GPLab: A Generative Agent-Based Framework for Policy Simulation and Evaluation

Real-world social policy problems invariably involve complex dynamics with cross-domain, multi-factor interactions, posing significant challenges to ex-ante policy simulation and impact evaluation. Traditional statistical modeling approaches often fail to adequately capture individual behavioral heterogeneity and system-level dynamic propagation effects, while conventional agent-based modeling (ABM) relies heavily on hand-crafted behavioral rules, which constrains its adaptability to realistic decision-making processes and diverse policy contexts. To address this challenge, we propose GPLab (Generative Policy Laboratory), a general-purpose framework for policy simulation and evaluation that integrates generative Large Language Models (LLMs). By leveraging LLM-based agents with bounded rationality, GPLab simulates the cognition and behavior of social individuals, overcoming critical limitations of rule-based ABM in policy semantic understanding and scenario adaptation. Furthermore, its modular architecture explicitly represents interconnected social subsystems, enabling heterogeneous agents to interact dynamically across multiple domains and produce contextually grounded behavioral responses. In two representative policy evaluation cases, we successfully captured policy intensity-dependent effects, group behavioral heterogeneity, and emotional opinion evolution patterns. Formal consistency metrics further confirmed the stability of agent behavior and coherence of individual traits throughout simulations. Extended experiments demonstrated that our framework achieved high rationality scores across five policy scenarios, validating its cross-domain transferability. This study presents a scenario-agnostic policy simulation platform that significantly advances computational social science applications in public decision support. The code is available at: https://github.com/SmartLegislation/GPLab