New APS method simulates 10M agents with 381x fewer LLM calls

Simulating large populations with LLM agents is computationally expensive: each agent queries the model every round, scaling linearly with size and horizon. A new paper from Quan Zheng and collaborators introduces APS (Adaptive Prototype Simulation), which reframes the problem as a recurrent oracle-allocation problem. Instead of querying every agent, APS selects a small set of adaptive prototype agents to generate responses, then propagates those responses to nearby agents via local response surfaces. To control approximation bias, it employs shadow-audit residual correction (estimating and correcting errors) and tail-protected singleton routing (directly querying isolated or high-curvature agents).

In a 10M-agent multi-round public-opinion simulation, APS achieved a 381.1-fold reduction in LLM calls while maintaining a final-round Jensen-Shannon divergence of just 0.094 compared to a full-LLM reference. The framework outperforms scale-oriented and same-budget baselines in distributional discrepancy. By treating bias control as a first-class concern, APS makes population-scale social simulations practical for policy modeling, epidemiology, and market research — without the astronomical costs of naive LLM-based agent simulations.