Stepwise Reasoning Enhancement for LLMs via External Subgraph Generation

2026-06-03

AI Summary

This work addresses the challenges large language models face in complex multi-step reasoning—namely logical inconsistency, insufficient factual grounding, and poor interpretability—by proposing the SGR framework. SGR innovatively integrates schema-guided subgraph retrieval from a Neo4j-based knowledge graph with a multi-path collaborative reasoning mechanism. It dynamically constructs query-relevant subgraphs to provide structured evidence, guiding the model through step-by-step reasoning while jointly leveraging direct Cypher-based inference and path validation to enhance answer reliability. Experimental results demonstrate that SGR significantly outperforms existing knowledge-augmented approaches on benchmark datasets including CWQ, WebQSP, GrailQA, and KQA Pro, achieving substantial gains in Hits@1. Ablation studies further confirm the effectiveness of each component within the framework.

Abstract

Large language models have shown strong performance in natural language generation and downstream reasoning tasks, but they still struggle with logical consistency, factual grounding, and interpretability in complex multi-step reasoning. To address these limitations, this paper proposes SGR, a stepwise reasoning enhancement framework that integrates large language models with external knowledge graphs through query-relevant subgraph generation. Given an input question, SGR first extracts key entities, relations, and constraints to construct a structured schema, then retrieves compact subgraphs from a knowledge graph using schema-guided querying. The generated subgraphs provide explicit relational evidence that guides the language model through step-by-step reasoning. In addition, SGR combines direct Cypher-based reasoning with collaborative reasoning integration, allowing candidate answers from multiple reasoning paths to be validated and aggregated according to both model confidence and graph consistency. Experiments on benchmark datasets including CWQ, WebQSP, GrailQA, and KQA Pro demonstrate that SGR improves reasoning accuracy and Hits@1 performance over standard prompting and several knowledge-enhanced baselines. Ablation studies further show that schema guidance and Neo4j-based retrieval are both crucial to the effectiveness of the framework. These results indicate that dynamically generated external subgraphs can improve the accuracy, robustness, and interpretability of LLM-based reasoning.

Problem

• logical consistency
• factual grounding
• interpretability
• multi-step reasoning
• large language models

Innovation

• stepwise reasoning
• knowledge graph
• subgraph retrieval
• schema-guided querying
• LLM reasoning enhancement