Natural Language Reasoning on BoolQ, ARC-e, ARC-c, WinoGrande, HellaSwag
75.2BoolQ AccuracyMoEITS
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MoEITSBase Model=Qwen1.5-2.7B, Reduction in expert parameters (P ↓)=52.92%2026.04 | 75.2 | 72.3 | 36.01 | 67.46 | 61.27 | 62.45 | |
| MoE-I^2Base Model=Qwen1.5-2.7B, Reduction in expert parameters (P ↓)=53.98%2026.04 | 75.08 | 71.68 | 41.13 | 66.54 | 53.08 | 61.5 | |
| MoE-PrunerBase Model=Qwen1.5-2.7B, Reduction in expert parameters (P ↓)=50%2026.04 | 69.14 | 52.02 | 29.1 | 59.12 | 42.99 | 50.47 | |
| MoPBase Model=Qwen1.5-2.7B, Reduction in expert parameters (P ↓)=33%2026.04 | 68.56 | 59.76 | 44.97 | 52.57 | 56.4 | 56.45 |