Failure Mode Prediction on ATBench
41AccuracyTRACES
Evaluation Results
| Method | Links | |
|---|---|---|
| TRACESBase Model=Llama3.1-8B-Instruct2026.05 | 41 | |
| TRACESBase Model=Qwen3-4B2026.05 | 34.2 | |
| AgentDoGBase Model=Qwen3-4B2026.05 | 28.8 | |
| AgentDoGBase Model=Llama3.1-8B-Instruct2026.05 | 26 | |
| Gemini-3-FlashBase Model=Original2026.05 | 22.4 | |
| GPT-5.2Base Model=Original2026.05 | 20.4 | |
| Gemini-3-ProBase Model=Original2026.05 | 17.6 | |
| QWQ-32BBase Model=Original2026.05 | 14.4 | |
| Qwen3-4BBase Model=Original2026.05 | 8.6 | |
| Llama3.1-8BBase Model=Original2026.05 | 6.8 |