A team of Chinese researchers has developed a framework called Light Society capable of simulating societies of more than one billion AI agents, each endowed with realistic demographic profiles and the capacity to display human-like social behaviors. The project, described in a paper submitted to arXiv, represents an approximately 100-fold improvement over previous agent-based simulations, which typically reached their limits at around 10 million agents due to computational constraints. The work arrives as China continues to prioritize large-scale AI research under its national technology strategy, channeling sustained investment into academic computing infrastructure and advanced model development.

Technical Architecture Behind Billion-Scale Simulation

The research team includes contributors from several of China's leading academic institutions: the University of Science and Technology of China, Tsinghua University, and Fudan University. Lead researchers Haoxiang Guan and Tie-Yan Liu built Light Society around two core technical innovations that make billion-scale simulation feasible.

The first is a mixture-of-models engine that distributes cognitive workload across multiple AI model types, rather than routing every agent decision through a single large language model. The second innovation involves knowledge-distilled surrogate models — smaller, faster models trained to replicate the behavior of larger, more sophisticated ones. Together, these approaches allow the framework to maintain realistic behavioral fidelity without the computational resources that would otherwise make billion-agent simulations impractical.

Each simulated agent is assigned a realistic profile derived from the World Values Survey, specifically Wave 7 of the long-running global research project. That survey covers beliefs, values, and demographic characteristics across dozens of countries, providing each simulated individual with a grounded identity rather than a randomly assigned one.

Flagship Experiments: Trust Games and Opinion Diffusion

The researchers conducted two flagship experiments to demonstrate Light Society's capabilities. The first was a series of Trust Games, a well-established experimental design from behavioral economics in which participants decide how much money to send to a stranger, knowing the stranger can either reciprocate or keep the funds. Running this across a billion-node network reveals patterns of cooperation and defection that cannot surface in smaller simulations.

The second experiment focused on opinion diffusion, modeling how ideas, beliefs, and sentiments propagate through massive social networks. The original source material explicitly notes that Light Society can be used to test methods of influencing public opinion — a capability with direct relevance to real-world questions about how narratives spread and evolve on platforms like WeChat, X, and other global social networks, where influence operations and viral content dynamics have become subjects of intense academic and policy interest.

How Light Society Advances Agent-Based Modeling

Agent-based modeling has existed for decades, with researchers applying it to study traffic patterns, disease transmission, and market behavior. During the COVID-19 pandemic, agent-based simulations informed policy decisions on lockdowns and mobility restrictions, though those models operated on far simpler agent logic than what LLM-powered systems now offer. Historically, these models depended on simple rule-based agents following predetermined scripts.

The integration of large language models into agent-based simulation is a relatively recent development. LLM-powered agents can engage in nuanced decision-making, adapt to novel situations, and produce emergent behaviors. Earlier LLM-based social simulations peaked at roughly 10 million agents before compute costs became prohibitive. Light Society's architectural choices — particularly the surrogate model approach — broke through that ceiling by two orders of magnitude.

The researchers have positioned the framework as enabling earth-scale "in silico" experiments, signaling an ambitious vision: using simulation as a laboratory for understanding human civilization itself. The concept parallels the broader research movement toward digital twins — virtual replicas of physical or social systems — which has drawn interest from governments and technology companies seeking to model urban planning, public health responses, and economic policy before implementing changes in the real world.