Researchers have developed CADRE, a framework for adapting medical vision-language models to new imaging modalities. This is important because updating deployed models can cause them to 'forget' previously learned information or drift towards shortcuts that compromise accuracy. CADRE uses low-rank adaptation and elastic weight consolidation to bound retained-competence loss and embedding drift. In experiments with breast cancer data across three different imaging types, CADRE achieved higher accuracy and lower forgetting rates compared to other adapting methods. --- Why it matters: This matters because it addresses a critical issue in medical AI: the need for models that can adapt to new situations without compromising their performance or safety. Engineers working on medical vision-language models will be interested in how CADRE's techniques can improve the stability and reliability of these systems. Source: https://arxiv.org/abs/2606.23487