Length Bias Mitigation on Length Bias
347Avg LengthPPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PPOCondition=Unmitigated2026.02 | 347 | 23.1 | |
| PPO W/KLCategory=Regularization Methods2026.02 | 268 | 22.8 | |
| FILTERING (R > mu + 2sigma)Category=Training Interventions2026.02 | 241 | 22.6 | |
| RM ENSEMBLECategory=Reward Model Improvements2026.02 | 224 | 23 | |
| INFORMCategory=Regularization Methods2026.02 | 208 | 23.2 | |
| ODINCategory=Regularization Methods2026.02 | 195 | 23.4 | |
| ARAFramework=Auditor-gated RLHF2026.02 | 162 | 24.1 | |
| SFT (NO RLHF)RLHF=None2026.02 | 148 | 21.4 |