Promptagent

# PromptAgent: Strategic planning optimization

**Date**: 2024  
**Institution**: Carnegie Mellon University, Microsoft Research

## Algorithm

PromptAgent discovered expert-level prompts through strategic exploration using Monte Carlo Tree Search for prompt space exploration. Employs error reflection analyzing model errors to generate insightful feedback with multi-step planning simulating future rewards.

Strategic innovation: First strategic planning approach to prompt optimization with human-like trial-and-error interaction and error-based learning. Consistently outperformed strong baselines across 12 tasks spanning BBH, domain-expert, and general NLU tasks.

Algorithmic framework: Tree search with expert-level knowledge injection through error feedback, though computationally intensive due to strategic exploration requirements.

## Comparison with Other Work

PromptAgent differs from local optimization methods like AutoPrompt and APO by using strategic planning rather than local search. Unlike continuous methods like Prefix-Tuning and Prompt Tuning, it uses discrete search with strategic exploration. Compared to LLM-as-optimizer methods like OPRO, PromptAgent uses tree search rather than population-based generation. Unlike reinforcement learning methods like RLPrompt and TEMPERA, it uses Monte Carlo Tree Search rather than policy gradients. PromptAgent is similar to evolutionary methods like EvoPrompt in using population-based search but with more sophisticated exploration strategies. It's more computationally intensive than gradient-based methods but can find higher-quality solutions through strategic exploration.