Code Generation on MBPP 3-shot (Flex and Reuse)
55.6Flex (%)Baseline
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BaselineBackbone=Dream2026.05 | 55.6 | — | |
| P+CBackbone=Dream2026.05 | 54.6 | — | |
| DARE-KVBackbone=Dream, Base Framework=P+C2026.05 | 54.4 | 74 | |
| DARE-KVBackbone=Dream, Base Framework=Fast-dLLM2026.05 | 54.2 | 57.4 | |
| Fast-dLLMBackbone=Dream2026.05 | 53 | — | |
| DARE-OBackbone=Dream, Base Framework=Fast-dLLM2026.05 | 51.8 | 92.9 | |
| DARE-OBackbone=Dream, Base Framework=P+C2026.05 | 51.2 | 90.8 | |
| BaselineBackbone=LLaDA-1.52026.05 | 39 | — | |
| DARE-KVBackbone=LLaDA-1.5, Base Framework=P+C2026.05 | 38.8 | 15 | |
| P+CBackbone=LLaDA-1.52026.05 | 38 | — | |
| DARE-OBackbone=LLaDA-1.5, Base Framework=P+C2026.05 | 37.8 | 65.9 | |
| Fast-dLLMBackbone=LLaDA-1.52026.05 | 35.8 | — | |
| DARE-OBackbone=LLaDA-1.5, Base Framework=Fast-dLLM2026.05 | 35 | 87.2 | |
| DARE-KVBackbone=LLaDA-1.5, Base Framework=Fast-dLLM2026.05 | 33.8 | 22.3 | |
| BaselineBackbone=LLaDA2026.05 | 30.4 | — | |
| P+CBackbone=LLaDA2026.05 | 27 | — | |
| DARE-KVBackbone=LLaDA, Base Framework=Fast-dLLM2026.05 | 26.6 | 26.9 | |
| DARE-OBackbone=LLaDA, Base Framework=Fast-dLLM2026.05 | 26.4 | 83.7 | |
| DARE-KVBackbone=LLaDA, Base Framework=P+C2026.05 | 26.2 | 34.5 | |
| Fast-dLLMBackbone=LLaDA2026.05 | 25.8 | — | |
| DARE-OBackbone=LLaDA, Base Framework=P+C2026.05 | 25.6 | 63.1 |