Reasoning on GPQA Diamond (P@1, P@32, #Tok)
28.41Precision@1TARPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TARPOBackbone=Qwen2.5-3B-Instruct2026.06 | 28.41 | 90.91 | 568.3 | |
| Entropy-RoutedBackbone=Qwen2.5-3B-Instruct2026.06 | 28.3 | 90.91 | 373.2 | |
| GRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 28.03 | 83.33 | 608.5 | |
| HRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 27.51 | 88.38 | 603.3 | |
| Pure LatentBackbone=Qwen2.5-3B-Instruct2026.06 | 27.26 | 37.37 | 405.6 | |
| CoTBackbone=Qwen2.5-3B-Instruct2026.06 | 26.83 | 92.93 | 410.6 |