Research map / Large language models / Language, safety and interpretability
Reasoned Safety Alignment: Ensuring Jailbreak Defense via Answer-Then-Check
Chentao Cao, Xiaojun Xu, Bo Han, Hang Li
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 Answer-Then-Check reasoning (ReSA-SFT) vs base model | Llama3.1-8B-Instruct · StrongREJECT jailbreaks avg (LlamaGuard) | defense success rate: 0.5008 -> 0.9046 | Improved |
| add RL stage after ReSA-SFT | Llama3.1-8B-Instruct · StrongREJECT jailbreaks avg (LlamaGuard) | defense success rate: 0.9046 -> 0.9932 | Improved |
| ReSA-SFT vs WJ-SFT against adaptive PAIR attack | Llama3.1-8B-Instruct · StrongREJECT PAIR | defense success rate: 0.3291 -> 0.6965 | Improved |
| ReSA-SFT safety tuning | Llama3.1-8B-Instruct · MATH500 | accuracy: 50.60% -> 49.00% | Hurt |
| ReSA-RL | Qwen2.5-7B-Instruct · MATH500 | accuracy: 77.00% -> 75.40% | Hurt |
| ReSA-SFT vs base on over-refusal | Llama3.1-8B-Instruct · over-refusal average (XSTest, OKTest, WJ-Eval) | over-refusal accuracy: 93.27% -> 95.02% | Improved |
| prompt engineering (goal priority) on deepseek-v3 | deepseek-v3 · XSTest | over-refusal accuracy: -13.60% | Hurt |
| train on only 0.5K ReSA samples (ablation) | Qwen2.5-7B-Instruct · jailbreak safety average | defense success | Improved |
| use ReSA data with only WildJailbreak prompts (63K) vs WJ-SFT (262K) (ablation) | Llama3.1-8B-Instruct · multiple jailbreak methods | defense success rate | Improved |
| add diverse jailbreak types to training data (full ReSA vs Only WJ) (ablation) | ReSA-SFT · TAP and ReNeLLM | defense success rate | Improved |
| adaptive Answer-Then-Check (skip check on normal queries) (ablation) | Qwen2.5-7B-Instruct · MATH500 runtime | runtime: 77s -> 62s | Improved |
| ReSA-SFT on harmful queries runtime | Qwen2.5-7B-Instruct · StrongREJECT runtime | runtime: 177s -> 46s | Improved |
Closest papers by meaning
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · similarity 0.91)
- Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs (NeurIPS 2025 · similarity 0.91)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.91)
- PANDAS: Improving Many-shot Jailbreaking via Positive Affirmation, Negative Demonstration, and Adaptive Sampling (ICML 2025 · similarity 0.91)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.90)
- Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training (ICLR 2026 · 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.
