Multi-hop Question Answering on 2WikiMultiHopQA (Speedup)
3.28Speedup FactorHybrid Verified Decoding
Evaluation Results
| Method | Links | |
|---|---|---|
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=SuffixDecoding2026.05 | 3.28 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Greedy Decoding2026.05 | 2.73 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Greedy Decoding2026.05 | 2.52 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=SuffixDecoding2026.05 | 2.49 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=Greedy Decoding2026.05 | 2.01 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=SuffixDecoding2026.05 | 1.65 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.4 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=EAGLE32026.05 | 1.19 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=EAGLE32026.05 | 1.04 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 0.98 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=EAGLE32026.05 | 0.96 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 0.87 |