General Speculative Decoding Performance on MT-Bench, HumanEval, and GSM8K Aggregate
6.52Average Acceptance Length (τ)PPOW
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PPOWModel=Q3-8B, Temperature=0.02026.05 | 6.52 | 3.39 | |
| PPOWModel=Q3-32B, Temperature=0.02026.05 | 6.44 | 3.92 | |
| PPOWModel=L31-8B, Temperature=0.02026.05 | 6.4 | 3.46 | |
| PPOWModel=L33-70B, Temperature=0.02026.05 | 6.29 | 4.36 | |
| EAGLE-3Model=Q3-8B, Temperature=0.02026.05 | 6.16 | 3.17 | |
| EAGLE-3Model=L31-8B, Temperature=0.02026.05 | 6.09 | 3.41 | |
| PPOWModel=L33-70B, Temperature=1.02026.05 | 6 | 4.21 | |
| EAGLE-3Model=Q3-32B, Temperature=0.02026.05 | 5.99 | 3.81 | |
| EAGLE-3Model=Qwen3.5-9B, m=15, Temperature (T)=0, Speculation Width (W)=42026.05 | 5.97 | 4.07 | |
| GRIFFINModel=L31-8B, Temperature=0.02026.05 | 5.95 | 3.31 | |
| EAGLE-3Model=L33-70B, Temperature=0.02026.05 | 5.94 | 4.15 | |
| GRIFFINModel=L33-70B, Temperature=0.02026.05 | 5.89 | 4.12 | |
| EAGLE-3Model=L33-70B, Temperature=1.02026.05 | 5.69 | 4 | |
| PPOWModel=Q3-32B, Temperature=1.02026.05 | 5.58 | 2.86 | |
| GRIFFINModel=L33-70B, Temperature=1.02026.05 | 5.52 | 3.9 | |
| PPOWModel=Q3-8B, Temperature=1.02026.05 | 5.48 | 2.72 | |
| PPOWModel=L31-8B, Temperature=1.02026.05 | 5.36 | 2.93 | |
| EAGLE-3Model=L31-8B, Temperature=1.02026.05 | 5.09 | 2.83 | |
| EAGLE-3Model=Q3-32B, Temperature=1.02026.05 | 5.09 | 2.67 | |
| EAGLE-3Model=Q3-8B, Temperature=1.02026.05 | 5.08 | 2.47 | |
| GRIFFINModel=L31-8B, Temperature=1.02026.05 | 5.06 | 2.81 | |
| EAGLE-3Model=Qwen3.5-9B, m=7, Temperature (T)=0, Speculation Width (W)=42026.05 | 5.05 | 4.14 | |
| EAGLE-3Model=Qwen3.5-4B, m=15, Temperature (T)=0, Speculation Width (W)=42026.05 | 4.9 | 3.33 | |
| EAGLE-3Model=Qwen3.5-4B, m=7, Temperature (T)=0, Speculation Width (W)=42026.05 | 4.4 | 3.61 | |
| EAGLE-3Model=Qwen3.5-9B, m=15, Temperature (T)=0, Speculation Width (W)=12026.05 | 4.23 | 2.88 | |
| EAGLE-3Model=Qwen3.5-9B, m=15, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.98 | 2.71 | |
| EAGLE-3Model=Qwen3.5-9B, m=7, Temperature (T)=0, Speculation Width (W)=12026.05 | 3.86 | 3.17 | |
| SPDModel=Qwen3.5-9B, n=16, Ls=2, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.83 | 3.83 | |
| EAGLE-3Model=Qwen3.5-9B, m=7, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.64 | 2.99 | |
| EAGLE-3Model=Qwen3.5-4B, m=15, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.58 | 2.44 | |
| EAGLE-3Model=Qwen3.5-4B, m=15, Temperature (T)=0, Speculation Width (W)=12026.05 | 3.51 | 2.39 | |
| EAGLE-3Model=Qwen3.5-9B, m=3, Temperature (T)=0, Speculation Width (W)=42026.05 | 3.46 | 3.16 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=2, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.44 | 3.44 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=4, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.43 | 3.43 | |
| EAGLE-3Model=Qwen3.5-4B, m=7, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.4 | 2.79 | |
| EAGLE-3Model=Qwen3.5-4B, m=7, Temperature (T)=0, Speculation Width (W)=12026.05 | 3.32 | 2.72 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=2, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.28 | 3.28 | |
| SPDModel=Qwen3.5-9B, n=16, Ls=2, Temperature (T)=0, Speculation Width (W)=42026.05 | 3.24 | 3.24 | |
| SPDModel=Qwen3.5-9B, n=16, Ls=2, Temperature (T)=0, Speculation Width (W)=12026.05 | 3.21 | 3.21 | |
| SPDModel=Qwen3.5-9B, n=8, Ls=4, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.15 | 3.15 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 3.13 | 3.13 | |
| EAGLE-3Model=Qwen3.5-4B, m=3, Temperature (T)=0, Speculation Width (W)=42026.05 | 3.01 | 2.75 | |
| EAGLE-3Model=Qwen3.5-9B, m=3, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.94 | 2.69 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=4, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.91 | 2.91 | |
| EAGLE-3Model=Qwen3.5-4B, m=3, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.88 | 2.63 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.88 | 2.88 | |
| EAGLE-3Model=Qwen3.5-9B, m=3, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.88 | 2.63 | |
| SPDModel=Qwen3.5-9B, n=16, Ls=2, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.87 | 2.87 | |
| EAGLE-3Model=Qwen3.5-9B, m=15, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.86 | 1.95 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=2, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.84 | 2.84 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=2, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.83 | 2.83 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=4, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.82 | 2.82 | |
| SPDModel=Qwen3.5-9B, n=8, Ls=4, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.82 | 2.82 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=2, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.81 | 2.81 | |
| EAGLE-3Model=Qwen3.5-9B, m=7, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.75 | 2.26 | |
| SPDModel=Qwen3.5-9B, n=8, Ls=4, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.74 | 2.74 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=2, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.72 | 2.72 | |
| EAGLE-3Model=Qwen3.5-4B, m=3, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.7 | 2.47 | |
| EAGLE-3Model=Qwen3.5-4B, m=15, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.65 | 1.8 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=4, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.65 | 2.65 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=1, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.64 | 2.64 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.62 | 2.62 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=4, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.56 | 2.56 | |
| SPDModel=Qwen3.5-9B, n=4, Ls=4, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.55 | 2.55 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=1, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.53 | 2.53 | |
| EAGLE-3Model=Qwen3.5-4B, m=7, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.51 | 2.06 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=2, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.49 | 2.49 | |
| SPDModel=Qwen3.5-9B, n=8, Ls=4, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.49 | 2.49 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=2, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.48 | 2.48 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=2, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.48 | 2.48 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.46 | 2.46 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=4, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.4 | 2.4 | |
| SPDModel=Qwen3.5-9B, n=4, Ls=4, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.38 | 2.38 | |
| PPSDModel=Qwen3.5-9B, n=8, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.35 | 1.88 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.33 | 2.33 | |
| SPDModel=Qwen3.5-4B, n=16, Ls=1, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.33 | 2.33 | |
| EAGLE-3Model=Qwen3.5-9B, m=3, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.33 | 2.13 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=4, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.29 | 2.29 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=2, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.28 | 2.28 | |
| SPDModel=Qwen3.5-9B, n=4, Ls=4, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.28 | 2.28 | |
| SPDModel=Qwen3.5-4B, n=8, Ls=1, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.24 | 2.24 | |
| EAGLE-3Model=Qwen3.5-4B, m=3, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.22 | 2.03 | |
| PPSDModel=Qwen3.5-9B, n=16, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.2 | 1.47 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=2, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.19 | 2.19 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=1, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.17 | 2.17 | |
| PPSDModel=Qwen3.5-9B, n=8, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.17 | 1.74 | |
| PPSDModel=Qwen3.5-9B, n=4, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.13 | 1.89 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=4, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.1 | 2.1 | |
| PPSDModel=Qwen3.5-9B, n=16, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.1 | 1.4 | |
| PPSDModel=Qwen3.5-4B, n=8, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.09 | 1.67 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 2.08 | 2.08 | |
| PPSDModel=Qwen3.5-9B, n=8, Ls=1, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.08 | 1.66 | |
| SPDModel=Qwen3.5-9B, n=4, Ls=4, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.08 | 2.08 | |
| PPSDModel=Qwen3.5-9B, n=4, Ls=1, Temperature (T)=0, Speculation Width (W)=42026.05 | 2.03 | 1.8 | |
| PPSDModel=Qwen3.5-4B, n=16, Ls=1, Temperature (T)=1, Speculation Width (W)=42026.05 | 2.02 | 1.35 | |
| SPDModel=Qwen3.5-4B, n=4, Ls=2, Temperature (T)=1, Speculation Width (W)=12026.05 | 2.02 | 2.02 | |
| PPSDModel=Qwen3.5-9B, n=4, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 1.98 | 1.76 | |
| PPSDModel=Qwen3.5-4B, n=8, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 1.93 | 1.54 | |
| PPSDModel=Qwen3.5-4B, n=16, Ls=1, Temperature (T)=0, Speculation Width (W)=12026.05 | 1.93 | 1.29 | |
| PPSDModel=Qwen3.5-9B, n=16, Ls=1, Temperature (T)=0, Speculation Width (W)=42026.05 | 1.92 | 1.28 |