Most existing medical benchmarks decompose clinical competence into isolated factual queries, single‑image tasks, or static case vignettes, scoring each answer separately but offering little insight into whether a model can fuse subspecialty knowledge with multimodal evidence and follow a patient throughout a care course. To address this gap we introduce NPCBench, a clinical apprenticeship benchmark for nasopharyngeal carcinoma. The benchmark mirrors specialist training from guideline and foundational learning, case planning, longitudinal review, CarePath integration, to open‑ended consultation. It comprises 4,777 component and consultation items and a 100‑patient CarePath cohort with 417 scored diagnostic and care tasks. Evaluation is episode‑level: a model must produce jointly correct diagnostic and care answers within the same patient pathway, testing grounding in visual evidence, interpretation of patient‑state changes, and compliance with clinical consultation requirements. Across 30 evaluated models we observe a composition gap—strong performance on individual tasks does not guarantee jointly correct answers for the same patient. Among 21 multimodal models, strict CarePath success peaks at 23%. Paired controls and physician calibration further probe evidence grounding, longitudinal revision, and consultation judgment. In sum, NPCBench expands medical AI assessment beyond static QA to an apprenticeship‑oriented evaluation of clinical workflows, offering a testbed for evidence‑based and temporally coherent reasoning in oncology care.
Review: NPCBench raises the bar for AI in real‑world clinical pathways, highlighting the need for multimodal evidence integration and patient‑centric longitudinal reasoning.