[Paper Review] Generative Agents: Interactive Simulacra of Human Behavior

Joon-Sung Park, Joseph C. O'Brien+4 more|arXiv (Cornell University)|Apr 7, 2023

Artificial Intelligence in Games163 citations

TL;DR

The paper introduces Generative Agents, a memory-driven architecture that augments large language models with memory streams, reflection, and planning to simulate believable long-term human behavior in an interactive environment. It demonstrates a 25-agent Smallville sandbox with emergent social dynamics and evaluates the architecture through ablation and end-to-end studies.

ABSTRACT

Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools. In this paper, we introduce generative agents--computational software agents that simulate believable human behavior. Generative agents wake up, cook breakfast, and head to work; artists paint, while authors write; they form opinions, notice each other, and initiate conversations; they remember and reflect on days past as they plan the next day. To enable generative agents, we describe an architecture that extends a large language model to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior. We instantiate generative agents to populate an interactive sandbox environment inspired by The Sims, where end users can interact with a small town of twenty five agents using natural language. In an evaluation, these generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentine's Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time. We demonstrate through ablation that the components of our agent architecture--observation, planning, and reflection--each contribute critically to the believability of agent behavior. By fusing large language models with computational, interactive agents, this work introduces architectural and interaction patterns for enabling believable simulations of human behavior.

Motivation & Objective

• Motivate the creation of believable proxies of human behavior for interactive systems and virtual environments.
• Propose an architecture that extends large language models with long-term memory and reasoning for coherent, evolving behavior.
• Showcase emergent social dynamics and daily-life routines in a small-town sandbox.
• Evaluate the causal impact of memory, reflection, and planning components on believability.

Proposed method

• Introduce memory stream as a comprehensive, natural-language long-term memory of agent experiences.
• Develop a memory retrieval model that balances relevance, recency, and importance to inform behavior.
• Add reflection to synthesize memories into higher-level inferences about self and others.
• Implement planning to translate reflections and current context into high-level plans and granular actions.
• Integrate memory, reflection, and planning with a large language model to drive believable behavior.
• Demonstrate in a 25-agent sandbox (Smallville) how agents remember, plan, act, and interact via natural language.

Experimental results

Research questions

• RQ1 How can an agent architecture maintain long-term coherence and memory in open-world simulations?
• RQ2 What is the causal impact of memory, reflection, and planning components on believability and behavior quality?
• RQ3 Can a small-language-model-powered agent population generate emergent social dynamics and coordinated activities over time?
• RQ4 What are the limitations and failure modes when using LLMs for believable agent behavior in interactive environments?

Key findings

• Generative agents produce believable individual behaviors and emergent social dynamics in an open-world sandbox.
• Each of memory retrieval, reflection, and planning contributes critically to believability, as shown by ablation experiments.
• Agents can coordinate complex tasks starting from minimal seed prompts (e.g., planning a party) without hand-coding all behaviors.
• End-to-end evaluation reveals emergent phenomena such as information diffusion and relationship memory within the agent society.
• The architecture reveals failure modes such as memory retrieval errors and over-formal language tendencies, informing future improvements.