Text Summarization on CNN/DM (MAT, Speedup)
4.23MATGraft-DFLASH(16)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Graft-DFLASH(16)Base Model=Qwen3-8B, Block Size=162026.05 | 4.23 | 2.68 | |
| DFLASH(16)Base Model=Qwen3-8B, Block Size=162026.05 | 3.83 | 2.23 | |
| LogitSpecBackbone=Vicuna 7B2025.07 | 3.28 | 2.61 | |
| Graft-EAGLE3(60)Base Model=Qwen3-8B, Budget=602026.05 | 2.91 | 2.17 | |
| LogitSpecBackbone=Vicuna 13B2025.07 | 2.9 | 2.17 | |
| EAGLE3(60)Base Model=Qwen3-8B, Budget=602026.05 | 2.83 | 1.71 | |
| LogitSpecModels=Llama 2 70B2025.07 | 2.67 | 2.1 | |
| LogitSpecBackbone=Vicuna 33B2025.07 | 2.66 | 2.01 | |
| PLDBackbone=Vicuna 7B2025.07 | 2.61 | 2.26 | |
| LogitSpecModels=Llama 2 13B2025.07 | 2.43 | 2.03 | |
| LogitSpecModels=Llama 2 7B2025.07 | 2.41 | 2.02 | |
| SpSBackbone=Vicuna 7B2025.07 | 2.34 | 1.58 | |
| PLDBackbone=Vicuna 13B2025.07 | 2.34 | 1.85 | |
| SpSBackbone=Vicuna 13B2025.07 | 2.18 | 1.48 | |
| PLDBackbone=Vicuna 33B2025.07 | 2.07 | 1.72 | |
| SpSModels=Llama 2 70B2025.07 | 2.01 | 1.71 | |
| SpSBackbone=Vicuna 33B2025.07 | 2.01 | 1.57 | |
| SpSModels=Llama 2 7B2025.07 | 1.99 | 1.46 | |
| PLDModels=Llama 2 70B2025.07 | 1.98 | 1.74 | |
| SpSModels=Llama 2 13B2025.07 | 1.95 | 1.34 | |
| PLDModels=Llama 2 7B2025.07 | 1.89 | 1.73 | |
| PLDModels=Llama 2 13B2025.07 | 1.89 | 1.52 | |
| RESTModels=Llama 2 7B2025.07 | 1.71 | 1.33 | |
| RESTModels=Llama 2 13B2025.07 | 1.71 | 1.34 | |
| RESTModels=Llama 2 70B2025.07 | 1.67 | 1.35 | |
| RESTBackbone=Vicuna 13B2025.07 | 1.65 | 1.22 | |
| RESTBackbone=Vicuna 33B2025.07 | 1.65 | 1.41 | |
| RESTBackbone=Vicuna 7B2025.07 | 1.64 | 1.19 | |
| LookaheadModels=Llama 2 7B2025.07 | 1.58 | 1.36 | |
| LookaheadModels=Llama 2 13B2025.07 | 1.56 | 1.18 | |
| LookaheadBackbone=Vicuna 7B2025.07 | 1.54 | 1.28 | |
| LookaheadModels=Llama 2 70B2025.07 | 1.53 | 1.28 | |
| LookaheadBackbone=Vicuna 33B2025.07 | 1.5 | 1.18 | |
| LookaheadBackbone=Vicuna 13B2025.07 | 1.46 | 1.12 |