Reproduce “Generative Agents: Interactive Simulacra of Human Behavior” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein

Generative agents use memory streams, reflection, and hierarchical plans to carry experience into actions in a simulated town. In Smallville, a seeded Valentine’s Day party intention, alongside Maria’s crush on Klaus, led to invitations and a gathering, with five of 12 invitees attending. Human-rated interviews favored the full architecture (μ=29.89) over no-reflection (μ=26.88), measuring believability, not real-world human behavior or agency.

Key insights

  • Augmenting Large Language Models (LLMs) with architectural components for dynamic memory, recursive reflection, and hierarchical planning is crucial for creating truly coherent and believable artificial agents.
  • Each architectural component (memory stream, reflection, planning) independently contributes to an agent's believability, with the full architecture significantly outperforming ablated versions.
  • Complex social behaviors, such as information diffusion, relationship formation, and coordinated group activities, can emerge autonomously in artificial societies composed of generative agents without explicit scripting.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.