摘要： 在多智能体博弈仿真中，大语言模型的能力已经被广泛研究，但其在模糊任务目标或不确定性环境中引导多智能体合作的决策能力往往出现“失灵”现象。针对这一问题，提出了一种基于分布式贝叶斯推断的多层级协同决策框架。该框架基于分布式贝叶斯推断方法，集成了“决策—互评—监管”三大功能模块，利用多个大语言模型进行协同决策，并在空间囚徒困境博弈中进行了实验验证。实验结果表明，该框架有效克服了大语言模型在模糊任务环境下的决策瓶颈，成功促进了多智能体合作行为的涌现。此外，通过对不同实验场景下模型决策能力的量化评估，发现模型的决策误差与模型规模不呈线性关系。在模糊任务指令下，LLAMA3（70B）模型的决策误差较LLAMA3（8B）模型高出16.6%，较LLAMA2（7B）模型高出7.2%，表明在更复杂环境中，单纯依赖模型规模的扩大未能显著提升决策性能。相反，大语言模型协同决策在提升决策一致性和有效性方面显示出显著优势。这些结果揭示了多模型协同在复杂决策环境中的关键作用，并为未来在不确定性任务下的智能体系统设计提供了重要参考。

Abstract: The capabilities of large language models (LLMs) in multi-agent game simulations have been widely studied, but their ability to guide cooperative decision-making among agents often fails in environments with fuzzy task objectives or uncertainty. To address this issue, a multi-level collaborative decision-making framework based on distributed Bayesian inference is proposed. This framework integrates three functional modules—decision-making, peer review, and supervision—utilizing multiple LLMs for collaborative decision-making. The framework is validated through experiments in spatial prisoner’s dilemma games. The results demonstrate that this framework effectively overcomes the decision-making bottlenecks of LLMs in fuzzy task environments, successfully promoting the emergence of cooperative behavior among agents. Additionally, a quantitative evaluation of decision-making capabilities across different experimental scenarios reveals that decision errors do not scale linearly with model size. Under fuzzy task instructions, the decision error of the LLAMA3 (70B) model is 16.6% higher than that of the LLAMA3 (8B) model and 7.2% higher than that of the LLAMA2 (7B) model, indicating that simply increasing model size does not significantly improve decision-making performance in more complex environments. In contrast, the collaborative decision-making of LLMs shows significant advantages in enhancing decision consistency and effectiveness. These findings highlight the crucial role of multi-model collaboration in complex decision environments and provide valuable insights for the design of intelligent agent systems in uncertain tasks.