Multi-turn Dialogue on MT-Bench (MAT and Speedup)
6.93MAT ScoreGraft
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GraftModel=Vicuna-13B2026.05 | 6.93 | 4.13 | |
| EAGLE3-2KModel=Vicuna-13B2026.05 | 6.81 | 3.85 | |
| GraftModel=LLaMA3.1-8B2026.05 | 6.49 | 3.88 | |
| EAGLE3-2KModel=LLaMA3.1-8B2026.05 | 6.42 | 3.61 | |
| ECHOModel=Vicuna-13B2026.05 | 6.35 | 4.05 | |
| ECHOModel=LLaMA3.1-8B2026.05 | 6.23 | 3.77 | |
| DDDModel=Vicuna-13B2026.05 | 6.07 | 3.99 | |
| Graft-DFLASH(16)Base Model=Qwen3-8B, Block Size=162026.05 | 4.54 | 2.93 | |
| DFLASH(16)Base Model=Qwen3-8B, Block Size=162026.05 | 4.39 | 2.69 | |
| GraftModel=Qwen3-8B2026.05 | 3.79 | 2.36 | |
| EAGLE3Model=Qwen3-8B2026.05 | 3.71 | 2.03 | |
| ECHOModel=Qwen3-8B2026.05 | 3.68 | 2.18 | |
| Graft-EAGLE3(60)Base Model=Qwen3-8B, Budget=602026.05 | 3.41 | 2.18 | |
| EAGLE3(60)Base Model=Qwen3-8B, Budget=602026.05 | 3.34 | 1.89 | |
| GraftModel=Qwen3-32B2026.05 | 3.15 | 2.46 | |
| EAGLE3Model=Qwen3-32B2026.05 | 3.02 | 2.03 | |
| ECHOModel=Qwen3-32B2026.05 | 2.96 | 2.27 | |
| MedusaModel=Vicuna-13B2026.05 | 2.61 | 2.13 | |
| GraftModel=Qwen3-235B2026.05 | 2.61 | 2.18 | |
| EAGLE3-2KModel=Qwen3-235B2026.05 | 2.56 | 1.81 | |
| ECHOModel=Qwen3-235B2026.05 | 2.38 | 2.02 | |
| TRModel=Vicuna-13B2026.05 | 2.37 | 1.82 | |
| TRModel=Qwen3-32B2026.05 | 2.37 | 1.83 | |
| TRModel=Qwen3-8B2026.05 | 2.36 | 1.86 | |
| SpsModel=Vicuna-13B2026.05 | 2.26 | 1.79 | |
| SAMDModel=Vicuna-13B2026.05 | 2.15 | 1.74 | |
| TRModel=Qwen3-235B2026.05 | 2.05 | 1.72 | |
| TRModel=LLaMA3.1-8B2026.05 | 1.87 | 1.76 | |
| LookaheadModel=Vicuna-13B2026.05 | 1.65 | 1.61 | |
| PLDModel=Vicuna-13B2026.05 | 1.55 | 1.6 |