Mathematical Reasoning on Minerva (Pass@1, Pass@32)
27.3Pass@1DynaMO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DynaMOBackbone=Qwen2.5-Math-7B2026.02 | 27.3 | 47.2 | |
| W-REINFORCEBackbone=Qwen2.5-Math-7B2026.02 | 26.4 | 49.5 | |
| Fork TokensBackbone=Qwen2.5-Math-7B2026.02 | 26.1 | 42.4 | |
| Clip-COVBackbone=Qwen2.5-Math-7B2026.02 | 25.4 | 45.9 | |
| KL-COVBackbone=Qwen2.5-Math-7B2026.02 | 24.5 | 39.9 | |
| Clip-HigherBackbone=Qwen2.5-Math-7B2026.02 | 24 | 46.1 | |
| Entropy AdvantagesBackbone=Qwen2.5-Math-7B2026.02 | 23.7 | 43.7 | |
| Entropy LossBackbone=Qwen2.5-Math-7B2026.02 | 23.3 | 45.9 | |
| GRPOBackbone=Qwen2.5-Math-7B2026.02 | 22.6 | 45.4 | |
| DynaMOBackbone=Qwen2.5-Math-1.5B2026.02 | 19.4 | 44 | |
| Entropy AdvantagesBackbone=Qwen2.5-Math-1.5B2026.02 | 18.2 | 43 | |
| W-REINFORCEBackbone=Qwen2.5-Math-1.5B2026.02 | 18.2 | 40.3 | |
| Entropy LossBackbone=Qwen2.5-Math-1.5B2026.02 | 17.6 | 43.6 | |
| GRPOBackbone=Qwen2.5-Math-1.5B2026.02 | 17.2 | 42.8 | |
| Clip-HigherBackbone=Qwen2.5-Math-1.5B2026.02 | 16.8 | 41.3 | |
| Fork TokensBackbone=Qwen2.5-Math-1.5B2026.02 | 16.6 | 36.8 | |
| Clip-COVBackbone=Qwen2.5-Math-1.5B2026.02 | 15.8 | 44.3 | |
| KL-COVBackbone=Qwen2.5-Math-1.5B2026.02 | 14.8 | 40.3 |