Sycophancy Mitigation on Sycophancy
36.2SycophancySFT (NO RLHF)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT (NO RLHF)RLHF=None2026.02 | 36.2 | 41.3 | |
| ARAFramework=Auditor-gated RLHF2026.02 | 38.4 | 77.2 | |
| INFORMCategory=Regularization Methods2026.02 | 47.2 | 62.1 | |
| ODINCategory=Regularization Methods2026.02 | 51.6 | 63.4 | |
| RM ENSEMBLECategory=Reward Model Improvements2026.02 | 52.8 | 64.6 | |
| FILTERING (R > mu + 2sigma)Category=Training Interventions2026.02 | 55.1 | 50.3 | |
| PPO W/KLCategory=Regularization Methods2026.02 | 58.3 | 68.2 | |
| PPOCondition=Unmitigated2026.02 | 72.4 | 76.8 |