Research map / Large language models / Language, safety and interpretability
Safety Instincts: LLMs Learn to Trust Their Internal Compass for Self-Defense
Guobin Shen, Dongcheng Zhao, Haibo Tong, Jindong Li, Feifei Zhao, Yi Zeng
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 entropy-minimizing reward (SIRL) instead of no safety training | Llama-3.1-8B-Instruct · JailbreakBench (DSR) | defense success rate (%): 84.3 -> 99.1 | Improved |
| add SIRL | Qwen2.5-7B-Instruct · JailbreakBench (DSR) | defense success rate (%): 82.8 -> 99.9 | Improved |
| add SIRL | Llama-3.2-3B-Instruct · JailbreakBench (DSR) | defense success rate (%): 95.6 -> 100.0 | Improved |
| use SIRL instead of SFT | Llama-3.1-8B-Instruct · AlpacaEval | win rate: 19.1 -> 51.9 | Improved |
| use SIRL instead of SFT | Qwen2.5-7B-Instruct · MATH-500 | accuracy (%): 31.8 -> 78.6 | Improved |
| add SIRL | Llama-3.1-8B-Instruct · MATH-500 | accuracy (%): 49.0 -> 51.2 | Improved |
| add SIRL | Llama-3.2-3B-Instruct · MATH-500 | accuracy (%): 42.2 -> 41.4 | Hurt |
| replace min-entropy reward with maximizing entropy (neg-SIRL) (ablation) | SIRL · JailbreakBench (DSR) | defense success rate (%): 99.1 -> 72.1 | Hurt |
| add entropy reward instead of random reward (ablation) | SIRL · JailbreakBench (DSR) | defense success rate (%): 85.2 -> 99.1 | Improved |
| replace entropy reward with minimizing perplexity (ablation) | SIRL · JailbreakBench (DSR) | defense success rate (%): 99.1 -> 98.7 | No clear effect |
| lower KL coefficient beta to 1e-4 (ablation) | SIRL · JailbreakBench (DSR) | defense success rate (%): 99.7% DSR; BBH drops 5.6 points and HumanEval 3.7 points | Mixed |
| raise KL coefficient beta to 1e-2 (ablation) | SIRL · JailbreakBench (DSR) | defense success rate (%): 93.6% DSR | Hurt |
Closest papers by meaning
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.92)
- Internalizing Safety Understanding in Large Reasoning Models via Verification (ICML 2026 · similarity 0.91)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.90)
- Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions (ICLR 2024 · similarity 0.90)
- The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions (ICML 2025 · similarity 0.90)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.90)
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.
