Personalized Foundation Model for Social Simulation

Presented by Zhongyu Wei

With the rapid advancement of large language models, large-scale social simulation frameworks—such as Generative Agents, AgentSociety, ElectionSim, and SocioVerse—have emerged as powerful testbeds for governance, public policy, and market analysis. However, the credibility of macro-level emergent phenomena fundamentally hinges on how faithfully agents model individual human behavior and preserve authentic population diversity. Off-the-shelf foundation models are not yet reliable simulators: uniform assistant alignment and post-training 'taxes' tend to homogenize personas and erase long-tail human variation. This talk presents a roadmap toward foundation models built explicitly for individual behavior simulation, progressing from compact parametric personalization methods like CURP and PHF to unified behavioral models. In particular, we introduce Socio-Foundation, a model organized around the five-dimensional FONTS capability framework and trained via hierarchical distillation to achieve robust, diverse, and consistent individual-level simulation across complex social environments.