Mathematical Reasoning on MATH-500 (val)
90.07Avg@5 AccuracyGRPO w/ Clip-higher
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO w/ Clip-higherBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 90.07 | |
| AsymGRPO w/ Clip-higherBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 89.73 | |
| AsymGRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 89.33 | |
| Dr.GRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 88.87 | |
| AsymGRPO (βpos = βneg)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 88.53 | |
| GRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 88.2 | |
| GRPO w/ Entro.RegularizationBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 88.2 | |
| Pos-Only Modulation (§ 3.1)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 87.13 | |
| Neg-Only Modulation (§ 3.1)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 87 | |
| GRPO w/ Entro.AdvBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 86.73 | |
| REINFORCEBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 86.6 | |
| Pass@K TrainingBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 86.33 | |
| EntIncrease (§ 3.2)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 85.6 | |
| EntDecrease (§ 3.2)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 83.73 | |
| Qwen3-4BTemperature=0.4, Top-p=1.02026.04 | 81.6 |