Harmfulness Evaluation on AgentHarm
62.42HS ScoreReAct
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ReActAgent Backbone=Average over four backbones (Qwen3-32B, Kimi-2.5, Gemini-2.5-Pro, GPT-5.1)2026.06 | 62.42 | 47.83 | |
| TRIAD + Qwen3.5-9BAgent Backbone=Average over four backbones (Qwen3-32B, Kimi-2.5, Gemini-2.5-Pro, GPT-5.1)2026.06 | 73.53 | 9.9 | |
| TRIAD + Tri-GuardAgent Backbone=Average over four backbones (Qwen3-32B, Kimi-2.5, Gemini-2.5-Pro, GPT-5.1)2026.06 | 80.92 | 13.05 |