Diffusion Language Model

A diffusion language model generates text by repeatedly refining a partially specified sequence rather than committing to every next token strictly from left to right. In a masked-token design, generation begins with placeholder positions and iteratively predicts which tokens belong there. Some systems refine an entire response; others generate and refine blocks in sequence.

That differs from the usual autoregressive Large Language Model (LLM), which conditions each new token on the preceding tokens and cannot revise an earlier token during the same decode pass. A diffusion model can use information from both sides of a position within the region being refined. The architecture and training objective determine how that refinement works; text diffusion is a family of approaches, not one fixed decoding algorithm.

Parallel updates may improve latency when a model runs at low concurrency on capable hardware. The trade-off is that several refinement passes can still be needed, and a faster single request does not necessarily imply better throughput or lower cost in a heavily batched service. Quality, controllability, streaming behavior, and the number of refinement steps should be evaluated on the intended workload rather than inferred from tokens-per-second alone.

This term is useful when comparing generation architectures and inference strategies. It is separate from model distillation, which transfers behavior between models, and from Mixture of Experts (MoE), which controls which parameters are active. A model can combine diffusion-style generation with an MoE backbone.

Google's DiffusionGemma technical introduction describes one blockwise implementation and the workload limits of its reported speed gains.