Mathematical Reasoning on (AIME25, AIME26, Olympiad, MATH-500, GSM8K, GPQA-D)
80AIME25 AccuracyMoE comp. (Fisher)
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| MoE comp. (Fisher)Backbone=Qwen3.5-35B-A3B, Compression Ratio (p)=50%2026.06 | 80 | 76.7 | 79.8 | 91.8 | 42.2 | 71.7 | 73.7 | |
| Expert-level FisherBackbone=Qwen3.5-35B-A3B, Compression Ratio (p)=50%2026.06 | 80 | 76.7 | 81.9 | 92.4 | 50.7 | 69.7 | 75.2 | |
| Fisher-IntDim-GBackbone=Qwen3.5-35B-A3B, Compression Ratio (p)=50%2026.06 | 76.7 | 73.3 | 79.5 | 92.4 | 62.6 | 70.2 | 75.8 | |
| Base ModelBackbone=Qwen3.5-35B-A3B, Compression Ratio (p)=0%2026.06 | 66.7 | 66.7 | 77.9 | 92.4 | 48.8 | 72.2 | 70.8 | |
| Fisher-IntDim-GBackbone=Qwen3-30B-A3B, Compression Ratio (p)=50%2026.06 | 60 | 56.7 | 58.6 | 74.6 | 81.4 | 50 | 63.5 | |
| MoE comp. (Fisher)Backbone=Qwen3-30B-A3B, Compression Ratio (p)=50%2026.06 | 46.7 | 50 | 55 | 76.4 | 92.3 | 38.9 | 59.9 | |
| Expert-level FisherBackbone=Qwen3-30B-A3B, Compression Ratio (p)=50%2026.06 | 36.7 | 53.3 | 54.9 | 76.2 | 93.4 | 49 | 60.6 | |
| Base ModelBackbone=Qwen3-30B-A3B, Compression Ratio (p)=0%2026.06 | 33.3 | 50 | 49.6 | 64.2 | 61.9 | 48.5 | 51.3 | |
| MoE-PrunerBackbone=Qwen3.5-35B-A3B, Compression Ratio (p)=50%2026.06 | 13.3 | 10 | 22.1 | 34.8 | 17.4 | 19.7 | 19.6 | |
| MoE-PrunerBackbone=Qwen3-30B-A3B, Compression Ratio (p)=50%2026.06 | 0 | 0 | 0.3 | 2.4 | 0.8 | 5.1 | 1.4 | |
| MoE comp.Backbone=Qwen3-30B-A3B, Compression Ratio (p)=50%2026.06 | 0 | 0 | 0.9 | 1.8 | 2.3 | 10.6 | 2.6 | |
| MoE comp. (activation)Backbone=Qwen3-30B-A3B, Compression Ratio (p)=50%2026.06 | 0 | 0 | 1.2 | 1.6 | 2.8 | 2.5 | 1.4 | |
| MoE comp.Backbone=Qwen3.5-35B-A3B, Compression Ratio (p)=50%2026.06 | 0 | 0 | 0 | 0.2 | 2.1 | 12.6 | 2.5 | |
| MoE comp. (activation)Backbone=Qwen3.5-35B-A3B, Compression Ratio (p)=50%2026.06 | 0 | 0 | 0.2 | 0.2 | 2.3 | 9.6 | 2 |