RLHF on HH-RLHF (held-out)
1.59Peak Gold RewardDRRO-RLHF
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DRRO-RLHFType=soft + dynamic2026.04 | 1.59 | 3.03 | 35.74 | |
| GRPO2026.04 | 1.35 | 3.06 | 44.83 | |
| DRRO-RLHFType=hard2026.04 | 1.29 | 2.87 | 32.33 | |
| BSPO2026.04 | 1.21 | 3.02 | 42.61 | |
| PPO2026.04 | 1.2 | 2.43 | 18.25 | |
| InfoRM2026.04 | 1.14 | 3.06 | 44.25 | |
| Ensemble-UWO2026.04 | 0.99 | 2.84 | 34.51 | |
| Ensemble-Mean2026.04 | 0.94 | 2.71 | 32.05 | |
| DRO-RLHF2026.04 | 0.54 | 1.72 | 16.54 |