Towards Faithful and Explainable Large Language Models: A Knowledge Graph-Guided Framework for Reasoning Enhancement and Hallucination Mitigation

Abstract

In recent years, Large Language Models (LLMs) have achieved remarkable performance across various natural language processing tasks. However, they still face significant challenges in complex reasoning scenarios, such as ambiguous inference paths and frequent hallucinations, which hinder their adoption in high-stakes domains like medicine, law, and education. To address these challenges, this paper proposes a novel unified framework that integrates Knowledge Graphs (KGs) to enhance the reasoning ability and factual consistency of LLMs. The framework is designed around two core objectives: reasoning enhancement and hallucination mitigation, and adopts a three-layer architecture to achieve these goals. Each layer is designed to interact through a structured semantic path, providing controllable reasoning guidance and factual grounding. Compared to traditional approaches like Retrieval-Augmented Generation (RAG), the proposed method introduces explicit control over reasoning chains, thereby enhancing interpretability and output fidelity. This paper further analyzes current limitations in generalizability and engineering deployment, and identifies potential directions for future research, such as multimodal knowledge integration and prototype system development. The proposed framework offers a novel perspective on constructing trustworthy and controllable LLM-based systems, paving the way for their broader application in critical domains.