Stepwise Reasoning Enhancement for LLMs via External Subgraph Generation
2026-06-03 • Computation and Language
Computation and Language
AI summaryⓘ
The authors developed a system called SGR to help large language models think through problems step-by-step using information from knowledge graphs. Their approach extracts important parts of a question to find related facts in a knowledge graph, which then guides the model’s reasoning. This method improves the model’s ability to answer complex questions more accurately and consistently. Tests showed that using SGR works better than other methods and that certain parts, like using schema guidance and graph retrieval, are especially important.
large language modelsknowledge graphsstepwise reasoningschema-guided queryingCypher query languagemulti-step reasoningbenchmark datasetsgraph consistencyNeo4jreasoning accuracy
Authors
Xin Zhang, Yang Cao, Baoxing Wu, Kai Song, Siying Li
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.