# Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations

**Authors:** Jaehun Jung, Lianhui Qin, Sean Welleck, Faeze Brahman, Chandra Bhagavatula, Ronan Le Bras, Yejin Choi  
**arXiv:** [2205.11822](https://arxiv.org/abs/2205.11822)  
**Date:** 2022-05-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.11822v1/

## Key points

- Researchers propose a few-shot inference method called Maieutic Prompting
- Inspired by the Socratic way of conversation
- Goal: Generate correct answers to questions from noisy and inconsistent generations of pre-trained language models (LMs)
- Achieved through generating a tree of explanations with logical relations and assigning truth values based on maximum satisfiability
- Tested on three challenging benchmarks for true/false question-answering tasks
- Compared performance with state-of-the-art prompting methods, Maieutic Prompting achieves up to 20% better accuracy
- Performs competitively with supervised models despite being fully unsupervised
- Improves robustness in inference and provides interpretable rationales
- Provides intrinsic interpretations over its inference process
- Empirical results and qualitative analyses support claims made by researchers
- Generation process involves recursively prompting the LM with generated explanations to validate correctness
- Nucleus sampling used to enhance diversity and coverage of generated explanations

## Summary

In this work, the researchers propose a novel few-shot inference method called Maieutic Prompting, inspired by the Socratic way of conversation. The goal is to generate correct answers to questions even from noisy and inconsistent generations of large pre-trained language models (LMs). They achieve this by systematically generating a tree of explanations that have logical relations between each other and assigning truth values to these explanations based on maximum satisfiability. The researchers test Maieutic Prompting on three challenging benchmarks that require complex commonsense reasoning for true/false question-answering tasks. They compare its performance with state-of-the-art prompting methods and find that Maieutic Prompting achieves up to 20% better accuracy. Despite being a fully unsupervised approach, it performs competitively with supervised models. Furthermore, they demonstrate that Maieutic Prompting improves robustness in inference while providing interpretable rationales. It also provides intrinsic interpretations over its inference process. The researchers provide empirical results and qualitative analyses to support their claims. They also discuss the generation process of the maieutic tree, which involves recursively prompting the LM with generated explanations to validate their correctness. They use nucleus sampling to enhance the diversity and coverage of generated explanations. Overall, this work presents a promising approach for improving reasoning capabilities in large pre-trained language models by generating correct explanations and framing inference as a satisfiability problem over these explanations and their logical relations.