CORTEX: A Structured Reasoning Benchmark for Trustworthy 3D Chest CT MLLMs

A research team has introduced CORTEX, a structured reasoning benchmark designed to make multimodal large language models more trustworthy when interpreting 3D chest CT scans, according to a preprint submitted in June 2026 [1]. The benchmark, formally named Clinically Organized Reasoning and sTructured EXplanation, addresses a gap in existing medical imaging datasets. Current chest CT question-answering resources often reduce expert radiology reports to answer-only pairs, discarding the reasoning that connects visual findings to diagnostic conclusions and omitting patient history [2]. As a result, the field has lacked both the structured supervision needed to train reasoning-capable models and a protocol to verify their outputs [2]. CORTEX restores this missing reasoning by constructing a four-stage diagnostic trace for each question, mirroring a radiologist’s workflow: task understanding, visual observation, diagnostic reasoning, and answer synthesis [3]. The traces are generated using frontier large language models with broad medical and general-domain knowledge, then filtered and verified through a stage-level evaluation protocol that combines automated rubric scoring with expert radiologist review [4]. Both the reasoning structure and the evaluation rubrics were designed in close collaboration with clinicians [5]. Built on CT-RATE, a large, publicly available chest CT dataset that previously lacked reasoning annotations, CORTEX comprises 76,177 validated reasoning traces spanning open-ended visual question answering, closed-ended visual question answering, and report generation [2]. The dataset and evaluation code will be made publicly available upon acceptance, the authors state [1]. The work was submitted to arXiv on June 25, 2026, by Anees Ur Rehman Hashmi and colleagues [1]. arXiv, an open-access repository that hosts preprints across physics, computer science, and related fields, does not itself peer-review submissions but provides rapid dissemination of research findings [9].