MedObvious: VLMs Fail at Basic Clinical Triage That Humans Find Trivial

arXiv·medium signal

Khan et al. expose a Medical Moravec's Paradox in Vision Language Models: VLMs that score well on complex medical reasoning benchmarks catastrophically fail on basic clinical triage tasks that any first-year medical student handles easily. The benchmark reveals that fluency in generating medical text masks fundamental failures in visual clinical assessment. Cautionary finding for anyone deploying VLMs in healthcare or safety-critical visual classification.