# Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models

**Authors:** Bilgehan Sel, Ahmad Al-Tawaha, Vanshaj Khattar, Lu Wang, Ruoxi Jia, Ming Jin  
**arXiv:** [2308.10379](https://arxiv.org/abs/2308.10379)  
**Date:** 2023-08-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.10379v1/

## Key points

- Recent literature focuses on surpassing the limitations of the "Chain-of-Thought" approach in enhancing reasoning capabilities of Large Language Models (LLMs)
- Existing methods often rely on external modus operandi, leading to increased costs, memory usage, and computational overheads
- Proposed strategy called Algorithm of Thoughts leverages algorithmic reasoning pathways to propel LLMs towards in-context learning
- Utilizing algorithmic examples taps into the innate recurrence dynamics of LLMs and expands idea exploration with just one or a few queries
- Technique outperforms earlier single-query methods and is comparable to a recent multi-query strategy using an extensive tree search algorithm
- Instructing an LLM using an algorithm can lead to performance surpassing that of the algorithm itself, suggesting LLMs have inherent intuition for optimized searches
- Efficacy of the method is delved into and its nuances in application are explored
- Related work includes input-output prompting and more advanced techniques proposed by Shao et al., Wei et al., and Lyu et al.
- Findings are summarized in conclusion
- Figure 2 provides an illustration outlining various strategies for tackling reasoning problems with LLMs, with green boxes indicating promising ideas identified by the LLM and red boxes representing less promising concepts.

## Summary

In recent literature, there has been a focus on surpassing the limitations of the "Chain-of-Thought" approach in enhancing the reasoning capabilities of Large Language Models (LLMs). To address the challenges posed by existing methods, which often rely on an external modus operandi that involves halting, modifying, and then resuming the generation process leading to increased costs, memory usage, and computational overheads due to the escalated number of query requests, we propose a novel strategy called the Algorithm of Thoughts. This approach leverages algorithmic reasoning pathways to propel LLMs towards in-context learning. By utilizing algorithmic examples, we tap into the innate recurrence dynamics of LLMs and expand their idea exploration with just one or a few queries. Our technique outperforms earlier single-query methods and is comparable to a recent multi-query strategy that employs an extensive tree search algorithm. Interestingly, our results indicate that instructing an LLM using an algorithm can lead to performance surpassing that of the algorithm itself. This suggests that LLMs have an inherent ability to weave their intuition into optimized searches. We delve into the efficacy of our method and explore its nuances in application. In subsequent sections, we situate our work within the existing literature and discuss our principal idea. We then present experimental results and investigate hypotheses related to this emerging capability of LLMs. Related work in this field includes standard prompting approaches such as input-output prompting which tends to yield lower performance compared to more advanced techniques proposed by Shao et al., Wei et al., and Lyu et al.. Finally, we conclude by summarizing our findings. Figure 2 provides an illustration outlining various strategies for tackling reasoning problems with LLMs. Each box represents a distinct thought or concept forming an incremental pathway to reasoning. Green boxes indicate promising ideas identified by the LLM while red boxes represent less promising concepts.