Researchers Introduce MedProb to Decode Medical Answers from Frozen VLMs

A new scientific preprint introduces MedProb, a lightweight probing framework designed for medical visual question answering (Med-VQA). The method predicts multiple-choice answers directly from frozen vision-language model (VLM) representations without relying on free-text generation.

Evaluated across the PATH-VQA, SLAKE, and VQA-RAD benchmarks, MedProb recovered significantly more answer-relevant signals than standard prompting, outperforming dedicated medical VLMs and complex multi-agent pipelines. The probing technique also narrowed the apparent performance gap between small and large models, indicating that compact architectures store more usable signal than generation-based metrics typically reveal.

Testing across 14 matched pairs of general-purpose and specialized VLMs showed that domain-specific adaptation does not consistently enhance linear decodability. Furthermore, the authors demonstrated that MedProb can be extended to open-ended tasks via a rejection-sampling scoring approach.