Reasoning on ARC Challenge (ACC, NLL, ECE, MCE)
88.2Accuracy (ACC)MAP
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| MAPMoE Architecture=Granite-MoE, Category=Base2026.03 | 88.2 | 0.923 | 20.1 | 42.8 | |
| MCDRMoE Architecture=Granite-MoE, Category=Weight2026.03 | 88 | 0.146 | 2.8 | 22.8 | |
| VGLR-FCMoE Architecture=Granite-MoE, Category=Ours2026.03 | 88 | 0.122 | 1.2 | 18.5 | |
| VGLR-MFMoE Architecture=Granite-MoE, Category=Ours2026.03 | 87.8 | 0.125 | 1.6 | 19.6 | |
| SWAGMoE Architecture=Granite-MoE, Category=Weight2026.03 | 87.2 | 0.138 | 3 | 26.6 | |
| VTSRMoE Architecture=Granite-MoE, Category=Ours2026.03 | 87.2 | 0.164 | 2 | 20.8 | |
| Temp-ScaleMoE Architecture=Granite-MoE, Category=Base2026.03 | 82.4 | 0.208 | 3.8 | 28.4 | |
| Temp-ScaleMoE Architecture=Qwen-MoE, Category=Base2026.03 | 76.2 | 1.422 | 10.2 | 30.4 | |
| VTSRMoE Architecture=Qwen-MoE, Category=Ours2026.03 | 76 | 0.924 | 3.8 | 28.6 | |
| MAPMoE Architecture=Qwen-MoE, Category=Base2026.03 | 75.8 | 1.852 | 25.2 | 62.4 | |
| VGLR-MFMoE Architecture=Qwen-MoE, Category=Ours2026.03 | 74.8 | 0.902 | 9.2 | 25.2 | |
| VGLR-FCMoE Architecture=Qwen-MoE, Category=Ours2026.03 | 74.8 | 0.892 | 2.2 | 20.4 | |
| SWAGMoE Architecture=Qwen-MoE, Category=Weight2026.03 | 74.6 | 1.122 | 20.2 | 38.7 | |
| MCDRMoE Architecture=Qwen-MoE, Category=Weight2026.03 | 73.2 | 1.134 | 19.8 | 44.6 | |
| BayesLoRAk=10, Params (M)=2.40–3.872025.06 | 69.46 | 1.43 | 17.49 | — | |
| LoRAParams (M)=4.482025.06 | 69.1 | 2.85 | 29 | — | |
| BayesLoRAr=8, k=10, Params (M)=4.482025.06 | 68.9 | 1.4 | 17.47 | — | |
| MCDParams (M)=4.482025.06 | 68.69 | 2.67 | 27.53 | — | |
| BayesLoRAr=8, k=0, Params (M)=4.482025.06 | 68.68 | 2.93 | 30.45 | — | |
| BayesLoRAk=0, Params (M)=2.40–3.872025.06 | 68.37 | 2.99 | 30.73 | — | |
| BBBParams (M)=9.982025.06 | 68.13 | 2.23 | 26.23 | — | |
| ENSParams (M)=44.802025.06 | 66.2 | 2.46 | 29.16 | — | |
| Temp-ScaleMoE Architecture=DeepSeek-MoE, Category=Base2026.03 | 59.2 | 1.242 | 18.8 | 64.6 | |
| VTSRMoE Architecture=DeepSeek-MoE, Category=Ours2026.03 | 58.4 | 0.672 | 17.8 | 54.2 | |
| MAPMoE Architecture=DeepSeek-MoE, Category=Base2026.03 | 58 | 1.462 | 23.7 | 82.6 | |
| VGLR-FCMoE Architecture=DeepSeek-MoE, Category=Ours2026.03 | 57.2 | 0.422 | 14.6 | 56.3 | |
| SWAGMoE Architecture=DeepSeek-MoE, Category=Weight2026.03 | 56.8 | 1.183 | 19.8 | 84.2 | |
| MCDRMoE Architecture=DeepSeek-MoE, Category=Weight2026.03 | 56.4 | 1.084 | 19.2 | 84.8 | |
| VGLR-MFMoE Architecture=DeepSeek-MoE, Category=Ours2026.03 | 56.2 | 0.724 | 15.8 | 56.2 |