Michael Ye

Collective behavior emerges from interactions among individuals with diverse beliefs, experiences, and decision-making strategies, yet most AI simulations represent agents using manually specified personas or fixed rules. We propose a framework that learns latent representations of individual agent states directly from behavioral data and uses these representations to simulate interactions at scale, to model collective cognition across a large number of individual agents. Rather than hand-crafting agent characteristics, the model infers latent behavioral embeddings that capture persistent differences between individuals and conditions language generation on these learned states. We investigate whether populations of such learned agents produce more realistic patterns of collective narrative formation, consensus, polarization, and emergent social behavior than existing LLM-based simulations. This work aims to establish a data-driven foundation for modeling collective cognition and understanding how individual-level behavioral diversity gives rise to macroscopic social dynamics.

Michael Ye is a researcher at UCBerkeley, and formerly he's been a prompt-engineering, product, & GTM consultant for VC-backed AI startups.