Speculative Decoding for Diffusion Language Models

This field studies how to reduce the sequential cost of diffusion language model generation while preserving the output quality of a verifier. Common approaches use a lightweight or self-speculative drafter followed by block verification, and extend drafting with trees, adaptive candidate budgets, or draft editing and refinement. Training methods align parallel predictions with left-to-right verification, while systems work reduces verifier calls and reuses computation across branches; evaluations often include reasoning and code benchmarks such as GSM8K, MATH, HumanEval, and MBPP, as well as throughput under concurrent serving. A central tension is that broader parallel exploration can increase drafting and verification costs or cause trajectory drift, so methods must maximize verified progress rather than draft length alone; multimodal applicability also remains unsettled.