A new arXiv preprint, "When Evidence Vanishes: Distilling Visual-Evidence Dependency in Medical Vision-Language Models," highlights a subtle failure mode in compact medical VLMs. These smaller models can give correct clinical answers without actually looking at the image, instead leaning on the wording of the question or statistical priors learned during training.

The authors argue that standard knowledge distillation is not enough: if the teacher itself relies on shortcuts, copying its outputs can pass those shortcuts on to the student. The paper therefore focuses on distilling visual-evidence dependency, aiming to make the student model's answers genuinely contingent on the image content.

As the source is an abstract, the full method and evaluation details are not yet available. The key claim, however, is that correctness alone is a misleading metric for medical VLMs, and that evidence use must be explicitly trained and measured.