Comprehensive Reasoning and Knowledge Benchmark (GSM8K, MMLU, HumanEval, ARC, GPQA)
95.41GSM8KReasoning
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| ReasoningModel Family=Qwen2.5-32B, Role=Source Model (QwQ-32B-Preview)2026.01 | 95.41 | 84.4 | 53.33 | 89.63 | 57.25 | 95.25 | 99.32 | 79.84 | 53.3 | 78.64 | |
| ReasonAnyMerging Method=ReasonAny2026.01 | 91.13 | 81 | 36.67 | 89.71 | 53.39 | 96.88 | 98.94 | 82 | 55.33 | 76.12 | |
| Task ArithmeticMerging Method=Task Arithmetic2026.01 | 89.84 | 73.6 | 20 | 83.54 | 25.02 | 95.59 | 98.77 | 81.76 | 36.36 | 67.16 | |
| DAREMerging Method=DARE2026.01 | 86.2 | 81 | 30 | 81.32 | 27.13 | 96.27 | 98.59 | 81.85 | 29.94 | 68.03 | |
| SafetyModel Family=Qwen2.5-32B, Role=Source Model (Safety-Instruct)2026.01 | 83.02 | 82.2 | 23.33 | 84.31 | 23.39 | 93.48 | 98.24 | 81.78 | 41.67 | 67.94 | |
| FuseLLMMerging Method=FuseLLM2026.01 | 81.73 | 79.8 | 26.67 | 71.34 | 39.85 | 95.25 | 98.59 | 80.8 | 29.94 | 67.11 | |
| LEDMerging Method=LED2026.01 | 69.75 | 78.4 | 20 | 84.31 | 48.6 | 95.59 | 98.41 | 80.83 | 52.28 | 69.8 | |
| TIESMerging Method=TIES2026.01 | 64.67 | 71.4 | 20 | 88.41 | 54.23 | 94.34 | 96.88 | 80.65 | 40.91 | 67.94 | |
| LinearMerging Method=Linear2026.01 | 28.28 | 15.2 | 10 | 89.02 | 24.74 | 95.93 | 98.94 | 81.85 | 45.18 | 54.35 |