Reproduce “Localized Zeroth-Order Prompt Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenyang Hu, Yao Shu, Zongmin Yu, Zhaoxuan Wu, Xiangqiang Lin, Zhongxiang Dai, See-Kiong Ng, Bryan Kian Hsiang Low
Hu et al. (2023) introduce Localized Zeroth-Order Prompt Optimization (ZOPO), an algorithm that efficiently tunes prompts for black-box LLMs by focusing on high-performing local optima. The method consistently achieves higher accuracy with fewer queries compared to existing approaches across various instruction induction and arithmetic reasoning tasks.
Key insights
- In prompt optimization, high-performing local optima are abundant and often perform comparably to the global optimum, suggesting that resource-intensive global search may be unnecessary.
- The input domain, specifically the choice of LLM for prompt generation and the prompt embedding method, significantly influences the optimization landscape and overall performance.
- Visualization of prompt performance distributions across tasks reveals numerous regions of good performance, supporting the efficiency of localized search.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.