Structured Suffix Modeling Gives Diffusion Language Models up to 72.81x Long-Sequence Speedup, Training-Free

arXiv 2608.23167·medium signalXBlueskyLinkedInCopy link

Diffusion language models decode many tokens per step but pay for interacting with all suffix tokens at every step, and existing fixes just keep a local suffix window while re-initializing suffix tokens identically each timestep. This method splits the suffix into local, middle, and tail regions and retains different token counts per region according to structural role, while carrying the previous step's decoding results into current suffix representations so they accumulate denoising information. It is training-free and orthogonal to parallel decoding and KV caching, reaching up to 72.81x speedup on long-sequence inference when combined with other acceleration techniques across three diffusion language models.