Reasoning Evaluation on ParaRev
63.7WR vs. BaseFull DRO
Evaluation Results
| Method | Links | |
|---|---|---|
| Full DROReward=DRO, Rollout-Group Rejection (Rubric)=true, Rollout-Group Rejection (Variance)=true2025.06 | 63.7 | |
| R3Reward=R3, Rollout-Group Rejection (Rubric)=true, Rollout-Group Rejection (Variance)=false2025.06 | 61.6 | |
| R3Reward=R3, Rollout-Group Rejection (Rubric)=false, Rollout-Group Rejection (Variance)=false2025.06 | 57.8 | |
| Rubric (RLER)Reward=Rubric (RLER), Rollout-Group Rejection (Rubric)=false, Rollout-Group Rejection (Variance)=false2025.06 | 55.9 | |
| Avg Prob (RLPR)Reward=Avg Prob (RLPR), Rollout-Group Rejection (Rubric)=false, Rollout-Group Rejection (Variance)=false2025.06 | 54.3 | |
| BaseReward=Base2025.06 | 50 | |
| Avg Logprob (VeriFree)Reward=Avg Logprob (VeriFree), Rollout-Group Rejection (Rubric)=false, Rollout-Group Rejection (Variance)=false2025.06 | 49.5 | |
| RL-F1Reward=RL-F1, Rollout-Group Rejection (Rubric)=false, Rollout-Group Rejection (Variance)=false2025.06 | 40.8 |