Mathematical Reasoning on Minerva (pass@4, pass@8, pass@16)
49.03Pass@4Qwen2.5-7B EP-GRPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen2.5-7B EP-GRPOModel=Qwen2.5-7B, Method=EP-GRPO2026.05 | 49.03 | 74.14 | 93.44 | |
| Qwen2.5-7B GRPOModel=Qwen2.5-7B, Method=GRPO2026.05 | 43.12 | 67.77 | 89.77 | |
| Qwen2.5-3B EP-GRPOModel=Qwen2.5-3B, Method=EP-GRPO2026.05 | 31.68 | 53.44 | 78.54 | |
| Qwen2.5-3B GRPOModel=Qwen2.5-3B, Method=GRPO2026.05 | 22.8 | 40.5 | 64.84 | |
| Qwen3-235B-A22B-InstructModel=Qwen3-235B-A22B-Instruct, Category=Commercial Models2026.05 | 21.44 | 23.4 | 25 | |
| Qwen2.5-3B GRPO (Higher Temp)Model=Qwen2.5-3B, Method=GRPO, Temperature=Higher2026.05 | 21.24 | 38.07 | 61.88 | |
| Qwen2.5-3B GRPO (More Rollouts)Model=Qwen2.5-3B, Method=GRPO, Rollouts=More2026.05 | 21.24 | 38.07 | 61.88 | |
| DeepSeek-R1-671B-0528Model=DeepSeek-R1-671B-0528, Category=Commercial Models2026.05 | 20.94 | 23.78 | 25 | |
| Qwen2.5-3B BaseModel=Qwen2.5-3B, Method=Base2026.05 | 11.95 | 22.54 | 40.2 | |
| Qwen2.5-7B BaseModel=Qwen2.5-7B, Method=Base2026.05 | 11.95 | 22.54 | 40.2 |