RAG hallucination detector models evidence as a graph

A new paper introduces Evidence Graph Consistency, a hallucination-detection framework for retrieval-augmented generation. Instead of comparing answers to retrieved passages with flat similarity scores, it builds a local graph of evidence and answer claims.

The authors evaluate the method on RAGTruth across six LLMs and report that graph-based consistency features expose model-family differences in hallucination behavior.

The work is useful because RAG systems often look grounded while still misusing or stitching evidence together incorrectly. Structural checks may catch failures that simpler citation or similarity metrics miss.