Definition
Plain language
A pattern where a model writes a draft, critiques it, and revises.
As stated in the literature
An inference-time technique in which the same model alternates between generating outputs, generating critiques, and revising based on the critique.
Why it matters: It's a cheap way to lift quality at inference time without retraining, though it depends on the model being able to spot its own flaws in the critique step.
For example, the model writes a draft email, then writes a critique pointing out it's too formal, then rewrites it in a friendlier tone.
Heard on the show
“And the named refinement agents — Self-Refine, and Reflexion, which extends it with a verbal-memory buffer of past failures — plus best-of-three sampling — actually went negative in this regime.”Episode 170 — When a One-Liner Beats Your Agent's Clever Verification Logic