A team of Chinese researchers has built a framework called Light Society that can simulate societies of over one billion AI agents, each equipped with realistic demographic profiles and the ability to exhibit human-like social behaviors. The project, detailed in a paper submitted to arXiv, represents a roughly 100x leap over previous agent-based simulations, which typically hit a wall at around 10 million agents due to computational bottlenecks.
How you simulate a billion people
The research team spans some of China’s most elite institutions: the University of Science and Technology of China, Tsinghua University, and Fudan University. Lead researchers Haoxiang Guan and Tie-Yan Liu designed Light Society around two core technical innovations that make billion-scale simulation feasible.
The first is a mixture-of-models engine, which essentially distributes the cognitive workload across multiple AI model types rather than routing every agent decision through a single massive language model. The second innovation involves knowledge-distilled surrogate models. In plain terms, the system trains smaller, faster models to mimic the behavior of larger, more sophisticated ones. This lets the framework maintain realistic behavioral fidelity without requiring the computational resources that would normally make billion-agent simulations impossible.
