Deception Evaluation on Sycophancy
89.77CoT Plan AccuracyGRPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GRPOBackbone=Llama-3.1-8B2026.03 | 89.77 | 72.27 | 74.54 | |
| GRPOBackbone=Qwen3-8B2026.03 | 66.59 | 37.95 | 72.5 | |
| CoT MonitorBackbone=Qwen3-8B2026.03 | 35.68 | 18.86 | 84.54 | |
| CoT MonitorBackbone=Llama-3.1-8B2026.03 | 19.31 | 72.5 | 30.91 | |
| SARBackbone=Qwen3-8B2026.03 | 16.81 | 9.77 | 89.31 | |
| SARBackbone=Llama-3.1-8B2026.03 | 14.31 | 35 | 70.9 |