Mechanism-Guided Generative Agent-Based Modeling for Scalable Population Simulation

Large language model (LLM) agents enable context-sensitive behavior in social and population simulation, but independently reasoning over every individual is costly and can produce unstable or poorly calibrated collective dynamics, limiting the reliability of large-scale generative agent-based models. We propose a cluster-level generative ABM that estimates transition dynamics over behaviorally coherent groups and subsequently realizes heterogeneous individual trajectories from these shared group-level estimates. We introduce ANCHOR, an LLM-agent-guided clustering method that groups agents by cross-context behavioral responses in addition to attributes and interaction structure. This enables functional behavioral groups characterized by how agents adapt under changing conditions rather than solely by demographic or network similarity. Within each group, state-specialized symbolic agents and a multimodal neural model provide complementary transition estimates that are combined through learned context-dependent fusion. Across social attention diffusion, financial sentiment diffusion, and epidemic diffusion over contact networks, with predictive evaluation up to $N=1{,}000$, our framework improves event-time alignment and probabilistic calibration over mechanistic, neural, and flat LLM-agent baselines while reducing LLM inference calls by $\sim$6--8$\times$. We further use Singapore's COVID-19 Circuit Breaker as a real-world policy-intervention stress test, showing coherent adaptation under an abrupt regime shift. Separate stress tests show approximately linear scaling of the non-LLM agent-realization stage up to $N=10{,}000$. These results support behaviorally structured group-level inference as a scalable and calibrated approach to heterogeneous LLM-based population simulation. Code: https://anonymous.4open.science/r/mech-gabm/