Sycophancy Mitigation on BCT Sycophancy (held-out)
100BRR RatioBase model
Evaluation Results
| Method | Links | |
|---|---|---|
| Base modelNote=no training2026.06 | 100 | |
| BCTTrain threat=Frustration2026.06 | 94 | |
| BCTTrain threat=Jailbreak2026.06 | 42 | |
| MLPCTTrain threat=Sycophancy2026.06 | 4 |