Project Grant 2551608

The National Science Foundation Division of Computing and Communication Foundations awarded the University of Pennsylvania one million dollars on September 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop uncertainty quantification methods for artificial intelligence decision-making systems.

The project will create model-free uncertainty quantification approaches centered on conformal prediction and calibration methods that provide valid coverage guarantees when prediction sets are used within decision-making pipelines. The work addresses three technical limitations of current approaches: designing algorithms with coverage guarantees conditional on downstream actions chosen using a prediction set, including settings where the action depends on the uncertainty estimate itself; optimizing the usefulness of prediction sets, including their size and value to risk-aware decision makers, while maintaining valid coverage constraints; and extending methods to distribution-shift and other challenging settings. The research focuses on making AI systems more dependable, efficient, and useful in high-stakes environments where mistakes carry significant cost or safety implications, such as autonomous vehicles, drones, and coding agents. Educational materials, software, and workshops developed through the project will be made broadly available.

The award obligates one million dollars with a period of performance from September 1, 2026, through August 31, 2030. Work is performed in Philadelphia, Pennsylvania.