Large Language Models as Optimizers

Source: arXiv Authors: Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, Xinyun Chen Date: 2023-09-07 Primary category: cs.LG All categories: cs.LG, cs.AI, cs.CL

Abstract / key passage

This paper introduces Optimization by PROmpting (OPRO), where a language model acts as a black-box optimizer that proposes new candidate solutions from a prompt containing previously tried solutions and their scores. In the prompt-optimization setting, the candidates are task instructions whose downstream task performance is evaluated and fed back into the next optimization step.

Harness takeaway

OPRO is the clean canonical reference for the “LLM as optimizer” branch. For harnesses it is useful whenever there is a scoreable artifact and little structured critique, but it also highlights how much optimization can proceed with almost nothing except candidate history and scalar value feedback.