Modeling Earth-Scale Human-Like Societies with One Billion Agents
Haoxiang Guan, Jiyan He, Liyang Fan, Zhenzhen Ren, Shaobin He, Xin Yu, Yuan Chen, Xueyin Xu, Shuxin Zheng, Yan Gao, Enhong Chen, Tie-Yan Liu, Zhen Liu
cs.MA, cs.AI, cs.CL, cs.CY, cs.SI
2025-06-07
Light Society models social processes as structured state transitions of LLM agents, using knowledge-distilled surrogates for most calls to reach one-billion-agent social simulation.
Simulating society on a computer has been stuck at two ends. Rule-based agent-based models (ABMs) use hand-coded behavior, too rigid to capture real human decisions. Recent LLM-powered agents behave realistically but are prohibitively expensive: prior systems need dozens of GPUs and weeks of compute to reach a million agents. Pushing toward Earth scale (a billion people) is computationally infeasible.
Light Society targets that scale bottleneck: keep the fidelity of LLM agents while reaching one billion of them.
The framework formalizes social simulation as a tuple: a timeline, agent state, environment state, events, an event queue, and a set of operations that drive state transitions. Agent state has three layers (a static demographic profile, internal status, external status); the environment has static and dynamic parts. All change runs through six LLM-powered operations: initialization, perception, policy, agent evolution, environment evolution, and update. This turns how people interact and how society shifts into controlled transitions of state.
The crux is saving compute. A mixture-of-models engine routes each operation: hard reasoning goes to a full LLM, routine decisions to a knowledge-distilled surrogate, with precomputed lookup tables as a backstop. The surrogate is an embedding-MLP distilled from LLM-generated interaction samples and replaces the large majority of calls. Four system optimizations stack on top: SHA-256-keyed LRU plus FAISS semantic caching to reuse prompts; billion-node graphs compressed into CSR format in HDF5; per-field numpy arrays for vectorized batching; and popping events of the same time and priority as concurrent batches.
Population data comes from World Values Survey (WVS) Wave 7, 96,125 cleaned profiles turned into natural-language personas carrying location, gender, age, education, employment, income, faith, ethnicity, and self-rated class, so agents have a real demographic structure.
Two flagship experiments. First, the Trust Game on gemini-2.0-flash-001. The single-round mean offer is 41/100, inside the 40% to 50% range typical of human play. Demographic patterns surface: upper-class agents send more than lower-class; postgraduate-educated agents trust more than those with little formal education; and the gap between the 16-to-34 and 55-plus groups widens and tightens its confidence interval as population grows. In a 10-round Ultimatum Game, the mean offer falls from 40.0 to 26.9, with the rejection rate peaking at 19.0% in round 7.
Second, opinion diffusion on a billion-agent BA network: the top 20% of high-degree nodes act as influencers, the bottom 80% as influencees. The real test is surrogate fidelity. Under the full-LLM baseline, per-round coefficient of variation (CV) across five runs sits at 0.0038% to 0.0060%; with 50% surrogate mixing, CV rises to the 0.14% to 0.34% range while trajectories stay qualitatively consistent. For surrogate selection they compared five architectures (Softmax Regression, LightGBM, MLP, Transformer, Qwen3-0.6B SLM): on the Martian-city topic macro-F1 clusters at 0.84 to 0.85, but the change-rate gap ranges from 8.6 pp (MLP) to 19.4 pp (Softmax Regression). They pick the embedding-MLP with the smallest gap (2.48%) over the highest-F1 checkpoint (7.14%).
On scale: prior comparable systems range from tens to roughly 10^7 agents; Light Society reaches 10^9.
The real contribution is engineering and methodological: a viable path to planet-scale LLM social simulation, and the central lever, distilled surrogates substituting for LLM calls, is transferable. Any large-scale simulation where each agent would invoke an LLM once but most calls are routine fits this mold. The surrogate-selection detail is especially worth remembering: pick surrogates by the distributional metric you ultimately care about, not by per-sample accuracy, because models with similar F1 can differ several-fold in distributional gap.
For researchers, this moves social simulation from demo-sized experiments to a scale where hypothesis testing is possible. The paper is also candid about which conclusions hold and which depend on model choice.
The authors flag several. First, behavioral baselines depend on the specific LLM: rerun the Trust Game with a different model and the send amounts drift, so absolute numbers cannot be read without the model context. Second, communication language moves results: the same topic run in Chinese versus French shifts stance-change rates by up to 4.7 percentage points. Third, surrogate selection carries risk; per-sample accuracy is not enough to judge distributional fidelity.
One thing to watch: the billion-agent opinion diffusion is driven largely by the surrogate, not by a billion genuine LLM calls. The surrogate was checked for qualitative consistency at small scale, but the paper gives little evidence on whether errors accumulate when extrapolated to a billion nodes over long horizons. And the framework is explicitly for studying social dynamics, not executing information operations, though a tool that simulates opinion diffusion at billion scale is plainly double-edged.