Building Human-Like Social Intelligence in Machines

Abstract

Social intelligence allows people to collaborate seamlessly and effectively at scale. As AI systems increasingly interact with people as assistants and collaborators, building social intelligence into machines has become a critical challenge. Yet today's frontier AI systems still lack core aspects of human social intelligence. This dissertation proposes a way to build such systems by scaling up cognitive models of human social cognition with foundation models, including large language models (LLMs) and vision-language models (VLMs). In particular, I develop a neurosymbolic architecture in which LLMs and VLMs parse unstructured inputs and orchestrate ad hoc cognitive models, such as Bayesian Theory of Mind (BToM), to support flexible and interpretable social inference. Across behavioral experiments, simulations, and real-world human--robot studies, I show that this approach showcases human-like social judgments and supports more effective human--AI collaboration, often outperforming larger end-to-end neural models. I conclude by discussing how these neurosymbolic agents can be further scaled to interact with and assist people in real-world settings.