Arithmetic Reasoning on GSM8K (4-shot CoT)
90.25AccuracyADAFUSE (Top-2 Base)
Evaluation Results
| Method | Links | |
|---|---|---|
| ADAFUSE (Top-2 Base)Base model selection strategy=Oracle top-2 base selection2026.01 | 90.25 | |
| Qwen3-8BModel Type=Base Model2026.01 | 89.99 | |
| LLaMA-3.1-8B-InstructModel Type=Base Model2026.01 | 81.05 | |
| ADAFUSE (Fixed Base)Base model selection strategy=Fixed base pair (LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3)2026.01 | 79.15 | |
| InternLM3-8B-InstructModel Type=Base Model2026.01 | 76.72 | |
| DEEPENBase model selection strategy=Fixed base pair (LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3)2026.01 | 67.63 | |
| UniTEBase model selection strategy=Fixed base pair (LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3)2026.01 | 64.59 | |
| SWEETSPANBase model selection strategy=Fixed base pair (LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3)2026.01 | 63.38 | |
| LLM-BLENDERBase model selection strategy=Fixed base pair (LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3)2026.01 | 62.55 | |
| Mistral-7B-Instruct-v0.3Model Type=Base Model2026.01 | 61.71 |