Safety Detection on XSTest benign safety-adjacent
23False Positive RateFinal-token probe
Evaluation Results
| Method | Links | |
|---|---|---|
| Final-token probeModel=Llama-3.1-8B2026.05 | 23 | |
| Final-token probeModel=Mistral-7B2026.05 | 32.5 | |
| Final-token probeModel=OLMo3-7B2026.05 | 36 |