Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
OPRO: Automatic Prompt Optimization
This talk introduces OPRO, a method where an LLM iteratively refines prompts using feedback, treating prompt design as a language-space optimization problem.
While large language models are incredibly capable, their performance can vary dramatically depending on how they’re prompted.
Prompting is still mostly manual: trial and error, slow, and hard to scale. Even small differences in phrasing or the order of examples can make a big difference in results.
OPRO, or Optimization by Prompting, changes that.
Instead of humans manually tuning prompts, an LLM can itself iteratively propose, test, and refine new prompts in natural language, using feedback to improve performance on a target task.
It treats prompt design as an optimization problem, but instead of gradient descent, it performs language-space optimization. Humans only need to define how to evaluate or score each prompt’s performance, and the model learns to improve by optimizing against that score.
In short, OPRO is a simple yet powerful approach to Automatic Prompt Optimization (APO).