Research
- Two Minds, One ModelNotes on HealthGPT: how one medical model learns to read scans and draw them, without the two jobs fighting.
- Building an RL Environment for Medical QATurning physician-written answers into a reward a language model can learn from.
- Grand Rounds for MachinesThree benchmarks for what medical exams miss: reasoning across a patient’s full record, changing course when new evidence arrives, and pushing back when a clinician is wrong.
- Teaching a Model to Read Brain ScansNotes on CortexMAE, our ICML 2026 paper: flat maps, masked autoencoders and an open benchmark for fMRI.