Mathematical Reasoning on MATH 500 (Accuracy, TNFT, Tokens, Delay)
64AccuracyOurs
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| OursBackbone=Qwen3-4B2026.05 | 64 | 0 | 742.26 | 26.51 | |
| VanillaBackbone=Qwen3-4B2026.05 | 60.8 | 103.16 | 1,363.95 | 45.82 | |
| BaseBackbone=Qwen3-4B2026.05 | 60 | 130.1 | 3,678.51 | 126.9 | |
| OursBackbone=Qwen3-1.7B2026.05 | 51.6 | 0 | 803 | 22.94 | |
| BaseBackbone=Qwen3-1.7B2026.05 | 48.4 | 130.1 | 3,229.16 | 88.15 | |
| VanillaBackbone=Qwen3-1.7B2026.05 | 48.2 | 103.16 | 1,612.76 | 43.14 |