Zero-shot Commonsense Reasoning Suite (PIQA, SIQA, HellaSwag, ARC-E, ARC-C, WinoGrande, LAMBADA)
66.92PIQA AccuracyReMoE
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| ReMoEModel Scale=Large, Routing Strategy=NG2026.05 | 66.92 | 38.79 | 32.66 | 55.93 | 24.57 | 51.41 | 27.1 | 42.48 | |
| DenseModel Scale=Large2026.05 | 66.87 | 37.77 | 32.92 | 56.27 | 24.74 | 51.7 | 29.34 | 42.8 | |
| DeepSeek-V3Model Scale=Large, Routing Strategy=GA2026.05 | 66.32 | 38.38 | 32.71 | 55.18 | 24.91 | 51.07 | 28.06 | 42.38 | |
| DECOModel Scale=Large, Routing Strategy=GA2026.05 | 66.14 | 38.74 | 32.69 | 56.41 | 23.72 | 51.78 | 26.57 | 42.29 | |
| BlockFFNModel Scale=Large2026.05 | 66.1 | 38.59 | 32.51 | 56.36 | 24.15 | 51.38 | 27.3 | 42.34 | |
| ReMoEModel Scale=Large, Routing Strategy=GA2026.05 | 66.05 | 38.69 | 32.08 | 55.26 | 24.32 | 52.33 | 27.19 | 42.27 | |
| DeepSeek-V3Model Scale=Large, Routing Strategy=NG2026.05 | 65.94 | 37.87 | 32.75 | 54.67 | 23.38 | 52.49 | 27.03 | 42.02 | |
| DECOModel Scale=Large, Routing Strategy=NG2026.05 | 65.94 | 39.61 | 32.78 | 57.11 | 25.43 | 53.28 | 25.5 | 42.81 | |
| TopPModel Scale=Large2026.05 | 64.53 | 36.44 | 31.42 | 50.55 | 22.53 | 51.07 | 14.13 | 38.67 | |
| DECOModel Scale=Small, MoE Normalization Variant=NG2026.05 | 60.94 | 36.85 | 27.96 | 45.84 | 20.22 | 50.67 | 15.8 | 36.9 | |
| BlockFFNModel Scale=Small2026.05 | 60.83 | 36.28 | 27.62 | 44.87 | 19.45 | 50.12 | 16.55 | 36.53 | |
| DenseModel Scale=Small2026.05 | 60.77 | 35.57 | 28.03 | 45.88 | 20.22 | 50.59 | 16.51 | 36.8 | |
| DeepSeek-V3Model Scale=Small, MoE Normalization Variant=NG2026.05 | 60.66 | 36.63 | 27.76 | 45.76 | 20.05 | 50.43 | 15.85 | 36.73 | |
| ReMoEModel Scale=Small, MoE Normalization Variant=GA2026.05 | 60.34 | 36.59 | 27.72 | 46.34 | 18.43 | 51.22 | 15.6 | 36.61 | |
| TopPModel Scale=Small2026.05 | 60.17 | 35.21 | 27.67 | 42.38 | 18.86 | 50.91 | 7.74 | 34.71 | |
| DeepSeek-V3Model Scale=Small, MoE Normalization Variant=GA2026.05 | 60.12 | 35.93 | 27.85 | 44.15 | 19.45 | 50.75 | 15.89 | 36.31 | |
| ReMoEModel Scale=Small, MoE Normalization Variant=NG2026.05 | 59.74 | 35.16 | 27.96 | 46.09 | 19.54 | 49.96 | 17.76 | 36.6 | |
| DECOModel Scale=Small, MoE Normalization Variant=GA2026.05 | 58.22 | 35.93 | 27.65 | 42.51 | 19.03 | 51.3 | 15.84 | 35.78 |