Reasoning on AIME 24 (Accuracy, Token Usage, and Speed)
73.3AccuracyLatentMAS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LatentMASMAS Setting=Hierarchical MAS Setting, Backbone=Qwen3-14B2025.11 | 73.3 | 10,230 | 1,089 | |
| TextMASMAS Setting=Hierarchical MAS Setting, Backbone=Qwen3-14B2025.11 | 70 | 29,025 | 5,718 | |
| LatentMASMAS Setting=Sequential MAS Setting, Backbone=Qwen3-14B2025.11 | 66.7 | 10,593 | 1,149 | |
| SingleMAS Setting=Sequential MAS Setting, Backbone=Qwen3-14B2025.11 | 63.3 | 11,263 | 1,018 | |
| TextMASMAS Setting=Sequential MAS Setting, Backbone=Qwen3-14B2025.11 | 63.3 | 32,092 | 4,554 | |
| SingleMAS Setting=Hierarchical MAS Setting, Backbone=Qwen3-14B2025.11 | 63.3 | 11,263 | 1,018 | |
| LatentMASMAS Setting=Sequential MAS Setting, Backbone=Qwen3-8B2025.11 | 56.7 | 8,953 | 688 | |
| TextMASMAS Setting=Sequential MAS Setting, Backbone=Qwen3-8B2025.11 | 53.3 | 38,596 | 2,808 | |
| TextMASMAS Setting=Hierarchical MAS Setting, Backbone=Qwen3-8B2025.11 | 53.3 | 42,629 | 3,132 | |
| LatentMASMAS Setting=Hierarchical MAS Setting, Backbone=Qwen3-8B2025.11 | 53.3 | 7,526 | 776 | |
| SingleMAS Setting=Sequential MAS Setting, Backbone=Qwen3-8B2025.11 | 50 | 12,891 | 421 | |
| SingleMAS Setting=Hierarchical MAS Setting, Backbone=Qwen3-8B2025.11 | 50 | 12,891 | 421 |