Modeling Earth-Scale Human-Like Societies with One Billion Agents

A research team has built Light Society, an agent-based simulation framework that models human-like societies with more than one billion agents, according to a paper posted on arXiv [1]. The system uses large language models to drive agent behavior and is calibrated with real-world demographic data [2]. The framework, detailed in a preprint last revised in June 2026, formalizes social processes as structured transitions of agent and environment states governed by LLM-powered simulation operations [1][2]. A mixture-of-models engine that combines full LLMs with distilled surrogates enables the system to reach the billion-agent scale while maintaining behavioral fidelity [2]. The agents are grounded in demographic profiles drawn from the World Values Survey, allowing researchers to run large-scale Trust Games and opinion-diffusion experiments that the authors say demonstrate both high fidelity and efficiency [2]. Traditional agent-based models have long been used to study social dynamics but are limited by simplified agent behaviors [2]. The emergence of generative AI, particularly large language models based on the transformer architecture, has made it possible to build agents with more sophisticated reasoning [7]. However, scaling such models to populations that approach the size of real societies has remained a challenge [2]. Light Society addresses that gap through joint algorithmic and system optimizations, the authors write [2]. The paper appeared on arXiv, the open-access e-print repository that hosts preprints across physics, computer science, and related fields [11]. As of late 2024, arXiv was receiving about 24,000 submissions per month and had passed two million total articles at the end of 2021 [11]. The Light Society manuscript was first submitted in June 2025 as a 1,475 KB file and updated a year later with a 3,923 KB version [1]. The work arrives amid broader efforts to build AI systems that can generalize across tasks. Artificial general intelligence, a hypothetical type of AI that matches or surpasses human capabilities across virtually all cognitive tasks, remains a stated goal of several major technology companies [5]. Simulation platforms that model collective human behavior at planetary scale could provide a testbed for studying emergent social phenomena, though they also raise questions about alignment — the challenge of ensuring AI systems pursue intended objectives rather than unintended ones [6]. Light Society is presented as a practical foundation for hypothesis testing and the study of emergent collective behaviors at planetary scale [2]. The authors have not yet described plans for public release or external validation of the framework.