Researchers have found a way to make diffusion language models fast without making them dumb.
The method, called the Discrete Average Generator, adapts an existing continuous-space technique known as MeanFlow for use with discrete text data. Instead of predicting one small step at a time, the model learns to predict an average transition over a whole time interval, then trains against a mathematical self-consistency check to keep that shortcut accurate. Tested on OpenWebText, a large web-text dataset, it produced the lowest perplexity (a standard measure of how well a model predicts text) among compared methods across 8 to 64 sampling steps, while cutting generation time by 16x. It also roughly matched existing methods on ImageNet, suggesting the trick isn't limited to text.
Diffusion language models have been pitched as a faster alternative to the step-by-step, left-to-right generation that powers most chatbots today. The catch has always been that going fast and staying coherent don't mix well - cut the steps, and output quality usually drops. This work attacks that tradeoff directly rather than just throwing more compute at it, which is the more common fix.
Worth noting: this is a research result on a dataset of web text, not a shipped product. A 16x speedup is the kind of number that looks great in a paper and often shrinks considerably once it meets a real serving stack and real-world prompts.