Autoregressive models write left-to-right, one token at a time. A diffusion language model starts from a fully masked canvas and denoises everywhere at once — committing the tokens it is most confident about, reconsidering the rest, conditioning on both directions. These are real recorded traces from a model we trained.
Each step, the model predicts every masked position, keeps its most confident predictions, and re-masks the rest (low-confidence remasking, as in LLaDA). Early steps lay down the punctuation and story skeleton; later steps fill in content — the order is emergent, not programmed.
Parallelism: many tokens commit per step (frontier diffusion LMs report ~1,000+ tokens/s this way). Infilling: in the infill trace, both sides of the gap are clamped — the model conditions on the future as easily as the past, which a left-to-right model fundamentally cannot. Hard constraints: clamped tokens are guaranteed verbatim, by construction — no prompt begging.