Exploring the opportunities and challenges of using large language models to represent institutional agency in land system modelling

Abstract. Public policy institutions play crucial roles in the land system, but modelling their policy-making processes is challenging. Large language models (LLMs) offer a novel approach to simulating many different types of human decision-making, including policy choices. This paper aims to investigate the opportunities and challenges that LLMs bring to land system modelling by integrating LLM-powered institutional agents within an agent-based land use model. Four types of LLM agents are examined, all of which, in the examples presented here, use taxes to steer meat production toward a target level. The LLM agents provide simulated reasoning and policy action output. The agents' performance is benchmarked against two baseline scenarios: one without policy interventions and another implementing optimal policy actions determined through a genetic algorithm. The findings show that, while LLM agents perform better than the non-intervention scenario, they fall short of the performance achieved by optimal policy actions. However, LLM agents demonstrate behaviour and decision-making, marked by policy consistency and transparent reasoning. This includes generating strategies such as incrementalism, delayed policy action, proactive policy adjustments, and balancing multiple stakeholder interests. Agents equipped with experiential learning capabilities excel in achieving policy objectives through progressive policy actions. The order in which reasoning and proposed policy actions are output has a notable effect on the agents' performance, suggesting that enforced reasoning both guides and explains LLM decisions. The approach presented here points to promising opportunities and significant challenges. The opportunities include, exploring naturalistic institutional decision-making, handling massive institutional documents, and human–AI cooperation. Challenges mainly lie in the scalability, interpretability, and reliability of LLMs.