# Generative Agents: Interactive Simulacra of Human Behavior

**Authors:** Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein  
**arXiv:** [2304.03442](https://arxiv.org/abs/2304.03442)  
**Date:** 2023-04-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.03442v2/

## Key points

- Generative agents are computational software agents that simulate believable human behavior
- These agents can perform various activities such as waking up, cooking breakfast, going to work, painting, writing, forming opinions, noticing others, initiating conversations, and remembering past events
- The authors propose an architecture that extends a large language model to enable the behavior of generative agents
- This architecture allows the agent to store a complete record of its experiences using natural language and synthesize those memories into higher-level reflections for planning behavior
- Generative agents are instantiated in an interactive sandbox environment inspired by The Sims video game
- Users can interact with a small town populated by twenty-five generative agents using natural language commands
- Evaluation shows that these generative agents exhibit believable individual and emergent social behaviors
- Agents autonomously spread invitations for a Valentine's Day party over two days and make new acquaintances while coordinating their actions to show up together at the right time for the party
- Ablation experiments demonstrate that observation, planning, and reflection are critical components contributing to believability of agent behavior
- Combining large language models with computational interactive agents like generative agents can create believable simulations of human behavior with applications in immersive environments, rehearsal spaces for interpersonal communication, and prototyping tools.

## Summary

In this paper titled "Generative Agents: Interactive Simulacra of Human Behavior," authors Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein introduce generative agents as computational software agents that simulate believable human behavior. These generative agents are capable of performing various activities such as waking up, cooking breakfast, going to work, painting, writing, forming opinions, noticing others, initiating conversations and remembering past events and planning for the future. To enable the behavior of these generative agents the authors propose an architecture that extends a large language model. This architecture allows the agent to store a complete record of its experiences using natural language and synthesize those memories over time into higher-level reflections which can be dynamically retrieved by the agent to plan its behavior. The authors instantiate generative agents in an interactive sandbox environment inspired by The Sims video game where users can interact with a small town populated by twenty-five generative agents using natural language commands. Through evaluation it is demonstrated that these generative agents exhibit believable individual and emergent social behaviors such as autonomously spreading invitations for a Valentine's Day party over two days or making new acquaintances and asking each other out on dates for the same event while coordinating their actions to show up together at the right time for the party. The authors further analyze their agent architecture through ablation experiments which demonstrate that observation, planning and reflection are critical components contributing to believability of agent behavior. By combining large language models with computational interactive agents like generative agents described in this work the authors introduce architectural and interaction patterns for creating believable simulations of human behavior with implications for various applications ranging from immersive environments to rehearsal spaces for interpersonal communication and prototyping tools.