← ML Research Wiki / 2304.03442

Generative Agents: Interactive Simulacra of Human Behavior

Joon Sung Park [email protected], Joseph C O'brien [email protected], Carrie J Cai [email protected], Meredith Ringel Morris [email protected], M.SPercy Liang [email protected], Michael S Bernstein, Bernstein, Generative, Stanford University Stanford USA, Stanford University Stanford USA, Google Research Mountain View CAUSA, Google DeepMind Seattle WAUSA, Stanford University Stanford USA, Stanford University Stanford USA, UIST '23 October 29-November 12023San FranciscoCAUSA, UIST '23 October 29-November 12023San FranciscoCAUSA, 2023San FranciscoCAUSA (2023)

Paper Information
arXiv ID
Venue
ACM Symposium on User Interface Software and Technology
Domain
artificial intelligence, human-computer interaction, social computing, virtual environments
SOTA Claim
Yes
Reproducibility
8/10

Abstract

Figure1: Generative agents are believable simulacra of human behavior for interactive applications.In this work, we demonstrate generative agents by populating a sandbox environment, reminiscent of The Sims, with twenty-five agents.Users can observe and intervene as agents plan their days, share news, form relationships, and coordinate group activities.

Summary

This paper introduces generative agents designed to simulate believable human behavior in an interactive environment, inspired by sandbox games like The Sims. The authors propose an architecture enabling agents to store, retrieve, reflect on memories, and interact with each other dynamically. The architecture consists of a memory stream, reflection, and planning components that collectively support believable behavior across various contexts. The authors conducted evaluations to demonstrate that these agents can engage in individual and emergent social behaviors. The findings suggest that generative agents can effectively manage complexities of human interactions and create immersive environments beneficial for various applications like social role-play and virtual environment simulations. The paper also discusses potential societal impacts, applications, future improvements, and ethical considerations surrounding the use of generative agents.

Methods

This paper employs the following methods:

  • Generative Models
  • Natural Language Processing

Models Used

  • ChatGPT

Datasets

The following datasets were used in this research:

  • None specified

Evaluation Metrics

  • None specified

Results

  • Agents produce believable behaviors
  • Emergent social dynamics observed
  • Successful planning and execution of events like a Valentine's Day party

Limitations

The authors identified the following limitations:

  • None specified

Technical Requirements

  • Number of GPUs: None specified
  • GPU Type: None specified

Keywords

generative agents interactive simulation human behavior modeling social dynamics large language models

Papers Using Similar Methods

External Resources