SGR: A Stepwise Reasoning Framework for LLMs with External Subgraph Generation

2026-05-15

AI Summary

This work addresses the susceptibility of large language models (LLMs) to noise and factual inconsistencies in complex reasoning, which often undermines their logical rigor. To mitigate this limitation, the authors propose SGR, a novel framework that integrates query-focused external knowledge subgraph generation with multi-step reasoning. SGR explicitly guides and constrains the LLM’s intermediate reasoning steps by leveraging the structural properties of knowledge graphs. Through prompt engineering, the method constructs structured subgraphs and aggregates multiple reasoning paths to enhance both factual consistency and logical reliability in the final answer. Experimental results demonstrate that SGR significantly outperforms existing approaches across multiple benchmark datasets, achieving notable improvements in reasoning accuracy and factual trustworthiness.

Abstract

Large Language Models (LLMs) have demonstrated strong capabilities across diverse NLP applications, such as translation, text generation, and question answering. Nevertheless, they remain limited in complex settings that demand deep reasoning and logical inference. Since these models are trained on large-scale text corpora, their generation process may still introduce irrelevant, noisy, or factually inconsistent content. To mitigate this problem, we introduce SGR, a stepwise framework that enhances LLM reasoning through external subgraph generation. SGR builds query-specific subgraphs from external knowledge bases and uses their semantic structure to support multi-step inference. By grounding intermediate reasoning steps in structured external knowledge, the framework helps the model concentrate on relevant entities, relations, and supporting evidence. In particular, SGR first constructs a subgraph tailored to the input question. It then guides the model to reason progressively over the generated structure and combines multiple reasoning trajectories to obtain the final prediction. Experimental results across several benchmark datasets show that SGR achieves consistent improvements over competitive baselines, highlighting its value for improving both reasoning accuracy and factual reliability.

Problem

• reasoning
• logical inference
• factually inconsistent content
• complex reasoning
• LLMs

Innovation

• stepwise reasoning
• external subgraph generation
• knowledge grounding
• large language models
• multi-step inference