Reasoning on Reasoning Suite (GSM8K, AIME, HumanEval, LiveCodeBench)
85.12GSM8KReasoning
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ReasoningMerging Method=Expert Model (DeepSeek-R1-Distill-Llama-8B)2026.01 | 85.12 | 84.4 | 33.33 | 76.69 | 29.91 | |
| ReasonAnyMerging Method=ReasonAny (ours)2026.01 | 77.77 | 52.4 | 16.67 | 59.51 | 25.51 | |
| LinearMerging Method=Linear Merging2026.01 | 75.28 | 37.4 | 6.67 | 32.52 | 2.81 | |
| Task ArithmeticMerging Method=Task Arithmetic2026.01 | 63.23 | 22.8 | 0 | 21.47 | 2.81 | |
| BiomedicineMerging Method=Expert Model (MMed-Llama-8B)2026.01 | 57.54 | 3 | 0 | 54.6 | 2.24 | |
| LEDMerging Method=LED Merging2026.01 | 56.33 | 6.4 | 0 | 60.12 | 24.33 | |
| TIESMerging Method=TIES Merging2026.01 | 49.73 | 15 | 0 | 39.26 | 11.34 | |
| FuseLLMMerging Method=FuseLLM2026.01 | 0.83 | 2.2 | 0 | 38.65 | 5.01 | |
| DAREMerging Method=DARE Merging2026.01 | 0.53 | 30.6 | 3.33 | 49.69 | 22.55 |