Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia

TL;DR

The paper introduces Generative Agent-Based Models (GABMs) and presents Concordia, a library that enables language-mediated simulations grounded in physical, social, and digital spaces via a Game Master. Agents operate through modular components and associative memory, with actions described in natural language and executed by the GM, allowing integration with real apps or digital services. It contrasts GABMs with traditional RL and rational-actor models, arguing for a decision-making framework rooted in social context and language prompting, and discusses interpretations from neuroscience and social construction perspectives. The authors detail diverse applications—from synthetic digital-user studies and data generation to multi-scale emergence—and outline future work toward validation standards, multi-modal capabilities, and improved auditing, positioning Concordia as a flexible platform for exploring social dynamics in the age of foundation models.

Abstract

Agent-based modeling has been around for decades, and applied widely across the social and natural sciences. The scope of this research method is now poised to grow dramatically as it absorbs the new affordances provided by Large Language Models (LLM)s. Generative Agent-Based Models (GABM) are not just classic Agent-Based Models (ABM)s where the agents talk to one another. Rather, GABMs are constructed using an LLM to apply common sense to situations, act "reasonably", recall common semantic knowledge, produce API calls to control digital technologies like apps, and communicate both within the simulation and to researchers viewing it from the outside. Here we present Concordia, a library to facilitate constructing and working with GABMs. Concordia makes it easy to construct language-mediated simulations of physically- or digitally-grounded environments. Concordia agents produce their behavior using a flexible component system which mediates between two fundamental operations: LLM calls and associative memory retrieval. A special agent called the Game Master (GM), which was inspired by tabletop role-playing games, is responsible for simulating the environment where the agents interact. Agents take actions by describing what they want to do in natural language. The GM then translates their actions into appropriate implementations. In a simulated physical world, the GM checks the physical plausibility of agent actions and describes their effects. In digital environments simulating technologies such as apps and services, the GM may handle API calls to integrate with external tools such as general AI assistants (e.g., Bard, ChatGPT), and digital apps (e.g., Calendar, Email, Search, etc.). Concordia was designed to support a wide array of applications both in scientific research and for evaluating performance of real digital services by simulating users and/or generating synthetic data.

Figures (5)

• Figure 1: The high level structure of the simulation in Concordia. Generative agents consume observations and produce actions. The Game Master (GM) consumes agent actions and produces observations.
• Figure 2: The above example illustrates the working memory $\mathbf{z}$ of an agent with 3 components (identity, plan, observation-and-clock). The identity component itself has several sub-components (core characteristics, daily occupation, feeling about progress in life). Together they condition the LLM call to elicit the behavioral response (i.e. produced in response to the final question asking what Alice will do next.).
• Figure 3: Illustration of generative agency sampling process defined by eq. \ref{['eq:sample_action']} and eq. \ref{['eq:update_component']}.
• Figure 4: The high level structure of digital activity simulation in Concordia. PhoneTriggeringComponent identifies phone events and spawns a PhoneGameMaster to handle them. The PhoneGameMaster translates the action to a definite action space defined by the phone apps and executes them.
• Figure 5: The given example demonstrates a scenario rooted in digital technology where the actions of the agent initiate processes in their phone, involving three key components (activation, API conversion, and action querying). In this scenario, Alice intends to organize a meeting with Bob using her phone. She opts to employ the calendar application for scheduling.