CAREER: Learning Like Humans: Structuring Intelligence through Concept Graphs and Agent Collaboration

This project promotes the progress of science by drawing on principles of human cognition and social behavior to develop artificial intelligence systems that are more interpretable, generalizable, and collaborative. Humans organize knowledge through structured concepts and hierarchies, reason through relational and logical patterns, and coordinate through effective social interactions. In contrast, current artificial intelligence systems often struggle with tasks requiring such structured, human-like understanding. This project uses graph structures as a unifying framework to bridge human cognition and artificial intelligence. It uses artificial intelligence models to uncover how knowledge is structured and processed in the brain and in learner interactions, and in turn uses principles of human cognition to guide the development of artificial intelligence systems with improved reasoning, interpretability, and collaboration. The project has potential benefits for neuroscience by providing computational tools to study cognitive organization and brain activity, for education by enabling personalized learning tools that map students' knowledge organization and support individualized feedback, and for artificial intelligence by advancing systems that are better aligned with the structural and collaborative principles that support human intelligence. To pursue this goal, the project is organized around three interconnected research tasks. First, the project uses artificial intelligence models to generate concept graphs aligned with brain activity and learner interactions, to study how knowledge is organized and processed and to support applications in brain decoding and personalized learning. Second, the project investigates how human-like structural principles, such as logical equivariance and compositionality, can be recognized and acquired by large language models, and develops targeted training strategies to improve reasoning and generalization. Third, the project develops graph-based models for multi-agent collaboration that represent argument structure, roles, and interaction patterns inspired by human teamwork, using debate and joint decision-making as settings for studying coordination and adaptive collaboration. The project also integrates research with education and outreach through undergraduate research mentoring, a recurring hands-on tutorial series and regional summer camp for high school students, and cross-institutional graduate activities in graph learning. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2543142 | Program: 01002930DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT,01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Yujun Yan | Institution: Dartmouth College, HANOVER, NH | Award Amount: $460,590 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2543142 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2543142.html