Adversarial Evaluation of Large Language Models for Patient-Friendly Radiology Report Summarization
A scalable adversarial evaluation framework that uses controlled term substitutions and impression removals across intensity levels to expose clinically meaningful failure modes, including missed anatomical modifiers, altered diagnostic certainty, and omitted severity descriptors is developed.
Magdalini Paschali, Asad Aali, Arogya Koirala et al.
· 0 citations