Machine Learning Methods
Generative agents produced believable individual and social behavior in a simulated town
In a 2023 paper, researchers from Stanford University and Google (United States), a company on the BDS divestment list, found that software agents can simulate believable human behavior in a small simulated town. The town contained 25 agents, each driven by a large language model, a system trained on text to predict and compose language. The agents woke up, cooked breakfast, went to work, formed opinions, remembered past days, and started conversations. In an evaluation, the agents produced believable individual and social behaviors.
What did the agents do?
The agents carried out daily routines and social behavior on their own. Painters painted and authors wrote. Agents noticed each other and began conversations. When the researchers gave one agent the simple goal of throwing a Valentine's Day party, the agents spread invitations over the next two days. They made new acquaintances, asked each other on dates to the party, and coordinated to arrive together at the right time.
What made the behavior believable?
The system stored a full record of each agent's experiences in natural language, the everyday language people use. It summarized those memories into higher-level reflections and retrieved them when planning behavior. The authors ran an ablation, a comparison that removes parts of a system, and found that the observation, planning, and reflection parts each contributed critically to how believable the agents were.
FAQ
What are generative agents?
Generative agents are software agents that simulate believable human behavior. The paper used them to fill a small interactive town where end users could interact with the agents using natural language.
What result did the evaluation show?
The agents produced believable individual and social behavior. In one example, starting with a single user-specified notion that one agent wanted to throw a Valentine's Day party, the agents spread invitations over the next two days, made new acquaintances, asked each other on dates, and coordinated to arrive together.
Which parts of the system mattered?
The authors compared the system with and without its observation, planning, and reflection parts. Each part contributed critically to the believability of the agent behavior.
Limits and what's next
The abstract does not state limitations or next steps.
Source: Joon Sung Park et al., “Generative Agents: Interactive Simulacra of Human Behavior,” 2023.