版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
1
arXiv:2608.01964v1[cs.CV]3Aug2026
LongHorizon-Harness:AdvancingLong-HorizonAgentsforReal-WorldTasks
ZiyuMa*,HailangHuang*,ShunZou*,YongWang*†,ShidongYang,YimingHu,FeiWei,XiangxiangChu
DreamXTeam,AlibabaGroup*Equalcontribution†Projectlead
Largelanguagemodel(LLM)agentsincreasinglyundertakelong-horizontasksthatrequiresustainedreasoning,tooluse,andrevisionacrossmanyinterdependentsteps.However,existingagentharnessesmaintaintaskexecution,taskstate,andcompletionassessmentwithinagrowingcontext,makingthestatedifficulttotrackandallowingincorrectself-assessmentstopropagateintolaterdecisions.Wereformulatelong-horizonexecutionasatask-statemanagementproblemandproposeLongHorizon-Harness,whichmaintainsthetaskstateexplicitlyoutsideexecutionandupdatesitonlywithfactsindependentlyverifiedfromtheenvironment.ItsManage-Execute-Audit(MEA)loopusesamanagertomaintainthetaskstateanddeterminethenextsubtask,afresh-contextexecutortoperformit,andaread-onlyauditortoverifytheresultingenvironmentstatebeforethenextround.AlightweightAgentAdaptersupportsinterchangeablemodelandharnessbackendswithoutmodifyingtheirnativeagentloops.LongHorizon-HarnessimprovesQwen3.7-Plusfrom51.8%to80.7%onWeaveBench,from69.7%to77.2%onTerminal-Bench2.1,andfrom2.8%to8.3%onOSWorld2.0.ItalsoraisesClaudeOpus4.7from20.0%to34.3%onanOSWorld2.0subset,demonstratingconsistentgainsacrossmodels,harnesses,andinteractiondomains.
Github:
/AMAP-ML/LongHorizon-Harness
Website:
https://lh-harness.pages.dev
Figure1Left:LongHorizon-Harnessimproveslong-horizonexecutionacrossbenchmarksandbackbones.Withthesamebackboneandexecutionbackend,itliftsWeaveBenchPassRatefrom51.8%to80.7%,Terminal-Bench2.1from69.7%to77.2%,andOSWorld2.0binarycompletionby3.0×,andthegainstransferfromQwen3.7-PlustoClaudeOpus4.7.Right:Auditedstatetransitions.Insteadofonecontinuouslygrowingsessionthatjudgesitsownprogress,amanagerre-plansthenextsubtaskfromauditedfacts,afresh-contextexecutorperformsit,andaread-onlyauditorcertifieswhatactuallychangedintheenvironment.Auditreportsaretheonlycross-roundmemory,andinterchangeablebackends(e.g.,ClaudeCode,Codex)serveeachrole.
2
1Introduction
Overthepastfewyears,largelanguagemodels(LLMs)haveevolvedfromconversationalmodelsintothedecision-makingcoreofautonomousagentsforsoftwareengineering(
Yangetal.
,
2024
;
Wangetal.
,
2025
;
Renetal.
,
2026
;
Dingetal.
,
2025
),general-purposeassistance(
Anthropic
,
2025
;
OpenAI
,
2025a
;
b
;
Anthropic
,
2026
),scientificdiscovery(
Sunetal.
,
2025
;
Weietal.
,
2025a
),computeruse(
Anthropic
,
2024
;
OpenAI
,
2025c
;
GoogleDeepMind
,
2025
;
Sageretal.
,
2026
;
Zhouetal.
,
2026
;
Zhengetal.
,
2026
),andmultimodalinteraction(
Zhangetal.
,
2023
;
Agasheetal.
,
2025
).Acrosstheseapplications,agentsincreasinglyfacelong-horizonexecution,whichrequiresrepeatedreasoning,tooluse,observation,andrevisionovermanyinterdependentsteps,sometimesacrossmultiplecontextwindowsorsessions.Thelengthoftasksthatagentscancompleteincreasinglydetermineshowmuchworkcanbedelegatedtothem.METR(
Kwaetal.
,
2026
)reportsthatthetask-completionhorizonofadvancedagentshasdoubledroughlyeverysevenmonths,withthetrendacceleratingtoaboutfourmonthsforrecentmodels.Frontiercodingagentscanalreadysustainhours-longworkonasingleproject(
OpenAI
,
2025a
;
Anthropic
,
2025
;
Yangetal.
,
2026a
).However,alongerhorizondoesnotbyitselfmakeexecutionreliable(
Dongetal.
,
2026
).
Thedifficultyoflong-horizonexecutionliesnotinanyindividualstep,butinsustainingcoherentprogressacrossalongsequenceofinterdependentactions.Acrosssystemsanddomains,threechallengesconsistentlyemerge:(i)Compoundingerrorsandgoaldrift.Errorsinearlieractionsordecisionsaccumulatealongthetrajectory,distortsubsequentchoices,andgraduallysteertheagentawayfromitsoriginalobjective(
Sun
etal.
,
2026
).(ii)Contextrot.Astheinteractionhistorygrows,relevantinformationbecomesincreasinglydifficulttoretrieveanduse,andagentperformancecandegradesharplyoncecontextutilizationcrossesacriticalthreshold(
Liuetal.
,
2024
;
Hongetal.
,
2025
).(iii)Task-stateloss.Long-horizontasksaredifficulttocompletewithoutanaccurateandup-to-datetaskstate(i.e.,therequirementstosatisfy,actionsalreadycompleted,artifactsproduced,andfactsdiscoveredfromtheenvironment),butagentsoftenfailtorecover,retain,andupdatethisstatethroughouttaskexecution.
Extensiveeffortshavestrengthenedagentsfromboththemodelandharnesssides.Frontiermodelscontinuetoscaleinsize,extendtheircontextwindows,andacquirestrongercodingandagenticcapabilitiesbytrainingonlarge-scale,high-qualitydata(
Weietal.
,
2025b
;
Chuetal.
,
2026
;
Lietal.
,
2025
;
Maetal.
,
2026
;
Jietal.
,
2025
;
Yangetal.
,
2026b
).Inparallel,agentharnessessuchasClaudeCode(
Anthropic
,
2025
),CodexCLI(
OpenAI
,
2025a
),andOpenClaw(
OpenClawContributors
,
2026
)havebecomethestandardsystemlayerfororganizingprompting,tooluse,contextmanagement,andmulti-stepexecutionaroundthemodel.Forlong-horizontasks,theseharnessesalreadysupportplanning,taskdecomposition,tooluse,andsubagentsthatexecuteorreviewworkinisolatedcontexts.However,existingharnessesstillfacetwostructurallimitations:(i)Taskexecutionandtask-statemanagementsharethesamegrowingcontext.Theagentusesthesamecontexttoexecutethetaskandmaintainitstaskstate,whilethegrowingexecutionhistorymakestaskstateincreasinglydifficulttotrack.(ii)Taskexecutionandcompletionassessmentremaincoupled.Theagentperformseachsubtaskandjudgeswhetherithasbeencompleted,withanincorrectjudgmentpotentiallybeingrecordedaspartofthetaskstateandusedasapremiseforsubsequentdecisions.
Toaddresstheselimitations,weproposeLongHorizon-Harness,aframeworkthatorganizeslong-horizonexecutionasasequenceofindependentlyauditedtask-statetransitions.Ourkeyideaistomaintainthetaskstateasanexplicitrecordoutsidetaskexecution,updateitonlywithfactsindependentlyverifiedfromtheenvironment,andderiveeachnextsubtaskfromthecurrentrecordandtheoriginalgoal.Specifically,LongHorizon-HarnessfollowsaManage-Execute-Audit(MEA)loopasshowninFig.
1
.Amanagerreadsthecurrenttaskstateanddefinesonesubtaskwithitsdependencies,constraints,andacceptancecriteria.Anexecutorperformsonlythissubtaskinafreshcontext,whilearead-onlyauditorindependentlyinspectstheenvironmenttodeterminewhatchanged,whatwascompleted,andwhatremainsunmet(
Zhugeetal.
,
2024
;
Zhengetal.
,
2023
).Themanagerupdatesthetaskstatefromtheauditresultandbeginsthenextround,whiletheexecutor’sinteractionhistoryisdiscardedaftereachroundsothatonlycompact,verifiedtaskstatepersistsacrossthetask.ThroughalightweightAgentAdapter,LongHorizon-Harnesspreservesthenativeagentloops(
Yaoetal.
,
2023
)ofexistingsystemsandsupportsinterchangeablebackendsforallthreeroles,spanningmodelssuchasClaudeOpus,GPT,andQwenandharnessessuchasCodexCLI(
OpenAI
,
2025a
),ClaudeCode(
Anthropic
,
2025
),OpenClaw(
OpenClawContributors
,
2026
),andHermesAgent(
Nous
Research
,
2026
).
3
Figure2OverviewofLongHorizon-Harness.LongHorizon-Harnessprocessesalong-horizontaskTthroughrepeatedManage-Execute-Auditrounds,shownbythedashedboxandnumberedflow❶–❸.ThemanagerreadstaskstateSiandconstructssubtaskcontractci,whichspecifiesthegoal,acceptancecriteria,boundaryconstraints,andrelevantpriorevidence(❶).Themanagermayinsteadrequestuserinformationorauthorizationthroughtheaskroute.TheselectedGUIorCLIexecutorperformsthesubtaskinafresh,budget-boundedcontextandmodifiestheenvironment(❷).Theauditorindependentlyinspectstheresultingenvironmentthroughread-onlytoolsandproducesauditreportvi(❸).Themanagerusesvitoupdatethetaskstatebeforethenextround.
WeevaluateLongHorizon-Harnessonthreerecentlong-horizonbenchmarks,WeaveBench(
Lietal.
,
2026
),OSWorld2.0(
Yuanetal.
,
2026
),andTerminal-Bench2.1(
Merrilletal.
,
2026
),whichtogethercovercross-interfacecoordination,desktopworkflows,andchallengingcommand-linetasks.Undermatchedmodelbackendsandevaluationprotocols,LongHorizon-HarnessraisesWeaveBenchPassRatefrom51.8%to80.7%usingQwen3.7-PlusandClaudeCode,nearlydoublingthestrongestofficiallyreportedresultof41.2%obtainedbyClaudeOpus4.7withClaudeCode.OnTerminal-Bench2.1,itimprovesperformancefrom69.7%to77.2%withQwen3.7-Plus.OnthefullOSWorld2.0benchmark,itraisesbinarycompletionfrom2.8%to8.3%withQwen3.7-Plus;ona34-tasksubset,itimprovestheresultfrom20.0%to34.3%withClaudeOpus4.7.Ourcontributionsaresummarizedasfollows:
•Wereformulatelong-horizonexecutionfromasinglegrowingtrajectoryintoatask-statemanagementproblem.Thetaskstateismaintainedexplicitlyoutsideexecution,updatedonlywithindependentlyverifiedfacts,andusedtodetermineeachnextsubtaskundertheoriginalgoal.
•WedesignLongHorizon-Harness,whichrealizesthisprinciplethroughaManage-Execute-Auditloop.Ineachround,themanagerdefinesonesubtaskfromthecurrenttaskstate,theexecutorperformsitinafreshcontext,andtheauditorindependentlyinspectstheenvironmentbeforethemanagerupdatesthestateandbeginsthenextround.
•WeevaluateLongHorizon-HarnessonWeaveBench,OSWorld2.0,andTerminal-Bench2.1.WithQwen3.7-Plus,itimprovesWeaveBenchPassRatefrom51.8%to80.7%,Terminal-Bench2.1from69.7%to77.2%,andOSWorld2.0binarycompletionfrom2.8%to8.3%,whilealsodeliveringconsistentgainswithClaudeOpus4.7.
4
2Method
2.1Overview
Givenalong-horizontaskTandacomputerenvironment,LongHorizon-Harnessexecutesthetaskthroughasequenceofdynamicallydeterminedroundsratherthanasinglecontinuouslygrowingsession.Theharnessmaintainsanexplicittaskstateoutsidetaskexecutionandadvancesitonlywithevidenceindependentlyverifiedfromtheenvironment.Acrossrounds,onlythetaskstateanditssupportingauditreportspersist;theexecutor’srawinteractiontrajectoryisdiscardedaftereachround.Fig.
2
providesanoverviewoftheframework.
EachroundfollowsaManage-Execute-Audit(MEA)loop.LetSidenotethetaskstateavailableatthebeginningofroundi,ei—1∈εthecurrentenvironmentstate,andVi—1=(v1,...,vi—1)theaccumulatedauditreports.Themanagerconstructsaboundedsubtaskcontractci.Afresh-contextexecutorperformsthecontract,transformstheenvironmentfromei—1toei,andreturnsanexecutionreportoi.Aseparateauditortheninspectseithroughread-onlytoolsandproducesauditreportvi.ThemanagerincorporatesviintothenexttaskstateSi+1beforedeterminingwhetheranotherroundisrequired.Theloopendswhentheauditedstatesatisfiestheoriginaltask,nopermittedsubtaskcanadvancetheremainingrequirements,userinputisrequired,ortheroundbudgetisexhausted.
2.2Manager
Themanagerownsthepersistenttaskstateanddetermineshowthetaskshouldproceed.IthasaccesstotheoriginaltaskT,thecurrenttaskstate,andallaccumulatedauditreports,buthasnodirectinterfacetothecomputerenvironment.Itcannotobserveapplicationstate,inspectworkspacecontents,invokeGUIorCLItools,ormodifytheenvironment.Itsdecisionsarebasedentirelyonthetaskstateandenvironmentevidencerecordedbytheauditors.
Afterroundi,themanagerupdatesthetaskstateandproducesthenextcontroldecision:
(Si+1,qi+1,ci+1)=Φmgr(T,Si,Vi),(1)
whereVi=(v1,...,vi)andqi+1isoneofExEcuTE,DoNE,BLockED,andAsk.Thecontractci+1isreturnedonlywhenfurtherexecutionisrequired.
Task-stateupdate.Thetaskstateisastructuredcollectionoftask-relevantrecords:arequirementrepresentsanobjectiveorconstraintderivedfromtheoriginaltask,anartifactrepresentsanoutputcreatedormodifiedduringexecution,andafactrecordsenvironmentinformationneededbysubsequentrounds.Eachrecordismarkedascompleted,pending,blocked,oruntrusted,andretainsreferencestotheauditevidencesupportingitscurrentstatus.TheinitialstateS1isconstructedfromTwithitsrequirementsmarkedaspending.Aftereachround,themanagerappliestheverifiedfindingsinvitoSi,addingorupdatingthecorrespondingrecordswhileleavingunresolvedonespending,blocked,oruntrustedasappropriate.Executorclaimsdonotdirectlychangethepersistentstate:arecordismarkedascompletedonlywhensupportedbycleanauditevidence.
Next-subtaskconstruction.ThemanagercomparesSi+1againsttheoriginaltask,selectsanunresolvedobjectivethatcanbeadvancedfromthecurrentstate,andchecksitsdependenciesandprerequisites.Itthenconstructsaboundedcontractci+1specifyingtheimmediategoal,acceptancecriteria,boundaryconstraints,andthetask-staterecordsandpriorauditreportsrelevanttoexecutionandverification.ThecontractisroutedtoaGUIorCLIexecutoraccordingtotheprimaryenvironmenttransitionitrequires.ThemanagerreturnsDoNEwhentheauditedstatesatisfiesTwithoutunresolvedintegrityviolations,BLockEDwhennopermittedactioncanadvancetheremainingrequirements,andAskwhenprogressrequiresuserinformationorauthorization;otherwise,itreturnsExEcuTEtogetherwithci+1.
2.3Executor
Theexecutorperformsthecontractselectedbythemanagerandistheonlyrolepermittedtointentionallymodifytheenvironment.Inroundi,itreceivestheoriginaltaskT,thecurrenttaskstateSi,thesubtask
5
contractci,andonlythepriorauditreportsreferencedbythecontract.Ittransformstheenvironmentfromei—1toei:
(ei,oi)=Φexec(T,Si,ci;ei—1),(2)
whereoisummarizestheactionsperformed,resultingstate,artifactsproducedormodified,andissuesencounteredduringexecution.Thereportoidescribestheexecutor’soutcomebutdoesnotestablishthatthecontracthasbeencompleted.
Fresh-contextexecution.Eachexecutorinvocationrunsasafresh,budget-boundedepisodecontainingonlytheinformationsuppliedforthecurrentround.Itdoesnotreceivetherawinteractiontrajectoriesofearlierrounds.Withintheepisode,theexecutormayperformmultiplecyclesofplanning,environmentinteraction,observation,andrevision.Whentheepisodeends,itsrawtrajectoryandinternalreasoningarediscarded;onlyoiisforwardedforauditing.
GUIandCLIcapabilityboundaries.GUIandCLIexecutorsoperatethroughdifferentenvironmentinterfaces.TheGUIexecutorreceivesscreen-orientedcapabilitiessuchastakingscreenshots,clicking,scrolling,andenteringtext,andisresponsiblefortransitionscenteredonapplicationandinterfacestate.TheCLIexecutorreceivescapabilitiessuchasshellexecution,fileediting,coding,andtesting,andisresponsiblefortransitionscenteredonworkspace,process,andprogramstate.
Theharnessexposesonlytheenvironmentinterfaceandtoolsetassignedtotheselectedexecutorrole.Capabilitiesoutsidethatroleareunavailableunlessexplicitlyprovidedbyitsconfiguration.Thisseparatesresponsibilityfordifferentclassesofstate-changingactionswhileallowingthemanagertoselecttheinterfaceappropriateforeachcontract.
Backendexecution.Executorsareinstantiatedthroughacommonagent-adapterinterface.Givenacontract,arole-specificenvironmentinterface,andanexecutionbudget,theadapterlaunchesanexistingbackendsuchasClaudeCode,CodexCLI,orOpenClawasoneboundedepisode.Thebackendretainsitsnativeplanningandtool-useloopandmaynormallyinvoketheshell,editfiles,writecode,runtests,orinteractwithapplicationswhenthosecapabilitiesareexposedbyitsassignedrole.Theharnessdoesnotreplacethebackend’sinternalexecutionprocess;itcontrolsthesuppliedcontext,availabletools,environmentpermissions,executionbudget,andreturnedreport.
2.4Auditor
Theauditorindependentlyverifiestheenvironmentstateproducedbytheexecutor.Afterexecution,itreceivestheoriginaltaskT,thetaskstateSi,thecontractci,thepriorauditreportsreferencedbythecontract,andtheexecutorreportoi.Itdoesnotreceivetheexecutor’srawinteractiontrajectoryorinternalreasoning.Theauditorinspectstheresultingenvironmenteiandproduces
vi=Φaud(T,Si,ci,oi;ei),(3)
whereviisappendedtothepersistentaudithistoryandreturnedtothemanagerbeforethenextround.
Independentenvironmentinspection.Theauditorstartsfromafreshcontextthatexcludestheexecutor’srawinteractiontrajectoryandinternalreasoning.Itmayuseoitolocaterelevantfiles,windows,logs,processes,orotheroutputs,butdeterminescompletionbyindependentlycomparingtheresultingenvironmentagainstthegoal,acceptancecriteria,andboundaryconstraintsinci.AGUIauditorexaminesapplicationandscreenstatethroughobservation-orientedinteractions,whereasaCLIauditorusesnon-mutatingcommandsandinspectiontoolstoexaminefiles,metadata,logs,processes,tests,andworkspacestate.Inbothcases,auditconclusionsmustbesupportedbyevidenceobtaineddirectlyfromtheenvironmentratherthanbytheexecutor’scompletionclaim.
6
Table1ResultsonWeaveBench.Grayrowsdenoteofficialresultsreportedby(
Lietal.
,
2026
),usingthebestreportedthinkingmodeforeachbackbone;modelsotherthanGPT-5.5andClaudeOpus4.7wereevaluatedonlywithOpENCLAw.BlackrowsdenoteourrunswithQwen3.7-Plus.Ourrunsuserootprivilegesinsidethetaskvirtualmachine,whereastheofficialresultsusearegularuseraccountandarethereforeincludedasreferencepointsratherthanmatchedcomparisons.PRdenotesfull-taskPassRate(%),andOveralldenotesthemeanscoreoverall114tasks.TheremainingcolumnsreportPassRateacrosstheeightbenchmarkdomains.
Model
Harness
PR↑
Overall↑
DSK
DOC
GAM
WEB
DAV
OPS
SPA
DES
ClaudeOpus4.7
CLAuDECoDE
41.2
0.532
55.6
47.1
23.5
53.3
23.1
50.0
33.3
40.0
OpENCLAw
35.1
0.482
55.6
29.4
23.5
66.7
15.4
41.7
16.7
20.0
HERmEsAGENT
28.1
0.516
33.3
47.1
11.8
26.7
30.8
50.0
8.3
10.0
CoDExCLI
13.2
0.378
16.7
11.8
11.8
6.7
7.7
25.0
16.7
10.0
GPT-5.5
CoDExCLI
35.1
0.499
38.9
29.4
23.5
53.3
15.4
50.0
58.3
10.0
OpENCLAw
33.3
0.466
38.9
35.3
35.3
21.4
23.1
38.5
33.3
40.0
HERmEsAGENT
31.6
0.466
55.6
29.4
35.3
40.0
7.7
25.0
25.0
20.0
CLAuDECoDE
14.9
0.299
33.3
11.8
11.8
0.0
15.4
16.7
25.0
0.0
GPT-5.4
OpENCLAw
22.8
0.465
55.6
35.3
5.9
0.0
23.1
23.1
8.3
20.0
GPT-5.3-codex
OpENCLAw
18.4
0.456
33.3
23.5
29.4
0.0
7.7
16.7
8.3
20.0
GPT-5.2-codex
OpENCLAw
6.1
0.321
5.6
11.8
0.0
0.0
15.4
16.7
0.0
0.0
GPT-5.1-codex
OpENCLAw
1.8
0.226
0.0
5.9
0.0
0.0
7.7
0.0
0.0
0.0
Gemini3.1pro
OpENCLAw
1.8
0.223
0.0
0.0
0.0
0.0
0.0
8.3
8.3
0.0
Qwen3.5-397B-A17B
OpENCLAw
0.9
0.318
0.0
0.0
0.0
0.0
0.0
8.3
0.0
0.0
Qwen3-VL-8B-Think
OpENCLAw
0.9
0.092
0.0
0.0
0.0
0.0
8.3
0.0
0.0
0.0
GUI-Owl-1.5-32B
OpENCLAw
0.0
0.065
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
Qwen3.7-Plus
ClaudeCode
51.8
0.702
83.3
76.5
29.4
46.7
53.8
66.7
16.7
20.0
LongHorizon-Harness(ClaudeCode)
80.7
0.835
88.9
100.0
58.8
73.3
84.6
91.7
66.7
80.0
Read-onlyauthority.Theauditormaychangeitsobservationviewwhennecessaryforinspectionbutcannotmodifytask-relevantenvironmentstate.Itcannotcreate,edit,overwrite,move,ordeleteprotectedartifacts,executestate-changingcommands,orperformGUIactionsthataltertheresultunderinspection.Theharnessmonitorstask-relevantworkspaceandartifactstatethroughoutauditing;anydetectedmutationisrecordedasanintegrityviolation,andtheresultingreportcannotsupportacompletedtask-staterecord.
Audit-resultconstruction.Theauditreportvirecordsthreeclassesoffindings.First,itassignsacompletionstatusofcomplete,incomplete,orblockedbyevaluatingthecontract’sacceptancecriteria.Second,itassignsanintegritystatusofclean,suspect,orviolationbycheckingworkspacemutations,artifactvalidityandprovenance,andrelevantdeletionconstraints.Third,itrecordsthetask-stateupdatessupportedbytheinspection,includingverifiedfacts,supportingevidence,andremaininggaps.
Theauditormayproposechangestorequirement,artifact,andfactrecords,butthemanagerdetermineshowthesefindingsareincorporatedintoSi+1.Consequently,oiremainsanunverifiedexecutionsummary,whereasviprovidestheenvironment-groundedevidencethatmayadvancethepersistenttaskstate.
3Experiments
3.1ExperimentalSetup
Comparedconfigurations.WeuseQwen-3.7-Plusastheprimarybackbonemodel,andadditionallyevaluateClaudeOpus-4.7toassessthegeneralityacrossbackbonemodels.Foreachbackbonemodel,wecomparetheoriginalClaudeCodeevaluationframework(
Anthropic
,
2025
)withLongHorizon-Harness,wherethelatterusesClaudeCodeasitsexecutionbackend.Unlessotherwisespecified,themanager,executor,andauditorareinstantiatedwiththesamebackbonemodel,ensuringthatthecomparisonreflectstheeffectoftask-statemanagementratherthandifferencesinmodelcapability.
7
Table2ResultsonOSWorld2.0.Grayrowsdenoteofficialresultsreportedby(
Yuanetal.
,
2026
)underbatched-andsingle-actionsettings,whiletheblackrowdenotesLongHorizon-HarnesswithQwen3.7-Plusinthehybridsetting.Binaryisthepercentageoftaskswhosefinalbenchmarkscoreequals1,andPartialisthemeanbenchmarkscoreoverall108tasks.Boldmarksthebestofficialresult.
Model
Harness/Mode
Binary↑
Partial↑
ClaudeOpus4.8
Batchedactions
20.6
54.8
ClaudeOpus4.7
Batchedactions
18.2
48.9
GPT-5.5
Batchedactions
13.0
49.5
ClaudeOpus4.8
Singleaction
18.5
49.3
ClaudeOpus4.7
Singleaction
13.9
49.1
ClaudeSonnet4.6
Singleaction
8.3
41.5
MiniMaxM3
Singleaction
4.6
22.3
Kimi2.6
Singleaction
4.6
22.1
Qwen3.7-Plus
Singleaction
2.8
21.5
Qwen3.7-Plus
LongHorizon-Harness(hybrid)
8.3
35.2
Implementationdetails.LongHorizon-HarnessintegratesexistingagentbackendsthroughaunifiedAgen-tAdapterinterface.Eachroleisassignedanindependentexecutionbudget:theexecutorislimitedto1800secondsperround,whileboththemanagerandtheauditorarelimitedto300seconds.WesetthemaximumnumberofMEAroundstoNmax=25.
Benchmarksandmetrics.WeevaluateLongHorizon-Harnessonthreerecentbenchmarksthatcovercomple-mentaryformsoflong-horizonexecution.
WeaveBench(
Lietal.
,
2026
)consistsof114tasksthatrequirecoordinatedGUIandCLIinteractionswithinthesameworkflow.WefollowthestandardWeaveBenchevaluationprotocol,usingthetaskdefinitions,workspaces,runtimeassets,andjudgetemplatesreleasedbytheofficialbenchmarkforalltasks.WereportPassRate,definedasthepercentageoffullypassedtasks,andOverall,definedastheaveragescoreacrossalltasks.Inaddition,wereportPassRateforeachoftheeightdomainscoveredbythebenchmark.
OSWorld2.0(
Yuanetal.
,
2026
)consistsof108desktopworkflowtasks,withamedianhumancompletiontimeofapproximately1.6hours.Weusetheofficialosworld-v2-2026.06.24releaseanditsstandardDocker-basedVMinfrastructure.Wereporttwometrics:BinaryandPartialAccuracy.BinaryAccuracycountsataskassuccessfulonlywhenitsfinalscoreis1,whilePartialAccuracyistheaveragescoreoveralltasks.
Terminal-Bench2.1(
Merrilletal.
,
2026
)isdesignedtoevaluatechallengingandrealisticcommand-linetasks.WefollowtheevaluationsetupusedfortheQwen-seriesmodels,andevaluateonTerminal-Bench2.1usingHarborwiththeDockerbackend.Finalresultsreporttheaveragescoreoverthreeindependentrunspertask,andarecomparedagainsttheofficialleaderboardresults.
3.2MainResults
Effectivenessonlong-horizoncomputeruse.Table
1
reportstheresultsonWeaveBench.TheClaudeCodeharnesswithQwen3.7-PlusachievesaPassRateof51.8%andameantaskscoreof0.702.UsingthesamemodelandretainingClaudeCodeastheexecutorbackend,LongHorizon-Harnessincreasestheseresultsto80.7%and0.835,respectively.Thismatchedcomparisonisolatesthecontributionoftheadditionaltask-statemanagementlayer.Performancealsoimprovesacrossalleighttaskdomains,showingthatthegainisnotconcentratedinaparticularapplicationcategory.Sincetheofficiallyreportedconfigurationsuseadifferentprivilegesetting,weincludethemasreferencepointsbutbaseourmainconclusiononthematchedClaudeCodecomparison.
8
Figure3ResultsonTerminal-Bench2.1.LongHorizon-HarnessretainsClaudeCodeasitsexecutorbackendandimprovessuccessratefrom69.7%to77.2%.Valuesmarkedwith∗areexternallyreportedmetrics.
Table3OSWorld2.0Opus4.7subset(34tasks).BothrowsuseClaudeOpus4.7asthebackbone.
Model
Harness/Mode
Binary↑
Partial↑
ClaudeOpus4.7
SINGLEAcTIoN
20.6
55.8
ClaudeOpus4.7
LongHorizon-Harness(hybrid)
35.3
66.9
Table
2
reportstheresultsonOSWorld2.0.WithQwen3.7-Plus,LongHorizon-Harnessincreasesbinarycompletionfrom2.8%to8.3%andpartialscorefrom21.5%to35.2%.Thehigherpartialscoreshowsthattheagentsatisfiesalargerfract
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 小学五年级科学下册《我们的空投包》项目化教学设计(大象版)
- 高中信息技术必修一“数字化学习与创新”教学设计
- 初中地理九年级中考一轮复习课地球在宇宙中的位置教学设计
- 初中七年级英语Unit 5 Fun Clubs Section A Pronunciation教学设计
- 小学五年级综合实践活动教学设计-探秘新能源
- 2026年洲际酒店集团(中国)秋招面试题及答案
- 电商仓储试题及答案
- 2027年中考历史复习七年级专题训练-宋元时期
- 技术服务协议
- 2026中国预制菜行业消费行为分析与渠道创新战略研究
- 浙江省用于社会福利事业彩票公益金使用管理办法
- 胖东来员工民主决策机制
- 财务管理期末试卷及答案5套
- 2026年六西格玛黑带考试试题及答案
- 2024统编版二年级道德与法治上册全册单元测试卷(含解析)
- 档案审核人员管理制度
- 《艺术展览叙事策略与观众体验提升:跨学科研究的创新实践》教学研究课题报告
- 过敏性紫癜的健康宣教
- 【全科医学概论5版】全套教学课件【694张】
- 腹部推拿课件
- CKD患者饮食依从性分期干预方案实施
评论
0/150
提交评论