Long-context Modeling on CNN/DailyMail
3.85SpeedupHybrid Verified Decoding
Evaluation results
| Method | Links | |
|---|---|---|
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=SuffixDecoding2026.05 | 3.85 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=SuffixDecoding2026.05 | 3.83 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=SuffixDecoding2026.05 | 2.75 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Greedy Decoding2026.05 | 2.43 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Greedy Decoding2026.05 | 2.33 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=Greedy Decoding2026.05 | 1.79 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.34 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.04 | |
| 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 | 0.99 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=EAGLE32026.05 | 0.94 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=EAGLE32026.05 | 0.94 |