超越个体智能：调查基于LLM的多智能体系统的协作、故障归因与自我进化

Shihao Qi†,1,2,Jie Ma†,B,1,3,Rui Xing†,1,4,Wei Guo†,1,5,Xiao Huang†,1,4Zhitao Gao2,6,Jianhao Deng2,Jun Liu1,2,6,Lingling Zhang1,2,6,Bifan Wei1,6Boqian Yang2,Pinghui Wang1,5,Jianwen Sun7,Jing Tao1,3Yaqiang Wu2,8,Hui Liu8,Yu Yao9,Tongliang Liu9 1MOE KLINNS Lab2School of Computer Science and Technology, Xi’an Jiaotong University3School of Cyber Science and Engineering, Xi’an Jiaotong University4School of Software Engineering, Xi’an Jiaotong University5School of Control Science and Engineering, Xi’an Jiaotong University6Shaanxi Provincial Key Laboratory of Big Data Knowledge Engineering7Laboratory for AI and New Forms of Education, Central China Normal University8Lenovo AI Technology Center, CTOO, Lenovo9Sydney AI Centre, The University of Sydney †Equal contributionBCorresponding author Abstract LLM-based autonomous agents have demonstrated strong capabilities in reasoning, planning, and tooluse, yet remain limited when tasks require sustained coordination across roles, tools, and environments.Multi-agent systems address this limitation through structured collaboration among specialized agents,but tighter coordination also amplifies a less explored risk:errors can propagate across agents andinteraction rounds, producing failures that are difficult to diagnose and, even once identified, rarelytranslate into structural self-improvement.Existing surveys have separately covered individual agentcapabilities, multi-agent collaboration, or agent self-evolution, but treat these topics in isolation—leaving thecausal dependencies among them largely unexamined. This survey provides a unified and comprehensivereview organized around four causally linked stages, which we term theLIFEprogression:Lay thecapability foundation,Integrate agents through collaboration,Find faults through attribution, andEvolvethrough autonomous self-improvement. We review the capability foundations of individual agents, theorganizational mechanisms of multi-agent collaboration, the methodological landscape of failure attribution,and the hierarchical design space of self-evolution. Throughout, we formally characterize the dependenciesbetween adjacent stages, revealing how each stage both depends on and constrains the next. Beyondsynthesizing existing work, we identify open challenges at the boundaries between LIFE stages and proposea cross-stage research agenda for closed-loop multi-agent systems capable of continuously diagnosingfailures, reorganizing collaborative structures, and refining agent behaviors, thereby extending currenthuman-engineered coordination frameworks toward more self-organizing and resilient forms of collectiveintelligence. By bridging these previously fragmented research threads into a coherent progression, thissurvey aims to offer both a systematic reference for current research and a conceptual roadmap towardautonomous, self-improving multi-agent intelligence.arXiv:2605.14892v2 [cs.AI] 15 May 2026 Keywords:large language model-based agents, multi-agent systems, multi-agent collaboration,failure attribution, self-evolution, survey. GitHub:§https://github.com/mira-ai-lab/awesome-mas-life Contents 1Introduction 2Individual Intelligence 2.2.1Input-Stage Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .92.2.2Reasoning-Process Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .102.2.3Output-Stage Regulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10 2.3Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.3.1Memory Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .132.3.2Memory Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .142.3.3Memory Retrieval and Utilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14 2.4Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.4.1Decomposition-Based Planning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .152.4.2Search-Based Planning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .17 2.5Tool Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.5.1Tool Capability Acquisition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .182.5.2Tool Invocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .