Safety Detection on SorryBench clean condition (held-out)
95Detection RateFinal-token probe
Evaluation Results
| Method | Links | |
|---|---|---|
| Final-token probeModel=Llama-3.1-8B2026.05 | 95 | |
| Final-token probeModel=OLMo3-7B2026.05 | 95 | |
| Final-token probeModel=Mistral-7B2026.05 | 94.6 |