Speculative Decoding on LiveCodeBench
7.16Speedup FactorTAPS
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| TAPSHardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-4B2026.05 | 7.16 | — | — | 8.4 | |
| TAPSHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-8B2026.05 | 6.74 | — | — | 8.08 | |
| TAPSHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-4B2026.05 | 6.67 | — | — | 7.93 | |
| TAPSHardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-8B2026.05 | 6.56 | — | — | 8.76 | |
| DFlashHardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-8B2026.05 | 5.51 | — | — | 6.74 | |
| DDTreeHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-4B2026.05 | 5.42 | — | — | 8.87 | |
| DFlashHardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-4B2026.05 | 5.41 | — | — | 6.38 | |
| DFlashHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-8B2026.05 | 4.93 | — | — | 6.69 | |
| DFlashHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-4B2026.05 | 4.9 | — | — | 6.5 | |
| TAPSHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 4.06 | — | — | 5.36 | |
| DDTreeHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-8B2026.05 | 3.99 | — | — | 9.15 | |
| DDTreeHardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-4B2026.05 | 3.59 | — | — | 9.46 | |
| TAPSHardware=NVIDIA A40-48GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 3.56 | — | — | 5.27 | |
| DDTreeHardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-8B2026.05 | 2.83 | — | — | 9.63 | |
| DFlashHardware=NVIDIA A40-48GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 2.77 | — | — | 3.84 | |
| DFlashHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 2.77 | — | — | 3.91 | |
| STANDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | 2.67 | 74.14 | 3.64 | — | |
| STANDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | 2.59 | 71.83 | 3.47 | — | |
| STANDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | 2.51 | 69.7 | 3.3 | — | |
| STANDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | 2.45 | 42.72 | 3.65 | — | |
| STANDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | 2.32 | 40.45 | 3.47 | — | |
| SAM+RecycleBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | 2.29 | 63.41 | 2.69 | — | |
| SAM+RecycleBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | 2.28 | 63.39 | 2.67 | — | |
| DDTreeHardware=NVIDIA A800-SXM4-80GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 2.27 | — | — | 6 | |
| SAM+RecycleBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | 2.24 | 62.2 | 2.68 | — | |
| STANDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | 2.22 | 38.71 | 3.29 | — | |
| RecycleBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | 2.21 | 61.36 | 2.74 | — | |
| RecycleBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | 2.2 | 60.93 | 2.73 | — | |
| RecycleBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | 2.18 | 60.62 | 2.73 | — | |
| SAM+RecycleBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | 1.97 | 34.31 | 2.65 | — | |
| SAM+RecycleBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | 1.96 | 34.15 | 2.65 | — | |
| SAM+RecycleBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | 1.96 | 34.19 | 2.66 | — | |
| RecycleBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | 1.95 | 34.05 | 2.72 | — | |
| RecycleBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | 1.95 | 33.98 | 2.72 | — | |
| RecycleBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | 1.95 | 33.96 | 2.72 | — | |
| EAGLE-3Hardware=NVIDIA A40-48GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 1.91 | — | — | 2.65 | |
| ANPDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | 1.76 | 48.81 | 1.96 | — | |
| SAMBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | 1.76 | 48.92 | 1.95 | — | |
| EAGLE-3Hardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-4B2026.05 | 1.73 | — | — | 2.91 | |
| ANPDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | 1.72 | 47.86 | 1.91 | — | |
| SAMBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | 1.7 | 47.24 | 1.89 | — | |
| ANPDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | 1.67 | 46.3 | 1.88 | — | |
| PLDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | 1.66 | 46.02 | 1.85 | — | |
| SAMBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | 1.64 | 45.63 | 1.85 | — | |
| EAGLE-3Hardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-4B2026.05 | 1.63 | — | — | 2.77 | |
| PLDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | 1.62 | 45.08 | 1.79 | — | |
| PLDBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | 1.59 | 44.01 | 1.73 | — | |
| EAGLE-3Hardware=NVIDIA A40-48GB GPU, Target Model=Qwen3-8B2026.05 | 1.57 | — | — | 2.65 | |
| EAGLE-3Hardware=NVIDIA A800-SXM4-80GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 1.57 | — | — | 2.22 | |
| EAGLE-3Hardware=NVIDIA A800-SXM4-80GB GPU, Target Model=Qwen3-8B2026.05 | 1.56 | — | — | 2.57 | |
| SAMBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | 1.53 | 26.67 | 1.96 | — | |
| DDTreeHardware=NVIDIA A40-48GB GPU, Target Model=LLaMA-3.1-8B-Instruct2026.05 | 1.53 | — | — | 6.02 | |
| ANPDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | 1.51 | 26.32 | 1.99 | — | |
| ANPDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | 1.47 | 25.63 | 1.93 | — | |
| SAMBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | 1.46 | 25.37 | 1.87 | — | |
| PLDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | 1.43 | 24.97 | 1.88 | — | |
| ANPDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | 1.42 | 24.78 | 1.87 | — | |
| SAMBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | 1.4 | 24.41 | 1.79 | — | |
| PLDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | 1.37 | 23.96 | 1.81 | — | |
| PLDBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | 1.34 | 23.36 | 1.74 | — | |
| Eagle-2Base Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | 1.26 | 21.89 | 2.51 | — | |
| Eagle-2Base Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | 1.26 | 21.89 | 2.51 | — | |
| Eagle-2Base Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | 1.26 | 21.89 | 2.51 | — | |
| Eagle-2Base Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | 0.99 | 27.61 | 2.13 | — | |
| Eagle-2Base Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | 0.99 | 27.61 | 2.13 | — | |
| Eagle-2Base Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | 0.99 | 27.61 | 2.13 | — | |
| PlainBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=42025.06 | — | 27.75 | — | — | |
| PlainBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=82025.06 | — | 27.75 | — | — | |
| PlainBase Model=DeepSeek-R1-Distill-Qwen-7B, Trajectories=162025.06 | — | 27.75 | — | — | |
| PlainBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=42025.06 | — | 17.43 | — | — | |
| PlainBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=82025.06 | — | 17.43 | — | — | |
| PlainBase Model=DeepSeek-R1-Distill-Qwen-14B, Trajectories=162025.06 | — | 17.43 | — | — |