Title: Boosting LLM Reasoning Through Dynamic External Subgraph Creation

Abstract: While large language models (LLMs) excel at natural language generation and various downstream reasoning applications, they frequently encounter difficulties maintaining logical consistency, factual accuracy, and interpretability when tackling complex, multi-step problems. To overcome these challenges, this study introduces SGR, a framework designed to enhance stepwise reasoning by connecting LLMs with external knowledge graphs via the generation of query-specific subgraphs. The process begins with SGR identifying key entities, relations, and constraints from an input question to build a structured schema. It then employs schema-guided queries to extract compact subgraphs from a knowledge base. These subgraphs supply explicit relational evidence, steering the language model through a sequential reasoning process. Furthermore, SGR merges direct reasoning via Cypher queries with collaborative integration techniques, enabling the validation and aggregation of candidate answers from diverse reasoning paths based on both graph consistency and model confidence. Evaluations on benchmark datasets such as CWQ, WebQSP, GrailQA, and KQA Pro reveal that SGR surpasses standard prompting methods and several knowledge-enhanced baselines in terms of Hits@1 scores and overall reasoning accuracy. Ablation analyses confirm that both the schema guidance mechanism and Neo4j-based retrieval are essential components for the framework’s success. Collectively, these findings suggest that the dynamic creation of external subgraphs significantly enhances the robustness, accuracy, and interpretability of reasoning performed by LLMs.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC