多智能体系统：从经典范式到大基础模型驱动的未来

Zixiang Wang, Mengjia Gong, Qiyu Sun, Jing Xu,Senior Member, IEEE, Shuai Mao, Xin Jin,Qing-Long Han,Fellow, IEEE, and Yang Tang,Fellow, IEEE MASs (CMASs) [9], [10]. CMASs rely on explicitly designedsystem models or task-specific learning mechanisms. Froma methodological perspective, existing CMASs research canbe broadly categorized into model-based and learning-basedapproaches. Model-based research has gradually establishedseveral classical problem domains and theoretical frameworks,including consensus control [11], formation control [12], taskscheduling [13], and bio-inspired optimization [14]. Thesestudies typically assume modelable systems and clear objec-tives to ensure provable stability and performance [9]. How-ever, in scenarios with unmodelable environments, unknownsystem dynamics and partial observability, reliance on explicitmodeling and control design is often limited. As a result,learning-basedmethods such as multi-agent reinforcementlearning (MARL) have emerged as an important alternative,enabling agents to learn coordinated policies through inter-action without accurate models [15]. While this paradigmpartially mitigates model dependency in complex settings,MARL still suffers from limitations in sample efficiency,stability, interpretability, and generalization [16].Thelimitations of CMASs motivate the exploration of Abstract—With the rapid advancement of artificial intelli-gence, multi-agent systems (MASs) are evolving from classicalparadigmstoward architectures built upon large foundationmodels (LFMs). This survey provides a systematic review andcomparative analysis of classical MASs (CMASs) and LFM-basedMASs(LMASs).First,within a closed-loop coordina-tion framework, CMASs are reviewed across four fundamentaldimensions: perception, communication, decision-making, andcontrol.Beyond this framework,LMASs integrate LFMs tolift collaboration from low-level state exchanges to semantic-levelreasoning,enabling more flexible coordination and im-proved adaptability across diverse scenarios. Then, a comparativeanalysis is conducted to contrast CMASs and LMASs acrossarchitecture, operating mechanism, adaptability, and application.Finally, future perspectives on MASs are presented, summarizingopen challenges and potential research opportunities. Index Terms—Artificial intelligence, Multi-agent System, Largefoundation model, Agentic AI. I. INTRODUCTION MUlti-agent systems (MASs) have become a core arti-ficial intelligence research paradigm with broad appli-cations in multiple disciplines, including robotics [1], socialintelligence [2], and satellite systems [3]. Inspired by biologi-cal swarms and functional requirements of complex distributedsystems [4], [5], MASs focus on how multiple autonomousagents achieve global coordination or collective intelligencethrough interaction [6]. Compared with single-agent systems,MASs provide a natural framework for modeling complexinteractionsand coordination among multiple autonomousentities in real-world environments [7], [8].In this survey, MASs that do not incorporate large foun- more general approaches with reasoning capabilities, leadingto the integration of large foundation models (LFMs) intoMASs [17]. In the context of MASs, LFMs serve as the cog-nitive core of agents. They enable agents to interpret unstruc-tured multimodal inputs, maintain contextual understanding,reason over complex tasks, and generate high-level actions orinteraction messages [18]. This shifts agent operation awayfrom predefined system models, handcrafted rules, or task-specific policies in CMASs toward semantic-level perceptionand language-based interaction, enabling more flexible coor-dination [19]. This evolution marks a fundamental paradigmshift from task-specific, environment-constrained CMASs tomore adaptive, general-purpose, and cognitively empoweredLFM-based MASs (LMASs). By leveraging the pretrainedknowledge and reasoning abilities of LFMs, these systems canperform complex multi-step planning, knowledge retrieval, andhigh-level decision-making [20], [21]. As illustrated in Fig. 1,unlike CMASs tailored to fixed environments, LMASs gener-alize well and accumulate experience across tasks, supportingflexible collaboration in open, dynamic scenarios [22], [23].The existing surveys on LMASs primarily focus on thearXiv:2604.18133v1 [cs.AI] 20 Apr 2026 dation models (LFMs) are collectively referred to as classical LFM-based paradigm and summarize system architectures,coordination mechanisms, and applications [8], [20], [24],[19], [25]. Unlike these surveys that examine LMASs inisolation, we propose a unified perspective bridging CMASsand LMASs. LMASs are not a replacement for CMASs, As shown in Fig. 2, this section introduces CMASs fromfouraspects:perception,communication,decision-making,and control. The first two address information acquisition anddissemination, while the latter enable distributed reasoningand coordinated actions under given objectives. This four-dimensional framew