Mathematical Reasoning on MATH (Pass@1, Pass@32)
70.25Pass@1HRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HRPOBackbone=Qwen2.5-7B-Instruct2026.06 | 70.25 | 88.2 | |
| TARPOBackbone=Qwen2.5-7B-Instruct2026.06 | 70.22 | 88.22 | |
| GRPOBackbone=Qwen2.5-7B-Instruct2026.06 | 70.05 | 88.02 | |
| CoTBackbone=Qwen2.5-7B-Instruct2026.06 | 61.16 | 87.3 | |
| HRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 59.97 | 85.06 | |
| GRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 59.86 | 84.72 | |
| TARPOBackbone=Qwen2.5-3B-Instruct2026.06 | 59.75 | 85.2 | |
| Pure LatentBackbone=Qwen2.5-7B-Instruct2026.06 | 56.78 | 68.66 | |
| CoTBackbone=Qwen2.5-3B-Instruct2026.06 | 55.64 | 85 | |
| Pure LatentBackbone=Qwen2.5-3B-Instruct2026.06 | 55.45 | 59.7 | |
| Entropy-RoutedBackbone=Qwen2.5-3B-Instruct2026.06 | 54.75 | 83.36 | |
| Entropy-RoutedBackbone=Qwen2.5-7B-Instruct2026.06 | 54.02 | 82.96 | |
| TARPOBackbone=Qwen2.5-1.5B-Instruct2026.06 | 49.71 | 82.28 | |
| HRPOBackbone=Qwen2.5-1.5B-Instruct2026.06 | 48.99 | 82.34 | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct2026.06 | 47.82 | 81.22 | |
| CoTBackbone=Qwen2.5-1.5B-Instruct2026.06 | 31.66 | 78.9 | |
| Pure LatentBackbone=Qwen2.5-1.5B-Instruct2026.06 | 25.29 | 37.2 | |
| Entropy-RoutedBackbone=Qwen2.5-1.5B-Instruct2026.06 | 25.29 | 72.04 |