Methods & resources
A benchmark designed to be inspected, not just ranked.
RADWORLD keeps model outputs, references, operation-specific metrics, and figure sources separate so every result can be traced to its protocol.
01
Protocol map
Assess → Read → Compare → Predict → Conclude → Advise. Each operation exposes a different part of a patient world model without pretending the metrics are interchangeable.
02
Fair comparisons
Patient-disjoint cohorts, frozen model releases, shared prompts, explicit interface eligibility, and operation-specific baselines.
03
Reader validation
Independent radiologist studies test whether report-derived targets remain recognizable under the rendered inputs given to models.
↓
Download and reproduce
The interactive site is generated from the same verified result files used by the manuscript.
Citation
RADWORLD
Do vision–language models have a world model of the patient?
@article{radworld2026,
title={RADWORLD: Do Vision--Language Models Have a World Model of the Patient?},
year={2026}
}