How American Universities Are Using AI Models From Meta to Accelerate Medical and Scientific Breakthroughs

California Institute of Technology

Our expanded partnership with Caltech tests whether large language models can function as instruments for original research in mathematics and physical sciences, help formulate and prove conjectures, predict previously unseen outcomes, and generate hypotheses in addition to summarizing existing knowledge to push the boundaries of what AI can do in rigorous scientific domains. The work spans five disciplines and is organized around open problems. For this work, we’re giving researchers access to our frontier Muse models through our first-party API.

Projects and Research Areas

• Mathematics: Research-level problem benchmark
  • Current benchmarks test models on problems similar to those that the model was trained on, making it difficult to distinguish genuine reasoning from pattern matching.
  • Mathematician Dr. Tony Yue Yu is building MathDB Frontier, a community-curated database and benchmark built on MathDB, which contains more than 86,900 research-level open problems. The work is designed to evaluate whether language models can engage with mathematical research at the frontier: formalizing conjectures, searching for lemmas, and generating machine-checkable proofs — rather than pattern-matching on problems similar to training data. This will provide a research-grade evaluation and experimentation platform to measure and improve formal mathematical reasoning and verifiable AI outputs.
• Physics: Constrained recontextualization
  • Language models often generate descriptions of physical systems that sound plausible but violate fundamental principles or lack grounding in measurable quantities.
  • Physicist Dr. Nai-Chang Yeh is developing a framework that forces language models to iteratively construct candidate physical representations under explicit constraints. The outcome will test whether models can formulate physical systems rather than generate plausible-sounding but ungrounded descriptions.
• Chemistry: Reaction prediction from lab notebooks
  • Because language models are trained on vast amounts of published text, it’s difficult to know whether they are predicting reaction outcomes or recalling them from training.
  • Chemist Dr. Sarah E. Reisman is creating a multi-product reaction benchmark using unpublished electronic lab notebook records. This will provide a test of whether models can predict reaction outcomes on genuinely novel data, and help accelerate capabilities of AI-driven chemistry tools.
• Biology and Applied Mathematics: Automated theory construction
  • Biological systems often exhibit complex collective behavior that resists simple modeling, and current approaches require extensive human insight to formulate the right equations.
  • Biologist Dr. Matt Thomson is building an "AI theorist" that automatically constructs and tests predictive Partial Differential Equation-based mathematical models for biological active matter. The work will evaluate whether language models can generate testable theoretical predictions in biological systems.
• Computer Science: Formally verifiable scientific artifacts
  • When models generate code, proofs, or technical specifications, users currently have no reliable way to verify correctness beyond manual inspection.
  • Computer Scientist Dr. Anima Anandkumar is developing TorchLean, a Lean 4 framework and related tooling that allows language-model-driven agents to construct, test, and repair formally checkable scientific and machine learning artifacts. This could help establish whether model outputs in technical domains can be made verifiable rather than merely plausible.

The Ohio State University College of Medicine

The Ohio State University College of Medicine’s research initiatives span three critical research domains, pairing several of the latest models from Meta with the university's clinical, radiology, and biomedical informatics expertise.

Projects and Research Areas

• Identifying heart failure risk earlier
  • Heart failure with preserved ejection fraction (HFpEF) accounts for roughly half of all heart failure cases, but it is hard to recognize before symptoms appear — its symptoms overlap with obesity and other common conditions, so formal evaluation usually begins only once suspicion has already developed. Meanwhile, abdominal CT scans ordered for unrelated reasons already capture body composition, liver fat, muscle mass, and vascular calcification — the cardiometabolic picture that underlies HFpEF.
  • Dr. Yuan Xue, an AI researcher and Assistant Professor of Biomedical Informatics, is leading a project with cardiologist Dr. Yuchi Han, and AI researcher Dr. Xia Ning, to investigate how open-weight Muse models could help identify patients at risk of heart failure before clinical diagnosis. The team will develop a multimodal AI approach that combines routine abdominal CT scans with longitudinal medical records and ECG data when available to predict future risk of heart failure with preserved ejection fraction (HFpEF). By making better use of information already collected during routine care, the project aims to provide clinicians with evidence-based risk assessments that could support earlier cardiovascular evaluation and intervention.
• Drafting radiology reports and understanding when radiologists should trust them
  • Demand for chest X-rays has outpaced the number of radiologists available to read them. Thoracic radiologists at Ohio State interpret about 200 chest X-rays a day, and at some institutions that figure reaches 1,000. Most published report-generation models look only at the image, while a radiologist reading the same study also draws on the reason for the visit, patient history, and prior exams. A second, less obvious gap is measurement — when a language model grades another model's report, the resulting accuracy numbers can themselves be biased.
  • Imaging AI researcher Dr. Dufan Wu and radiologists Dr. Chiemezie Amadi and Dr. Paul Lewis will use Muse models to draft chest X-ray reports using multimodal input and comparison against prior exams. One aspect of the project is to statistically correct the bias in LLM-as-evaluator metrics using a small fixed set of human-annotated reports, so deployment decisions rest on reliable numbers. The other, led with psychologist Dr. Peter Kvam, is a reader study in a simulated clinical environment examining how radiologists actually rely on an AI draft — distinguishing under-reliance, where correct content is rejected, from over-reliance, where incorrect content is accepted — and correlating that behavior with the specific kinds of errors the model makes.
• Clinician-guided agents for Alzheimer's diagnosis
  • Alzheimer's disease and related dementias affect an estimated 7.4 million Americans aged 65 and older, yet diagnosis often follows years of decline. The diagnostic process is inherently sequential: a clinician searches the record, weighs competing hypotheses, chooses the next most informative test, and revises. Electronic health records are not built to support that loop — they cannot easily reconstruct how a patient's cognition has changed over time or suggest which assessment would most reduce uncertainty.
  • Dr. Xia Ning will lead a project, together with Dr. Yuan Xue and cognitive neurologist Dr. Douglas Scharre, to develop a clinician-in-the-loop multi-agent system built on open-weight Muse models as a locally deployable baseline. The multi-agent system will aim to analyze and assess patient data related to cognitive history, reason over clinical evidence, identify diagnostic uncertainties, and recommend the next most informative clinical actions — all while keeping clinicians at the center of decision-making. The work draws on curated records from Ohio State’s Memory Disorders Clinic, which sees over 5,600 patient visits a year.

Louisiana State University

The Louisiana State University research initiatives span multiple campuses and include clinical, surgical, and methodological research. The projects selected for support will work with a range of open-weight AI models from Meta, including computer vision models like SAM and DINO, as well as wearables from Meta.

Projects and Research Areas

• Multi-institutional neurosurgical outcomes registry
  • Building registries that track patient outcomes after neurosurgical procedures manually is labor intensive and inconsistent, limiting the data available for quality improvement and research.
  • Dr. Subhajit Chakrabarty is establishing a standardized neurosurgical outcomes registry across multiple institutions, using open-weight AI models from Meta for automated registry curation from clinical text. The project will use SAM for segmentation and DINO, a self-supervised model for images, to conduct radiomics feature extraction from medical images. The intent is for the registry to track outcomes across several clinical areas and to support real-time decision-making using Muse Glimmer and Muse Spark Open — our open-weight models that can run on institutional infrastructure to handle sensitive patient data on site.
• Non-contact monitoring for infant tracheostomy
  • When tracheostomy tubes become dislodged in infant patients, it’s a life-threatening emergency compounded by the fact that babies with tracheostomies can’t cry out for help. And continuous direct observation by trained staff isn’t always possible.
  • Dr. Michael Dunham is leading the Eyes-On project, which is developing a computer vision system to continuously detect tracheostomy tube displacement and signs of distress in infants who cannot reliably vocalize. The system is exploring usage of SAM models for tracking and is designed for integration with bedside cameras and edge computing. A proof-of-concept has been published, with more work to come.
• AI coaching for medical training in virtual reality
  • Effective training on medical procedures traditionally requires expert supervision, which limits how much practice trainees can get and how quickly they can receive feedback.
  • Emergency medicine physician Dr. Peter DeBlieux is working on an AI-powered coaching platform to guide healthcare professionals through procedural training in virtual reality. Built on Meta’s open-weight AI models with retrieval-augmented generation (RAG) and running on Meta headsets, the system offers real-time, hands-free coaching inside immersive simulations — with plans to integrate SeamlessM4T to support interaction in nearly 100 languages. A working prototype exists and Dr. DeBlieux looks forward to sharing user studies in the future.
• Privacy-preserving pose estimation validation
  • Validating camera-free, radio frequency-based pose estimation systems currently requires expensive motion-capture equipment.
  • Dr. Subhajit Chakrabarty is aiming to validate camera-free radio frequency (RF) based pose estimation systems, which can monitor patients through walls without a camera in the room. The work will use open-weight SAM models to generate reference skeletal labels from synchronized RGB images, removing the need for expensive motion-capture equipment. This could help enable privacy-preserving fall detection and health monitoring.
• Causal AI for head and neck cancer
  • Standard predictive models often function as black boxes, generating predictions without explaining which factors are driving them or which might be changed to improve outcomes.
  • Machine learning researcher Dr. Supratik Mukhopadhyay aims to build causally grounded, clinician-interpretable AI for head and neck cancer that predicts and explains outcomes driving morbidity, cost, and mortality. The approach uses language models from Meta for extracting variables from clinical text and requires each model to identify modifiable drivers behind predictions. Four retrospective cohorts have been completed with manuscripts in preparation.

Indiana University

The Indiana University research initiatives span healthcare delivery in resource-limited settings, cognitive science, and assistive technology. The projects share a focus on real-world deployment constraints: devices that must work without reliable internet connectivity, interactions that must accommodate users with cognitive differences, and research that requires understanding how humans actually learn. The projects selected for support will work with open-weight Muse models, as well as computer vision models such as Segment Anything Model (SAM), V-JEPA, and DINO, and SeamlessM4T, a multilingual communication model from Meta.

Projects and Research Areas

• Multilingual AI for maternal and newborn care
  • Kangaroo Mother Care is a method of newborn care involving skin-to-skin contact that reduces infant mortality. Effective implementation requires guidance and monitoring that trained healthcare workers are not always available to provide.
  • Dr. Saptarshi Purkayastha leads NeoSmartML+, which aims to add on-device AI-powered caregiver guidance, multilingual voice and text interaction using SeamlessM4T and NLLB-200, and SAM 3-based image assessments for monitoring infant positioning. All components are designed to run offline on low-cost Android devices, addressing settings in low- and middle-income countries where reliable connectivity and specialized medical staff are often unavailable.
• Multimodal AI for early language development
  • AI systems typically require far more training data than human children need to learn similar concepts, suggesting that something important about human learning remains poorly understood.
  • Computer Scientist Dr. David Crandall is studying how children learn words and visual concepts from natural everyday experience. Crandall uses DINO and V-JEPA, which are models designed to learn visual representations without explicit labels, to analyze what children see and hear as they acquire new words. The goal is to identify principles of efficient learning that could inform more data-efficient AI training while advancing understanding of human cognitive development.
• Conversational AI for dementia care
  • People with dementia often experience communication difficulties that standard conversational AI systems handle poorly, leading to frustration rather than support.
  • This project, led by Dr. Crandall, aims to develop multimodal AI methods for analyzing conversations involving older adults living with dementia, and to use this understanding to help conversational agents — including social robots — adapt their interactions to individual users. The research tests whether AI systems can be made sensitive to the communication patterns of people with cognitive differences.

UTHealth Houston

UTHealth Houston is using Muse Glimmer, one of our smallest open-weight models, to support projects that focus on clinical agent workflows and agentic benchmarking in healthcare settings. Future collaboration could explore AI-enabled drug discovery and development, including antibody design and optimization.

Projects and Research Areas

• Local deployment and agentic benchmarking
  • Deploying large models locally introduces computational constraints that cloud deployment avoids. It remains unclear how these constraints affect model usefulness for clinical tasks.
  • Biomedical Informaticist Dr. Xiaoqian Jiang is deploying and optimizing Muse Glimmer within UTHealth Houston’s secure GPU infrastructure for privacy-compliant local inference for drug discovery and bioinformatics research. The work will evaluate computational efficiency, latency, and token reduction strategies, and will benchmark multi-step planning and tool-calling accuracy across healthcare interfaces.
  • Early internal testing has compared workflows across different model configurations, though these remain preliminary tests.
• Clinical agent reasoning
  • For glioblastoma patients, relevant information is scattered across record systems in different formats, including imaging, pathology, treatment records, and clinical notes. This makes it difficult for clinicians to maintain a comprehensive view of disease progression.
  • The ClinAgent GBM project from Dr. Jiang aims to adapt agentic reasoning workflows to handle complex multimodal electronic health records for glioblastoma patients. The work involves imaging follow-up summaries from paired MRI visits and will evaluate multi-agent reasoning to support brain imaging segmentation.
  • Pilot work with clinical partners is underway, with early tests across a sample of 65 patient cases. These represent initial workflow validation rather than clinical outcome measures.
• Point-of-care prototyping
  • Clinicians spend substantial time on documentation that could otherwise go to patient care. This research project, fully led by UTHealth Houston with funding from Meta, aims to explore initial proof-of-concept workflows for hands-free healthcare applications. The goal is to improve efficiencies and decrease lost time due to handoffs by using Meta Glasses and Muse Glimmer, which enable local environments to preserve patient privacy.

What’s Next

We’re excited to support these five universities as they use our models to accelerate the important work they’re doing. We look forward to sharing more in the future.