LLM-based Agents in Supply Chain Games: The Role of Incomplete Information and Model Heterogeneity

Effective collaboration is essential for mitigating market volatility, yet complete information sharing among partners is often impractical. By employing diverse Large Language Models as autonomous agents, we design controlled experiments in which information is shared only among subsets of enterprises, approximating realistic business environments. Our results reveal a counterintuitive finding: partial information sharing can generate system level benefits comparable to those achieved under full transparency. We further compare agent behavior and identify differences in decision stability. DeepSeek exhibiting the most consistent performance, followed by Qwen and Llama. Finally, experiments within a Llama based environment show that introducing a higher capability model can improve both stability and aggregate performance. Overall, our study provides a scalable experimental framework for artificial society modeling and demonstrates the potential of LLM-based agent simulations for investigating complex socio economic systems.