Rex Ying wins NSF CAREER Award to build more powerful and efficient AI
For his work on rethinking the mathematical foundations of artificial intelligence, Rex Ying has won a Faculty Early Career Development (CAREER) Award from the National Science Foundation (NSF). Ying, assistant professor of computer science, will use the five-year grant to develop a new class of AI models that process and represent data in flexible, curved mathematical spaces. It’s an approach that could make AI systems simultaneously more powerful, more efficient, and easier to understand.
The NSF CAREER Award is a prestigious honor for young faculty members, supporting the early career activities of teachers and scholars considered most likely to become the academic leaders of the future.
Most AI systems today, including the large language models behind tools like ChatGPT, store and process information using Euclidean geometry, the same flat, grid-like mathematics taught in high school. For many kinds of data, that works well enough. But relationships between entities like scientific documents, molecules, and social networks often have a naturally curved structure that flat geometry struggles to capture accurately. Forcing that kind of data into flat space introduces distortions, the same way flattening a globe onto a map warps the size of continents.
“Foundation models often use the default Euclidean geometry as the embedding space to train on data that requires a curved space to model the distribution,” Ying said. “If you give the model a space that actually fits the structure of the data, you can build smaller models that can achieve even better performance than current larger models.”
Ying's approach centers on developing a fully flexible model architecture in which different pieces of data can each be embedded in whatever curved space best fits their structure, rather than forcing everything into a single fixed geometry. The model is designed to work across multiple data types at once, integrating a curved-space graph model with large language models to handle both text and relational data together.
A distinct emphasis of the project is interpretability. Because curved spaces carry inherent geometric properties, the resulting models can offer clearer insight into why a particular output was produced, a meaningful advantage in fields where understanding a model's reasoning is as important as its answer.
"In many fields, it's not enough for the model to get the right answer,” Ying said. “A clinician or a scientist needs to understand why. The geometric properties of these spaces give us a natural way to dig under the hood."
The research has direct applications across medicine, biology, chemistry, and the study of how scientific knowledge develops and spreads, fields where relationships between entities are complex and the cost of a flawed model can be significant. The project involves collaborators at Yale School of Medicine, the Broad Institute of MIT and Harvard, Google Research, Snap Inc., and Amazon Web Services.
All models, benchmarks, and datasets produced through the project will be released publicly. The project also includes an educational component, with plans to engage graduate, undergraduate, and high school students through research experiences and new coursework.
More Details
Published Date
Jul 20, 2026
