Math Reasoning on MATH 500 (Accuracy, Avg, Delta)
97.6AccuracyPreMoE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PreMoEBase Model=DeepSeek-R1, Sparsity=50%2025.05 | 97.6 | 76.52 | 1.01 | |
| FullBase Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 97.4 | 79.25 | — | |
| EASY-EPBase Model=DeepSeek-R1, Sparsity=50%2025.05 | 97.2 | 74.69 | -0.82 | |
| Act-LogitsBase Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 97.2 | 76.91 | -2.34 | |
| FullBase Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 97.2 | 80.69 | — | |
| EASY-EPBase Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 96.82 | 76.45 | -2.8 | |
| PreMoEBase Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 96.8 | 78.45 | -0.8 | |
| FullBase Model=DeepSeek-R1, Sparsity=50%2025.05 | 96.6 | 75.51 | — | |
| FrequencyBase Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 96.4 | 69.67 | -9.58 | |
| PreMoEBase Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 96.4 | 79.92 | -0.77 | |
| SEER (G)Base Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 96 | 70.93 | -8.32 | |
| SEER (L)Base Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 95.2 | 71.29 | -7.96 | |
| FrequencyBase Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 90.4 | 54.26 | -32.76 | |
| FrequencyBase Model=DeepSeek-R1, Sparsity=50%2025.05 | 88.8 | 49.93 | -25.58 | |
| Prefix-RFTBase Model=Qwen2.5-Math-7B2025.07 | 88.4 | — | — | |
| Act-LogitsBase Model=DeepSeek-R1, Sparsity=50%2025.05 | 88.2 | 62.87 | -12.64 | |
| LUFFYBase Model=Qwen2.5-Math-7B2025.07 | 87.6 | — | — | |
| ReLIFTBase Model=Qwen2.5-Math-7B2025.07 | 87.4 | — | — | |
| SFT+RFTBase Model=Qwen2.5-Math-7B2025.07 | 87.2 | — | — | |
| RandomBase Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 85 | 50.13 | -29.12 | |
| RFTBase Model=Qwen2.5-Math-7B2025.07 | 84.4 | — | — | |
| UFTBase Model=Qwen2.5-Math-7B2025.07 | 83.8 | — | — | |
| SFTBase Model=Qwen2.5-Math-7B2025.07 | 82.6 | — | — | |
| RL w/ SFT LossBase Model=Qwen2.5-Math-7B2025.07 | 80.4 | — | — | |
| Oat-ZeroBase Model=Qwen2.5-Math-7B2025.07 | 78 | — | — | |
| SimpleRL-ZeroBase Model=Qwen2.5-Math-7B2025.07 | 76 | — | — | |
| EASY-EPBase Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 58.62 | 54.26 | -26.43 | |
| SEER (L)Base Model=DeepSeek-R1, Sparsity=50%2025.05 | 54.8 | 23.56 | -51.95 | |
| RandomBase Model=DeepSeek-R1, Sparsity=50%2025.05 | 54 | 26.69 | -48.82 | |
| SEER (G)Base Model=DeepSeek-R1, Sparsity=50%2025.05 | 53 | 24.31 | -51.2 | |
| RandomBase Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 44.8 | 13.07 | -67.62 | |
| Qwen2.5-Math-7B2025.07 | 43.6 | — | — | |
| All-LogitsBase Model=openPangu-Ultra, Sparsity=31.25%2025.05 | 12 | 5.39 | -73.86 | |
| All-LogitsBase Model=DeepSeek-R1, Sparsity=50%2025.05 | 3.6 | 5.46 | -70.05 | |
| All-LogitsBase Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 1.6 | 0.35 | -80.34 | |
| SEER (L)Base Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 1.6 | 0.64 | -80.05 | |
| Act-LogitsBase Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 1.4 | 0.82 | -79.87 | |
| SEER (G)Base Model=Qwen3-30B-A3B, Sparsity=50%2025.05 | 0.2 | 0.34 | -80.35 |