Real-world Harm Prediction on ATBench
39AccuracyTRACES
Evaluation Results
| Method | Links | |
|---|---|---|
| TRACESBase Model=Qwen3-4B2026.05 | 39 | |
| TRACESBase Model=Llama3.1-8B-Instruct2026.05 | 36.4 | |
| AgentDoGBase Model=Qwen3-4B2026.05 | 35.6 | |
| Gemini-3-FlashBase Model=Original2026.05 | 34.8 | |
| QWQ-32BBase Model=Original2026.05 | 34.8 | |
| AgentDoGBase Model=Llama3.1-8B-Instruct2026.05 | 34.4 | |
| Gemini-3-ProBase Model=Original2026.05 | 32 | |
| GPT-5.2Base Model=Original2026.05 | 30.8 | |
| Qwen3-4BBase Model=Original2026.05 | 16.4 | |
| Llama3.1-8BBase Model=Original2026.05 | 15.8 |