HaloGuard: A Practical Hybrid Framework for Hallucination Detection in Retrieval-Augmented Generation Systems : Combining Multi-Judge Agreement, Uncertainty Quantification, and Lexical Features Without Embedding Dependencies
The issue of hallucination detection for Retrieval-Augmented Generation (RAG) systems is an ongoing reliability challenge that arises with the scale up of LLM deployment. We introduce the HaloGuard: a systems-integration study that integrates three types of signals, multi-judge LLM consensus, uncertainty quantification...