understudy / universitysnacks
info snack

Recursive Language Model

A model loop that improves output through propose, critique, revise, and verify steps.

companion card

A Recursive Language Model improves a task by looping: propose, critique, revise, and stop when the output clears a rubric.

What it means

A recursive language model pattern runs a model multiple times around the same task. One pass proposes an answer, another critiques it, another revises it, and a stopping rule decides when the output is good enough.

Why product teams care

Recursive loops can raise quality on complex tasks, but they add latency and cost. They need rubrics, limits, and failure handling.

Understudy angle

Understudy can measure whether recursive loops are worth it by comparing quality lift against extra tokens, time, and cost.

take this with you

Recursion is useful when the task benefits from explicit revision, not just one-shot generation.

Try one propose-critique-revise loop on a hard example and score the delta.