Structure Guided Prompt: Instructing Large Language Model in Multi-Step Reasoning by Exploring Graph Structure of the Text

TL;DR

The paper addresses the challenge of multi-step reasoning in large language models by introducing Structure Guided Prompt, a zero-shot framework that converts unstructured text into a graph and guides LLMs to navigate this graph using task-specific strategies. The method comprises three stages—graph construction, planning, and execution—applied across diverse reasoning tasks (relation/entity prediction, graph sorting, graph querying, and logical inference). Empirical results across CLUTRR, BBH, HotpotQA, and Entailment Bank (with GPT-3.5 and GPT-4) show consistent gains over zero-shot chain-of-thought prompts, especially in dynamic KG settings and other graph-oriented tasks, though some complex entailment scenarios remain challenging due to graph construction and rule ordering. The work highlights the potential of graph-based prompting to broaden LLM reasoning capabilities and informs future research on robust graph-grounded reasoning and verification methods.

Abstract

Although Large Language Models (LLMs) excel at addressing straightforward reasoning tasks, they frequently struggle with difficulties when confronted by more complex multi-step reasoning due to a range of factors. Firstly, natural language often encompasses complex relationships among entities, making it challenging to maintain a clear reasoning chain over longer spans. Secondly, the abundance of linguistic diversity means that the same entities and relationships can be expressed using different terminologies and structures, complicating the task of identifying and establishing connections between multiple pieces of information. Graphs provide an effective solution to represent data rich in relational information and capture long-term dependencies among entities. To harness the potential of graphs, our paper introduces Structure Guided Prompt, an innovative three-stage task-agnostic prompting framework designed to improve the multi-step reasoning capabilities of LLMs in a zero-shot setting. This framework explicitly converts unstructured text into a graph via LLMs and instructs them to navigate this graph using task-specific strategies to formulate responses. By effectively organizing information and guiding navigation, it enables LLMs to provide more accurate and context-aware responses. Our experiments show that this framework significantly enhances the reasoning capabilities of LLMs, enabling them to excel in a broader spectrum of natural language scenarios.

Figures (11)

• Figure 1: GPT-4’s performance using 0-shot chain-of-thought (0-CoT) (represented as the orange bars) compared to the results of Structure Guided Prompt (represented as the blue bars) across a variety of tasks. It is evident that Structure Guided Prompt consistently and significantly outperforms the approach with 0-CoT.
• Figure 2: An example illustrating how humans manage multi-step questions. Our objective is to deduce the relationship between two individuals, Michelle and Stanley, highlighted in red, from a given story. Given that the story involves various individuals, humans typically first create a graph to clearly visualize the relationships among them. Then, they infer the relationship step by step, based on the graph.
• Figure 3: The graph representation of a story from CLUTRR dataset. our objective is to determine the family relationship between two nodes, Seth and Jeremy, which are highlighted in red.
• Figure 4: The bridging question in HotpotQA. It relies on multi-hop sequential reasoning to answer the question.
• Figure : Relation Prediction