From static answers to medical AI that can reason, adapt, and be tested safely.
Medicine is not a single question with a single answer. Evidence arrives over time, decisions change the next observation, and people operate with incomplete information and real constraints. Our research develops the benchmarks, reasoning systems, and interactive environments needed to study AI under those conditions.
Benchmark
What should a model know, show, and withstand before we trust its output?
Final-answer accuracy can conceal weak evidence, brittle reasoning, and failures that matter clinically. We design benchmarks and evaluation methods that surface these gaps—from hidden patient concerns to long-form answers and reasoning under compression.
Reason
How should medical AI reason when the patient—and the evidence—keeps changing?
Clinical decisions unfold across encounters, tests, interventions, and uncertainty. We develop systems that retrieve from long patient histories, retain useful diagnostic experience, internalize clinical guidance, and update their reasoning as new evidence arrives.
Simulate
Can we test clinical agents in worlds that react to what they do?
Static datasets omit the feedback loops of care. We study simulated patients, clinical teams, and care environments in which information is partial, actions have consequences, and agents must observe, decide, communicate, and reassess over time.
What ties these programs together
We prioritize evaluations that reveal clinically meaningful failure, not only aggregate performance.
We model medicine as an evolving process rather than a collection of isolated prompts.
We turn guidelines, evidence, and experience into reasoning systems that can support concrete decisions.
Have a hard medical-AI problem?
We are interested in clinically grounded collaborations, shared evaluation challenges, and partnerships that connect research to real care settings.