Glossary · Term

LLaDA

← all terms

Definition

Plain language

A language model that generates text by repeatedly filling in masked-out words rather than writing left to right.

As stated in the literature

A discrete diffusion language model that works directly on tokens via iterative masking and infilling, contrasted with continuous latent diffusion approaches.

Why it matters: It demonstrates that a full language model can be built on the masking-and-infilling approach, offering an alternative to writing text strictly left to right.

For example, rather than writing a reply word by word, it starts from a sentence full of blanks and fills them in over several passes.

Heard on the show

“There's discrete diffusion — you might have heard of LLaDA — which works directly on tokens: mask words out, learn to fill them back in, iterate.”
Episode 127 — What Diffusion Language Models Were Missing: A Map, Not an Algorithm

Mentioned in 1 episode

  1. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm

Related terms