# Answering Questions by Meta-Reasoning over Multiple Chains of Thought

**Authors:** Ori Yoran, Tomer Wolfson, Ben Bogin, Uri Katz, Daniel Deutch, Jonathan Berant  
**arXiv:** [2304.13007](https://arxiv.org/abs/2304.13007)  
**Date:** 2023-04-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.13007v1/

## Key points

- Multi-Chain Reasoning (MCR) is an approach to multi-hop question answering that prompts large language models to meta-reason over multiple chains of thought.
- MCR examines different reasoning chains, mixes information between them, and selects the most relevant facts in generating an explanation and predicting the answer.
- MCR outperforms strong baselines on seven multi-hop QA datasets and its explanations exhibit high quality, enabling humans to verify its answers.
- The process of generating reasoning chains involves a decomposition model that generates intermediate questions based on the original question and previous reasoning steps, and a retriever component that retrieves relevant information from external sources.
- The generated reasoning chains are then fed into a prompted LLM which generates multiple reasoning chains by interleaving generated intermediate questions with retrieved contexts.
- The main contribution of this work is introducing a second LLM that is prompted to meta-reason on multiple reasoning chains, collecting evidence facts as its explanation and generating the final answer.
- This approach considers the relations between intermediate steps across chains and provides a unified explanation for the predicted answer.

## Summary

The paper introduces Multi-Chain Reasoning (MCR), an approach to multi-hop question answering (QA) that prompts large language models to meta-reason over multiple chains of thought, rather than simply aggregating their answers. Unlike existing approaches that break down questions into a sequence of reasoning steps and discard intermediate steps, MCR examines different reasoning chains, mixes information between them, and selects the most relevant facts in generating an explanation and predicting the answer. The authors demonstrate that MCR outperforms strong baselines on seven multi-hop QA datasets and that its explanations exhibit high quality, enabling humans to verify its answers. The paper describes the process of generating reasoning chains using a decomposition model and a retriever component. The decomposition model generates intermediate questions based on the original question and previous reasoning steps while the retriever component retrieves relevant information from external sources. The generated reasoning chains are then fed into a prompted LLM which generates multiple reasoning chains by interleaving generated intermediate questions with retrieved contexts. The main contribution of this work is introducing a second LLM that is prompted to meta-reason on multiple reasoning chains, collecting evidence facts as its explanation and generating the final answer. This approach considers the relations between intermediate steps across chains and provides a unified explanation for the predicted answer. Overall, this paper presents an innovative approach to multi-hop QA that improves performance by considering multiple reasoning chains and providing more comprehensive explanations for predicted answers. It enables humans to verify its answers through high quality explanations generated by MCR's meta-reasoning process over multiple reasoning chains.