Mathematical Reasoning on AIME 2025 (Avg@10 accuracy)
36Avg@10 AccuracyAsymGRPO w/ Clip-higher
Evaluation Results
| Method | Links | |
|---|---|---|
| AsymGRPO w/ Clip-higherBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 36 | |
| GRPO w/ Clip-higherBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 32.33 | |
| Pass@K TrainingBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 31 | |
| Dr.GRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 30 | |
| AsymGRPO (βpos = βneg)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 29.33 | |
| AsymGRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 28.67 | |
| GRPO w/ Entro.RegularizationBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 28.33 | |
| Pos-Only Modulation (§ 3.1)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 28 | |
| GRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 27.33 | |
| Neg-Only Modulation (§ 3.1)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 27 | |
| GRPO w/ Entro.AdvBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 25.33 | |
| REINFORCEBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 24.67 | |
| EntDecrease (§ 3.2)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 23.67 | |
| EntIncrease (§ 3.2)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 23.33 | |
| Qwen3-4BTemperature=0.4, Top-p=1.02026.04 | 20 |