Code Generation on HumanEval (P@1, P@32, #Tok)
63.62Pass@1TARPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TARPOBackbone=Qwen2.5-3B-Instruct2026.06 | 63.62 | 88.41 | 256.1 | |
| Pure LatentBackbone=Qwen2.5-3B-Instruct2026.06 | 60.19 | 82.32 | 261.4 | |
| CoTBackbone=Qwen2.5-3B-Instruct2026.06 | 59.68 | 90.24 | 271.4 | |
| HRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 59.51 | 90.85 | 340.5 | |
| GRPOBackbone=Qwen2.5-3B-Instruct2026.06 | 58.86 | 92.07 | 355.6 | |
| Entropy-RoutedBackbone=Qwen2.5-3B-Instruct2026.06 | 58.31 | 90.24 | 260.5 |