[Re] Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents

Abstract

Large Language Models (LLMs) are increasingly used in strategic decision-making environments, including game-theoretic scenarios where multiple agents interact under predefined rules. One such setting is the common pool resource environment. In this study, we build upon Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (Piatti et al., 2024), a framework designed to test cooperation strategies among LLM agents. We begin by replicating their results to a large degree to validate the framework, reproducing the original claims regarding model scale in their simulation environment. Then, we extend the analysis to include models that represent the recent reasoning paradigm: Phi-4, DeepSeek-R1, and one of the distilled variants, which show improvements over their baseline counterparts but come at a higher computational cost. Here, we identify a notable trend: specialized models with reasoning-oriented training outperform general-purpose models of similar scale in this environment. Finally, we investigate the impact of different experiments, including the veil of ignorance mechanism and other prompting strategies based on universalization principles with varying levels of abstraction. Our results suggest that older models benefit significantly from explicit boundary conditions, whereas newer models demonstrate greater robustness to implicit constraints.