Deception Detection on 7 non-control benchmark families GEMMA4-26B-A4B-IT
91.1AUROCSTATEWITNESS
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| STATEWITNESS2026.06 | 91.1 | 35.8 | 57.6 | |
| Best black-boxSelection=Best configuration2026.06 | 83.3 | 69.7 | 69.7 | |
| Best linear probeSelection=Best configuration2026.06 | 76.2 | 75.2 | 90.8 |