Research map / Large language models / Agents, code and math
More Sail than Ballast: Addressing Harmful Knowledge Leakage in the Expansive Reasoning Space of LRMs
Qibing Ren, Xinhao Song, Ke Fan, Lijun Li, Zhanpeng Zhou, Gongshen Liu, Junchi Yan, Lizhuang Ma, Jing Shao
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 faithfulness and helpfulness rewards to safety-only reward (ablation) | OpenSafeRL on LLaMA-3-8B · XSTest | over-refusal rate (%): 50.80 -> 2.80 | Improved |
| add faithfulness reward to safety + helpfulness (ablation) | OpenSafeRL on LLaMA-3-8B · AutoDAN | attack success rate (%): 37.0 -> 14.0 | Improved |
| add helpfulness reward to safety-only reward (ablation) | OpenSafeRL on LLaMA-3-8B · AutoDAN | attack success rate (%): 2.0 -> 37.0 | Hurt |
| train on synthesized open-ended data instead of plain data (ablation) | OpenSafeRL on LLaMA-3-8B · AutoDAN | attack success rate (%): 36.0 -> 14.0 | Improved |
| add safety principle system prompt to RL training (ablation) | OpenSafeRL on LLaMA-3-8B · XSTest | over-refusal rate (%): 29.20 -> 2.80 | Improved |
| add safety principle prompt to vanilla model without RL (ablation) | LLaMA-3-8B-Instruct · AutoDAN | attack success rate (%): 45.5 -> 22.0 | Improved |
| insert wait token at test time (ablation) | LLaMA-3-8B-Instruct vanilla · XSTest | over-refusal rate (%): 6.80 -> 11.60 | Hurt |
| insert wait token at test time (ablation) | OpenSafeRL on LLaMA-3-8B · AutoDAN | attack success rate (%): 14.0 -> 2.0 | Improved |
| raise safety weight alpha from 0.5 to 0.7 (ablation) | OpenSafeRL on LLaMA-3-8B · AutoDAN / GSM8K | ASR and accuracy: AutoDAN 14.0 -> 12.0; GSM8K 74.08 -> 73.00 | Mixed |
| use OpenSafeRL instead of vanilla | Qwen-2.5-7B-Instruct · OpenLeakBench | harmfulness rate (%): 62% -> 23.5% | Improved |
| use OpenSafeRL instead of Circuit Breakers | Qwen-2.5-7B-Instruct · over-refusal average | over-refusal rate (%): 31.07 -> 2.02 | Improved |
| apply gradient ascent alignment | Qwen-2.5-7B-Instruct · over-refusal average | over-refusal rate (%): 7.55 -> 70.35 | Hurt |
Closest papers by meaning
- Finding and Reactivating Post-Trained LLMs' Hidden Safety Mechanisms (NeurIPS 2025 · similarity 0.91)
- SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment (NeurIPS 2025 · similarity 0.89)
- Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention (ICLR 2026 · similarity 0.89)
- CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning (NeurIPS 2025 · similarity 0.88)
- AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.88)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 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.
