# Automating Thought of Search: A Journey Towards Soundness and Completeness

**Authors:** Daniel Cao, Michael Katz, Harsha Kokel, Kavitha Srinivas, Shirin Sohrabi  
**arXiv:** [2408.11326](https://arxiv.org/abs/2408.11326)  
**Date:** 2024-08-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.11326v1/

## Key points

- Authors explore the use of large language models (LLMs) in planning and search tasks
- Traditional LLMs prioritize flexibility over soundness in defining search spaces
- Introduction of ToS method involves human collaboration to create sound successor function and goal test, achieving 100% accuracy in solving datasets
- AutoToS is an automated version that eliminates human intervention, guiding LLMs towards generating sound and complete search components through feedback from unit tests
- AutoToS achieves 100% accuracy across various domains with minimal feedback iterations
- Automation streamlines the process and showcases significant progress in LLMs for complex reasoning tasks
- Study highlights potential of leveraging LLMs for automating planning tasks efficiently with high levels of accuracy and completeness

## Summary

In their paper titled "Automating Thought of Search: A Journey Towards Soundness and Completeness," authors Daniel Cao, Michael Katz, Harsha Kokel, Kavitha Srinivas, and Shirin Sohrabi delve into the realm of large language models (LLMs) and their application in planning and search tasks. Traditionally, LLMs have been utilized as world models to define search spaces, prioritizing flexibility over soundness. However, a recent approach known as  introduced a novel method by defining the search space with code generated by language models. ToS involves human collaboration to create a sound successor function and goal test, resulting in solving datasets with 100% accuracy. Building upon the success of ToS, the authors introduce , an automated version that eliminates human intervention in solving planning problems. AutoToS guides LLMs step by step towards generating sound and complete search components through feedback from both generic and domain-specific unit tests. The results are impressive, achieving 100% accuracy across various domains with minimal feedback iterations. This automation not only streamlines the process but also showcases the significant progress made by LLMs in code generation and refinement for complex reasoning tasks. Overall, this study highlights the potential of leveraging LLMs for automating planning tasks efficiently while maintaining high levels of accuracy and completeness. The collaborative efforts between humans and machines pave the way for advancements in artificial intelligence research and applications in diverse domains.