General Reasoning on ARC-C (Accuracy, Token usage, Speed)
95.5AccuracyLatentMAS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LatentMASBackbone=Qwen3-14B, MAS Setting=Hierarchical2025.11 | 95.5 | 295 | 1,090 | |
| TextMASBackbone=Qwen3-14B, MAS Setting=Hierarchical2025.11 | 95.3 | 2,167 | 4,283 | |
| LatentMASBackbone=Qwen3-8B, MAS Setting=Hierarchical2025.11 | 93.9 | 344 | 714 | |
| TextMASBackbone=Qwen3-8B, MAS Setting=Hierarchical2025.11 | 93.3 | 2,854 | 2,034 | |
| SingleBackbone=Qwen3-14B, MAS Setting=Hierarchical2025.11 | 92.6 | 773 | 338 | |
| TextMASBackbone=Qwen3-4B, MAS Setting=Hierarchical2025.11 | 92.5 | 2,674 | 1,275 | |
| LatentMASBackbone=Qwen3-4B, MAS Setting=Hierarchical2025.11 | 91.7 | 447 | 299 | |
| SingleBackbone=Qwen3-8B, MAS Setting=Hierarchical2025.11 | 91 | 846 | 266 | |
| SingleBackbone=Qwen3-4B, MAS Setting=Hierarchical2025.11 | 89.2 | 913 | 97 |