Detecting Hallucinations for Large Language Model-based Knowledge Graph Reasoning

The article presents LUCID, a novel hallucination detection method specifically designed for large language model (LLM)-based knowledge graph (KG) reasoning frameworks. LUCID integrates LLM attention scores, KG semantics, and structural information using a graph neural network, addressing the limitations of existing methods that overlook KG structure. Experimental results demonstrate that LUCID outperforms 15 baseline methods across nine benchmark datasets, highlighting its importance for enhancing the reliability of LLM outputs in knowledge-driven applications.