Predictive Entropy Links Calibration and Paraphrase Sensitivity in Medical Vision-Language Models
Published in arXiv preprint, 2026
We show that predictive entropy connects two problems that are usually studied separately: calibration and paraphrase sensitivity in medical Vision-Language Models (VLMs). Samples where the model is poorly calibrated are also the samples most likely to flip when the question is reworded, which suggests a single uncertainty signal can flag both risks before deployment.
Recommended citation: Sadanandan, B., & Behzadan, V. (2026). Predictive Entropy Links Calibration and Paraphrase Sensitivity in Medical Vision-Language Models. arXiv preprint arXiv:2604.08941.
Download Paper