Sabotage Detection on TRACE
90.1Average PrecisionMonitor
Evaluation Results
| Method | Links | |
|---|---|---|
| MonitorTPC=50, Model=Qwen3.52026.04 | 90.1 | |
| MeerkatTPC=10, Model=Qwen3.52026.04 | 88.2 | |
| MeerkatTPC=100, Model=Qwen3.52026.04 | 88.1 | |
| MeerkatTPC=50, Model=Qwen3.52026.04 | 87.4 | |
| MonitorTPC=10, Model=GLM-52026.04 | 86.1 | |
| MonitorTPC=100, Model=Qwen3.52026.04 | 84.9 | |
| MonitorTPC=10, Model=Qwen3.52026.04 | 84.3 | |
| MonitorTPC=25, Model=GLM-52026.04 | 83 | |
| MonitorTPC=10, Model=GPT-5.4m2026.04 | 82.2 | |
| Naive AgentTPC=50, Model=Qwen3.52026.04 | 81.9 | |
| MonitorTPC=25, Model=Qwen3.52026.04 | 80.8 | |
| MeerkatTPC=25, Model=Qwen3.52026.04 | 77.2 | |
| MonitorTPC=50, Model=GLM-52026.04 | 75.4 | |
| Naive AgentTPC=10, Model=Qwen3.52026.04 | 75.1 | |
| Naive AgentTPC=25, Model=Qwen3.52026.04 | 74.3 | |
| MeerkatTPC=10, Model=GPT-5.4m2026.04 | 71.7 | |
| MeerkatTPC=10, Model=GLM-52026.04 | 71.4 | |
| MeerkatTPC=100, Model=GLM-52026.04 | 70 | |
| MonitorTPC=100, Model=GLM-52026.04 | 70 | |
| MeerkatTPC=25, Model=GLM-52026.04 | 68 | |
| MonitorTPC=50, Model=GPT-5.4m2026.04 | 66.1 | |
| MeerkatTPC=50, Model=GLM-52026.04 | 63.8 | |
| MonitorTPC=25, Model=GPT-5.4m2026.04 | 63.4 | |
| Naive AgentTPC=100, Model=Qwen3.52026.04 | 62.1 | |
| MeerkatTPC=50, Model=GPT-5.4m2026.04 | 59.3 | |
| MonitorTPC=100, Model=GPT-5.4m2026.04 | 56.4 | |
| MeerkatTPC=100, Model=GPT-5.4m2026.04 | 54.6 | |
| MeerkatTPC=25, Model=GPT-5.4m2026.04 | 53.9 |