Reproduce “RePrompt: Planning by Automatic Prompt Engineering for Large Language Models Agents” 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
Weizhe Chen, Sven Koenig, Bistra Dilkina
Researchers at the University of Southern California developed REPROMPT, an automatic prompt engineering method for Large Language Model (LLM) agents that refines step-by-step instructions by utilizing intermediate feedback from agent interactions, bypassing the need for a final solution checker. The system demonstrated improved performance across diverse planning and reasoning tasks, including a reduction in PDDL generation errors and an increase in travel planning success rates.
Key insights
- Automatic prompt optimization for LLM agents is feasible and effective using only intermediate feedback, addressing the common challenge of lacking a final ground-truth checker in real-world scenarios.
- Systematically optimizing the step-by-step instructions within prompts based on aggregated interaction history leads to more robust and reliable LLM agent reasoning and higher first-round success rates.
- The proposed method, REPROMPT, is generalizable across various types of intermediate feedback (e.g., accurate expert annotations, noisy self-reflection logs) and diverse LLM agent environments, demonstrating adaptability to different noise levels and feedback structures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.