Reasoning on GSM8K, Math, AIME2024, HumanEval, and LiveCodeBench
87.23GSM8KReasoning
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ReasoningBackbone=Qwen2.5-7B2026.01 | 87.23 | 86.2 | 60 | 76.63 | 30.37 | |
| ReasonAnyBackbone=Qwen2.5-7B2026.01 | 86.28 | 69.4 | 33.33 | 64.65 | 26.71 | |
| LEDBackbone=Qwen2.5-7B2026.01 | 72.48 | 60.6 | 10 | 52.91 | 24.51 | |
| SafetyBackbone=Qwen2.5-7B2026.01 | 69.42 | 74 | 13.33 | 50.64 | 12.43 | |
| Task ArithmeticBackbone=Qwen2.5-7B2026.01 | 62.17 | 42.8 | 6.67 | 41.35 | 16.93 | |
| LinearBackbone=Qwen2.5-7B2026.01 | 50.42 | 43.8 | 0 | 23.14 | 10.38 | |
| FuseLLMBackbone=Qwen2.5-7B2026.01 | 1.81 | 0.2 | 0 | 1.23 | 4.43 | |
| TiesBackbone=Qwen2.5-7B2026.01 | 0.83 | 2.6 | 6.67 | 0 | 10.38 | |
| DAREBackbone=Qwen2.5-7B2026.01 | 0.53 | 1 | 0 | 0 | 3.38 |