protologue

Diffusion Language Model

Also called diffusion LM, DLM, dLLM, masked diffusion language model, continuous diffusion language model, CDLM.

A diffusion language model generates text by starting from a fully corrupted sequence, such as all mask tokens or noise over token embeddings, and refining the whole sequence over a series of denoising steps, instead of producing one token at a time from left to right.

Description

There are two main families. Discrete diffusion models corrupt the tokens themselves, most often by replacing them with a mask token, and learn to restore them. Continuous diffusion models first map tokens to real-valued embeddings and add Gaussian-style noise to those. Diffusion-LM was an early continuous model aimed at controllable generation. LLaDA, a masked diffusion model trained from scratch at scale, performed comparably to autoregressive baselines, which suggested that strong language-model abilities do not depend on left-to-right generation. Because a diffusion model updates many positions in each step, it can produce several tokens per step and condition on text both before and after a position.

Sources

  1. Li et al. (2022). Diffusion-LM Improves Controllable Text Generation.
  2. Nie et al. (2025). Large Language Diffusion Models.
  3. Dieleman (2026). Continuous diffusion language models.

Cite this entry

Protologue. (2026). Diffusion Language Model. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0228). https://protologue.com/t/diffusion-language-model/

BibTeX
@misc{protologue_diffusion_language_model,
  title = {Diffusion Language Model},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0228},
  url = {https://protologue.com/t/diffusion-language-model/}
}

Markdown JSON