Med-StepBench: A Hierarchical Reasoning Framework for Evaluating Hallucinations in Medical Vision-Language Models
MedStepBench is a new large-scale benchmark for finding errors (hallucinations) in medical vision-language models by testing their reasoning across four sequential clinical steps. It mimics how a radiologist thinks, breaking down complex tasks like analyzing PET/CT scans into mapping anatomy, identifying lesions, characterizing features, and synthesizing a diagnosis.