Glossary · Term

REPA

← all terms

Definition

Plain language

A training trick that teaches a model's compressed text space to be organized by meaning by copying the layout from a model that already understands language.

As stated in the literature

Representation alignment — a loss term that pulls a VAE encoder's internal features toward those of a frozen pretrained model, shaping latent geometry to be navigable for diffusion without improving reconstruction.

Also called: representation alignment

Why it matters: It shapes the compressed space to be easy for a diffusion model to navigate, which can be the key ingredient that makes generation work well.

For example, it teaches a model's compressed text space to arrange related sentences near each other by copying the sense of meaning from a model that already understands language.

Heard on the show

“Which brings us to the thing the triage notes call the load-bearing innovation — REPA.”
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 concepts

Related terms