Mathematical Reasoning on GSM8K 500-example (test)
84.8Pass@1 FinalAH-GRPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AH-GRPOModel=3B, alpha=0.5, Steps=1502026.06 | 84.8 | 86.2 | 0.063 | 0.0051 | |
| GRPOModel=3B, alpha=0.0, Steps=1502026.06 | 84.6 | 85.2 | 0.0885 | 0.0054 | |
| SA-AH-GRPOModel=3B, alpha=0.5, Steps=1802026.06 | 84.6 | 85.8 | 0.0246 | 0.0079 | |
| BaseModel=3B, Evaluation Protocol=zero-shot2026.06 | 83.1 | — | — | — | |
| SA-AH-GRPOModel=1.5B, alpha=0.5, Steps=1802026.06 | 68.6 | 68.6 | 0.1718 | 0.0099 | |
| GRPOModel=1.5B, alpha=0.0, Steps=1502026.06 | 66 | 66.8 | 0.0885 | 0.0054 | |
| AH-GRPOModel=1.5B, alpha=0.5, Steps=1502026.06 | 65 | 67.6 | 0.063 | 0.0051 | |
| BaseModel=1.5B, Evaluation Protocol=zero-shot2026.06 | 63.7 | — | — | — |