Sabotage detection on MLE-Sabotage
0.87log-AUROCAction-only
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Action-onlyMonitor Model=GPT-4.12026.01 | 0.87 | — | |
| Extract-and-EvaluateMonitor Model=Claude-3.5-Haiku2026.01 | 0.81 | — | |
| Extract-and-EvaluateMonitor Model=GPT-4.12026.01 | 0.8 | — | |
| CoT-onlyMonitor Model=Claude-3.5-Haiku2026.01 | 0.72 | — | |
| CoT+actionMonitor Model=GPT-4.12026.01 | 0.58 | — | |
| CoT-onlyMonitor Model=GPT-4.12026.01 | 0.57 | — | |
| Action-onlyMonitor Model=Claude-3.5-Haiku2026.01 | 0.54 | — | |
| CoT+actionMonitor Model=Claude-3.5-Haiku2026.01 | 0.49 | — |