Mathematical Reasoning on GSM8K (PR, Faithfulness, AUROC, Step Analysis)
0.6Performance Ratio (PR)GRPO Baseline
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GRPO BaselineCondition=Baseline, Decoding strategy=greedy@12026.05 | 0.6 | — | 99.4 | 77.8 | 0.924 | 2,351 | |
| ProFILCondition=ProFIL, Decoding strategy=greedy@12026.05 | 0 | 100 | 100 | 83 | — | 1,895 |