Code Generation on MBPP+ (Accuracy, Token Usage, Speed)
0.738AccuracyLatentMAS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LatentMASBackbone=Qwen3-14B, MAS Setting=Hierarchical2025.11 | 0.738 | 1,187 | 2,507 | |
| TextMASBackbone=Qwen3-14B, MAS Setting=Hierarchical2025.11 | 0.73 | 7,458 | 9,162 | |
| LatentMASBackbone=Qwen3-8B, MAS Setting=Hierarchical2025.11 | 0.722 | 1,264 | 1,387 | |
| TextMASBackbone=Qwen3-8B, MAS Setting=Hierarchical2025.11 | 0.719 | 7,703 | 3,898 | |
| LatentMASBackbone=Qwen3-4B, MAS Setting=Hierarchical2025.11 | 0.706 | 1,339 | 489 | |
| TextMASBackbone=Qwen3-4B, MAS Setting=Hierarchical2025.11 | 0.693 | 6,782 | 1,766 | |
| SingleBackbone=Qwen3-14B, MAS Setting=Hierarchical2025.11 | 0.685 | 1,858 | 2,410 | |
| SingleBackbone=Qwen3-8B, MAS Setting=Hierarchical2025.11 | 0.648 | 2,053 | 1,064 | |
| SingleBackbone=Qwen3-4B, MAS Setting=Hierarchical2025.11 | 0.635 | 1,634 | 523 |