Code Generation on SWE-bench OpenHands
3.7SpeedupHybrid Verified Decoding
Evaluation Results
| Method | Links | |
|---|---|---|
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Greedy Decoding2026.05 | 3.7 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Greedy Decoding2026.05 | 3.31 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=SuffixDecoding2026.05 | 3.21 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=SuffixDecoding2026.05 | 2.6 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=Greedy Decoding2026.05 | 2.5 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=EAGLE32026.05 | 2.11 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=SuffixDecoding2026.05 | 1.52 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.28 | |
| Hybrid Verified DecodingTarget Model=Llama3.1-8B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 1.26 | |
| Hybrid Verified DecodingTarget Model=Qwen3-8B, Reference Baseline=EAGLE32026.05 | 1.09 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=EAGLE32026.05 | 1.08 | |
| Hybrid Verified DecodingTarget Model=Qwen3-4B, Reference Baseline=Frequency-Based Hybrid Decoding2026.05 | 0.88 |