Overconfidence and Calibration in Medical VQA: Empirical Findings and Hallucination-Aware Mitigation
The research addressed overconfidence in medical AI by introducing Hallucination-Aware Calibration (HAC). While standard calibration methods improved reliability, they didn't boost prediction accuracy. HAC incorporated vision-grounded hallucination signals to enhance both calibration and discriminative quality, significantly improving AUROC, especially for open-ended questions without losing reliability.