Few-shot adaptation of medical vision-language models (VLMs) normally assumes each class has at least a few labeled examples. In practice, low-prevalence classes can be left with no labeled support at all, which is the "zero-support" setting this paper targets. The authors of the arXiv preprint describe this as a realistic scenario for medical VLM deployment.
The paper notes that an existing training-free solver, SS-Text+, still functions in this zero-support regime. The authors then propose Gated Prototype Transfer, a method aimed directly at this gap, though the available abstract text does not specify its internal workings.
Because this digest draws on only the arXiv announcement, there are no independent results or competing claims to compare; SS-Text+ is described as operational in the zero-support setting, and the new method is introduced as the paper's response to it.