Large language models and vision-language models used in medicine are usually frozen once deployed. That means they do not learn from the cases they solve, even though medicine is a field where new clinical evidence continually emerges.
A new arXiv preprint takes on this problem with a model-agnostic framework. The authors argue that deployment experience should feed back into the model, allowing expertise to evolve without being tied to a specific architecture. The abstract does not give details of the method, but the framing suggests a general solution for multimodal medical AI.
The paper is a preprint and has not yet been peer-reviewed. If the approach holds up, it could help medical AI systems stay current with evolving practice rather than remaining stuck at the moment of their training.