Mathematical Reasoning on Minerva Math (test)
55.5Best-of-8 Accuracypass@8 (Upper Bound)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| pass@8 (Upper Bound)Policy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=pass@8, Evaluation Seeds=32026.05 | 55.5 | — | — | — | — | |
| Implicit PRM (CE)Policy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 47.4 | — | — | — | — | |
| Math-Shepherd-PRM-7BPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 47.3 | — | — | — | — | |
| RLHFlow-PRM-Deepseek-8BPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 47.2 | — | — | — | — | |
| Qwen2.5-Math-7B-PRMPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 47.2 | — | — | — | — | |
| Implicit PRM (DPO)Policy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 47.2 | — | — | — | — | |
| Qwen2.5-Math-7B-PRM800KPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 47.1 | — | — | — | — | |
| RLHFlow-PRM-Mistral-8BPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 46.9 | — | — | — | — | |
| sPRMPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 46.7 | — | — | — | — | |
| uPRMPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 46.7 | — | — | — | — | |
| Skywork-PRM-7BPolicy Model=Qwen2.5-Math-7B-Instruct, Decoding Strategy=Best-of-8, Evaluation Seeds=32026.05 | 46.6 | — | — | — | — | |
| BaseModel=Qwen2.5-1.5B2026.05 | — | — | — | 13.97 | 6.99 | |
| DAPO+IMAXModel=Qwen3-4B2026.05 | — | — | — | 23.9 | 10.85 | |
| DAPO+IMAXModel=Qwen3-8B2026.05 | — | — | — | 23.16 | 11.4 | |
| DEERTarget Model=Qwen3-30B-A3B, Temperature=0.62025.12 | — | 2.02 | 2.31 | — | — | |
| EAGLE3Target Model=Qwen3-30B-A3B, Temperature=0.62025.12 | — | 1.91 | 2.07 | — | — | |
| GRPO-fullModel=Qwen2.5-1.5B2026.05 | — | — | — | 18.01 | 10.94 | |
| GRPO-fullModel=Qwen3-8B2026.05 | — | — | — | 29.41 | 22.33 | |
| GRPO+IMAXModel=Qwen2.5-1.5B2026.05 | — | — | — | 18.38 | 9.47 |