NPCBench: A Clinical Apprenticeship Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma

NPCBench. A Clinical Apprenticeship Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma. By Pengkai Wang and fifteen others, from the Hong Kong Polytechnic University, Infix AI, and three other institutions. Published on the fifth of November.

Most medical benchmarks take clinical competence apart before testing it. They ask isolated factual questions, pose single-image tasks, or present a static case vignette, and they score each answer on its own. That reveals whether a model knows things. It reveals much less about whether a model can combine subspecialty knowledge with multimodal evidence, and follow one patient's condition as it changes over a course of care. Care does not happen one question at a time. The decisions for a patient have to fit together, and they have to keep fitting as the patient's state changes.

NPCBench is built to test that second ability, for one disease: nasopharyngeal carcinoma, or NPC. It frames evaluation as a clinical apprenticeship. The stages mirror how a specialist is trained, moving from guideline and foundational knowledge, to planning a case, to reviewing it over time, to CarePath integration, and finally to open-ended consultation. Each stage builds on the one before, the way a trainee's responsibilities grow. Along the way the benchmark asks three things of a model: whether it grounds its decisions in visual evidence, whether it interprets changes in patient state, and whether it meets clinical requirements when consulting.

There are four thousand seven hundred and seventy-seven component and consultation items. Alongside them sits a CarePath cohort of one hundred patients, carrying four hundred and seventeen scored diagnostic and care tasks. Stepwise tasks are linked to each patient's care trajectory, and the evaluation runs at the level of the whole episode. The question it asks is not whether a diagnosis is right, or whether a care answer is right, but whether both are right together, within the same pathway, for the same patient. A correct diagnosis followed by the wrong care step does not count as a success, and neither does the reverse.

Thirty models were evaluated, and the central observation is what the authors call a composition gap. Doing well on individual tasks does not ensure that the diagnostic and care answers for one patient come out jointly correct. Among the twenty-one multimodal models, strict CarePath success reaches at most twenty-three percent. Put another way, even the best of them meets the strict standard less than a quarter of the time. Paired controls and physician calibration were then used to look more closely at three things: evidence grounding, longitudinal revision, and how consultation answers are judged.

A model can pass each step and still fail the patient, and that is the distinction this benchmark is designed to expose. By moving from static question answering toward an apprenticeship-style assessment of clinical workflow, NPCBench offers a testbed for reasoning in oncology care that is grounded in evidence and coherent over time.