Code Generation on HumanEval (Accuracy and Throughput)
172.25Throughput (tokens/s)OSDT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OSDTBackbone=LLaDA-8B, Decoding Strategy=One-Shot Dynamic Thresholding2025.11 | 172.25 | 40.85 | — | |
| Fast-dLLMBackbone=LLaDA-8B, Decoding Strategy=Fixed-threshold (tau = 0.9)2025.11 | 152.51 | 39.63 | — | |
| Fast-dLLMBackbone=LLaDA-8B, Decoding Strategy=Factor-based2025.11 | 114.71 | 43.29 | — | |
| MoE-SpAcBase Model=DeepSeek-V2-Lite2026.02 | 47.26 | — | 9.46 | |
| llama.cpp-w/ SDBase Model=DeepSeek-V2-Lite, Speculative Decoding=true2026.02 | 30.49 | — | 17.85 | |
| HybriMoEBase Model=DeepSeek-V2-Lite2026.02 | 17.95 | — | 24.16 |