A diffusion LLM starts from a fully masked or noised sequence and refines every position over several denoising steps, so tokens are produced in parallel rather than strictly left to right. That trades the autoregressive one-token-at-a-time bottleneck for a fixed number of parallel passes, and lets earlier tokens be revised as later context resolves.
Worked example: a language model that generates text by DIFFUSION — start from noise/masked tokens and iteratively denoise the whole sequence in parallel — instead of autoregression’s strict left-to-right one-token-at-a-time. Gotcha: parallel refinement can be faster and lets the model revise earlier tokens (autoregression cannot un-commit), but matching autoregressive quality and handling variable length is hard, and most production LLMs are still autoregressive; diffusion-LLMs are a promising alternative decoding paradigm, not yet the default.