Long-context Generation (Reasoning) on AIME25
49.96TPTKnapSpec
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| KnapSpecBackbone=Qwen3-4B, Layers (L)=362026.02 | 49.96 | 1.35 | 89.3 | |
| DELBackbone=Qwen3-4B, Layers (L)=362026.02 | 40.6 | 0.97 | 22.3 | |
| ARBackbone=Qwen3-4B, Layers (L)=362026.02 | 36.85 | 1 | — | |
| KnapSpecBackbone=Qwen3-8B, Layers (L)=362026.02 | 36.81 | 1.26 | 88.5 | |
| CLaSpBackbone=Qwen3-4B, Layers (L)=362026.02 | 32.26 | 1.16 | 91.2 | |
| KnapSpecBackbone=Qwen3-32B, Layers (L)=642026.02 | 32 | 1.42 | 93.1 | |
| KnapSpecBackbone=Qwen3-14B, Layers (L)=402026.02 | 31.06 | 1.27 | 88.9 | |
| SWIFTBackbone=Qwen3-4B, Layers (L)=362026.02 | 30.23 | 1.11 | 89 | |
| DELBackbone=Qwen3-8B, Layers (L)=362026.02 | 28.71 | 0.83 | 0 | |
| ARBackbone=Qwen3-8B, Layers (L)=362026.02 | 28.66 | 1 | — | |
| CLaSpBackbone=Qwen3-8B, Layers (L)=362026.02 | 23.94 | 1.15 | 90.5 | |
| DELBackbone=Qwen3-14B, Layers (L)=402026.02 | 23.25 | 0.83 | 0 | |
| ARBackbone=Qwen3-14B, Layers (L)=402026.02 | 23.17 | 1 | — | |
| SWIFTBackbone=Qwen3-14B, Layers (L)=402026.02 | 21.86 | 1.2 | 85.4 | |
| SWIFTBackbone=Qwen3-8B, Layers (L)=362026.02 | 21.65 | 0.9 | 71 | |
| CLaSpBackbone=Qwen3-14B, Layers (L)=402026.02 | 21.42 | 1.2 | 92.5 | |
| CLaSpBackbone=Qwen3-32B, Layers (L)=642026.02 | 21.3 | 1.29 | 95.1 | |
| SWIFTBackbone=Qwen3-32B, Layers (L)=642026.02 | 20.73 | 1.22 | 80.4 | |
| DELBackbone=Qwen3-32B, Layers (L)=642026.02 | 19.95 | 0.88 | 2.8 | |
| ARBackbone=Qwen3-32B, Layers (L)=642026.02 | 19.72 | 1 | — |