Social World Models

Paul Liang (MIT) https://simons.berkeley.edu/talks/pau... Topics in Intelligence: World Models and Social Reasoning Humans navigate complex social environments by continuously predicting the behavior, intentions, and beliefs of others. Inspired by this ability, Social World Models extend existing physical world models to the social domain, capturing not only physical dynamics but also the latent mental states and strategies of interacting agents. This talk will explore how these models can be learned, represented, and leveraged to build towards socially intelligent AI. We will present our recent work in multimodal modeling of social behaviors, social reasoning benchmarks and models, and long-horizon self-evolving social agents.