Mathematical Reasoning on MATH 500 (avg@1)
81.6Avg@1 ScoreVIMPO
Evaluation Results
| Method | Links | |
|---|---|---|
| VIMPOReward condition=Clean, Training steps=Full training2026.06 | 81.6 | |
| Naive GRPOReward condition=Clean, Training steps=Full training2026.06 | 80.4 | |
| GRPOReward condition=Clean, Training steps=Full training2026.06 | 79.6 | |
| VIMPOReward condition=25% reward flipping, Training steps=200 steps2026.06 | 78.2 | |
| GRPOReward condition=25% reward flipping, Training steps=200 steps2026.06 | 75.2 | |
| Qwen3-4B-BaseReward condition=Clean, Training steps=Full training2026.06 | 54 |