Joon Sung Park

Joon Sung Park is a computer scientist and co-founder and CEO of Simile. His collaborative research introduced generative agents: language-model actors that remember experiences, reflect on them, and plan behavior. He subsequently helped develop simulations grounded in interviews and surveys of real people, testing whether agents could reproduce the responses of the individuals they represent.[4][5][7]

The research behind Simile began during Park's Stanford doctorate, which he recalls starting in 2020. Earlier work on social prototypes supplied a setting for testing simulated interactions before the team built a town of agents with persistent memories.[1][8][5]

Park trained as an oil painter, initially studying in a hyper-realism studio under Hae-Kwang Jeong. He recalls treating art as a prospective profession and becoming interested in computation as a medium while in college. He earned a B.A. in computer science from Swarthmore College, an M.S. from the University of Illinois at Urbana–Champaign, and a Ph.D. in computer science from Stanford.[3][1]

At Stanford, Park contributed to the interaction section of the 2021 report On the Opportunities and Risks of Foundation Models. In his August 2026 interview, he recalled asking with Michael Bernstein and Percy Liang which application would matter most when viewed from ten years in the future. They considered both simulation and personalized assistants. Park favored simulation partly from fascination with creating worlds and partly from the belief that useful automation first needed accurate models of its users. His hypothetical assistant ordering Hawaiian pizza for someone who dislikes pineapple illustrated how completing a task can still fail the person.[3][1]

In the 2022 Social Simulacra collaboration, Park worked with Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Liang, and Bernstein on populated prototypes for social computing systems. Designers supplied community goals, rules, and member personas; a language model generated thousands of members and interactions, including posts, replies, and antisocial behavior. The technique let designers explore changes and interventions before recruiting a large community. The team's evaluations found that designers could use these simulations to refine their designs.[8]

Park coauthored the April 2023 Generative Agents: Interactive Simulacra of Human Behavior paper with Joseph C. O'Brien, Cai, Morris, Liang, and Bernstein. Their architecture recorded experiences in natural language, synthesized higher-level reflections, and dynamically retrieved memories to plan behavior. Park later recalled that the team chose text memory over a knowledge graph or bespoke model because language models could already understand and reason over text. He also acknowledged that retrieving and interpreting very large memory collections remained difficult.[5][1]

The team placed 25 agents in a town inspired by The Sims. Starting with one user-specified intention to hold a Valentine's Day party, agents spread invitations over two simulated days, made acquaintances, arranged dates, and coordinated attendance. The paper reported that observation, planning, and reflection each contributed critically to behavioral believability. This tested whether the agents behaved plausibly; the later individual-simulation work tested fidelity to particular real people. Generative Agents won a UIST 2023 Best Paper Award.[5][6][3]

The November 2024 Generative Agent Simulations of 1,000 People paper, coauthored by Park and eight collaborators, modeled 1,052 real individuals using qualitative interviews about their lives. It reported that agents reproduced General Social Survey responses 85% as accurately as participants reproduced their own answers two weeks later. That figure was a relative survey-consistency benchmark, rather than 85% accuracy across all human behavior. The paper also reported personality and experimental-outcome predictions and reduced accuracy biases across racial and ideological groups compared with demographic descriptions.[6]

The June 2026 revision, LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals, compared agents built from two-hour interviews, structured surveys, or both. On held-out General Social Survey items, it reported 83%, 82%, and 86% of participants' own two-week test-retest consistency, respectively, versus 74% for demographics-only agents. Combining interviews and surveys performed best, though gains over either alone were modest. The study tested prediction across outcomes without task-specific training data.[7]

Park's June 2025 doctoral dissertation was titled Generative Agent Simulations of Human Behavior. In June 2026, he announced that it had received Stanford's Arthur Samuel Award for Best Ph.D. Dissertation in Computer Science.[3]

Park founded Simile with Bernstein, Liang, and Lainie Yallen. In August 2026, he argued that prompting alone could not supply all the behavioral knowledge missing from general-purpose models: training also needed to capture what he called “social physics.” His objective was to reproduce people's preferences, biases, and mistakes. As he put it, “if I make some mistakes, the model has to make the same mistake.” He illustrated this with his preference for walking home from Stanford to think, even when faster transport was available.[1]

Park described Simile's inputs as interviews, observational behavior such as transactions, and experimental data about causal mechanisms. He emphasized participant consent and decisions with real stakes—for example, delivering items purchased during an experimental shopping task. He said the company trained distinct population-level and individual-level models. He also distinguished an open-science follow-up model trained using randomized controlled trials from the company's commercial offering, saying that research model was not commercially served.[1]

For Park, simulation's value lies in comparing interventions and understanding paths to an outcome. He described Wealthfront's product-testing work as extending beyond questions to multimodal inputs and agents using mockups or websites. His broader aspiration was to make people's perspectives available in decisions affecting them when repeated consultation would otherwise be costly or impractical.[1]

In the same interview, Park reported preliminary signs that more human data and compute produced predictable improvements in simulation performance. Simulating eight billion people was his longer-term vision. He also offered a personal comparison with his earlier artistic work: “The best paintings teach you something deep about the subject that you’re trying to represent.” He saw simulation similarly—as an imperfect representation that could reveal important features of individuals and their collective lives.[1]