




版權(quán)說(shuō)明:本文檔由用戶(hù)提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請(qǐng)進(jìn)行舉報(bào)或認(rèn)領(lǐng)
文檔簡(jiǎn)介
arXivv[cs.CL]17AprarXivv[cs.CL]17Apr2023YujiaQinShengdingHuYankaiLinWeizeChenNingDing,GanquCui111,ZheniZeng1,YufeiHuang1,ChaojunXiao1,ChiHan3,YiRenFung3,YushengSu1,HuadongWang1,ChengQian1,RunchuTian1,KunlunZhu8,ShihaoLiang8,XingyuShen1,BokaiXu1,ZhenZhang1,YiningYe1,BowenLi1,ZiweiTang5,JingYi1,YuzhangZhu1,ZhenningDai1,LanYan1,XinCong1,JasonPhang4,ChengYang5,TongshuangWu6,HengJi3,ZhiyuanLiu1*,MaosongSun1*qyj20@AbstractHumanspossessanextraordinaryabilitytocreateandutilizetools,allowingthemtofulptinsurentedropriatetoolsWealsodiscusshowtotrainmodelsforimprovedtoolusecapabilitiesandfacilitatethegeneralizationintoollearning.Consideringthelackofaya1/OpenBMB/BMTools2Contents1Introduction42Background6 ive 2.3ParadigmShift 8 ing ToolorientedLearning 123ToolLearning14 3.2.2PlanningwithReasoning 17elsforImprovedToolLearning LearningfromFeedback 21 4ApplicationandExperiment24 4.2Experiments 275Discussion28 lMakerAIsEvolutionaryRole dgeConictsinToolLearning s 6Conclusion36Contributions36ACaseStudy51A.13DModels 513A.2Stock 53A.3MakingSlides 54A.4MovieHunter 59A.5SearchEngine 60A.6ChemicalsMining 61A.7CookingAssistant 62A.8AIPainting 63vigatingKnowledgeGraphs A.10ALFWorld 65ACalculator 67ather AOnlineShopping 68A.14Map 69 A.16Translation 72A.17Wikipedia 734+otenign+SEearchnoisMuofdfeiDl+ngRobCookiIwouldrecommend:model.sliceswithyogurt.+otenign+SEearchnoisMuofdfeiDl+ngRobCookiIwouldrecommend:model.sliceswithyogurt.1Introductioninhumanactivity.Sincethedawnofcivilization,toolshavebeenintegraltotheveryessenceofouromerydbettedusingshallowstatisticalmodelsoritysikeurrentlyspecializedtoolscanbePleasePleaserecommendsomebooksonpersonalfinanceandinvesting.PleasePleasemakeabananayogurtforme.PleasePleasedrawawatercolorstyleEiffelTowerpainting.2/Blendthebanana1/2/BlendthebananaStableDiffusion2.SimpleWealth1.The2.SimpleWealth3.…HereisawatercolorHereisawatercolorgeneratedbydiffusion5asanetalketalcyframeworkfortooluseinAIsystems(§2.1),followedbyacategorizationoftoolsfromtheperspectiveoftheuserinterface(§2.2).ThenwereviewtheAIparadigmshiftbroughtbyfoundationmodelsandhighlightthehensiveliteraturereviewforexistingexplorationintoollearningwhichisdividedintotwongftoollsvelyinallywediscussotherimportantresearchtopicsforapplyingourgeneralframeworktorealworldsllearningcon?ictsintoollearning,whichcanleadtoinaccurateandunreliablemodelpredictions.Weidentifytwowe62Backgroundussing2.1CognitiveOriginsofToolUsedtxtensionsofhumanbeingsjustasancientghtingmpleondtheanteriorsupramarginalgyrusactivationofobservingtooluseistypicalofhumansubjects,ofwhichrelatedtotheoriginsofcumulativetechnologicalevolution(e.g.,theimprovementintheef?ciencyandolApartfromtoolsinthephysicalworld,wecanalsoturntomoreabstracttoolbehavior.Takecognitivetools(Heyes,2018)asanexample:itreferstoanauxiliarytoolforthinkingandhelpslearnersachieven72.2ToolCategorization:AUser-InterfacePerspectiveThegrowingnumberandcomplexityoftoolsinourworldmakeitincreasinglyimportanttounderstandbenetsandpotentialforgrowthInthispaperweareparticularlyinterestedinthosetoolsthatcanbetools.ednglaninteractiveinterfaceievisualngcurvefornontechnicalusers.Fromthisviewpoint,toollearningwithfoundationmodelssharethesameprimarygoal,reusuallyattheyake8(b)GUI-basedToolsGraphicalUserInterface(GUI)(c)Program-basedToolsProgrammingInterface(a)(b)GUI-basedToolsGraphicalUserInterface(GUI)(c)Program-basedToolsProgrammingInterface(a)PhysicalInteraction-basedToolsPhysicalToolsObservationSoftwares/SDKDeveloperAgentRealworldKnowledgeGraphDatabaseObservationVisualoperationtoprogrammingoperationPhysicalworldtovirtualworldUserPhotoshopWebPhotoshopWebsndofelxiblyexecutingtoolsofdifferenttypesInthispaperwee2.3ParadigmShiftnNenkovaMcKeownAlthoughtrainedonmassivecorpora,fromwhichgenerallinguisticabilityandworldknowledgearelearned.Thisropriatetextualpromptscanyieldthedesiredoutputfornewureandnaturallanguageasthemediumtouniformlyperformvarious9ancepplicationsandschedulingmeetingsmodelsserveastranslatorsmakingcomplextasksmoreaccessibletoindividualswithoutspecializedlfromtheuseoftoollearningintherealrceof2.4ComplementaryRolesofSpecializedToolsandFoundationModelssfortoollearning:(1)MitigationforMemorization.Althoughfoundationmodelshavedemonstratedangeverytharelativelyshortcontextlyhatarecictasks,suchasWolfram3forscienti?ccalculation,throughtheutilizationoftailoredalgorithms.Insteadetoolstoationssuchashealthcareornancewhereexplainabilityiscritical2/en-us/microsoft-3653/istanttoadversarialattacksOverallincorporatingtoolsintotheolsMakingandReasoningAbilitiesFoundationmodelsaretrainedonvastamountsofdata,enablingthemtoacquireworldknowledgeacrossawiderangeofdomains.Ifproperlysteered,suchknowledgecannizesse2.5LiteratureReviewofToolLearning(i.e.,AIfortool).2.5.1Tool-augmentedLearningnspecicknowledgeandimprovetheirgenerationqualityResearchinthisareahasoraugmentationisthetextretrievertoolwhichdevelopsfromtheearlysparseretrieverSparckJonesRobertsonetal,1995)totheknowledgeretriever.Forinstance,kNN-LM(Khandelwaletal.,2020)combinesaPLMandak-nearestPLMinInstructionIAnswerTwitterTrendingis1.HotInstructionIAnswerTwitterTrendingis1.HotWeather2.NBAPLAYOFFS3.…Tweet“ILOVEPIZZA”Tool1.Openbrowser2.LoginTwitter3.Createatweet(a)Tool-AugmentedLearningTellTellTwitterTrendingFounFoundationResult1.HotWeather2.NBAPLAYOFFS3.…(b)Tool-OrientedLearningPlanning1.Openbrowser.2.LoginTwitter.3.Createatweet.InstructionundationInstructionundationeinratethenalanswerTheintuitionisthatinmOtherToolsApartfromthetextorimageretrieverresearchershaveexploredemcalculatortoperformbasicarithmeticteachPLMsbetterutilizeatool.Toolformer(Schicketal.,2023)extendstheideaofParisietal.(2022)translationsystem,Wikipediasearchingtool,andcalendar.HereweprovideanexemplarygenerationheetalbGaoetal.(2022)proposetoaugmentPLMswithPythoninterpreters.Speci?cally,givenTable1:Representativeworksoftool-augmentedlearning.Foreachwork,wespecifythetoolusedforName&RefAugmentationAugmentationMethodwaletallTOOLFORMERSchicketal23)MIND’SEYE(Liuetal.,2022)SHOWYOURWORK(Nyeetal.,2021)2.5.2Tool-orientedLearningedloitingfoundationmodelsvastEmbodiedRoboticLearning.Themostrepresentativeapplicationoftool-orientedlearningisroboticsendonsAsthecenterpieceforplanningandreasoning,PLMsarelimitedtoprocessingtextualinputs.Fortoolsmultimodalcapabilitieswithout?ne-tuning.Incontrast,othersexploredbuildingmultimodalfoundationAutomationforOtherTools.Besidesroboticlearning,tool-orientedlearninghasbeenappliedtootherscenarios,including(1)websearchautomation:WebGPT(Nakanoetal.,2021)interactswithasearchdrecordingimportantinformationToachievethistherchgivenhumaninstructions;(3)dialogue-basedimagedrawingandediting:toenableunderstandingandnctexistingonKimetalproposetopromptlargelanguagewithtoollearningGivenatoolexecutioncouldsolvethetask.Bothstreamsshareacommongoal,i.e.,toleveragethestrengthsofspecializedtoolsandfoundationmodelsfortargettasks(§2.4).Whileresearchintoollearninghas4/en-us/bing/apis/bing-web-search-api5https://huggingface.co6/Torantulino/Auto-GPT3ToolLearningtddiscuss3.1ComponentsofToolLearningplexhother3.1.1UnderstandingtheComponentsAPIerwithsanningristhenpassedtothecontrollertoassistitsdecision-making.Byobservingthisfeedback,thecontrolleriple3.1.2ConnectingtheComponentstenatingftandettoformxtormodeledwithcomplexneuralPerceivernedbackPerceivernedbackntrollerningFouningModelantructionHuHumancutionnmentedbackFigure4:Illustrationofthetoollearningframework,wherewedisplaythehumanuserandfourcorespC(at)=p9C(at|xt,Ht,q),(1)canp9C(at|xt,Ht,q)=p9C(at|Ti,xt,Ht,q)×p9C(Ti|xt,Ht,q),(2)JieJennarytongistondanactionsequenceatthatultimatelyfulllsthetask3.2TheGeneralProcedure:FromIntenttoPlan3.2.1UnderstandingIntentandToolsspectssunderstandingcapabilities,challengesstillexistinreal-worldtoollearningscenarios:(1)Understandingllengeisdealingwiththeinherentvaguenessandambiguityintheuseryngthemodelsgeneralizetomorediverseuserinstructions.Astheintentspaceistheoreticallyin?nite,itisalmostimpracticalforurposengofvelusageidesconcretetooluseZeroZero-shotPrompting:Hereweprovideatool(API)"forecast_weather(city:str,N:int)",whichcouldforecasttheweatheraboutacityonaspeci?cdate(afterNdaysfromtoday).ThereturnedryoutoanswerQuestion:"What’sthetemperatureinShanghaitomorrow?"returnforecast_weather("Shanghai",1)["temperature"]foriinrange(2):ifforecast_weather("London",i+1)["precipitation"]>0:returnTruereturnFalseitelApotentialsolutionistoaddanintermediatestageoftoolselection,which?rstretrievesasmallsetoftaremostsuitableforthetaskathandAnothersolutionisnetuningwhichoptimizesmodelseedforincorporatingtooldenitionsintheinputwhichshrinkstheinputlengthandacceleratesmodel3.2.2PlanningwithReasoningingptoacertainsize(Weietal.,2022b).Inparticular,foundationmodelswithtensorhundredsofbillionsofietalbElicitingReasoninginFoundationModels.DespitetheextensivestudyoftheconceptofreasoninginWasonKelleythenotionofreasoningasappliedtofoundationIntrospectiveIntrospectiveReasoningControllerFoundationModelEnvControllerFoundationModelEnv&HumanMulti-IterationExtrospectiveReasoningToolSetControllerFoundationModelEnv&HumanToolControllerFoundationModelEnv&HumanToolSetestheillustrationforsimplicity.gThevanillafew-shotpromptlearning(Brownetal.,2020),wherebymodelsareprovidedwithaprompte(CoT)prompting.Unlikevanillafew-shotpromptlearning,CoTadditionallyinsertsthereasoningtracerequiredtoderivethe?nalanswerforeachexampleintheprompt.Inthisway,CoTpromptsmodelstoiderangeoftasksincludingarithmeticreasoningessfultiveTheformerinvolvesgeneratingastaticplanoftoolusewithoutgtsIntrospectiveReasoningThiskindofreasoningdirectlygeneratesmulti-stepplansfortoolusewithoutramsforvisionmodelsAlthoughVisualChatGPThastheformofiterativereasoningineachintermediatestep,bleediateexecutionresultsAmorerationalapproachtoplanningistakingtheenvironmentEinto
溫馨提示
- 1. 本站所有資源如無(wú)特殊說(shuō)明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請(qǐng)下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請(qǐng)聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶(hù)所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁(yè)內(nèi)容里面會(huì)有圖紙預(yù)覽,若沒(méi)有圖紙預(yù)覽就沒(méi)有圖紙。
- 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
- 5. 人人文庫(kù)網(wǎng)僅提供信息存儲(chǔ)空間,僅對(duì)用戶(hù)上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對(duì)用戶(hù)上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對(duì)任何下載內(nèi)容負(fù)責(zé)。
- 6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容,請(qǐng)與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時(shí)也不承擔(dān)用戶(hù)因使用這些下載資源對(duì)自己和他人造成任何形式的傷害或損失。
最新文檔
- 家長(zhǎng)實(shí)踐活動(dòng)心得體會(huì)
- 一年級(jí)數(shù)學(xué)計(jì)算題專(zhuān)項(xiàng)練習(xí)集錦
- 四年級(jí)數(shù)學(xué)(小數(shù)加減運(yùn)算)計(jì)算題專(zhuān)項(xiàng)練習(xí)與答案匯編
- 管理超市的經(jīng)驗(yàn)分享簡(jiǎn)短
- 醫(yī)美廠家合同范例
- 第21課 教學(xué)設(shè)計(jì)-七年級(jí)上學(xué)期體育與健康
- 2025年有關(guān)教師培訓(xùn)心得總結(jié)材料
- 臨時(shí)購(gòu)貨合同范例
- 2025年教師學(xué)期工作總結(jié)范文
- 中石油柴油合同范例
- 河南省洛陽(yáng)市瀍河回族區(qū)2023-2024學(xué)年九年級(jí)上學(xué)期期末語(yǔ)文試題
- 2024年3月時(shí)政熱點(diǎn)(一)
- 2024年吉林省中考語(yǔ)文試卷答案解讀及復(fù)習(xí)備考指導(dǎo)
- 2024天津市衛(wèi)生健康委員會(huì)所屬天津市衛(wèi)生健康促進(jìn)中心選聘3人(高頻重點(diǎn)復(fù)習(xí)提升訓(xùn)練)共500題附帶答案詳解
- 咨詢(xún)行業(yè)人工智能應(yīng)用研究
- SLT 478-2021 水利數(shù)據(jù)庫(kù)表結(jié)構(gòu)及標(biāo)識(shí)符編制總則
- GB/T 44251-2024腿式機(jī)器人性能及試驗(yàn)方法
- MIL-STD-1916抽樣計(jì)劃表(抽樣數(shù))大
- 馬克思主義基本原理概論400道(考點(diǎn)提分)
- 研究生調(diào)劑合同
- 【翻譯知識(shí)】新聞標(biāo)題翻譯
評(píng)論
0/150
提交評(píng)論