From Data to Theory with Foundation Models

Foundation models are transforming science by offering powerful ways to generalize across domains, scale learning, and generate meaningful structure from data. In this talk, I explore how these models can be used not only to describe patterns in data but also to construct executable, generalizable theories of complex systems. I focus on human behavior as a test case, showing how tools from statistics and the behavioral sciences can be combined with large language models to reverse-engineer cognitive processes. I will present a cognitive foundation model trained on over 200 canonical psychology experiments that predicts human behavior across tasks and generalizes to entirely new ones. This model can be used for scientific discovery via methods of regret minimization. I will also showcase results from a model that goes one step further, mapping from behavioral data directly to the code of underlying cognitive models, allowing for transparent, testable theory generation, and outline a vision of a model that maps directly from data to the best hypothesis generating observations. Taken together, this line of work outlines a path toward scientific theory discovery powered by foundation models, using behavioral science as a proving ground for a broader approach to modeling the world.