OPRO

Overview

OPRO is the optimization-by-prompting method introduced in Large Language Models as Optimizers (2023). It treats a language model as a black-box optimizer that proposes new candidate solutions from a history of previously tried candidates and their scores.

Why it matters

It matters because it is the cleanest reference for the “LLM as optimizer” branch. For harnesses, it is useful whenever there is a scoreable artifact but little structured critique, making candidate history itself a reusable optimization substrate.

Distinctive trait

Its distinctive trait is in-context black-box optimization over candidate/value histories rather than tree search, population evolution, or explicit prompt-program structure.

Relationships

Read OPRO with promptbreeder, promptagent, rlprompt, and prompt-optimizer-regimes-for-harnesses. It is a natural contrast class for textgrad, where the update signal is richer than scalar score history.