Long-context Generation (Reasoning) on MMLU-Pro
13.67TPTSWIFT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SWIFTBackbone=Qwen3-14B, Layers (L)=402026.02 | 13.67 | 0.74 | 53.6 | |
| CLaSpBackbone=Qwen3-14B, Layers (L)=402026.02 | 15.82 | 1.08 | 87 | |
| SWIFTBackbone=Qwen3-8B, Layers (L)=362026.02 | 20.32 | 1.12 | 93 | |
| SWIFTBackbone=Qwen3-32B, Layers (L)=642026.02 | 21.75 | 0.9 | 46.8 | |
| CLaSpBackbone=Qwen3-8B, Layers (L)=362026.02 | 22.75 | 1.16 | 90.8 | |
| ARBackbone=Qwen3-32B, Layers (L)=642026.02 | 23.15 | 1 | — | |
| DELBackbone=Qwen3-32B, Layers (L)=642026.02 | 23.2 | 0.87 | 0 | |
| SWIFTBackbone=Qwen3-4B, Layers (L)=362026.02 | 23.74 | 1 | 7.5 | |
| CLaSpBackbone=Qwen3-32B, Layers (L)=642026.02 | 23.9 | 1.09 | 92 | |
| ARBackbone=Qwen3-14B, Layers (L)=402026.02 | 24.06 | 1 | — | |
| DELBackbone=Qwen3-14B, Layers (L)=402026.02 | 24.61 | 0.85 | 17.2 | |
| ARBackbone=Qwen3-8B, Layers (L)=362026.02 | 24.97 | 1 | — | |
| CLaSpBackbone=Qwen3-4B, Layers (L)=362026.02 | 25.68 | 1.15 | 89.9 | |
| DELBackbone=Qwen3-8B, Layers (L)=362026.02 | 27.91 | 0.88 | 23.9 | |
| ARBackbone=Qwen3-4B, Layers (L)=362026.02 | 30.93 | 1 | — | |
| KnapSpecBackbone=Qwen3-32B, Layers (L)=642026.02 | 34.62 | 1.2 | 85.1 | |
| KnapSpecBackbone=Qwen3-14B, Layers (L)=402026.02 | 35.51 | 1.16 | 88 | |
| DELBackbone=Qwen3-4B, Layers (L)=362026.02 | 38.36 | 0.99 | 66.1 | |
| KnapSpecBackbone=Qwen3-8B, Layers (L)=362026.02 | 40.01 | 1.33 | 91.2 | |
| KnapSpecBackbone=Qwen3-4B, Layers (L)=362026.02 | 45.25 | 1.29 | 90.2 |