Mathematical Reasoning on AMC23 (Pass@1, Pass@32)
74.4Pass@1DynaMO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DynaMOBackbone=Qwen2.5-Math-7B2026.02 | 74.4 | 92.9 | |
| Clip-COVBackbone=Qwen2.5-Math-7B2026.02 | 72.7 | 89.3 | |
| W-REINFORCEBackbone=Qwen2.5-Math-7B2026.02 | 72.5 | 89.8 | |
| Fork TokensBackbone=Qwen2.5-Math-7B2026.02 | 71 | 87.3 | |
| KL-COVBackbone=Qwen2.5-Math-7B2026.02 | 70.6 | 88.5 | |
| GRPOBackbone=Qwen2.5-Math-7B2026.02 | 68.3 | 90.8 | |
| Entropy AdvantagesBackbone=Qwen2.5-Math-7B2026.02 | 67.9 | 85.2 | |
| Clip-HigherBackbone=Qwen2.5-Math-7B2026.02 | 67.8 | 89.9 | |
| Entropy LossBackbone=Qwen2.5-Math-7B2026.02 | 66 | 87 | |
| DynaMOBackbone=Qwen2.5-Math-1.5B2026.02 | 63.6 | 91.9 | |
| W-REINFORCEBackbone=Qwen2.5-Math-1.5B2026.02 | 63 | 85.7 | |
| Entropy AdvantagesBackbone=Qwen2.5-Math-1.5B2026.02 | 62 | 86.4 | |
| Clip-COVBackbone=Qwen2.5-Math-1.5B2026.02 | 59.5 | 89.7 | |
| GRPOBackbone=Qwen2.5-Math-1.5B2026.02 | 56 | 90 | |
| KL-COVBackbone=Qwen2.5-Math-1.5B2026.02 | 55.8 | 91.3 | |
| Entropy LossBackbone=Qwen2.5-Math-1.5B2026.02 | 55.6 | 86.9 | |
| Fork TokensBackbone=Qwen2.5-Math-1.5B2026.02 | 52.5 | 85.6 | |
| Clip-HigherBackbone=Qwen2.5-Math-1.5B2026.02 | 50.6 | 89.9 |