Research map / Large language models / Alignment and preferences
Mitigating the Safety–Utility Trade-off in LLM Alignment via Adaptive Safe Context Learning
Yanbo Wang, Minzheng Wang, Jian Liang, Lu Wang, Yongcan Yu, Ran He
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 inverse-frequency advantage reweighting (IFPO) in place of standard GRPO (ablation) | ASCL post-training · Qwen3-4B, H-avg | H-avg: 89.21 -> 89.82 | Improved |
| add inverse-frequency reweighting (IFPO) in place of GRPO (ablation) | ASCL post-training · Qwen3-4B, general reasoning | reasoning score: 76.49 -> 79.06 | Improved |
| add a small rule-call penalty to GRPO reward (ablation) | GRPO · Qwen3-4B, H-avg | H-avg: 89.21 -> 87.89 | Hurt |
| add a 0.1 rule-call penalty compared with IFPO (ablation) | GRPO w/ penalty · Qwen3-4B, general reasoning | reasoning score: 79.06 -> 74.51 | Hurt |
| add RL with IFPO after behavior cloning (ablation) | BC · Qwen3-4B, over-refusal average | over-refusal rate (%): 16.33 -> 6.00 | Improved |
| add RL with IFPO after behavior cloning (ablation) | BC · Qwen3-4B, safety average | safety rate (%): 95.98 -> 98.87 | Improved |
| add post-training (BC and IFPO) on top of zero-shot ASCL (ablation) | ASCL-ZS · Qwen3-14B, over-refusal average | over-refusal rate (%): 28.41 -> 5.84 | Improved |
| add safety-rule retrieval tool (ASCL-ZS) versus pure prompt (ablation) | Pure · Qwen3-4B, over-refusal average | over-refusal rate (%): 6.04 -> 23.42 | Hurt |
| add a one-sentence safety reminder (short) to the system prompt (ablation) | Pure · Qwen3-4B, safety average | safety rate (%): 65.66 -> 77.85 | Improved |
| add IFPO RL stage versus BC on general reasoning H-avg (ablation) | BC · Qwen3-14B, H-avg | H-avg: 87.45 -> 91.43 | Improved |
| add BC/IFPO post-training versus zero-shot ASCL on general reasoning | ASCL-ZS · Qwen3-14B, MATH avg@16 | MATH accuracy: 96.75 -> 92.46 | Hurt |
| train on surrogate CoT data with RL only (no BC, no retrieval) (ablation) | ASCL · three Qwen3 sizes, safety | safety rate | Hurt |
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.
- STAR-1 — STAR-1: Safer Alignment of Reasoning LLMs with 1K Data (AAAI 2026) · better on 6 of 6 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
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.90)
- Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention (ICLR 2026 · similarity 0.89)
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · similarity 0.89)
- Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models (NeurIPS 2026 · similarity 0.89)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.89)
- SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization (CVPR 2026 · similarity 0.89)
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.
