Federal Grant Award Summary

The University of Pennsylvania received a $354,579 Project Grant award from the National Science Foundation's Division of Computing and Communication Foundations under the Computer and Information Science and Engineering (CISE) program (CFDA 47.070), effective June 1, 2026 through May 31, 2031. This CAREER award funds research to develop a comprehensive science of artificial intelligence (AI) reliability by identifying failure modes in AI systems and designing principled methods to enhance their dependability in both autonomous and human-AI collaborative settings. The project delivers scientific foundations and open-source evaluation tools for safer AI deployment in critical applications such as medical diagnosis, while advancing understanding of why transformer-based AI systems learn brittle shortcuts rather than robust algorithms.

The research produces both theoretical and practical outputs across two primary areas: at the model level, the project develops interventions in training data, training algorithms, and inference mechanisms—including representative training sets for robust algorithmic behavior, sample-efficient multi-step reasoning methods, and formal safeguards against adversarial manipulation. At the interaction level, the research develops theory and protocols for effective human-AI collaboration under conditions of complementary information and imperfect alignment. Beyond technical deliverables, the award extends impact through expansion of the Learning Theory Alliance Mentorship Program, promotion of computationally efficient research, and development of new graduate and undergraduate courses.