Research map / Large language models / Language, safety and interpretability
SARSteer: Safeguarding Large Audio Language Models via Safe-Ablated Refusal Steering
Weilin Lin, Jianze Li, Hui Xiong, Li Liu
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 |
|---|---|---|---|
| add text-derived refusal steering with safe-space ablation (SARSteer) | Qwen2-Audio · Figstep-audio (harmful) | ASR: 51.60 -> 10.80 | Improved |
| add SARSteer | Kimi-Audio · Figstep-audio (harmful-safe) | BRR: 61.40 -> 88.80 | Improved |
| vanilla audio-derived steering (MDSteer-h2s, MDSteer-c2r) added to LALM | Qwen2-Audio · Figstep-audio (harmful) | ASR: 51.60 -> 84.00 (h2s), 90.80 (c2r) | Hurt |
| prompt-based defense FSD added | Qwen2-Audio · Figstep-audio (harmful-safe) | BRR: 70.20 -> 63.20 | Hurt |
| prompt-based defenses (AdaShield, FSD) added | Kimi-Audio · AirBench avg score | average score: 7.20 -> 6.72 (AdaShield), 6.83 (FSD) | Hurt |
| add PCA safe-space ablation to the refusal vector (V1 vs full) (ablation) | SARSteer · Figstep-audio | BRR: 70.10 -> 79.95 (Qwen2-Audio) | Improved |
| use PCA decomposed safe-space ablation instead of mean-projection ablation (V2) (ablation) | SARSteer · Figstep-audio | ASR: 30.40 -> 10.80 (Qwen2-Audio) | Improved |
| compute refusal direction from safe data instead of harmful data (ablation) | SARSteer · Figstep-audio | ASR: Qwen2-Audio ASR 10.80 -> 27.60; Kimi-Audio 10.00 -> 6.80 | Mixed |
| use longer AdaShield-style refusal prompt instead of the simple one (ablation) | SARSteer · Figstep-audio | ASR / BRR: Qwen2-Audio ASR 10.80 -> 2.80, BRR 79.95 -> 70.00 | Mixed |
| vary number of top principal components k from 5 to 45 (ablation) | SARSteer · Figstep-audio | ASR/BRR | No clear effect |
| vary sample number n from 10 to 100 (ablation) | SARSteer · Figstep-audio | ASR/BRR | No clear effect |
| apply SARSteer to pre-trained base LALM | Qwen2-Audio-Base · Figstep-audio (harmful) | ASR: 62.80 -> 15.20 | Improved |
Closest papers by meaning
- SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering (AAAI 2025 · similarity 0.89)
- Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation (ICLR 2025 · similarity 0.89)
- Just Enough Shifts: Mitigating Over-Refusal in Aligned Language Models with Targeted Representation Fine-Tuning (ICML 2025 · similarity 0.88)
- ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio–Language Models (NeurIPS 2025 · similarity 0.88)
- Refusal Direction is Universal Across Safety-Aligned Languages (NeurIPS 2025 · similarity 0.88)
- OR-Bench: An Over-Refusal Benchmark for Large Language Models (ICML 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.
