Reproducibility Study of "Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents"

TL;DR

This reproducibility study confirms GovSim's core claims: only the largest LLMs can sustain sustainable cooperation in resource-sharing games, and universalization generally increases survival time for several models. It validates the original methodology while extending it with diverse architectures (e.g., DeepSeek-V3, Qwen variants) and new settings (Japanese language prompts, inverse-harm resource, heterogeneous agents). The findings show high-performing models can positively influence weaker agents, suggesting pathways to more efficient multi-agent AI systems, but also raise ethical concerns about manipulation and resource framing. Overall, the work demonstrates GovSim's adaptability across models and languages, providing actionable insights for deploying cooperative AI in complex, resource-constrained environments.

Abstract

This study evaluates and extends the findings made by Piatti et al., who introduced GovSim, a simulation framework designed to assess the cooperative decision-making capabilities of large language models (LLMs) in resource-sharing scenarios. By replicating key experiments, we validate claims regarding the performance of large models, such as GPT-4-turbo, compared to smaller models. The impact of the universalization principle is also examined, with results showing that large models can achieve sustainable cooperation, with or without the principle, while smaller models fail without it. In addition, we provide multiple extensions to explore the applicability of the framework to new settings. We evaluate additional models, such as DeepSeek-V3 and GPT-4o-mini, to test whether cooperative behavior generalizes across different architectures and model sizes. Furthermore, we introduce new settings: we create a heterogeneous multi-agent environment, study a scenario using Japanese instructions, and explore an "inverse environment" where agents must cooperate to mitigate harmful resource distributions. Our results confirm that the benchmark can be applied to new models, scenarios, and languages, offering valuable insights into the adaptability of LLMs in complex cooperative tasks. Moreover, the experiment involving heterogeneous multi-agent systems demonstrates that high-performing models can influence lower-performing ones to adopt similar behaviors. This finding has significant implications for other agent-based applications, potentially enabling more efficient use of computational resources and contributing to the development of more effective cooperative AI systems.

Figures (9)

• Figure 1: Example of a conversation between two agents in the MultiGov scenario. John (DeepSeekV3) notes that Luke (GPT-4o-Mini) overfished and warns that widespread imitation could harm the lake. He proposes a 10-ton per person limit and asks Luke for his thoughts.
• Figure 2: Sustainability test results for the multi-agent fishery scenario with multi-agent default scenario. The plots show the available resources after harvesting (line) and the collected resources in each month by each agent (columns). The captions show the agent combination used in each experiment.
• Figure 3: Sustainability test results for the homogeneous-agent fishery default scenario, showing available resources after collection each month for GPT (\ref{['subfig:gpt']}), Llama-3 (\ref{['subfig:llama3']}), Llama-2 (\ref{['subfig:llama2']}), Mistral (\ref{['subfig:mistral']}), DeepSeek-V3 (\ref{['subfig:deepseek']}), and Qwen (\ref{['subfig:qwen']}) models. Models pass the sustainability test if resources remain above zero for the full 12-month simulation. Failure typically occurs when the first harvest exceeds 70% of the available resource, leading to resource collapse and survival times of 1-2 months. This behavior is observed in GPT-3.5, GPT-4o-mini, Mistral-7B, all Llama, and Qwen models. In contrast, initial harvests below 50% enable cooperation and sustainable resource extraction, resulting in 12-month survival. Models achieving this include GPT-4o, GPT-4o-Turbo, and DeepSeek-V3.
• Figure 4: Sustainability test results for the homogeneous-agent fishery universalization scenario, showing available resources after collection in each month for different model families. In this scenario, the universalization principle is communicated to each agent: when deciding how many resources to collect, agents consider the possibility that others will do the same. The Llama-2 family models and the Qwen-2.5-7B model showed no improvement over the default scenario. As expected, the DeepSeek-V3, GPT-4o, and GPT-4o-Turbo models passed the sustainability test, as they did in the default scenario. The Mistral-7B, Llama-3-8B, Llama-3-70B, Qwen-2.5-0.5B, GPT-3.5, and GPT-4o-mini models showed significant improvements, increasing their survival time compared to the default scenario. Notably, only the Llama-3 family models improved, while the Llama-2 family models did not.
• Figure 5: Sustainability test results for the homogeneous-agent fishery default scenario, using Japanese prompts, showing available resources after collection in each month for different model families. The DeepSeek-V3 and GPT-4o models passed the sustainability test with a 12-month survival time, while GPT-4o-mini failed. These results, which are similar to those obtained for English prompts in \ref{['fig:graph_fishery_default']}, indicate that language does not affect the models' behavior.