Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection

In three linesNew Evidence Graph Consistency (EGC) method to detect hallucinations in RAG systems. Tested on 5,767 responses from 6 LLMs (Llama-2, GPT-4, GPT-3.5, Mistral-7B), it reveals that structural consistency of evidence graphs works inversely across model families, invalidating its use as a universal signal.

Summary generated by Claude — human-verified