Research map / Large language models / Language, safety and interpretability
WALKSAFE: Risk-aware Graph Random Walk with Bi-GRPO for LLM Safety
Shilong Pan, Zhiliang Tian, Wanlong Yu, Zhen Huang, Qingyu Qiu, Zihan Chen, Zhonghao Sun, Minlie Huang, Dongsheng 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 entity-relation graph structure (ablation) | WalkSafe, Llama-3.2-3B · TET | avg harmlessness (lower better): 6.70 -> 3.25 | Improved |
| add entity-relation graph structure (ablation) | WalkSafe, Qwen2.5-3B · TET | helpfulness: 1.40 -> 2.58 | Improved |
| add risk-aware random walk instead of empirical threshold (ablation) | WalkSafe, Qwen3-4B · TET | avg harmlessness (lower better): 18.24 -> 15.63 | Improved |
| add risk-aware random walk instead of empirical threshold (ablation) | WalkSafe, Llama-3.2-3B · TET | avg harmlessness (lower better): 6.62 -> 3.25 | Improved |
| use Bi-GRPO instead of vanilla GRPO (ablation) | WalkSafe, Qwen2.5-3B · TET | helpfulness: 1.02 -> 2.58 | Improved |
| use Bi-GRPO instead of vanilla GRPO (ablation) | WalkSafe, Llama-3.2-3B · TET | avg harmlessness (lower better): 3.41 -> 3.25 | Improved |
| use Bi-GRPO instead of vanilla GRPO (ablation) | WalkSafe, Qwen3-4B · TET | avg harmlessness (lower better): 18.06 -> 15.63 | Improved |
| apply WalkSafe post-training | Llama-3.2-3B-Instruct · TET | avg harmlessness (lower better): 11.22 -> 3.25 | Improved |
| apply WalkSafe post-training | Qwen3-4B · TET | avg harmlessness (lower better): 24.73 -> 15.62 | Improved |
| apply WalkSafe post-training | Qwen2.5-3B-Instruct · TET | helpfulness: 0.04 -> 2.58 | Improved |
| apply STAIR safety training | Llama-3.2-3B-Instruct · TET | helpfulness: -0.12 -> -0.49 | Hurt |
| apply STAIR safety training | Qwen2.5-3B-Instruct · TET | avg harmlessness (lower better): 16.17 -> 7.09 | Improved |
Closest papers by meaning
- Safety Pretraining: Toward the Next Generation of Safe AI (NeurIPS 2025 · similarity 0.90)
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · similarity 0.90)
- Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM (AAAI 2025 · similarity 0.90)
- On Prompt-Driven Safeguarding for Large Language Models (ICML 2024 · similarity 0.89)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge (NeurIPS 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.
