Safety Detection on SorryBench wrapped with HarmBench templates (held-out)
86.9Detection RateFinal-token probe
Evaluation Results
| Method | Links | |
|---|---|---|
| Final-token probeModel=Mistral-7B2026.05 | 86.9 | |
| Final-token probeModel=Llama-3.1-8B2026.05 | 72.2 | |
| Final-token probeModel=OLMo3-7B2026.05 | 63.6 |