# Emergent Analogical Reasoning in Large Language Models

**Authors:** Taylor Webb, Keith J. Holyoak, Hongjing Lu  
**arXiv:** [2212.09196](https://arxiv.org/abs/2212.09196)  
**Date:** 2022-12-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.09196v1/

## Key points

- Large language models like GPT-3 can exhibit human-like cognitive abilities
- GPT-3 has a strong ability for abstract pattern induction and performs on par with or surpasses humans in most settings
- GPT-3 can solve tasks involving analogical reasoning, including text-based matrix reasoning tasks
- GPT-3 has the capacity to find zero-shot solutions to a wide range of analogy problems
- Prompting participants to consider previously presented stories improves their performance in solving complex analogy problems
- This research provides valuable insights into the cognitive abilities of large language models and their potential applications in solving naturalistic problems.

## Summary

The recent emergence of large language models such as GPT-3 has sparked a debate about whether they can exhibit human-like cognitive abilities. Of particular interest is their capacity to reason by analogy and solve novel problems without any direct training. A study comparing GPT-3 with human reasoners found that it demonstrated a surprisingly strong ability for abstract pattern induction and performed on par with or even surpassed humans in most settings. This included tasks involving analogical reasoning, including a text-based matrix reasoning task inspired by Raven's Progressive Matrices. These findings suggest that large language models like GPT-3 have acquired an emergent capacity to find zero-shot solutions to a wide range of analogy problems. Additionally, the researchers explored how participants' performance improved when prompted to consider previously presented stories in solving complex analogy problems. This research provides valuable insights into the cognitive abilities of large language models and their potential applications in solving naturalistic problems.