Deception Detection on 7 non-control benchmark families GPT-OSS-20B
0.92AUROCSTATEWITNESS
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| STATEWITNESS2026.06 | 0.92 | 29.3 | 68.1 | |
| Best black-boxSelection=Best configuration2026.06 | 0.809 | 30.7 | 54.7 | |
| Best linear probeSelection=Best configuration2026.06 | 0.705 | 0.6 | 4 |