Kuleshov Group publishes a full build guide for diffusion language models

Kuleshov Group·medium signalXBlueskyLinkedInCopy link

The Cornell group behind several open diffusion LLMs posted a long-form guide on 2026-08-30 adapted from their ICLR 2026 and MLSS 2026 talks, walking from Gaussian diffusion through simple masked diffusion to the pieces that make a real model work: block diffusion for variable-length generation, encoder/decoder/encoder-decoder architecture choices, iterative refinement with built-in error correction, distillation for sampling speed, and post-training. It frames diffusion against the three structural limits of autoregression, namely no revision of emitted tokens, one step per token, and backward-only attention. This is the reference to hand someone who wants to actually build one rather than read another survey.