# The Art of SOCRATIC QUESTIONING: Zero-shot Multimodal Reasoning with Recursive Thinking and Self-Questioning

**Authors:** Jingyuan Qi, Zhiyang Xu, Ying Shen, Minqian Liu, Di Jin, Qifan Wang, Lifu Huang  
**arXiv:** [2305.14999](https://arxiv.org/abs/2305.14999)  
**Date:** 2023-05-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.14999v1/

## Key points

- The authors propose a new algorithm called Socratic Questioning for improving the thinking process of large-scale language models (LLMs) in solving complex reasoning problems.
- They highlight two important capabilities of humans in problem-solving: self-questioning and recursive thinking.
- The existing Chain-of-Thought (CoT) approach used by LLMs is limited as it follows a greedy thinking process and does not consider alternative starting points or explore multiple paths of thinking.
- The Socratic Questioning algorithm introduces a Self-Questioning module that employs a large-scale language model to propose sub-problems related to the original problem as intermediate steps, and recursively backtracks and answers these sub-problems until reaching the original problem.
- The effectiveness of the approach is evaluated on the visual question-answering task, showing significant performance improvement over baselines on three public benchmark datasets.
- Qualitative analysis demonstrates that the intermediate thinking steps elicited by Socratic Questioning are similar to humans' recursive thinking process when solving complex reasoning problems.
- The approach allows LLMs to actively search for relevant information and tackle simpler sub-tasks, distinguishing it from previous approaches like CoT and instruction tuning.
- Overall, Socratic Questioning enhances the problem-solving capabilities of large-scale language models by incorporating self-questioning and recursive thinking processes.

## Summary

In this work, the authors propose a new algorithm called Socratic Questioning that aims to improve the thinking process of large-scale language models (LLMs) when solving complex reasoning problems. They highlight two important capabilities of humans in problem-solving: self-questioning and recursive thinking. Self-questioning involves proactively asking questions from different angles to gain a better understanding of the problem, while recursive thinking involves breaking down complex problems into simpler sub-problems and solving them recursively. The authors argue that the existing Chain-of-Thought (CoT) approach used by LLMs is limited as it follows a greedy thinking process and does not consider alternative starting points or explore multiple paths of thinking. The proposed Socratic Questioning algorithm addresses these limitations by introducing a Self-Questioning module that employs a large-scale language model to propose sub-problems related to the original problem as intermediate steps. The algorithm then recursively backtracks and answers these sub-problems until it reaches the original problem. To evaluate the effectiveness of their approach, the authors apply it to the visual question-answering task and compare it with several baselines on three public benchmark datasets. They observe a significant performance improvement over all baselines on almost all datasets. The authors also conduct qualitative analysis to demonstrate that the intermediate thinking steps elicited by Socratic Questioning are similar to humans' recursively thinking process when solving complex reasoning problems. They emphasize that their approach allows LLMs to actively search for relevant information related to a complicated task and tackle simpler sub-tasks, distinguishing it from previous approaches like CoT and instruction tuning. Overall, this work presents an innovative algorithm, Socratic Questioning, which enhances the problem-solving capabilities of large-scale language models by incorporating self-questioning and recursive thinking processes. The results show promising performance improvements in visual question-answering tasks, indicating its potential for broader applications in complex reasoning problems.