Mathematical Reasoning on AIME 25 (Acc, Tok, ReL)
36.7AccuracyMixed Attention Module
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mixed Attention ModuleSeries=Qwen2.5-7B, beta=0.4, alpha=0.32026.02 | 36.7 | 13,310 | 442.2 | |
| OursBackbone=Qwen2.5-7B2026.02 | 33.3 | 13,191 | 440.6 | |
| Qwen2.5-7B (Base)Series=Qwen2.5-7B, beta=0.4, alpha=0.32026.02 | 30 | 13,204 | 548.1 | |
| Ours (Base)Backbone=Qwen2.5-7B2026.02 | 30 | 13,204 | 548.1 | |
| INFTYTHINKBackbone=Qwen2.5-7B2026.02 | 30 | 15,448 | 530.5 | |
| LightThinkerBackbone=Qwen2.5-7B2026.02 | 26.7 | 13,250 | 443.9 | |
| OursBackbone=Qwen2.5-1.5B2026.02 | 23.3 | 12,910 | 416 | |
| H2OBackbone=Qwen2.5-7B2026.02 | 23.3 | 13,204 | 443.7 | |
| SepLLMBackbone=Qwen2.5-7B2026.02 | 23.3 | 12,972 | 435.9 | |
| Ours (Base)Backbone=Qwen2.5-1.5B2026.02 | 16.7 | 12,684 | 504.7 | |
| INFTYTHINKBackbone=Qwen2.5-1.5B2026.02 | 16.7 | 15,020 | 501 | |
| SepLLMBackbone=Qwen2.5-1.5B2026.02 | 13.3 | 12,690 | 423.9 | |
| LightThinkerBackbone=Qwen2.5-1.5B2026.02 | 13.3 | 12,934 | 434.6 | |
| H2OBackbone=Qwen2.5-1.5B2026.02 | 10 | 12,802 | 428.6 |