Multi-Agent LLM Simulation of Stakeholder Interactions in Complex Policy Decision-Making: Evidence from Dutch Infrastructure Projects

Modern governance processes face increasing complexity due to diverse stakeholders, competing priorities, and interdependent resources. While traditional research methods offer valuable insights into these dynamics, they often struggle to capture the informal, iterative nature of stakeholder interactions and decision-making processes comprehensively. This research introduces an innovative methodological framework that leverages Large Language Models (LLMs) to simulate and analyze complex multi-stakeholder governance processes. Our study addresses a fundamental question: How can multi-agent simulations using LLMs effectively replicate stakeholder interactions and decision-making processes in complex policy environments, and what factors influence the fidelity and outcomes of these simulations compared to historical cases? Through this investigation, we examine three critical dimensions: the effectiveness of multi-agent simulations in replicating both outcomes and processes, the manifestation of power dynamics and equity in simulated interactions, and the influence of structural and contextual factors on stakeholder strategies. The research employs a novel simulation framework that integrates network governance theory with LLM-based agent modeling. By conditioning simulations on well-documented actor characteristics and arena structures, we enable dynamic emergence of interactions and outcomes that reflect real-world governance processes. We apply this framework to a detailed case study of the Rotterdam Harbor Expansion, specifically examining the VERM and PMR rounds, which provide rich historical data for validation and comparison. Our methodology addresses key limitations of traditional research approaches by enabling real-time analysis of stakeholder dynamics, reducing dependence on retrospective data, and facilitating the exploration of alternative scenarios. The framework specifically examines how power imbalances manifest in simulated environments, whether the system replicates historical power dynamics, and how varying contextual factors influence negotiation outcomes. The findings demonstrate the potential of LLM-based simulations to uncover hidden patterns in governance processes, identify systemic biases, and reveal opportunities for more equitable stakeholder participation. Through careful comparison with historical records, we evaluate the framework's ability to replicate documented processes while providing new insights into the mechanisms driving policy outcomes. This research contributes to both theoretical understanding and practical applications in governance studies. It advances the field by introducing a novel methodological approach that combines the structural insights of network governance theory with the adaptive capabilities of LLMs. The framework offers policymakers a powerful tool for exploring governance scenarios, understanding stakeholder dynamics, and developing more inclusive decision-making processes. The implications of this study extend beyond theoretical advancement, offering practical insights for policymakers and stakeholders engaged in complex governance processes. By enabling the simulation and analysis of multi-stakeholder interactions, our framework provides a foundation for developing more effective and equitable approaches to collaborative governance in an increasingly complex policy landscape.