Trajectory-safety diagnosis on ATBench-F
49.2R.S. ScoreQwen2-7B-Instruct + FATE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen2-7B-Instruct + FATEModel Type=Ours2026.05 | 49.2 | 18.4 | 43.1 | |
| AgentDoG-Qwen2-7BModel Type=Guard models2026.05 | 46.8 | 16.5 | 40.6 | |
| GPT-4Model Type=Closed-source models2026.05 | 33.6 | 13.5 | 30.2 | |
| GPT-4oModel Type=Closed-source models2026.05 | 29.5 | 12 | 26.8 | |
| Gemini-1.5-ProModel Type=Closed-source models2026.05 | 24.8 | 12.6 | 18.5 | |
| Gemini-1.5-FlashModel Type=Closed-source models2026.05 | 18.4 | 8.3 | 15 | |
| QwQ-32B-PreviewModel Type=Open-source models2026.05 | 15.8 | 9.4 | 22.9 | |
| Qwen2.5-72B-InstructModel Type=Open-source models2026.05 | 7.7 | 3.6 | 6.8 | |
| Qwen2.5-32B-InstructModel Type=Open-source models2026.05 | 7 | 11.6 | 26.6 | |
| Qwen2.5-7BModel Type=Open-source models2026.05 | 6.6 | 3 | 8.2 | |
| Llama-3.1-8B-InstructModel Type=Open-source models2026.05 | 6.2 | 5.8 | 15.5 | |
| Qwen2.5-7B-InstructModel Type=Open-source models2026.05 | 5.3 | 6 | 15.5 | |
| Qwen2-7BModel Type=Open-source models2026.05 | 4.4 | 8.2 | 18.3 | |
| Qwen2-7B-InstructModel Type=Open-source models2026.05 | 1 | 9.6 | 21.2 |