FSM: A Finite State Machine Based Zero-Shot Prompting Paradigm for Multi-Hop Question Answering

TL;DR

This paper introduces FSM, a finite state machine prompting paradigm that enhances large language models' multi-hop reasoning by decomposing questions into sub-questions, self-correcting, and reducing hallucinations, especially on complex datasets.

Contribution

The paper presents a novel FSM-based prompting method that improves multi-hop question answering accuracy and trustworthiness over chain-of-thought prompting, particularly on challenging datasets.

Findings

FSM outperforms baseline on complex datasets like Musique.

FSM reduces hallucination and improves answer correctness.

FSM enhances adherence to output format requirements.

Abstract

Large Language Models (LLMs) with chain-of-thought (COT) prompting have demonstrated impressive abilities on simple nature language inference tasks. However, they tend to perform poorly on Multi-hop Question Answering (MHQA) tasks due to several challenges, including hallucination, error propagation and limited context length. We propose a prompting method, Finite State Machine (FSM) to enhance the reasoning capabilities of LLM for complex tasks in addition to improved effectiveness and trustworthiness. Different from COT methods, FSM addresses MHQA by iteratively decomposing a question into multi-turn sub-questions, and self-correcting in time, improving the accuracy of answers in each step. Specifically, FSM addresses one sub-question at a time and decides on the next step based on its current result and state, in an automaton-like format. Experiments on benchmarks show the effectiveness of our method. Although our method performs on par with the baseline on relatively simpler datasets, it excels on challenging datasets like Musique. Moreover, this approach mitigates the hallucination phenomenon, wherein the correct final answer can be recovered despite errors in intermediate reasoning. Furthermore, our method improves LLMs' ability to follow specified output format requirements, significantly reducing the difficulty of answer interpretation and the need for reformatting.