Leaderboard
Follow one patient across three or more real radiology reports, tracking how each finding changes over time and between scans, and write up the full history.
Read the findings of a real radiology report and name the radiologist’s diagnosis in your own words, with no options given, scored on both the top answer and the top three.
Read a real radiology report with the impression removed and reach the same clinical conclusion the radiologist wrote.
Read a real breast imaging report with the conclusion removed and assign the BI-RADS category, where every case is a probably-benign BI-RADS 3 that is easy to over- or under-call.
Answer clinical questions about real radiology reports that span many organ systems, testing whether a model draws the right conclusion from the findings.
Answer clinical questions about real orthopedic radiology reports, drawing the conclusion a reader would from the bone and joint findings.
Choose the right diagnosis or next workup step from a real radiology report, using only questions that can’t be answered without reading the report.
The same diagnosis and workup task as AXIOM-Diagnosis over nearly twice as many questions, for a tighter comparison between models.
Methodology
Real records, clinicians in the loop, intervals on every score.
01
De-identified reports, not exam questions or synthetic vignettes.
02
Every question is developed with clinicians and validity-checked.
03
Questions answerable without the record are removed.
04
Every score carries a 95% interval. Close scores are ties.
Contact us. We run the full suite privately, including pre-release checkpoints.
Yes. The data behind AXIOM is available to labs under a data agreement.
Models that helped select a set’s questions are not ranked on it.
A gap of a few points is often noise.
As new frontier models ship. Current results are from November 2026.
Unlocking the World's Healthcare Data.