面向大模型多智能体系统的多维评估方法  

A multi-dimensional evaluation method for large language model-powered multi-agent systems

作　　者：董之南 张勤学 胡进 仵志鹏 卢志龙 DONG Zhinan;ZHANG Qinxue;HU Jin;WU Zhipeng;LU Zhilong(No.8511 Research Institute of CASIC,Nanjing 210007,China)

机构地区：[1]中国航天科工集团八五一一研究所,江苏南京210007

出　　处：《指挥控制与仿真》2025年第2期121-131,共11页Command Control & Simulation

摘　　要：大模型驱动的多智能体系统在增强人工智能水平方面具有巨大的潜力,为军事智能化提供了创新解决方案。但是,目前的大模型驱动的多智能体系统能够独立完成任务目标的自主性受任务复杂程度的影响较大,且系统处理结果与初始目标之间的一致性程度较差,有必要对大模型驱动的多智能体系统的自主性和一致性进行评估分析。已有研究尚未对大模型多智能体系统的自主性和一致性水平进行全面评估。提出了一种多维评估方法,能够分析提取大模型驱动的多智能体整体架构的自主性和一致性,得到系统的整体性能评估结果和具体改进方法。通过对7个已有选定系统的实验分析,研究验证了多维评估方法在实际应用中的可行性。The multi-agent system driven by large models has great potential in enhancing the level of artificial intelligence,providing innovative solutions for military intelligence.However,the degree of autonomy of current large model driven multi-agent systems in independently completing task objectives is greatly affected by task complexity,and the consistency between system processing results and initial objectives is poor.It is necessary to evaluate and analyze the autonomy and consistency of large model driven multi-agent systems.Previous studies have not yet comprehensively evaluated the autonomy and consistency levels of large model multi-agent systems.This article proposes a multidimensional evaluation method that can analyze and extract the autonomy and consistency of the overall architecture of multi-agent systems driven by large models,and obtain the overall performance evaluation results and specific improvement methods of the system.Through experimental analysis of 7 selected systems,the feasibility of multidimensional evaluation methods in practical applications has been verified.

关 键 词：大模型 多智能体 系统评估 多智能体协同 自主性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] E919[自动化与计算机技术—控制科学与工程]