusal with Conditional Activation Steering (ICLR 2025 · similarity 0.88)
- Graph-Regularized Sparse Autoencoders for LLM Safety Steering (NeurIPS 2026 · similarity 0.88)
- Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models (NeurIPS 2026 · similarity 0.87)
- LLMs Encode Harmfulness and Refusal Separately (NeurIPS 2025 · similarity 0.87)
- Learning to Steer: Input-dependent Steering for Multimodal LLMs (NeurIPS 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.
