Common Sense Reasoning on ARC-E, ARC-C, BoolQ, WinoGrande, HellaSwag, MMLU, PIQA
78.9ARC-E AccuracyPuzzleMoE
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| PuzzleMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=25%2025.11 | 78.9 | 51.6 | 88.2 | 70.4 | 58.3 | 76.6 | 79.3 | — | 71.9 | |
| PuzzleMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=37.5%2025.11 | 78.5 | 51.4 | 87.2 | 70.5 | 57.4 | 75.8 | 78.9 | — | 71.4 | |
| PuzzleMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=50%2025.11 | 78.5 | 51 | 88 | 70.1 | 57.1 | 75.1 | 78.9 | — | 71.2 | |
| PuzzleMoEBase Model=Deepseek-MoE-16b, Sparsity=25%2025.11 | 75.7 | 44 | 73.1 | 70.6 | 57.2 | 37.2 | 78.7 | — | 62.4 | |
| PuzzleMoEBase Model=Deepseek-MoE-16b, Sparsity=37.5%2025.11 | 75.3 | 44.2 | 73.3 | 69.9 | 56.2 | 36.9 | 78.7 | — | 62.1 | |
| PuzzleMoEBase Model=Deepseek-MoE-16b, Sparsity=50%2025.11 | 75.2 | 43 | 74.5 | 70.3 | 56.3 | 36.9 | 78.4 | — | 62.1 | |
| PuzzleMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=66.7%2025.11 | 74.4 | 43.7 | 86.9 | 69.9 | 53.9 | 63.8 | 76.2 | — | 67 | |
| HC-SMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=25%2025.11 | 70.8 | 40.1 | 82.4 | 63.2 | 48.1 | 60.8 | 71.2 | — | 62.4 | |
| PuzzleMoEBase Model=Deepseek-MoE-16b, Sparsity=66.7%2025.11 | 70.1 | 36.9 | 73.1 | 70.1 | 49.8 | 29.6 | 75.6 | — | 57.9 | |
| HC-SMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=37.5%2025.11 | 67.8 | 37.6 | 81.2 | 60 | 46.2 | 53.7 | 71 | — | 59.6 | |
| PuzzleMoEBase Model=Deepseek-MoE-16b, Sparsity=75%2025.11 | 65.3 | 33.4 | 62.2 | 66.4 | 54.1 | 24.9 | 72.2 | — | 54.1 | |
| HC-SMoEBase Model=Deepseek-MoE-16b, Sparsity=25%2025.11 | 65.1 | 36.7 | 66.4 | 65.7 | 44.3 | 24.5 | 73.1 | — | 53.7 | |
| PuzzleMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=75%2025.11 | 63.2 | 33.7 | 80.1 | 69.1 | 46.8 | 48.1 | 72.5 | — | 59.1 | |
| subnet-selectionParams=6.06B, Latency (ms)=87.38, Model Size Range=5-6B2026.06 | 61.41 | 37.12 | 70.43 | 65.59 | 59.96 | 30.98 | 70.46 | 56.56 | — | |
| HC-SMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=50%2025.11 | 60.9 | 34.3 | 80.8 | 54 | 44.6 | 39.6 | 69.8 | — | 54.9 | |
| ours-no-quantParams=6.06B, Latency (ms)=87.38, Model Size Range=5-6B2026.06 | 59.18 | 42.32 | 72.54 | 65.51 | 60.77 | 49.07 | 70.82 | 60.03 | — | |
| HC-SMoEBase Model=Deepseek-MoE-16b, Sparsity=37.5%2025.11 | 58.9 | 31.2 | 65.1 | 63.4 | 41.3 | 24.7 | 72 | — | 50.9 | |
| LoNAS*Params=6.84B, Latency (ms)=99.00, Model Size Range=5-6B2026.06 | 42.59 | 31.83 | 57.77 | 63.12 | 51 | 24.26 | 64.42 | 47.86 | — | |
| HC-SMoEBase Model=Deepseek-MoE-16b, Sparsity=50%2025.11 | 41.9 | 22.3 | 62.3 | 55.3 | 31.2 | 23 | 62.3 | — | 42.6 | |
| ours-no-quantParams=4.14B, Latency (ms)=53.93, Model Size Range=4-5B2026.06 | 41.33 | 29.79 | 62.69 | 56.27 | 41.98 | 26.9 | 62.68 | 45.93 | — | |
| HC-SMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=66.7%2025.11 | 40.2 | 28.9 | 78.6 | 53 | 35.7 | 33.4 | 60.1 | — | 47.1 | |
| HC-SMoEBase Model=Deepseek-MoE-16b, Sparsity=66.7%2025.11 | 38.6 | 20.5 | 62 | 54.3 | 26.7 | 24.5 | 60.5 | — | 41 | |
| ours-no-quantParams=3.37B, Latency (ms)=40.90, Model Size Range=3-4B2026.06 | 38.13 | 23.21 | 60.55 | 52.72 | 29.21 | 25.86 | 56.53 | 40.89 | — | |
| ours-no-quantParams=2.27B, Latency (ms)=43.14, Model Size Range=2-3B2026.06 | 34.18 | 23.98 | 60.83 | 50.67 | 27.43 | 25.38 | 55.28 | 39.68 | — | |
| subnet-selectionParams=4.32B, Latency (ms)=57.45, Model Size Range=4-5B2026.06 | 31.99 | 26.88 | 37.86 | 49.49 | 31.59 | 25.07 | 57.73 | 37.22 | — | |
| HC-SMoEBase Model=Qwen3-MoE-30B-A3B, Sparsity=75%2025.11 | 31.2 | 25.6 | 76.4 | 52.5 | 28.8 | 27.2 | 54.3 | — | 42.3 | |
| subnet-selectionParams=3.28B, Latency (ms)=81.31, Model Size Range=3-4B2026.06 | 28.2 | 25.77 | 48.78 | 37.95 | 27.15 | 24.49 | 54.03 | 35.19 | — | |
| HC-SMoEBase Model=Deepseek-MoE-16b, Sparsity=75%2025.11 | 27.1 | 19.9 | 59.8 | 50.8 | 26.3 | 23.6 | 54.3 | — | 35.7 | |
| subnet-selectionParams=2.54B, Latency (ms)=70.95, Model Size Range=2-3B2026.06 | 25.66 | 26.62 | 49.88 | 38.5 | 26.38 | 25.72 | 53.76 | 35.22 | — | |
| LoNAS*Params=4.00B, Latency (ms)=73.72, Model Size Range=4-5B2026.06 | 25.25 | 26.11 | 46.96 | 56.67 | 26.82 | 23.74 | 52.84 | 36.91 | — | |
| LoNASParams=3.28B, Latency (ms)=81.31, Model Size Range=3-4B2026.06 | 24.49 | 27.47 | 49.72 | 38.69 | 26.58 | 24.65 | 51.41 | 34.72 | — | |
| LoNAS*Params=2.67B, Latency (ms)=69.84, Model Size Range=2-3B2026.06 | 24.45 | 25.77 | 48.86 | 61.87 | 26.05 | 22.95 | 52.23 | 37.45 | — |