NOVA: NOise-aware Verbal Confidence CAlibration for Robust Large Language Models in RAG Systems

The article introduces NOVA, a noise-aware calibration framework designed to enhance the confidence assessment of large language models (LLMs) in retrieval-augmented generation (RAG) systems. NOVA employs NOise-Aware Verbal Confidence CAlibration Rules and is fine-tuned on approximately 2,000 HotpotQA examples, leading to improvements in expected calibration error (ECE) scores by 10.9% in-domain and 8.0% out-of-domain. This advancement is significant for practitioners as it addresses the challenge of overconfidence in LLMs when faced with noisy or contradictory evidence, thereby improving the reliability of model outputs in critical applications.