Tre Tomaszewski's Preliminary Exam
PhD student Tre Tomaszewski will present their preliminary proposal, "The hidden lives of social machines: How cognitive system designs influence agent behavior in synthetic societies." Tomaszewski's preliminary examination committee includes Associate Professor Jessie Chin (chair); Professor Dong Wang; ChengXiang Zhai, Donald Biggar Willett Professor for the Siebel School of Computing and Data Science; and Marcel Binz, research scientist and deputy head of the Institute for Human-Centered AI at Helmholtz Munich
Abstract
As synthetic agents equipped with large language models (LLMs) rapidly integrate into social systems, their capacity for coherent, adaptive interaction invites questions beyond mere behavioral mimicry. While recent efforts have focused on whether LLM agents can simulate human behavior, this study investigates a more foundational issue: how different configurations of memory and reasoning modules shape autonomous agent behavior over time. Rather than treating agents as proxies for human participants, this work frames them as evolving cognitive entities embedded within social environments.
Drawing on cognitive architectures, agent-based modeling, and natural language interaction, this study examines nine agent configurations derived from a 3-by-3 matrix of memory (contextual, vector-based, symbolic) and reasoning (reflexive, chain-of-thought, tree-of-thought) systems. Agents operate within a simulated microblogging platform, interacting through posts, replies, and follows across repeated runs. Key observables include interaction patterns, semantic convergence, belief dynamics, and emergent communication traits.
By systematically varying internal architectures and observing the resulting social phenomena, this research aims to clarify the causal links between cognition and sociality in autonomous agents. The findings are intended not only to inform future design of intelligent systems but also to provide a methodological foundation for studying nonhuman social cognition at scale.