Mathematical Reasoning on GSM8K (Acc, Time, Peak)
91.28AccuracyVanilla + SepLLM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Vanilla + SepLLMBackbone=Qwen2.5-7B Series2026.06 | 91.28 | 49.03 | 1,024 | |
| Vanilla + H2OBackbone=Qwen2.5-7B Series2026.06 | 91.05 | 37.98 | 1,024 | |
| VanillaBackbone=Qwen2.5-7B Series2026.06 | 90.83 | 22.19 | 1,559 | |
| HybridThinkerBackbone=Qwen2.5-7B Series2026.06 | 90.6 | 20.85 | 864 | |
| VanillaBackbone=Llama3.1-8B Series2026.06 | 89.31 | 27.89 | 1,702 | |
| Vanilla + H2OBackbone=Llama3.1-8B Series2026.06 | 89.31 | 48.08 | 1,024 | |
| CoTBackbone=Qwen2.5-7B Series2026.06 | 88.32 | 3.49 | 519 | |
| Vanilla + SepLLMBackbone=Llama3.1-8B Series2026.06 | 88.32 | 27.92 | 1,024 | |
| HybridThinkerBackbone=Llama3.1-8B Series2026.06 | 87.34 | 25.5 | 878 | |
| LightThinkerBackbone=Qwen2.5-7B Series2026.06 | 87.26 | 23.05 | 684 | |
| LightThinkerBackbone=Llama3.1-8B Series2026.06 | 85.82 | 25.87 | 664 | |
| CoTBackbone=Llama3.1-8B Series2026.06 | 64.52 | 3.63 | 497 | |
| Distill-R1Backbone=Qwen2.5-7B Series2026.06 | 60.58 | 6.54 | 522 | |
| Distill-R1Backbone=Llama3.1-8B Series2026.06 | 57.62 | 7.2 | 524 |