LUMINA: Detecting Hallucinations in RAG System with Context-Knowledge Signals
This addresses hallucinations in RAG-based LLMs, which is a critical issue for reliable AI applications, though it is incremental as it builds on existing utilization-based detection approaches.
The paper tackles the problem of hallucinations in Retrieval-Augmented Generation (RAG) systems, proposing LUMINA to detect them using context-knowledge signals, and it achieves up to +13% AUROC improvement over prior methods on benchmarks like HalluRAG.
Retrieval-Augmented Generation (RAG) aims to mitigate hallucinations in large language models (LLMs) by grounding responses in retrieved documents. Yet, RAG-based LLMs still hallucinate even when provided with correct and sufficient context. A growing line of work suggests that this stems from an imbalance between how models use external context and their internal knowledge, and several approaches have attempted to quantify these signals for hallucination detection. However, existing methods require extensive hyperparameter tuning, limiting their generalizability. We propose LUMINA, a novel framework that detects hallucinations in RAG systems through context-knowledge signals: external context utilization is quantified via distributional distance, while internal knowledge utilization is measured by tracking how predicted tokens evolve across transformer layers. We further introduce a framework for statistically validating these measurements. Experiments on common RAG hallucination benchmarks and four open-source LLMs show that LUMINA achieves consistently high AUROC and AUPRC scores, outperforming prior utilization-based methods by up to +13% AUROC on HalluRAG. Moreover, LUMINA remains robust under relaxed assumptions about retrieval quality and model matching, offering both effectiveness and practicality.