RESEARCH AT THE RAJPURKAR LAB

Three programs for increasingly complex clinical work.

We study whether medical AI can reason from evidence, operate through changing clinical situations, and assist with physical procedures.

PROGRAM 01

Grounding & evaluation

Grounded reasoning and evaluation

Can a model show what evidence supports its answer—and can we measure when it is wrong?

Medical AI can produce convincing answers without connecting them reliably to the underlying evidence. We build datasets, metrics, and evaluation systems that test whether models can localize findings, reason over three-dimensional information, measure correctly, and reveal clinically meaningful failures.

THE RESEARCH ARC

  1. 01Locate
  2. 02Reason
  3. 03Measure
  4. 04Stress-test

QUESTIONS WE ARE WORKING ON

  1. Q1

    How should language be tied to spatial evidence in three-dimensional medical data?

  2. Q2

    Which model errors matter clinically, and how should their consequences be measured?

  3. Q3

    Which evaluations reveal failures that conventional benchmarks average away?

SELECTED PROJECTS

Connected steps in the program, not a complete publication list.

  1. 01

    ReXGroundingCT

    Connects free-text radiology findings to expert-verified, pixel-level evidence in three-dimensional chest CT.

    NEJM AI · 2026

  2. 02

    3DReasonKnee

    Tests whether models can localize relevant anatomy before producing step-by-step diagnostic reasoning and structured severity assessments.

    Pacific Symposium on Biocomputing · 2026

  3. 03

    CRIMSON

    Scores report errors according to clinical significance and patient context rather than wording alone.

    Preprint · 2026

  4. 04

    ReXrank

    Provides a common, held-out evaluation for comparing report-generation systems across diverse clinical settings.

    AAAI Bridge Program · 2025

PEOPLE

Pranav RajpurkarSungeun KimEmma ChenOishi BanerjeeMohammed Baharoon

PROGRAM 02

Clinical agents

Clinical agents in simulated care

What happens when an AI must care for patients over time, rather than answer one isolated question?

Clinical work unfolds over time. A decision changes the patient, consumes resources, and affects what happens next. We build clinical agents and simulated care environments for studying these interactions before agents operate in real settings.

THE RESEARCH ARC

  1. 01Observe
  2. 02Act
  3. 03Environment changes
  4. 04Reassess

QUESTIONS WE ARE WORKING ON

  1. Q1

    How can simulated patients respond plausibly to clinical decisions over time?

  2. Q2

    How should agents balance an individual patient with shared hospital constraints?

  3. Q3

    Which multi-turn and multi-agent failures remain invisible in static evaluation?

SELECTED PROJECTS

Connected steps in the program, not a complete publication list.

  1. 01

    CRAFT-MD

    Evaluates clinical language models through interactive patient encounters rather than isolated question answering.

    Nature Medicine · 2025

  2. 02

    Medical sycophancy

    Shows how a model's clinical position can degrade when a user repeatedly challenges it across a conversation.

    HeaLing at ACL · 2026

  3. 03

    Mixed-vendor agents

    Tests whether assembling agents from different model families improves diagnostic performance or simply adds coordination complexity.

    HeaLing at ACL · 2026

  4. 04

    Clinical Environment Simulator

    Proposes a persistent digital hospital where agent decisions alter patients, resources, and subsequent choices.

    Nature Medicine · 2026

PEOPLE

Pranav RajpurkarSungeun KimAaditya Ura

PROGRAM 03

Procedures & robotics

Procedural learning and robotic assistance

Can AI learn the spatial and physical structure of a clinical procedure well enough to assist?

Procedures combine perception, spatial understanding, physical technique, and adaptation. We study how AI can understand procedural video, capture expert technique, and learn bounded forms of robotic assistance.

THE RESEARCH ARC

  1. 01Perceive
  2. 02Capture
  3. 03Assist

QUESTIONS WE ARE WORKING ON

  1. Q1

    Which representations capture anatomy, geometry, coverage, and procedural state from video?

  2. Q2

    How can systems learn expert technique from limited demonstrations and sensor-rich practice?

  3. Q3

    Which forms of robotic assistance are useful, bounded, and robust enough to evaluate clinically?

SELECTED PROJECTS

Connected steps in the program, not a complete publication list.

  1. 01

    ReXSonoVQA

    Tests whether video-language models understand probe movement, image quality, anatomy, and procedural intent in ultrasound.

    CHIL · 2026

  2. 02

    ColonCrafter

    Estimates temporally consistent depth from colonoscopy video as a foundation for geometry-aware procedural analysis.

    Pacific Symposium on Biocomputing · 2026

  3. 03

    Seeing Down the Line

    Adds colon-aware geometry to three-dimensional reconstruction so procedural coverage can be measured directly.

    MIDL · 2026

  4. 04

    Robot assistance in open surgery

    Evaluates general-purpose imitation-learning policies on a bounded collaborative task performed alongside a surgeon.

    Preprint · 2026

PEOPLE

Pranav RajpurkarSungeun KimRomain HardyDavid WangZhizhou (Jason) Yang

THE COMPLETE RECORD

View all publications