Generative agents are reshaping the paradigm of social-science simulation and group behavior prediction, offering a new microscopic perspective for understanding opinion dynamics and crowd mobility. Nevertheless, existing studies remain constrained by two major bottlenecks. First, full-scale large language models (LLMs) incur prohibitive computational and communication overhead in thousand-agent simulations. Second, conventional data-driven approaches are prone to fitting spurious correlations, which limits their ability to support robust simulation-based counterfactual reasoning and intervention-oriented analysis. To address these challenges, we propose a general prediction-intervention framework based on a hierarchical hybrid architecture and structural causal modeling ideas. The framework adopts a Teacher-Student dual-system design: a lightweight network handles high-frequency physical interactions, while a Teacher-LLM is responsible for complex social semantics. Moreover, an entropy-constrained semantic routing module (Information Bottleneck) is introduced to compress redundant information, and counterfactual reasoning is integrated to support intervention-oriented simulation and utility analysis. Experiments on two benchmark modalities, including Twitter public-opinion data and ETH trajectories, demonstrate that the proposed model achieves strong predictive performance (Macro-F1 = 0.87/Average Displacement Error (ADE) = 0.20 m) while improving inference speed by 4.5× compared with full-scale LLMs and reducing per-agent communication volume by 68%. Compared with standard supervised and representative reference baselines, the model also shows favorable robustness and and achieves simulation-based counterfactual intervention gains up to 24.2%. This work provides an effective computational paradigm for building low-cost and interpretable large-scale social digital-twin systems.
Xiao et al. (Fri,) studied this question.