Paper deep dive
Distributed Agent System: Fault-Tolerant Collaboration Among Embodied Agents
Kai Yu, Lu Chen, Hanqi Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/18/2026, 1:30:56 PM
Summary
The paper introduces the Distributed Agent System (DAS), a device-edge-cloud framework designed for fault-tolerant collaboration among heterogeneous embodied agents in industrial settings. It redefines agent reliability as system-level fault tolerance rather than single-turn accuracy, proposing a two-layer architecture: single-agent execution reliability via fault-tolerant alignment (handling uncertainty through refusal, clarification, and routing) and cross-agent communication reliability via semi-formal language protocols (ensuring verifiable, executable communication rules).
Entities (7)
Relation Signals (6)
Distributed Agent System â employs â Fault-Tolerant Alignment
confidence 95% · This paper proposes Distributed Agent System (DAS)... present a two-layer fault-tolerance architecture: single-agent execution reliability via fault-tolerant alignment
Distributed Agent System â employs â Semi-Formal Language Protocols
confidence 95% · cross-agent communication reliability via semi-formal language protocols
Distributed Agent System â supports â Embodied Agents
confidence 92% · framework for fault-tolerant collaboration among heterogeneous agents... tailored to collaborative tasks among heterogeneous agents... embodied agents deployed on end devices
Fault-Tolerant Alignment â addresses â Single-Agent Execution Reliability
confidence 90% · single-agent execution reliability via fault-tolerant alignment
Semi-Formal Language Protocols â addresses â Cross-Agent Communication Reliability
confidence 90% · cross-agent communication reliability via semi-formal language protocols
DiSRouter â ispartof â Distributed Agent System
confidence 85% · The second is the DiSRouter (Distributed Self-Router) mechanism... gives agents the ability to assess their own competence
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:AI engineering is shifting from passive text generation by large language models (LLMs) to agent-driven task execution, creating new reliability challenges for long-horizon tasks under resource constraints and environmental uncertainty. Conventional error-elimination optimization strategies fail to address cumulative error propagation. This paper proposes Distributed Agent System (DAS), a device-edge-cloud framework for fault-tolerant collaboration among heterogeneous agents. We redefine agent reliability as system-level fault tolerance rather than single-turn zero-error accuracy, and present a two-layer fault-tolerance architecture: single-agent execution reliability via fault-tolerant alignment, and cross-agent communication reliability via semi-formal language protocols. This framework provides a practical engineering pathway for reliable heterogeneous embodied agents collaboration in industrial scenarios.
Tags
Links
- Source: https://arxiv.org/abs/2607.10811v1
- Canonical: https://arxiv.org/abs/2607.10811v1
Trouble viewing inline? Open PDF directly â
Full Text
19,951 characters extracted from source content.
Expand or collapse full text
DistributedAgentSystem:Fault-TolerantCollaborationAmong EmbodiedAgents KaiYu*,LuChen,HanqiLi a X-LANCELab,SchoolofComputerScience,ShanghaiJiaoTongUniversity,Shanghai,200240,China b JiangsuKeyLabofLanguageComputing,Suzhou,215028,China * Correspondingauthor.E-mail:kai.yu@sjtu.edu.cn,Tel:13917575615,Address:800DongchuanRoad,MinhangDistrict,Shanghai,200240,China ABSTRACT AIengineeringisshiftingfrompassivetextgenerationbylargelanguagemodels(LLMs)toagent-driven taskexecution,creatingnewreliabilitychallengesforlong-horizontasksunderresourceconstraintsand environmentaluncertainty.Conventionalerror-eliminationoptimizationstrategiesfailtoaddress cumulativeerrorpropagation.ThispaperproposesDistributedAgentSystem(DAS),adevice-edge-cloud frameworkforfault-tolerantcollaborationamongheterogeneousagents.Weredefineagentreliabilityas system-levelfaulttoleranceratherthansingle-turnzero-erroraccuracy,andpresentatwo-layerfault- tolerancearchitecture:single-agentexecutionreliabilityviafault-tolerantalignment,andcross-agent communicationreliabilityviasemi-formallanguageprotocols.Thisframeworkprovidesapractical engineeringpathwayforreliableheterogeneousembodiedagentscollaborationinindustrialscenarios. KEYWORDS DistributedAgentSystem;Fault-TolerantCollaboration;EmbodiedAgents;ReliabilityAlignment;Semi- FormalLanguageProtocol;Multi-agentCollaboration 1.Introduction:FromGenerativeAItoAgenticTaskExecution AIengineeringisundergoingafundamentalshiftfrompassivetextgenerationdominatedbylarge languagemodels(LLMs)towardsagent-driven,end-to-endtaskexecution[1].Unlikeearlierapplication modesthatrelyonstagedhumaninterventionorstraightforwardinstructionmatching,next-generation agentscancompleteaclosedloopofenvironmentalperception,reasoning,planning,decision-making,and physicalexecution,therebysupportinglong-horizonautonomousoperationsinindustrialsettings[2,3]. ThisparadigmshiftalsochangeshowAIreliabilityshouldbeevaluated.Q&A-basedLLMinteractionis typicallyevaluatedbysingle-turnaccuracy,whereasautonomousindustrialembodiedagentsmustbe evaluatedbysystem-levelstabilityandreliabilityacrosscontinualmulti-steptasks. Long-horizonautonomousexecutionposesnewreliabilitychallengesandhasbecomeacorebottleneckfor industrialAIdeployment.End-deployedlightweightembodiedagentsfacestrictconstraintsoncomputing power,storage,andbandwidth,withlimitedmodelcapacitycomparedtocloudfrontierLLMs.Dynamic disturbancesinopenindustrialenvironmentsleadtoinherentuncertainty,whichcannotbefullyeliminated bymodelscalingordataiteration.Critically,minorlocalerrorsaccumulateandamplifyalonglongtask chains,potentiallycausingoveralltaskfailure. Currentmainstreamoptimizationmethods,includingtool-parametercalibration,modelconfidence estimation,andmulti-turnself-reflection,canmitigatemodelhallucinationsandexecutionerrorsonlyin single-turnorlocalscenarios[4,5,6,7].Theycannotsuppressthesystematicamplificationoferrorsinlong- horizontasks.Meanwhile,mostmulti-agentcollaborationresearchisbuiltonhomogeneouscloud-hosted LLMsforvirtualscenarios,prioritizingsingle-turnaccuracy,andcannotadapttoheterogeneousindustrial end-devicedeploymentandphysicalexecution[8].Tobridgethisgap,weproposeadistributedfault- tolerantcollaborationframeworksuitableforend-deviceembodiedagents,replacingerroreliminationwith system-levelfault-tolerancemanagement. 2.DistributedAgentSystem:DefinitionandScope ADistributedAgentSystem(DAS)isadistributedframeworkforcollectiveintelligence,builtfor heterogeneousend-edge-cloudagentdeployment.Itbalanceslocalautonomousoperationsonend deviceswithglobalcoordinationandschedulingonthecloud,andistailoredtocollaborativetasks amongheterogeneousagentscharacterizedbylong-horizondecisionsequences,highreliability requirementsandcostsensitivityconstraints.TheprimarydeploymentunitsofDASareembodied agentsdeployedonenddevices,includingindustrialrobots,environmentalsensingandcontrolequipment, IoTterminals,andotherphysicalhardwarethattightlycoupleintelligentdecision-makingwithphysical execution. DASemployslayeredcoordinationwithdistinctresponsibilities.Lightweightend-deviceagentshandle real-timeperceptionandbasicautonomousactions.Theedgeaggregatesregionaldataandcoordinates localtasks.Thecloudsupportsglobalplanningandoptimization.Crucially,alltiersrunindistributed modeandmakeautonomousdecisionsindependentlyânotcentrallydispatchedbythecloudâachievinga dynamicbalancesuitableforcomplexindustrialworkflows. DASdiffersfundamentallyfromcloudLLM-basedmulti-agentsystemsacrossfivecoredimensions.In architecture,itisbuiltonheterogeneousdevice-edge-cloudagentsratherthanhomogeneousclustersof cloudfoundationmodels.Intermsofresourceassumptions,itoperatesunderstrictend-deviceconstraints oncomputation,storage,andcommunication,insharpcontrasttotheabundantcomputingpowerand comprehensiveknowledgebasesavailableincloud-nativesystems.Regardingoptimizationobjectives,it prioritizessystem-levellong-horizondecisionsequencereliabilitythroughproactiveerroraccumulation managementinsteadofpursuingperfectsingle-turnoutputaccuracy.Forcorechallenges,ittackleserror accumulationandcoupledrisksbetweendecision-makingandphysicalexecution,ratherthantherelatively low-risktasksofquestionansweringandsoftwaredecompositionfacedbycloud-basedsystems.Finally,in applicationscenarios,itfocusesonphysicallong-horizonoperationsinindustrialsettingsasopposedto purevirtualsoftwareenvironments[9,10]. 3.RedefiningReliability:FromErrorEliminationtoFault-ToleranceManagement Insufficientreliabilityistheprimarybottleneckpreventinglarge-scaleindustrialdeploymentofembodied agents.CurrentacademicapproachestoLLMuncertaintygenerallyfollowthesamelogic:uselarger models,moretrainingdata,high-precisionconfidenceestimation,anditerativereflectiontoeliminate hallucinationsandexecutionerrorsentirely.Thislogicmaybefeasibleincompute-rich,cloud-hosted scenarioswithshortinteractioncycles,suchasquestionansweringandsoftwaretooluse.However,itdoes notfitresource-constrainedindustrialdeploymentonenddeviceswithlong-horizonautonomousexecution, wheretheoperatingconditionsarefundamentallydifferent. Westartfromabasicinsight:uncertaintyintroducedbymodelsandenvironmentsisnotasystem defectthatmustbecompletelyeradicated.Instead,itisaninherentpropertyofhuman-AIandAI- AIinteractionandphysicaloperationunderincompleteinformation;itneednotandcannotbefully eliminated.Fromtheperspectiveofcognitivelinguistics,ambiguityinnaturallanguageandinexchanged informationisanimportantfeatureofefficientcollaboration[11,12,13].Humansresolvesuchambiguity dynamicallythroughmulti-turninteractionandadapttocomplexscenarios.Thismetacognitiveregulatory mechanismhasbeenusedinPOMDP-basedspokendialoguesystems[21,22]andcanbetransferredtothe optimizationofagentsystems. Industrialend-deviceconstraintsfurthersupportthis:hardwareresourcelimitspreventunboundedmodel scaling;openenvironmentsinvolveunseenoperatingconditions;andlong-horizontasksamplifysmall errorsintotaskfailure.Therefore,reliabilityoptimizationforindustrialagentsmustshiftitscorelogic from"errorelimination"to"uncertaintymanagementanderror-propagationsuppression".Basedonthis insight,weredefineagentreliabilityasfollows: Reliabilityisthecomprehensivesystem-levelcapabilityofanagenttocompleteaclosed-looptask processstably,consistently,andcontrollablyundergiventaskandenvironmentalconstraints, ensuringsafebehavior,trustworthyoutcomes,predictabledecisions,andfault-toleranthandlingof exceptions.Itisessentiallyautilitytrade-offbetweentask-executionbenefitsandthecostsof hardware,time,andhumanintervention. Thisdefinitiondiffersfromtraditionalevaluationframeworks:itnolongerjudgesreliabilitybyaccuracy orsuccessinasingleinteraction,butinsteadtreatsend-to-endfault-tolerancemanagement,error- amplificationsuppression,andstableglobaltaskcompletionasthecorecriteria.Smalllocalerrorscanbe graduallyresolvedthroughmulti-turninteraction,dynamicinformationupdating,andthecomplementary capabilitiesofmultiplenodes. Onthebasisofthisnewreliabilitydefinition,DASforindustriallong-horizonoperationshouldadopta two-layerfault-tolerancearchitecturetomanagelocalerrorsandsuppresstheirpropagationatthe systemlevel.Thefirstlayerissingle-agentexecutionfaulttolerance:fault-tolerantactionssuchas proactiverefusal,taskrouting,anduserclarificationareusedtomanageexecutionerrorsinindividual embodiedagentsandreduceharmfulhallucinatedoutputsatthesource.Thesecondlayerismulti-agent communicationfaulttolerance:standardizedsemi-formalprotocolsareusedtounifycross-agent communicationrules,blockthecascadingpropagationofsemanticdeviationsandinformationambiguity, andintroducedynamicformalverificationandauditingduringexecution.Thetwolayersworktogetherto systematicallyaddressreliabilitydegradationinDAS. 4.ParadigmOne:ReliableSingle-AgentExecutionviaFault-TolerantReliabilityAlignment Toaddressagents'inherentmodeluncertainty,tool-callinghallucinations,andincorrecttaskassessment, weproposeafault-tolerantreliabilityalignmentparadigm.Thisparadigmreconstructstheoptimization objectivestomeetthefault-tolerancerequirementsoflightweightagents.Itscoreinnovationisto acknowledgetheinherentnatureofmodeluncertainty,moveawayfromtheobjectiveofabsolutezero error,andachievereliableexecutionunderuncertainconditionsthroughdiversefault-tolerantactions[4,5]. TraditionalLLMoftentreatsnon-responseinunknownscenariosanderroneousoutputsasequivalent negativesamples,therebyneglectingthevalueofadaptivefault-tolerantbehaviorunderuncertainty. Consequently,theoutputactionspaceoftraditionalLLMagentsisusuallydominatedbyconcretetask- executionactionssuchasdirectresponses;fewpoliciesexplicitlydistinguishfault-tolerantcommunicative actionssuchasrefusal,clarification,confirmationandtaskdelegationetc.Whensuchagentsaredeployed onlightweightenddevices,theymayeasilyproduceoutputsbeyondtheircapabilityboundaries,inducing tool-callinghallucinationsandcapabilityoverreach,andthereforeleadingtounreliableindustrialoperation. Thefault-tolerantreliabilityalignmentparadigmreframesmodeloptimizationaroundacomprehensive utilitytrade-off:maximizingtaskusefulnesswhileminimizinginteractioncost.Itaccountsforend-device computingoverhead,thecostoferroneousexecution,andthecostofhumanintervention,makingitbetter suitedtotheresourceconstraintsoflightweightagents.Itscentralbreakthroughistoexpandtheaction spaceforuncertaintyhandlingandbreakthebinarylogicof"answerorerror."Inadditiontostandardtask- executionactions,theparadigmintroducesanindecisiveactionspaceconsistingofadaptivefault-tolerant actionssuchasproactiverefusal,intentclarification,leadingquestionsandcross-layertaskrouting.Italso adoptsanend-to-endtrainableoptimizationframeworkandusesreinforcementlearningforcontinuous iteration.Thisdistinguishesitfrompassive"confidencedetectionplusfixedrules"methodsandenables data-drivendynamicadaptationtocomplexandchangingindustrialconditions. Thisparadigmcanbeimplementedthroughtwopaths.Thefirstisexplicitmodelingofindecisivefault- tolerantactionsatknowledgeboundaries.Throughdedicatedalignmenttraining,lightweightagentslearn toidentifytheirowncapabilityboundariesandchoosefault-tolerantactionsratherthangivingforced answerorexecution,whentaskparametersaremissing,scenariosareunseen,orataskfallsoutsidetheir knowledgescope[4,5].ThesecondistheDiSRouter(DistributedSelf-Router)mechanism[14],which givesagentstheabilitytoassesstheirowncompetenceandautonomouslydecidewhethertoexecuteatask locallyordelegateitupwardtohigher-capabilityagents,therebyavoidingexecutionfailurescausedby capabilitymismatch.Refusalandroutingarerepresentativefault-tolerantactions.Intentclarification, uncertaintyannotation,andotherdiverseindecisiveactionsremainimportantresearchdirectionsfor single-agentfault-toleranceoptimization. 5.ParadigmTwo:ReliableCross-AgentCommunicationviaSemi-FormalLanguageProtocols Single-agentlocalreliabilitydoesnotguaranteereliableglobalcollaborationacrossdistributedagents. Currentmulti-agentcollaborationframeworksembedalltaskconstraintsandhandoffrulesentirelywithin natural-languageprompts,whichproducesvagueandunverifiablegovernancerules.Minorsemantic deviationsatanysinglenodecanpropagateandamplifyalongthecollaborationchainandeventuallylead toglobaltaskfailure.Thisisacorechallengethatcannotbeaddressedbysingle-nodefaulttolerancealone. Therootcauseofagentcommunicationerrorsistheinherenttrade-offbetweennaturallanguage expressivenessandformalprotocolverifiability.FromtheperspectiveofChomsky'slanguagehierarchy [19,20],agentcommunicationalwaysfacesatrade-offbetweenexpressivenessandverifiability.Pure naturallanguageisflexiblebutambiguous;pureformallanguagesarerigorousbutlimitedinsemantic expressivenessfordynamicindustrialscenarios.LLMsâparsingperformancealsodegradeswithincreasing promptcomplexity.Reliableorchestrationthereforerequiresexternalprotocol-basedcontrolmechanisms, notjustinternalmodelcapabilities. Wearguethatreliablecross-agentcollaborationshouldbebuiltaroundanexecutableprotocol programmingsystembasedonsemi-formallanguageprotocols[15].Itscentralgoalistoblockerror propagationalongthecommunicationchainandensuresystem-levelcommunicationstability.Thekey designideaistoconvertimplicitrulesembeddedinnatural-languageprompts,suchastaskconstraints, transitionpolicies,andhandoffresponsibilities,intostandardizedformalprotocolstructuresthatare checkable,executable,andauditable.Thisseparatesexplicitconstraintauditingfromimplicitgeneral reasoningwhilepreservingthegeneralsemanticreasoningcapabilityofnaturallanguage,thereby balancingexpressiveflexibilitywithmachineverifiability. Thisparadigmbuildsadualfault-toleranceassurancemechanismforagentcommunication.Atcompile time,syntacticverification,logicchecking,andparametervalidationdetectandblockstructuralerrors beforeexecution.Atruntime,guardmechanismsconstrainagentinteractionbehaviorinrealtime,block ambiguousinformationandinvalidorrule-violatingoperations,andkeepthecollaborativeprocesson reliableexecutionpathstopreventerrorpropagation.Thisparadigmcomplementssingle-agentfault tolerance,addressesthesystemicproblemof"locallyreliablebutgloballyfailing"collaboration,and supportsstablelong-horizonoperationofDAS. 6.IndustrialDeploymentChallengesandFutureOutlook Thedistributedagentsystemframeworkisstillinthetransitionfromtheorytoindustrialpractice,with fourcorebottlenecksrestrictinglarge-scaledeployment:missingstandardizedinterfacesforheterogeneous agents;highoverheadofexistingfault-tolerantalgorithmsandprotocolverification;absenceofunified industrialstandardsforsemi-formalcommunicationprotocols;andunderdevelopedindustrialreliability evaluationsystemsforlong-horizondistributedagentcoordination. InresponsetothesebottlenecksandinlinewiththedevelopmentoftrustworthyindustrialAI,future researchshouldfocusonsixdirections:1)dynamicreliabilityevolutionmechanismsforheterogeneous end-deviceagents;2)protocol-drivendistributedcollaborationdeploymentframeworks;3)lightweight fault-toleranceandverificationalgorithms;4)industrystandardsforsemi-formalprotocols;5)open-source evaluationecosystemsforend-deviceagents;andmulti-levelfault-toleranceassuranceforindustrialsafety [16,17,9]. Meanwhile,thetheoreticalframeworkproposedhereisbroadlyadaptable.Beyondheterogeneousend- devicescenarios,itcanalsobeextendedtolong-horizoncollaborationamongmulti-agentsystemsbuilt aroundcloud-hostedfrontierLLMs.Fault-tolerantcollaborationunderhybridcloud-hosteddeployments thatcombinesmallandlargemodelsisaparticularlyvaluablefutureresearchdirection[18]. 7.ConcludingRemarks WeintroducetheDistributedAgentSystem(DAS)conceptandproposeafault-tolerantcollaboration paradigmforembodiedagentsbasedontherecognitionthatuncertaintyisobjectivelyunavoidable.It replacesthepursuitofsingle-turnzeroerrorwithsystem-levelfault-tolerancemanagement,therebyfilling acriticalgapbetweenresearchonmulti-agentsystemsbuiltaroundcloud-hostedfrontierLLMsandagent collaborationinconstrainedindustrialend-devicescenarios.Byredefiningagentreliability,itbreakswith thetraditionalLLM-agentoptimizationlogicwhichoveremphasizesaccuracyanderrorelimination. Single-agentfault-tolerantreliabilityalignmentaddresseserrorsattheirsourceinindividualend-device agents,whilesemi-formallanguageprotocolsaddresserrorpropagationinmulti-agentcommunication. Together,thetwoparadigmsformanintegratedfault-toleranceframeworkfortrustworthyindustrialAI.As lightweightalgorithmdesign,protocolstandardization,andevaluation-systemstandardizationcontinueto advance,thisframeworkcanbroadlysupportinterconnectionamongintelligentenddevices,industrialIoT, andothercorescenarios,whileprovidingnewtheoreticalsupportandengineeringpathwaysforreliable long-horizonmulti-agentcollaborationacrossbroaderapplicationcontexts. References [1]LiX,ShiW,ZhangH,PengC,WuS,TongW.TheAgentic-AICore:AnAI-Empowered,Mission-OrientedCoreNetworkforNext-Generation MobileTelecommunications.Engineering2026;56(1):104-119. [2]MaN,PanJ,LiuY,YangY,HanY,GuoJ,WuZ,YangZ,YangZ,LiD.EmbodiedInteractiveIntelligenceTowardsAutonomousDriving. Engineering2026;59(4):337-351. [3]FanD,LiuM,ShaoY,YangL,LiuY,ZhangY,RenY,WangZ.Domain-SpecificLargeLanguageModelforMaintenanceDecision-Makingon WindFarmsbyLabeled-Data-SupervisedFine-Tuning.Engineering2026;60(5):343-361. [4]XuH,ZhuZ,PanL,WangZ,ZhuS,MaD,CaoR,ChenL,YuK.Reducingtoolhallucinationviareliabilityalignment.In:Proceedingsofthe 42ndInternationalConferenceonMachineLearning;2025Jul13-19;Vancouver,Canada.p.69992-70006. [5]XuH,ZhuZ,ZhangS,MaD,FanS,ChenL,YuK.Rejectionimprovesreliability:trainingLLMstorefuseunknownquestionsusingRLfrom knowledgefeedback.In:Proceedingsofthe1stConferenceonLanguageModeling;2024Oct7-9;Philadelphia,PA,USA. [6]XuH,YangZ,ZhuZ,LanK,WangZ,WuM,JiZ,ChenL,FungP,YuK.Delusionsoflargelanguagemodels.2025.arXiv:2503.06709. [7]LiuY,ZhouY,LiuY,XuZ,HeY.IntelligentFaultDiagnosisforCNCThroughtheIntegrationofLargeLanguageModelsandDomain KnowledgeGraphs.Engineering2025;53(10):311-322. [8]ChenJ,ChenY,PhamDT,SongY,WuJ,XingL,ChenY.ALargeLanguageModel-basedMulti-AgentFrameworktoAutonomouslyDesign AlgorithmsforEarthObservationSatelliteSchedulingProblem.Engineering.Inpress. [9]Camarinha-MatosLM.CollaborativeNetworksinIndustry4.0andIndustry5.0.Engineering.Inpress. [10]ChenZ,YuW,WuC,ChenF,WangZ,ZhouC,YouY,LiS,ZhuQ,MaN,SunY,LiD,FanadyB,JiangS,YanZ,ZhouS,LiL,HsiehC,Bai Y,XiaoL,ChungC,ChanC,CuiZ,GrĂ€tzelM,ZhaoH.AgenticRoboticBoxesforPerovskiteSolarCellFabricationwithRecipeLanguageModel. Engineering2026;61(6):365-374. [11]PiantadosiST,TilyH,GibsonE.Thecommunicativefunctionofambiguityinlanguage.Cognition2012;122(3):280-291. [12]ZipfGK.HumanBehaviorandthePrincipleofLeastEffort:AnIntroductiontoHumanEcology.Addison-WesleyPress,1949. [13]RenJ,SunY,DuH,YuanW,WangC,WangX,ZhouY,ZhuZ,WangF,CuiS.GenerativeSemanticCommunication:Architectures, Technologies,andApplications.Engineering2026;56(1):45-61. [14]ZhengH,XuH,LinYK,FanS,ChenL,YuK.DiSRouter:distributedself-routingforLLMselections.In:Proceedingsofthe14thInternational ConferenceonLearningRepresentations;2026Apr23-27;RiodeJaneiro,Brazil. [15]LiH,PengJ,WangZ,ChenL,YuK.XFlow:anexecutableprotocolprogrammingsystemforreliablemulti-agentworkflows.2026. arXiv:2606.14790. [16]HuangK,LiangC,HuaM,ZhanX,GeM.Hierarchicalthree-dimensionaloutputtrackingfornetworkeduncertainroboticsystemswith Byzantinefault-tolerantcapability.NonlinearDynamics2026,114(9):627. [17]WuQ,DuS.Fixed-timefault-tolerantconsensusofuncertainmultiagentsystemsviafullydistributedcontrol.IETControlTheory& Applications2026;20(1):e70123. [18]WuF,ShenT,BĂ€ckT,ChenJ,HuangG,JinY,KuangK,LiM,LuC,MiaoJ,WangY,WeiY,WuF,YanJ,YangH,YangY,ZhangS,Zhao Z,ZhuangY,PanY.Knowledge-Empowered,Collaborative,andCo-EvolvingAIModels:ThePost-LLMRoadmap.Engineering2025;44:87-100. [19]ChomskyN.Threemodelsforthedescriptionoflanguage.IRETransactionsonInformationTheory1956;2(3):113-124. [20]ChomskyN.TheMinimalistProgram.Cambridge,MA:MITPress;2014. [21]YoungS.CognitiveUserInterfaces.IEEESignalProcessingMagazine2010;27(3):128-140. [22]YoungS,GasiÄM,KeizerS,MairesseF,SchatzmannJ,ThomsonBandYuK.TheHiddenInformationStateModel:apracticalframeworkfor POMDP-basedspokendialoguemanagement.ComputerSpeechandLanguage2010;24(2):150-174.