What Happens When You Give a Billion AI Agents an Opinion and Let Them Talk
- David Borish
- 2 minutes ago
- 6 min read
Most large language model social simulations top out around a million agents, and even that scale typically requires dozens of GPUs running for weeks. A team spanning Zhongguancun Academy, the Zhongguancun Institute of Artificial Intelligence, Shenzhen University, Shanghai University of Finance and Economics, and Tsinghua University set out to push past that ceiling by three orders of magnitude. Their framework, Light Society, ran a social network simulation with over one billion LLM-driven agents, and the researchers used it to study two classic problems in social science: trust and reciprocity, and opinion propagation through influence networks.
The paper, posted to arXiv in June 2025 and revised in June 2026, argues that the bottleneck in LLM-based social simulation has never really been about behavioral realism. Language models can already generate agents with personality traits, memory, and context-sensitive emotional responses. The bottleneck is computational: every additional agent means more LLM calls, more storage, more latency. Light Society's contribution is a set of engineering choices aimed at collapsing that cost curve without collapsing the fidelity of what gets simulated.
How Light Society Works
The framework formalizes a simulation as a tuple of agents, an environment, an event queue, and a set of six LLM-powered operations that govern how the system evolves. Each agent carries three layers of state: a static profile that never changes (demographics, personality), an internal status that evolves over time (memory, beliefs, emotion), and an external status that reflects observable conditions like location or social ties. The environment similarly splits into static components, such as spatial layout, and dynamic ones, such as weather or disaster events.
Everything that happens in the simulation, an agent sending a message, forming a relationship, relocating, gets represented as an event with a timestamp and a priority level, sitting in a queue that resolves interactions in temporally consistent order. Six operations drive the whole system: initialization from a seed dataset, perception, decision-making, agent state evolution, environment evolution, and an update step that applies events back into the simulation.
None of that architecture alone explains how the team reached one billion agents. The efficiency comes from a separate stack of optimizations layered on top. Semantic prompt caching converts each LLM query into a vector representation and searches a vector database for structurally similar prompts that have already been answered, avoiding redundant calls.
Knowledge distillation trains compact surrogate models on live simulation data to approximate LLM outputs for routine decisions, and those surrogates get periodically retrained to track behavioral drift. A mixture-of-models architecture then routes each simulation operation to whichever backend, full LLM, distilled surrogate, or a different provider's model, fits the fidelity and resource tradeoff at that moment. At the system level, a distributed execution engine and compressed graph representations handle state updates and neighbor queries across a billion-node network, while the event queue supports batching similar events together and running independent events concurrently.
Testing Trust Across a Billion-Agent Population
The first demonstration used the World Values Survey Wave 7 dataset to build 96,125 agent profiles spanning country, age, education, income, and subjective social class, then dropped those profiles into a classic trust game. A trustor decides how much of a $10 endowment to send to an anonymous trustee, the amount triples in transit, and the trustee decides how much to send back. Every profile played the trustor role once and the trustee role separately under ten different received amounts, using two models, Gemini 2.0 Flash and GPT-4.1 nano.
Trust behavior tracked socioeconomic status in both models. Agents assigned upper-class identities and higher education levels sent noticeably larger amounts than those assigned lower-class or less-educated profiles. GPT-4.1 nano behaved more conservatively than Gemini 2.0 Flash across most demographic groups, sending smaller amounts on average even though the directional patterns matched. Trustee return amounts rose roughly linearly with what they received, and trustors ended up with positive net profit at most transfer levels, a pattern the researchers describe as evidence that the agents reproduce reciprocal norms rather than pure self-interest.
The more interesting result showed up when the team varied population size directly. The gap in trust between younger (16 to 34) and older (55-plus) trustors widened as the simulated population grew larger, with confidence intervals narrowing across eight trials per population size. In other words, patterns that look noisy or ambiguous in a small sample became sharper and more stable as the agent count scaled up, which is close to the opposite of what happens in most statistical sampling, where larger samples just narrow the error bars around an already-visible effect. Here, scale appeared to reveal structure that smaller runs didn't have the statistical power to show clearly.
Watching Opinions Spread Through a Billion-Node Network
The second study is where the billion-agent claim gets stress-tested. The team generated a synthetic Barabási–Albert network of one billion nodes, a scale-free topology meant to mirror the uneven connectivity of real social networks, where a small number of highly connected people carry disproportionate influence. Each agent drew its identity from a pool of 10,000 World Values Survey profiles. The top 20 percent of nodes by connection count were designated influencers, and the remaining 80 percent influencees.
All agents were given the same seed statement, "AI automation will lead to mass unemployment," and assigned an initial stance of agree, disagree, or neutral. Influencer opinions were configured under three regimes: roughly equal disagree and neutral (1D1N), roughly equal agree and neutral (1A1N), and a fully randomized distribution matching the general population. Each round, 1 percent of influencers were sampled and interacted with their immediate neighbors.
The seeding regime determined the outcome almost entirely. Under 1A1N, the influencee population shifted toward agreement over successive rounds. Under 1D1N, it shifted toward disagreement. Under the random seeding, where influencer opinions already matched the broader population, the shift was minimal and trended toward neutral rather than toward any particular stance. The researchers frame this as evidence that influence concentrated in a small, opinion-skewed group can steer a population even when only a fraction of influencers act each round, while influencers who reflect the population's existing balance tend to moderate rather than polarize it.
Individual attributes also shaped who won these exchanges. Higher-education agents were both more persuasive and more resistant to being persuaded themselves, an asymmetry the paper connects to real-world patterns where access to education and income correlates with communicative authority. Influence success rose with both education and income, peaking among agents with postgraduate education and high income.
What the Efficiency Numbers Actually Show
To reach billion-agent scale, the team trained a multilayer perceptron surrogate model on 100,000 LLM-generated interaction samples, then used that surrogate to replace most LLM calls during the full simulation. They validated this substitution on a smaller 10,000-agent network, comparing opinion trajectories at 0, 25, 50, 75, and 100 percent surrogate usage. The trajectories stayed qualitatively consistent across substitution levels, including the timing of inflection points, which the paper treats as evidence that the surrogate preserves the underlying decision dynamics rather than just approximating aggregate outcomes.
The revised version of the paper adds a reproducibility check: across five independent runs of the full billion-agent simulation, the round-by-round coefficient of variation in opinion fractions stayed under 0.01 percent. That figure comes from the authors' own reported analysis rather than independent replication, and it describes stability within a fixed experimental setup rather than robustness across different setups. The same revision notes that changing the underlying language model, the communication language used in agent interactions, or the network topology can shift outcomes measurably, which the authors treat as a design parameter worth studying rather than noise to average away.
Open Questions
The paper is explicit that this is a demonstration of feasibility rather than a validated model of any specific real-world society. The agent profiles come from survey data rather than tracking of actual individuals, and the trust game and opinion propagation setups are simplified enough to isolate specific mechanisms rather than capture the full texture of social behavior. The compute infrastructure used to run the billion-agent simulation isn't detailed in the paper, and none of the efficiency or reproducibility figures have been independently verified outside the authors' own reporting. The team lists richer memory mechanisms, longitudinal behavior modeling, and integration with real-world data as next steps toward simulations that could inform actual policy questions rather than just demonstrate scale.
David Borish is the author of The Tony Hawk Paradox, forthcoming. More of his writing on frontier AI research and its downstream effects can be found at davidborish.com.
