NOVA: NOise-aware Verbal Confidence CAlibration for Robust Large Language Models in RAG Systems
NOVA introduces a framework to fix Large Language Models' overconfidence when using Retrieval-Augmented Generation (RAG) systems with noisy data. The method uses three specific rules—Conflict Independence, Noise Invariance, and Parametric Fallback—to teach models how to ignore bad evidence and rely on internal knowledge instead. This results in significantly more reliable answers and lower false confidence scores.