PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization

Source: arXiv Authors: Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Haotian Luo, Jiayou Zhang, Nebojsa Jojic, Eric P. Xing, Zhiting Hu Date: 2023-10-25 Primary category: cs.CL

Abstract / key passage

PromptAgent treats prompt optimization as a strategic planning problem. It reflects on model errors, explores prompt states with a Monte-Carlo-tree-search-style procedure, and searches for high-reward edit paths rather than only proposing independent prompt candidates.

Harness takeaway

PromptAgent is the clearest planning-based prompt optimizer in this cluster. It matters for harnesses because it assumes richer failure traces than scalar-only search and therefore sits closer to systems that already log structured error evidence, tool traces, or evaluator feedback.