MIRAGE: Noise-Aware Bayesian Calibration with Mutual Information for Reliable RAG

Retrieval-Augmented Generation (RAG) applications, from high-stakes medical diagnosis and financial decision-making to content recommendation, require reliable uncertainty calibration for trustworthy outputs. However, existing RAG exhibits poor calibration due to noisy retrieval content and multi-stage error propagation. Current methods assume high-quality contexts, ignoring retrieval noise’s impact on uncertainty estimation. We propose MIRAGE, a robust uncertainty calibration framework for RAG. MIRAGE employs mutual information to quantify noisy context impact on model robustness for retrieval uncertainty estimation, then applies a Bayesian calibrator to capture interactions between retrieval uncertainty and raw uncertainty for structured calibration. As a plug-in module, MIRAGE can seamlessly integrate with existing RAG architectures without structural modifications. Extensive experiments across general, finance, biomedical, and movie domains demonstrate improved uncertainty calibration and correlation with diverse RAG architectures.