Long-context language generation on GovReport
3.36Speedup (×)Hybrid Verified Decoding
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=SuffixDecoding2026.05 | 3.36 | — | — | |
| LongSpecBackbone=LC-13B, Temperature=02025.02 | 2.67 | 3.58 | 76.26 | |
| LongSpecBackbone=V-13B, Temperature=02025.02 | 2.49 | 3.31 | 71.08 | |
| LongSpecBackbone=LC-7B, Temperature=02025.02 | 2.41 | 3.59 | 101.43 | |
| LongSpecBackbone=V-7B, Temperature=02025.02 | 2.23 | 3.57 | 102.23 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=SuffixDecoding2026.05 | 2.08 | — | — | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Greedy Decoding2026.05 | 2.03 | — | — | |
| PLDBackbone=LC-7B, Temperature=02025.02 | 1.63 | 2.1 | 68.66 | |
| PLDBackbone=V-7B, Temperature=02025.02 | 1.62 | 2.2 | 73.91 | |
| LongSpecBackbone=L-8B, Temperature=02025.02 | 1.59 | 3.25 | 84.57 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=Greedy Decoding2026.05 | 1.52 | — | — | |
| PLDBackbone=L-8B, Temperature=02025.02 | 1.46 | 2.08 | 77.45 | |
| MagicDecBackbone=V-13B, Temperature=02025.02 | 1.34 | 2.95 | 38.24 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=EAGLE32026.05 | 1.31 | — | — | |
| PLDBackbone=LC-13B, Temperature=02025.02 | 1.24 | 1.67 | 35.35 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.18 | — | — | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.15 | — | — | |
| PLDBackbone=V-13B, Temperature=02025.02 | 1.13 | 1.37 | 32.1 | |
| MagicDecBackbone=LC-13B, Temperature=02025.02 | 1.1 | 2.4 | 31.37 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.02 | — | — | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=EAGLE32026.05 | 1.01 | — | — | |
| Vanilla FABackbone=V-7B, Temperature=02025.02 | 1 | 1 | 45.76 | |
| Vanilla FABackbone=V-13B, Temperature=02025.02 | 1 | 1 | 28.52 | |
| Vanilla FABackbone=LC-7B, Temperature=02025.02 | 1 | 1 | 42.14 | |
| Vanilla FABackbone=LC-13B, Temperature=02025.02 | 1 | 1 | 28.56 | |
| Vanilla FABackbone=L-8B, Temperature=02025.02 | 1 | 1 | 53.14 | |
| MagicDecBackbone=LC-7B, Temperature=02025.02 | 0.99 | 2.26 | 41.9 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=SuffixDecoding2026.05 | 0.98 | — | — | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=EAGLE32026.05 | 0.95 | — | — | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Greedy Decoding2026.05 | 0.95 | — | — | |
| MagicDecBackbone=V-7B, Temperature=02025.02 | 0.91 | 2.23 | 41.68 | |
| MagicDecBackbone=L-8B, Temperature=02025.02 | 0.68 | 2.04 | 36.14 | |
| Vanilla HFBackbone=V-7B, Temperature=02025.02 | — | 1 | 25.25 | |
| Vanilla HFBackbone=V-13B, Temperature=02025.02 | — | 1 | 17.25 | |
| Vanilla HFBackbone=LC-7B, Temperature=02025.02 | — | 1 | 25.27 | |
| Vanilla HFBackbone=LC-13B, Temperature=02025.02 | — | 1 | 17.72 | |
| Vanilla HFBackbone=L-8B, Temperature=02025.02 | — | 1 | 21.59 |