# Self-Discover: Large Language Models Self-Compose Reasoning Structures

**Authors:** Pei Zhou, Jay Pujara, Xiang Ren, Xinyun Chen, Heng-Tze Cheng, Quoc V. Le, Ed H. Chi, Denny Zhou, Swaroop Mishra, Huaixiu Steven Zheng  
**arXiv:** [2402.03620](https://arxiv.org/abs/2402.03620)  
**Date:** 2024-02-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.03620v1/

## Key points

- The study introduces the framework of SELF-DISCOVER for LLMs to autonomously uncover task-specific reasoning structures
- LLMs select atomic reasoning modules and combine them into an explicit structure to guide their decoding process
- Significant improvements in performance on challenging benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH compared to existing methods like Chain of Thought (CoT)
- Outperforms inference-intensive approaches while requiring less compute
- Self-discovered structures have demonstrated universality across different model families and exhibit similarities with human reasoning patterns

## Summary

The study titled "Self-Discover: Large Language Models Self-Compose Reasoning Structures" introduces the framework of SELF-DISCOVER for LLMs to autonomously uncover task-specific reasoning structures. This aims to address complex reasoning problems that are challenging for conventional prompting methods. The core concept involves a self-discovery process where LLMs select atomic reasoning modules and combine them into an explicit structure to guide their decoding process. In the first stage, relevant modules are selected, adapted, and implemented into a structured plan for solving the task. This approach has shown significant improvements in performance on challenging benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH compared to existing methods like Chain of Thought (CoT). It also outperforms inference-intensive approaches while requiring less compute. The self-discovered structures have demonstrated universality across different model families and exhibit similarities with human reasoning patterns. Overall, SELF-DISCOVER presents a promising approach for enhancing LLMs' ability to tackle intricate reasoning tasks through autonomous discovery of task-specific structures.