Sycophancy Evaluation on SycophancyEval
54.2Sycophancy RateLag-DPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Lag-DPOBase Model=Qwen3-4B-Instruct-25072026.04 | 54.2 | 45.8 | |
| VanillaBase Model=Qwen3-4B-Instruct-25072026.04 | 87.2 | 12.9 | |
| SFT-CombinedBase Model=Qwen3-4B-Instruct-25072026.04 | 87.6 | 12.5 | |
| SafeRLHFBase Model=Qwen3-4B-Instruct-25072026.04 | 87.8 | 12.2 | |
| Multi-Neg DPOBase Model=Qwen3-4B-Instruct-25072026.04 | 89.4 | 10.6 | |
| PCGrad-DPOBase Model=Qwen3-4B-Instruct-25072026.04 | 89.6 | 10.4 | |
| MODPOBase Model=Qwen3-4B-Instruct-25072026.04 | 96.2 | 3.8 | |
| SACPOBase Model=Qwen3-4B-Instruct-25072026.04 | 97.4 | 2.6 | |
| SFT-AnchorBase Model=Qwen3-4B-Instruct-25072026.04 | 97.6 | 2.4 |