Mathematical Reasoning on GSM8K (Pass@k Metrics)
90.13Pass@1HRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HRPOBackbone=Qwen2.5-7B-Instruct2026.06 | 90.13 | 97.42 | |
| TARPOBackbone=Qwen2.5-7B-Instruct2026.06 | 89.94 | 97.27 | |
| GRPOBackbone=Qwen2.5-7B-Instruct2026.06 | 89.87 | 97.42 | |
| Pure LatentBackbone=Qwen2.5-7B-Instruct2026.06 | 88.4 | 90.9 | |
| Entropy-RoutedBackbone=Qwen2.5-7B-Instruct2026.06 | 88.07 | 97.04 | |
| CoTBackbone=Qwen2.5-7B-Instruct2026.06 | 87.97 | 97.73 | |
| TARPOBackbone=Qwen2.5-3B-Instruct2026.06 | 83.23 | 96.97 | |
| HRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 82.75 | 97.42 | |
| GRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 82.55 | 97.57 | |
| Entropy-RoutedBackbone=Qwen2.5-3B-Instruct2026.06 | 73.91 | 95.53 | |
| Pure LatentBackbone=Qwen2.5-3B-Instruct2026.06 | 73.84 | 78.92 | |
| CoTBackbone=Qwen2.5-3B-Instruct2026.06 | 73.5 | 96.97 | |
| TARPOBackbone=Qwen2.5-1.5B-Instruct2026.06 | 70.76 | 96.97 | |
| HRPOBackbone=Qwen2.5-1.5B-Instruct2026.06 | 69.71 | 96.66 | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct2026.06 | 69.04 | 96.66 | |
| CoTBackbone=Qwen2.5-1.5B-Instruct2026.06 | 60.89 | 95.53 | |
| Entropy-RoutedBackbone=Qwen2.5-1.5B-Instruct2026.06 | 60.66 | 93.1 | |
| Pure LatentBackbone=Qwen2.5-1.5B-Instruct2026.06 | 60.53 | 65.5 |