Generative Agents: Interactive Simulacra of Human Behavior
(2304.03442)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 LLMs with computational, interactive agents, this work introduces architectural and interaction patterns for enabling believable simulations of human behavior.
Overview
-
The paper discusses advanced computational agents that simulate human behavior dynamically and coherently over time.
-
It introduces a sophisticated agent architecture that incorporates memory, reflection, and planning for reactive and long-term decision-making.
-
Agents are implemented in the Smallville sandbox environment, which tests their ability to navigate and respond in a human-like manner.
-
Evaluations show that agents can recall experiences, exhibit self-knowledge, plan, and show emergent social dynamics like relationship formation.
-
The potential use of generative agents in interactive systems and ubiquitous computing is highlighted, alongside the need for ethical considerations.
Introduction
The notion of computational agents capable of reliably simulating human behaviors has long been a topic of interest across various domains, ranging from gaming to social science research. A pivotal development in this realm is the construction of agents that can engage in not just isolated actions, but believable, dynamic sequences that maintain coherence over time while reacting to complex environmental inputs.
Generative Agent Architecture
Key to these agents is an advanced architecture that allows for the interplay of memory, reflection, and planning. The concept of a memory stream stands central to this structure, providing a recorded log of past experiences in natural language. This log is strategically accessed to inform the agent's immediate and longer-term decision-making processes. Critical to this architecture is the retrieval system that selects pertinent memory fragments to guide present action. Moreover, reflections offer high-level inferences distilled from these memories, feeding back into the memory stream to craft even richer behavioral patterns over time.
Implementation and Sandbox Environment
Implementing the architecture reflects a mixture of challenges and strategic optimization. A prime optimization is the use of cached summaries that streamline prompt responses, vital for reducing computational load and improving real-time interactions. The Smallville sandbox environment serves as a test bed, where agents navigate and respond to a structure that mirrors a human-socialized setting. Actions within the Smallville are grounded through a transformation process that translates structured environmental data into natural language, aligning the artificial behavior more closely with human social behavior.
Evaluation
Through controlled evaluations comprising interviews probing an array of agent capabilities, it has been demonstrated that these generative agents reliably recall and synthesize past experiences to guide their behavior, displaying self-knowledge and the ability to plan. Their capacity to react to environmental changes and evolve subsequent plans was noted, with reflection playing a pivotal role in enriching behavioral complexity. In an end-to-end review, agents exhibited emergent social dynamics such as information diffusion, relationship formation, and coordinated activities—hallmarks of advanced, interactive agents.
Considerations and Future Directions
The advances in generative agent architecture emphasize the potential for agents to serve as proxies in interactive systems, enabling more personalized and effective use of technology in daily life. Moreover, this line of innovation raises implications for areas like ubiquitous computing, where generative agents can facilitate the design processes through human behavior modeling. Future work would benefit from exploring the robustness of these agents and optimizing their underlying models. Finally, ethical and societal considerations underscore the necessity of ensuring that generative agents are used in a manner that complements, rather than replaces, human interaction and input.