Fragile Cooperation: Endogenizing Climate Treaty Participation with LLM-Driven Regional Agents

Climate cooperation is reversible and unstable, and who participates in climate agreements strongly shapes mitigation costs and climate outcomes. However, integrated assessment models treat treaty participation as an exogenous scenario input. We propose to endogenize participation by coupling the 2023 version of the Dynamic Integrated Climate-Economy model (DICE-2023) with twelve regional large language model agents, one per region of the 2010 Regional Integrated model of Climate and the Economy (RICE-2010), through the global participation rate, updated in five-year steps. Unlike static equilibrium or language-free learning approaches, this framework gives agents memory and political identity, so that withdrawal contagion, erosion cycles, and negotiation effects can emerge. The framework lets negotiators stress-test agreement architectures and gives modelers carbon-price and warming ranges that internalize political erosion risk.