Accuracy on GPQA Diamond (Scientific Reasoning)
40.4AccuracyAMARIS
Evaluation Results
| Method | Links | |
|---|---|---|
| AMARISRubric Strategy=per-instance rubric2026.05 | 40.4 | |
| AMARISRubric Strategy=global rubric2026.05 | 39.9 | |
| RubricHub (RuRL)RL Training Approach=RuRL2026.05 | 38.8 | |
| RuscaRLRL Training Approach=RuscaRL2026.05 | 38.5 | |
| RaRRL Training Approach=RaR2026.05 | 37.6 | |
| Naive (no RL)RL Training Approach=None2026.05 | 35 |