Reducing LLM Hallucinations with Knowledge Graphs: A Survey

Abstract

Large language models (LLMs) have achieved remarkable performance in natural language processing tasks, yet they still suffer from hallucinations, i.e., the generation of factually incorrect, inconsistent, or nonsensical information. Hallucinations undermine trust and limit the applicability of LLM-based systems especially in high-risk scenarios, such as personal healthcare or legal support. To address this issue, knowledge graphs (KGs) have increasingly been adopted as external sources of trustworthy knowledge for detecting and mitigating hallucinations throughout the LLM lifecycle. In recent years, the number of proposed approaches for reducing hallucinations with KG support have rapidly increased, generating a vast landscape of solutions, emerging trends, and opportunities for future research. In this survey, we present a literature review on reducing LLM hallucination with KGs, presenting a comprehensive taxonomy to organize existing methods according to the stage at which the KG intervention occurs: during model training, at inference time, or after generation. We discuss approaches for hallucination detection and mitigation, as well as KG-based benchmarks for evaluating and assessing LLM hallucinations. Finally, we highlight open challenges (e.g., KG incompleteness, cultural and language coverage, efficiency) and future research directions towards the development of more reliable, explainable, and factuality-aware KG-supported LLMs.