Mathematical Reasoning on AMC 2023 (Avg@10 accuracy)
83.25Avg@10 AccuracyAsymGRPO w/ Clip-higher
Evaluation Results
| Method | Links | |
|---|---|---|
| AsymGRPO w/ Clip-higherBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 83.25 | |
| AsymGRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 81 | |
| GRPO w/ Clip-higherBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 78.5 | |
| AsymGRPO (βpos = βneg)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 78.5 | |
| GRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 78.25 | |
| Dr.GRPOBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 78.25 | |
| Neg-Only Modulation (§ 3.1)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 78 | |
| GRPO w/ Entro.AdvBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 77.25 | |
| Pos-Only Modulation (§ 3.1)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 76.75 | |
| GRPO w/ Entro.RegularizationBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 75.5 | |
| Pass@K TrainingBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 74 | |
| REINFORCEBase Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 73.75 | |
| EntIncrease (§ 3.2)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 71.75 | |
| EntDecrease (§ 3.2)Base Model=Qwen3-4B, Temperature=0.4, Top-p=1.02026.04 | 71.5 | |
| Qwen3-4BTemperature=0.4, Top-p=1.02026.04 | 63.75 |