New AI Trick Lets Huge Simulated Crowds Act More Human

Scientists are using AI agents to run simulations of markets, cities, and even whole economies. Imagine a video game where every resident has their own life and makes their own decisions — that's the goal. But there's a practical problem: each AI agent is very smart and uses a lot of computing power. To run thousands of them at once, researchers have to compress what each agent knows, which loses the small details that make people behave differently. As a result, the agents slowly start acting like clones, and the simulation becomes less realistic.

The new paper, accepted at a top AI conference, proposes a clever fix. Instead of squeezing everything into one compressed bundle, they split each agent's brain into two parts. First, a lightweight "Prospect State" tracks the agent's emotional and psychological traces — like fear, greed, or hesitation — using simple math that is very cheap to run. Second, a full "Semantic State" lets the agent use the power of large language models for complex reasoning, planning, and decisions. The two parts work together, letting simulations include far more diverse agents without blowing up costs.

Why does this matter? Economic simulations are a big use case. Real markets are driven by thousands of individual decisions, each shaped by the psychological states of investors and consumers. If those AI simulations become more realistic and can scale larger, they could help policymakers and companies test things like interest rate changes, tax policies, or new product launches before spending real money or causing real-world disruption.

There is still a catch. The current test is academic, not a commercial product yet. And while the psychological state makes agents more varied, it's still a rough approximation of real human emotion. But the direction is clear: smarter scaling of AI societies means better predictions for the rest of us, from housing bubbles to supply chain surprises.