| MMBench: | | | Is Your | | Multi-modal | | Model | | an All-around | | -------- | --- | --- | ------- | --- | ----------- | --- | ----- | --- | ------------- | Player? YuanLiu1,∗,HaodongDuan1,∗,‡,YuanhanZhang2,∗,BoLi2,∗,SongyangZhang1,∗, WangboZhao4,YikeYuan5,JiaqiWang1,ConghuiHe1,ZiweiLiu2,†,KaiChen1,† 4202 guA 02 ]VC.sc[ 5v18260.7032:viXra DahuaLin1,3,† | | | 1ShanghaiAILaboratory | | | 2NanyangTechnologicalUniversity | | | | | | --- | ------------------------------- | --------------------- | --- | --- | ------------------------------- | ------------------------------ | --- | --- | --- | | | 3TheChineseUniversityofHongKong | | | | | 4NationalUniversityofSingapore | | | | 5ZhejiangUniversity | | | ∗EqualContribution | | | ‡ProjectLead | | †CorrespondingAuthor | | | | --- | --- | ------------------ | --- | --- | ------------ | --- | -------------------- | --- | --- | Abstract Largevision-languagemodels(VLMs)haverecentlyachievedremarkableprogress, | | exhibiting | | impressive | multimodal | perception | | and reasoning | | abilities. However, | | --- | ------------------------------ | --- | ---------- | ---------- | ------------------------------------ | ------- | ------------- | ---------- | ------------------- | | | effectively | | evaluating | these | large VLMs | remains | a major | challenge, | hindering | | | futuredevelopmentinthisdomain. | | | | TraditionalbenchmarkslikeVQAv2orCOCO | | | | | Captionprovidequantitativeperformancemeasurementsbutlackfine-grainedabil- | | ityassessmentandrobustevaluationmetrics. | | | | | | Meanwhile,subjectivebenchmarks, | | | | --- | ---------------------------------------- | --- | --- | --- | --- | --- | ------------------------------- | --- | --- | suchasOwlEval,offercomprehensiveevaluationsofamodel’sabilitiesbyincor- poratinghumanlabor,whichisnotscalableandmaydisplaysignificantbias. In responsetothesechallenges,weproposeMMBench,abilingualbenchmarkfor | | assessingthemulti-modalcapabilitiesofVLMs. | | | | | | MMBenchmethodicallydevelops | | | | --- | ------------------------------------------ | --- | ----------------------------------------------------------- | --- | --------- | --------- | --------------------------- | --- | --- | | | acomprehensiveevaluation | | | | pipeline, | primarily | comprisedofthefollowingkey | | | | | features: | | 1. MMBenchismeticulouslycuratedwithwell-designedqualitycon- | | | | | | | trolschemes,surpassingexistingsimilarbenchmarksintermsofthenumberand | | varietyofevaluationquestionsandabilities;2. | | | | | | MMBenchintroducesarigorous | | | | --- | ------------------------------------------- | --- | --- | --- | --- | --- | -------------------------- | --- | --- | CircularEvalstrategyandincorporateslargelanguagemodelstoconvertfree-form predictionsintopre-definedchoices,whichhelpstoyieldaccurateevaluationre- | | sults | for | models with | limited | instruction-following | | capabilities. | | 3. MMBench | | --- | ----- | --- | ----------- | ------- | --------------------- | --- | ------------- | --- | ---------- | incorporatesmultiple-choicequestionsinbothEnglishandChineseversions,en- ablinganapples-to-applescomparisonofVLMs’performanceunderabilingual | | context. | | Tosummarize,MMBenchisasystematicallydesignedobjectivebench- | | | | | | | | --- | ----------------------------------------------------------- | --- | ----------------------------------------------------------- | --- | --- | --- | --- | --- | ------ | | | markforarobustandholisticevaluationofvision-languagemodels. | | | | | | | | Wehope | MMBenchwillassisttheresearchcommunityinbetterevaluatingtheirmodelsand | | facilitatefutureprogressinthisarea. | | | | TheevalutationcodeofMMBenchhasbeen | | | | | | --- | ----------------------------------- | --- | --- | --- | ---------------------------------- | --- | --- | --- | --- | | | integratedintoVLMEvalKit[14]. | | | | 1 | | | | | 1 Introduction Recently,notableprogresshasbeenachievedwithintherealmoflargelanguagemodels(LLMs). For instance, the latest LLMs, such as OpenAI’s ChatGPT and GPT-4 [37], have demonstrated remarkablereasoningcapabilitiesthatarecomparableto,andinsomecases,evensurpasshuman capabilities. DrawinginspirationfromthesepromisingadvancementsinLLMs,largevision-language models (LVLMs) have also experienced a revolutionary transformation. Notable works, such as 1ThisisarevisedversionreleasedinApril2024. ItdescribesMMBenchv1.1, arefinedversionofthe MMBench(withbetterdataquality).Pleaserefertohttps://arxiv.org/pdf/2307.06281v3forthe previousversion,whichisreleasedinAugust2023. Technicalreport Future Identity Prediction Function Reasoning Reasoning Celebrity Image Recognition Emotion Image Attribute Quality Recognition GPT4-V Image Attribute Gemini-Pro-V Scene Comparison Qwen-VL-Max Image Action InternLM-XComposer2 Style Recognition LLaVA-v1.5-13B Image Structuralized CogVLM-Chat-17B Topic Image-Text Understanding Yi-VL-34B Natural MiniCPM-V Relation Spatial Relationship Object Social Localization Relation OCR Physical R P e h l y a s t i i c o a n l Property Figure1: Resultsofeightrepresentativelargevision-languagemodels(VLMs)acrossthe20 abilitydimensionsdefinedinMMBench-test. GPT-4v [37], Gemini-Pro-V [44] and LLaVA [33], have demonstrated enhanced capabilities in imagecontentrecognitionandreasoningwithinthedomainofvision-languagemodels,exhibiting superior performance compared to earlier works. Nevertheless, a large proportion of the early studies [18, 56, 33] tend to emphasize showcasing qualitative examples rather than undertaking comprehensiveandquantitativeexperimentstothoroughlyassesstheirmodelperformance. Thelack ofquantitativeassessmentposesaconsiderablechallengeforcomparingvariousmodels. Recent studieshaveprimarilyexploredtwoapproachestoconductquantitativeevaluations.Thefirstapproach involvesutilizingexistingpublicdatasets[19,9]forobjectiveevaluation,whilethesecondapproach employshumanannotators[49,48]toperformsubjectiveevaluations. However,itisworthnoting thatbothapproachesexhibitsomeinherentlimitations. A multitude of public datasets, such as VQAv2 [19], COCO Caption [9], GQA [23], and OK- VQA[35],havelongservedasvaluableresourcesforthequantitativeevaluationofVLMs. These datasetsofferobjectivemetrics,includingaccuracy,BLEU,CIDEr,etc. However,whenemployedto evaluatemoreadvancedLVLMs,thesebenchmarksencounterthefollowingchallenges. 1. False NegativeIssues: Mostexistingevaluationmetricsrequireanexactmatchbetweentheprediction and the reference target, leading to potential limitations. For instance, in the VQA task, even if thepredictionis“bicycle”whilethereferenceansweris“bike”,theexistingmetricwouldassign a negative score to the prediction, resulting in a considerable number of false-negative samples. 2. LackingFinegrainedAnalysis: Currentpublicdatasetspredominantlyfocusonevaluatinga model’sperformanceonspecifictasks,offeringlimitedinsightsintothefine-grainedcapabilities ofthesemodels. Thus,theyprovideinsufficientfeedbackregardingpotentialdirectionsforfuture improvements. Giventheaforementionedchallenges,recentstudies,suchasOwlEval[49]andLVLM-eHub[48] proposehuman-involvedsubjectiveevaluationstrategies,aimingtoaddressexistingmethods’limita- tionsbyincorporatinghumanjudgmentandperceptionintheevaluationprocess. OwlEvalartificially constructs82open-endedquestionsbasedonimagesfrompublicdatasetsandemployshumananno- tatorstoassessthequalityofVLMpredictions. Similarly,inspiredbyFastChat[53],LVLM-eHub developsanonlineplatformwheretwomodelsarepromptedtoanswerthesamequestionrelatedto animage. Aparticipantthencomparestheanswersprovidedbytwomodels. Subjectiveevaluation strategiesoffernumerousbenefits. Theseincludeaccuratematching,wherehumanscanprecisely correlateapredictionwiththetarget,evenwhenexpressedindifferentwords,andcomprehensive assessment,wherehumansareinclinedtojuxtaposetwopredictionsconsideringmultiplefacets. The ultimatescoreiscomputedasthemeanscoreacrossdiverseabilities,facilitatingaholisticevaluation ofthemodel’scapabilities. WhilesubjectiveevaluationallowsforamorecomprehensiveassessmentofaVLM,italsointroduces newchallenges. Firstly,humanevaluationsareinherentlybiased. Consequently,itbecomeschalleng- ingtoreproducetheresultspresentedinaworkwithadifferentgroupofannotators. Also,existing subjectiveevaluationstrategiesfacescalabilityissues. Employingannotatorsformodelevaluation 2 aftereachexperimentisanexpensiveendeavor. Moreover,evaluationdatasetsofsmallsizescan resultinstatisticalinstability. Toensurearobustevaluation,collectingmoredatabecomesnecessary, whichinturndemandsasignificantamountofhumanlabor. Inlightofthechallengesfacedbyconventionalobjectiveandsubjectivebenchmarks,wepropose MMBench,asystematicallydesignedobjectiveevaluationbenchmarktorobustlyevaluatedifferent abilitiesoflargevision-languagemodels. Currently,MMBenchcontainsover3000multiple-choice questionscovering20differentabilitydimensions,suchasobjectlocalizationandsocialreasoning, forevaluatingvision-languagemodels. Eachabilitydimensionencompassesover125questions, with the quantity of questions per ability maintained at a roughly equal level. The distribution facilitatesabalancedandthoroughassessmentoftheseabilities. SincesomeexistingVLMshave limitedinstruction-followingcapabilityandcannotdirectlyoutputchoicelabels(A,B,C,etc.) for MMBenchquestions,theevaluationbasedonexactmatchingmaynotyieldaccurateandreasonable conclusions. Inordertoreducethenumberoffalse-negativesamplesduringanswermatching,we employGPT-4tomatchamodel’spredictiontocandidateschoicesinamulti-choicequestionand then output the label for the matched choice. We conduct a comparison between GPT-4-based choicematchingandhumanevaluations,anddiscoveredthatGPT-4canaccuratelymatchhuman assessmentsin91.5%ofcases,demonstratingitsgoodalignmentandrobustnessasachoiceextractor. Tomaketheevaluationmorerobust,weproposeanovelevaluationstrategy,namedCircularEval (detailsinSec.4.3). Wecomprehensivelyevaluate21well-knownvision-languagemodels(across differentmodelarchitecturesandscales)onMMBenchandreporttheirperformanceondifferent abilitydimensions. Theperformancerankingoffersadirectcomparisonbetweenvariousmodels andprovidesvaluablefeedbackforfutureoptimization. Insummary, ourmaincontributionsare three-fold: • Systematically-constructedDataset: TothoroughlyevaluatethecapacityofaVLM,wecarefully curatedadatasetcomprisingatotalof3,217meticulouslyselectedquestions, coveringadiverse spectrumof20fine-grainedskills. • RobustEvaluation: Weintroduceanovelcircularevaluationstrategy(CircularEval)toimprove the robustness of our evaluation process. After that, GPT-4 is employed to match the model’s predictionwithgivenchoices, whichcansuccessfullyextractchoicesevenfrompredictionsofa VLMwithpoorinstruction-followingcapability. • AnalysisandObservations: Weperformacomprehensiveevaluationofaseriesofwell-known vision-language models using MMBench, and the evaluation results can provide insights to the researchcommunityforfutureimprovement. 2 RelatedWork 2.1 MultimodalDatasets Large-scale VLMs have shown promising potential in multimodal tasks such as complex scene understanding and visual question answering. Though qualitative results so far are encouraging, quantitative evaluation is of great necessity to systematically evaluate and compare the abilities ofdifferentVLMs. Recentworkshaveevaluatedtheirmodelsonnumerousexistingpublicmulti- modalitydatasets. COCOCaption[9],Nocaps[3],andFlickr30k[51]providehuman-generated image captions and the corresponding task is to describe the image content in the form of text. Visualquestionansweringdatasets,suchasGQA[23],OK-VQA[35],VQAv2[19],andVizwiz[20], containquestion-answerpairsrelatedtothegivenimage,usedtomeasurethemodel’sabilityonvisual perceptionandreasoning. Somedatasetsprovidemorechallengingquestion-answeringscenariosby incorporatingadditionaltasks. Forexample,TextVQA[42]proposesquestionsabouttextshownin theimage,thusinvolvingtheOCRtaskinquestion-answering. ScienceQA[34]focusesonscientific topics,requiringthemodeltointegratecommonsenseintoreasoning. Youcook2[55]replacesimages withvideoclips,introducingadditionaltemporalinformation. However,theaforementioneddatasets aredesignedonspecificdomains,andcanonlyevaluatethemodel’sperformanceononeorseveral tasks. Besides,differentdataformatsandevaluationmetricsacrossdatasetsmakeitmoredifficultto comprehensivelyassessamodel’scapability. Yeetal.[49]constructedOwlEval,anevaluationset encompassingavarietyofvisual-relatedtasks,albeitofalimitedsize. Fuetal.[17]introducedMME, whichassessesaVLM’scapabilitiesfromvariousperspectivesatasmallscale. Divergingfromprior 3 works,inthispaper,wepresentanovelmultimodalbenchmark,MMBench. Wealsodeviseasuiteof evaluationstandardsaimedatensuringthestabilityandaccuracyoftheevaluationresults. 2.2 MultimodalModels BuildinguponthesuccessofLargeLanguageModels(LLMs)suchasGPTs[41,7,40],LLaMA[46], andVicuna[53],recentadvancementshavebeenmadeinmultimodalmodels. Flamingo[4],anearly attemptatintegratingLLMsintovision-languagepretraining,hasmadesignificantstrides. Tocondi- tioneffectivelyonvisualfeatures,itincorporatesseveralgatedcross-attentiondenseblockswithin pretrainedlanguageencoderlayers. OpenFlamingo[4]offersanopen-sourceversionofthismodel. BLIP-2 [28] introduces a Querying Transformer (Q-former) to bridge the modality gap between thefrozenimageencoderandthelargelanguageencoder. Subsequently,InstructBLIP[11]extends BLIP-2[28]withvision-languageinstructiontuning,achievingsuperiorperformance. MiniGPT- 4 [56] attributes the prowess of GPT-4 [37] to advanced LLMs and proposes the use of a single projectionlayertoalignthevisualrepresentationwiththelanguagemodel. LLaVA[33]alsoutilizes GPT-4togenerateinstruction-followingdataforvision-languagetuning. Thelearningparadigmand themultimodalinstructiontuningcorpusproposedbyLLaVAarewidelyadoptedbysubsequent works[32,8,2,10]. Duringtheinstructiontuning,Low-RankAdaptation(LoRA[22])hasbeen adopted by recent works [49, 12, 10] on language models to achieve better performance on mul- timodalunderstanding. Intherealmofproprietarymodels,theAPIsofmultiplepowerfulVLMs havealsobeenmadepubliclyavailabletoprosperdownstreamapplications,includingGPT-4v[37], Gemini-Pro-V[44],andQwen-VL-Max[6]. Afterconductingathoroughevaluationofthesemodels ontheproposedMMBench,weofferinsightsforfuturemultimodalresearch. 3 TheconstructionofMMBench ThreecharacteristicsdifferentiateMMBenchfromexistingbenchmarksformulti-modalityunder- standing: i)MMBenchadoptsimages/problemsfromvarioussourcestoevaluatediversifiedabilities inahierarchicaltaxonomy;ii)MMBenchperformsrigorousqualitycontroltoensurethecorrectness andvalidityoftestingsamples;iii)MMBenchisabilingualmulti-modalbenchmarkandenablesan apple-to-applecomparisonofVLMperformanceunderEnglishandChinesecontexts. Belowwewill delveintomoredetailsoftheconstructionofMMBench. 3.1 TheHierachicalAbilityTaxonomyofMMBench Humanpossessremarkableperceptionandreasoningcapabilities. Theseabilitieshavebeencrucial inhumanevolutionandserveasafoundationforcomplexcognitiveprocesses. Perceptionrefersto gatheringinformationfromsensoryinputs,whilereasoninginvolvesdrawingconclusionsbasedon thisinformation. Together,theyformthebasisofmosttasksintherealworld,includingrecognizing objects,solvingproblems,andmakingdecisions[36,16]. Inpursuitofgenuinegeneralartificial intelligence(AGI),vision-languagemodels(VLMs)arealsoexpectedtoexhibitstrongperception andreasoningabilities. Therefore,weadoptPerceptionandReasoningaslevel-1(L-1)abilitiesin ourtaxonomy. Afterthat,weincorporatemorefine-grainedabilitydimensionsintothetaxonomy, andcategorizethemintosixL-2andtwentyL-3abilitydimensions. Wedisplaytheabilitytaxonomy inFigure2andyoucanfinddetaileddefinitionsofeachfine-grainedabilityintheAppendix. 3.2 DataCollectionandQualityControl QuestionCollection. InMMBench, wecollectvision-languageQAsintheformatofmultiple- choiceproblemsforeachL-3ability. AproblemP correspondstoaquadruple(Q ,C ,I ,A ). Q i i i i i i denotesthequestion,C representsasetwithn(2≤n≤4)choicesc ,c ,...,c ,I correspondsto i 1 2 n i theimageassociatedwiththequestion,andA isthecorrectanswer. Thedata—includingimages, i choices,andquestions—aremanuallycollectedfrommultiplesourcesbyagroupofvolunteers. For eachL-3ability,wefirstsetanexamplebycompiling10∼20multiple-choicequestions. Thenwe enlistthevolunteers,allofwhomareundergraduateorgraduatestudentsfromvariousdisciplines,to expandtheproblemset. Theexpansionisbasedontheabilitydefinitionandpotentialdatasources, which include both public datasets and the Internet. According to the statistics, more than 80% of questions in MMBench are collected from the Internet. For the remaining 20% samples, the 4 d g [1 T Im 5 op 0 a P [ i 1 ] g c r P o e 3 h p 2 y e ] s r i t c y a [ l 1 R 5 e 0 F la u ] t n io c n tion A [1 t 2 R R 7 t ea ] e r so a i n I b d i s n e [ u n g o 4 ti t n ty 0 e i n 9 g ] [1 2 5 ] P r e d ictio R n F u t R u r e e L e [ a a o 3 s s g 0 o o ic S n t r 8 u ct u n r ali a z e i ] n l I m i a g e - T e xt n g U n d e rst a n d g i n [ 1 8 3] R [4 e 3 a R s 2 S e o o ] l c n i a R al e i t l a n t i i o [ o g 1 n 5 n 3] Phy R s e ic la a [ t 1 l io 2 n 5 N ] R at e u l [ a r 1 t R a A 5 i l o e t 4 n c t ] o [ r 1 g ib n 4 u i 9 t t e i ] on [1149] Q I [ m 2 u E I a 0 m a m l g 8 [ i a t 1 e o ] g y t 5 e i 0 on ] I m S ag ce e 2 n 4 e 0 e ] sr n a o o i C t p e c ] r 2 e 0 P 9 [ Pe [ [C F r 2 i r c P n 0 o e e e s 6 r - p s [ c - g 4 t 8 e i r 4 i p n a ] o t 6 s i n t i n ] o a e n n d ce] ] 0 2 ] e 7 c [ n a t s ni - n e o l g i t n p i S e d c [ e r n e [ i P a 1 r g 5 - e O ni C F R Re [ C c 2 o e 4 g L le 6 n o b ] i c r t [ O a i 1 i o t b l y 7 n i j z e 5 a c t ] t ion [ 0 R ] I m a ge St yle [ 1 5 4] ] 5 9 1[ n oiti n g o c e R n oit c A [1 2 6 ] C o m p a r iso n A ttr ib u te [1 2 e 5 la ] tio n S s p h a ip tia l Figure 2: Ability dimensions in MMBench. Currently, MMBench incorporates three levels of abilitydimensions,encompassing20distinctleafabilities. (a) Multi-Modal Questions (b) Filter TEXT-ONLY Filter WRONG Text-Only Inference Multi-Modal Inference with SOTA LLMs with SOTA VLMs GPT-4 Qwen-Max GPT-4v LLaVA-v1.5 Gemini-Pro Gemini-Pro-Vision … English Version (Original) ChineseVersion (Translated) QUESTION. Think about the magnetic force between the QUESTION. 考虑每对磁铁之间的磁力。 If the majority is Correct IfallVLMs are Wrong magnets in each pair. Which of the following statements is true? 以下哪个陈述是正确的？ A. The magnitude of the magnetic force is smaller in Pair 2. A. 第二对磁铁之间的磁力大小较小。 Human Verification B C pa . . i T T r h s h . e e m m a a g g n n i i t t u u d d e e o o f f t t h h e e m m a a g g n n e e t t i i c c f f o o r r c c e e i i s s t sm he a s ll a e m r e in i n P a b i o r t 1 h . B C . . 第 两 一 对 对 磁 磁 铁 铁 之 之 间 间 的 的 磁 磁 力 力 大 大 小 小 相 较 同 小 。 。 Figure3: TheconstructionofMMBench. (a). ThequalitycontrolstrategiesadoptedinMMBench; (b)AnillustrationofquestionsinMMBench-CN. imagesaregatheredfromthevalidationsetofpublicdatasets(iftheyexist)whilethequestionsare self-constructed,whichisnotsupposedtobeusedfortraining. IntheAppendix,welistdatasources usedincollectionandprovidevisualizationofsamplescorrespondingtoeachL-3ability. QualityControl. Rawdatacollectedfromvolunteersmayincludewrongorunqualifiedsamples. Duringinvestigation, wefindthatthereexisttwomajorpatternsforsuchsamples: i)theanswer tothequestioncanbeinferredwithtext-onlyinputs,whichmakesitinappropriateforevaluating themultimodalunderstandingcapabilityofVLMs; ii)thesampleissimplywrong,eitherwitha flawedquestion,choices,oranincorrectanswer. Wedesigntwostrategiestofilterthoselow-quality samples,whichisvisualizedinFigure3(a). Weadopt‘majorityvoting’todetecttext-onlysamples: datasamplesareinferredwithstate-of-the-artLLMs(GPT-4[37],Gemini-Pro[44],etc.). Ifmore thanhalfoftheLLMscananswerthequestioncorrectlywithtext-onlyinputs,thequestionwillbe manuallyverifiedandthenremovedifitisunqualified. Todetectwrongsamples,wealsoimplement anautomaticfilteringmechanism. Weselectseveralstate-of-the-artVLMs(includingbothopen- 5 Ground-Truth Answer MiniGPT-4-13B InstructBLIP-13B VisualGLM-6B Distribution (Rolling) Prediction Distribution Prediction Distribution Prediction Distribution A：26.4% A：23.2% A：36.0% A：31.5% B A B：26.4% B A B：42.1% A B：39.6% B A B：28.4% D C：25.1% D C：19.2% B D C：11.1% D C：19.8% C C C C D：22.1% D：15.5% D：13.3% D：20.3% Figure4: Thechoicedistributionofground-truthanswersandpredictionsofsampleVLMs(all CircularEvalrecords). Sincethereexistquestionswithonly2/3choicesinMMBench,thechoice distributionofground-truthisnotexactlyeven. source and proprietary ones), to answer all questions in MMBench . If all VLMs fail to answer thequestioncorrectly, weconsiderthisquestionpotentiallyproblematic. Suchquestionswillbe manuallycheckedandexcludediftheyareactuallywrong. Thequalitycontrolparadigmhelpsusto constructhigh-qualitydatasetsandcanalsobeusedtocleanotherexistingbenchmarks. MMBench-CN. We further convert the curated MMBench into a Chinese version. During the process, all content in questions and choices are translated to Chinese based on GPT-4, except forpropernouns,symbols,andcode. Allthosetranslationsareverifiedbyhumanstoensurethe validity. MMBench-CNenablesanapple-to-applecomparisonofVLMperformanceunderEnglish andChinesecontexts. AnexampleinMMBench-CNisillustratedinFigure3(b). 3.3 MMBenchStatistics DataStatistics. Inthepresentstudy,wehavegatheredatotalof3,217datasamplesspanningacross 20distinctL-3abilities. Wedepicttheproblemcountsofallthe3levelsofabilitiesinFigure2. To ensure a balanced and comprehensive evaluation for each ability, we try to maintain an even distributionamongproblemsassociatedwithdifferentabilitiesduringdatacollection,withatleast 125samplesforeachL-3category. DataSplits. Wefollowthestandardpracticeinpreviousworks[35]tosplitMMBenchintodev andtestsubsetsataratioof4:6. Forthedevsubset,wemakealldatasamplespubliclyavailable alongwiththegroundtruthanswersforallquestions. Forthetestsubset,onlythedatasamplesare released,whilethegroundtruthanswersremainconfidential. Toobtainthetestsubsetevaluation results,oneneedstosubmitthepredictionstoMMBenchevaluationserver. 4 EvaluationStrategy InMMBench,weproposeanewstrategythatyieldsrobustevaluationresultswithaffordablecosts. Todealwiththefree-formoutputsofVLMs,weproposeutilizingstate-of-the-artLLMsasahelper for choice extraction. We conduct extensive experiments to study the LLM-involved evaluation procedure. TheresultswellsupporttheeffectivenessofGPT-4asachoiceextractor. Wefurtheradopt anewevaluationstrategynamedCircularEval,whichfeedsaquestiontoaVLMmultipletimes (withshuffledchoices)andchecksifaVLMsucceedsinallattempts. WithCircularEval,wedeliver arigorousevaluationandmoreeffectivelydisplaytheperformancegapbetweenVLMs. 4.1 LLM-involvedChoiceExtraction InourinitialattemptstoevaluateonMMBenchquestions,weobservedthattheinstruction-following capabilitiesofVLMscanvarysignificantly. Thoughproblemsarepresentedasclearmultiple-choice questions with well-formatted options, many VLMs still output the answers in free-form text2, especiallyforVLMsthathavenotbeentrainedwithmultiple-choicequestionsorproprietaryVLMs forgeneralpurposes(GPT-4v,Qwen-VL-Max,etc.). Extractingchoicesfromfree-formpredictionsis straight-forwardforhumanbeings,butmightbedifficultwithrule-basedmatching. Tothisend,we designauniversalevaluationstrategyforallVLMswithdifferentinstruction-followingcapabilities: 2Forexample,themodeloutputcanbethe meaningofchoice“A” ratherthan “A”. 6 Perfectly Aligned w. Human Mis-Aligned w. Human Closed-Source LLMs Open-Source LLMs Table1: StatisticsofIFcapabilitiesofVLMs. Wereport theheuristicmatchingsuccessrateofVLMs,andtheaccuracy beforeandafterLLM-basedchoiceextraction. In‘X+Y’,X denotesthematching-basedaccuracy,Yindicatesthegainof usingLLMasthechoiceextractor. ModelName MatchRate DEVAcc ModelName MatchRate DEVAcc MiniGPT4-7B 85.7 47.9+8.8 MiniGPT4-13B 84.8 52.1+8.7 InstructBLIP-7B 93.6 57.1+4.3 InstuctBLIP-13B 93.7 58.4+5.6 Figure 5: Alignment rates be- IDEFICS-9B-Instruct 96.6 58.4+1.5 Qwen-VL-Chat 93.8 73.3+3.6 tween human and different MiniCPM-V 95.2 70.9+4.5 VisualGLM-6B 64.8 39.9+23.2 LLMs. ‘chatgpt’ is ‘gpt-3.5- GPT-4v 91.8 81.5+3.6 GeminiProVision 97.5 81.8+0.8 turbo’. Open-source LLMs are Qwen-VL-Plus 77.4 64.5+15.0 Qwen-VL-Max 96.0 82.0+3.2 ‘chat’variants. The original VL problem: Circular Evaluation Q: How many apples are there in the image? A.4; B. 3; C. 2; D. 1 GT: A 4 Passes in Circular Evaluation (choices with circular shift): 1. Q: How many apples are there in the image? Choices: A. 4; B. 3; C. 2; D. 1. VLM prediction: A. GT: A ✔ 2. Q: How many apples are there in the image? Choices: A. 3; B. 2; C. 1; D. 4. VLM prediction: D. GT: D ✔ 3. Q: How many apples are there in the image? Choices: A. 2; B. 1; C. 4; D. 3. VLM prediction: B. GT: C ✖ 4. Q: How many apples are there in the image? Choices: A. 1; B. 4; C. 3; D. 2. VLM prediction: B. GT: B ✔ VLM failed at pass 3. Thus wrong. Figure6: CircularEvalstrategy. InCircularEval,aproblemistestedmultipletimeswithcircular shiftedchoicesandtheVLMneedstosucceedinalltestingpasses. Inthisexample,theVLMfailed inpass3andthusconsideredfailedtheproblem. Step1. MatchingPrediction. Initially,weattempttoextractchoicesfromVLMpredictionsusing heuristicmatching. Weaimtoextractthechoicelabel(e.g.,A,B,C,D)fromtheVLM’soutput. If successful,weusethisastheprediction. Ifnot,weattempttoextractthechoicelabelusinganLLM. Step2.MatchingLLM’soutput. Ifstep1fails,wetrytoextractthechoicewithLLMs(gpt-4-0125 bydefault). WefirstprovideChatGPTwiththequestion,choices,andmodelprediction. Then,we requestittoalignthepredictionwithoneofthegivenchoices,andsubsequentlyproducethelabelof thecorrespondingoption. IftheLLMfindsthatthemodelpredictionissignificantlydifferentfrom allchoices,weaskittoreturnapseudochoice‘Z’.Inexperiments,wefindthatforalmostallcases weencountered,theLLMcanoutputavalidchoiceaccordingtotheinstruction. Foreachsample,we comparethemodel’slabelprediction(afterGPT’ssimilarityreadout)withtheactualgroundtruth label. Ifthepredictionmatchesthelabel,thetestsampleisconsideredcorrect. 4.2 LLMastheChoiceExtractor: AFeasibilityAnalysis Instruction following (IF) capabilities of VLMs vary a lot. We conduct pilot experiments to study the effectiveness of LLMs as the choice extractor. As a first step, we perform single-pass inferenceonallMMBenchquestionswithVLMsinourevaluationcoreset(definedinSec.5.2). WhilethereexistVLMsthatperfectlyfollowthemultiple-choiceformatandachievehighsuccess rates(> 99%)inheuristicmatching, allproprietarymodelsandasignificantproportionofopen- source VLMs failed to generate well-formatted outputs. In Table 1, we list the success rates of differentVLMsinheuristicmatching3. AmongallVLMs,VisualGLMachievesthelowestmatching successrate,whichismerely65%. ForthoseVLMs,incorporatingLLMsasthechoiceextractor leadstosignificantchangeinthefinalaccuracy. AnothernoteworthythingisthattheIFcapability and the overall multimodal understanding capability is not necessarily correlated. For example, OpenFlamingov2[4]demonstratestopIFcapabilityamongallVLMs,whilealsoachievingoneof theworstperformancesonMMBench(Table3). 3VLMsthatachieve>99%matchingratesarenotlisted,includingLLaVAseries,Yi-VLseries,mPLUG- Owl2,OpenFlamingov2,andCogVLM-Chat. 7 QualityandstabilityofLLMChoiceExtractors. ForVLMpredictionsthatcannotbeparsedby heuristicmatching,weadoptGPT-4asthechoiceextractor. Tovalidateitsefficacy,wefirstbuild asubsetoftheinferencerecords. EachiteminthesetisapairofquestionsandVLMpredictions, whichcannotbeparsedbystep-1matching. Wesample10%ofthosehardexamples(∼420samples), andaskvolunteerstoperformmanualchoiceextractiononthesedatasamples. Suchannotations enableustovalidatethechoiceextractionofLLMs,bymeasuringtheiralignmentrateswithhumans. Figure5reportsthealignmentrates(extractedchoicesareexactlythesame)betweenLLMsand humans. We find that a great number of LLMs can complete the task well and achieve decent alignmentratewithhuman. AmongproprietaryLLMs,GPT-4achievesthehighestlevelofalignment rate,whichis91.5%,whileGPT-3.5-TurboandQwen-Maxachievearound85%. Open-sourceLLMs achievemorediversifiedperformanceonthechoicematchingtask. InternLM2-7B[45]achievesan 87%alignmentrateandsignificantlyoutperformsotheropen-sourceLLMsandGPT-3.5-Turbo. In thefollowingexperiments,weadoptgpt-4-0125asthechoiceextractorduetoitssuperioralignment capability. Meanwhile,wealsonotethattheslightdifferenceintop-performingLLMs’alignment rateshaslittleeffectonthequantitativeperformanceofVLMs. 4.3 CircularEvalStrategy InMMBench,theproblemsarepresentedasmultiple-choicequestions. Suchformulationposesan evaluationchallenge: randomguessingwillleadto∼25%Top-1accuracyfor4-choicequestions, potentiallyreducingthediscernibleperformancedifferencesamongVLMs. Besides, wenoticed thatVLMsmayprefertopredictacertainchoiceamongallgivenchoices(Figure4),whichfurther amplifiesthebiasinevaluation. Tothisend,weintroduceamorerobustevaluationstrategytermed CircularEvaluation(orCircularEval). Underthissetting,eachquestionisfedtoaVLMN times (N isthenumberofchoices). Eachtime,circularshiftingisappliedtothechoicesandtheanswerto generateanewpromptforVLMs(exampleinFigure6). AVLMisconsideredsuccessfulinsolving aquestiononlyifitcorrectlypredictstheanswerinallcircularpasses. Inpractice,onceaVLM failsonacircularpasses,thereisnoneedtoinfertheremainingpasses,whichmakestheactualcost ofCircularEvallessthanN×underpracticalscenarios. CircularEvalcanachieveagoodtrade-off betweenrobustnessandcost. 5 EvaluationResults 5.1 ExperimentalSetup Forthemainresults,weevaluatevariousmodelsbelongingtothreemajorcategoriesonMMBench: (a)Text-OnlyGPT-4[37];(b)Open-SourceVLMsincludingmodelvariantsofOpenFlamingo[4], MiniGPT4[56],InstructBLIP[11],LLaVA[32],IDEFICS[26],CogVLM[47],Qwen-VL[6],Yi- VL[2],mPLUG-Owl[50],InternLM-XComposer[12],andMiniCPM-V[39];(c)ProprietaryVLMs includingQwen-VL-[Plus/Max][6],Gemini-Pro-V[44],andGPT-4v[37]. Forafaircomparison,we adoptthezero-shotsettingtoinferMMBenchquestionswithallVLMs,basedonthesameprompt. ForallVLMs,open-endedgenerationisadoptedtoobtaintheprediction,and‘gpt-4-0125’isusedas thechoiceextractor. IntheAppendix,weprovidedetailedinformationregardingthearchitectureand theparametersizeforallOpen-SourceVLMsevaluatedinthispaper,aswellasadditionalresultsfor moreVLMsundervarioussettings. WeconductalltheevaluationwithVLMEvalKit[14]. 5.2 MainResults CircularEval vs. VanillaEval. Before delving deeper into concrete evaluation results, we first compare our CircularEval (infer a question over multiple passes, consistency as a must) with VanillaEval (infer a question only once). In Table 2, we present the results with two evaluation strategiesonMMBench-dev. FormostVLMs,switchingfromVanillaEvaltoCircularEvalleads toasignificantdropinmodelaccuracy. Ingeneral,comparisonsunderCircularEvalcanreveala moresignificantperformancegapbetweendifferentVLMs. LLaVA-v1.5-13Boutperformsits7B counterpartby2.1%Top-1accuracyunderVanillaEval,whileamuchlargerperformancegap(4.7% Top-1)isobservedunderCircularEval. Asaspecialcase,theperformanceofOpenFlamingov2drops from36.7%toonly2.6%whenwemovefromVanillaEvaltoCircularEval. CircularEvalissucha challengingsettingthatitevenmakesstate-of-the-artproprietaryVLMs(GPT-4v,Qwen-VL-Max, 8 Table2: CircularEvalvs.VanillaEval. WereporttheCircularEvalTop-1accuracyandaccuracy drop(comparedtoVanillaEval)ofallVLMsonMMBench-dev. VLM Circular AccChange VLM Circular AccChange VLM Circular AccChange MiniGPT4-7B 32.7% -24.1% MiniGPT4-13B 37.5% -23.2% Yi-VL-6B 65.6% -9.8% InstructBLIP-7B 37.4% -24.0% InstructBLIP-13B 40.9% -23.0% Yi-VL-34B 68.2% -9.5% LLaVA-v1.5-7B 62.5% -11.2% LLaVA-v1.5-13B 67.2% -8.6% MiniCPM-V 64.8% -10.6% IDEFICS-9B-Instruct 37.2% -22.6% LLaVA-InternLM2-20B 72.8% -7.0% Qwen-VL-Plus 62.9% -16.6% VisualGLM-6B 36.1% -27.0% CogVLM-Chat-17B 62.4% -15.6% Qwen-VL-Max 76.4% -8.7% Qwen-VL-Chat 59.5% -17.4% mPLUG-Owl2 63.5% -8.7% Gemini-Pro-V 70.9% -11.7% OpenFlamingov2 2.6% -34.1% InternLM-XComposer2 79.1% -4.7% GPT-4v 74.3% -10.8% etc.) sufferfrom∼10%Top-1accuracydrops. Inthefollowingexperiments,weadoptthemore rigorousandwell-definedCircularEvalasourdefaultevaluationparadigm. WeexhaustivelyevaluateallVLMsonallexistingleafabilitiesofMMBench. InTable3,wereport themodels’overallperformanceandtheperformanceinsixL-2abilitiesonthetestsplit,namely Coarse Perception (CP), Fine-grained Perception (single-instance, FP-S; cross-instance, FP-C), Attribute Reasoning (AR), Logic Reasoning (LR), and Relation Reasoning (RR).4 The results offervaluableinsightsintotheindividualstrengthsandlimitationsofeachVLMinmulti-modal understanding. PerformanceonMMBench-test. WefirstconductasanitycheckbyinferringMMBenchques- tionswithGPT-4,usingtext-onlyinputs. Afterconductingtherigorousqualitycontrolparadigmin Sec.3.2,GPT-4demonstratesarandom-leveloverallaccuracy.Amongopen-sourceVLMs,InternLM- XComposer2[12]achievesthebestperformanceandsurpassotheropen-sourceorproprietarymodels byalargemargin,w.r.t.theoverallscore,demonstratingitssuperiorabilityinmultimodalunder- standing. After that, models adopting the architecture of LLaVA [33] (LLaVA series and Yi-VL series)alsoshowcasestrongoverallperformance,whichisjustinferiortothestate-of-the-artclosed- sourceGPT-4vandQwen-VL-Max. Withasmallparametersize(≤3B),MiniCPM-Vachievesover 60%Top-1accuracy,highlightingthepotentialofsmall-scaleVLMs. ModelsincludingMiniGPT, IDEFICS,VisualGLM,andInstructBLIPdemonstratesignificantlyinferiorperformancecomparedto otherVLMs,whileOpenFlamingov2showsrandom-levelperformanceduetothelackofinstruction tuning. LLMplaysavitalrole. Fromtheevaluationresults,wefindthatthelargelanguagemodel(LLM) adoptedplaysavitalroleintheVLMperformance. Forinstance,allLLaVAseriesVLMs(v1.5-7B, v1.5-13B,InternLM2-20B)adoptthesamevisionbackboneandaretrainedwiththesamemultimodal corpus,whileswitchingtheLLMfromVicuna-v1.5[53]tothemorepowerfulInternLM2-20B[45] leadstosteadyimprovementacrossallL-2capabilities(especiallysignificantforreasoningtasks). ThescalingalsoholdsforvariantswithdifferentsizesfromthesameLLMfamily. Byadoptingthe 13BvariantofVicunaratherthanthe7Bvariant,VLMsintheMiniGPT,InstructBLIP,andLLaVA v1.5seriesoutperformtheir7Bcounterpartsby8.3%,1.5%,and3.5%overallTop-1accuracieson theMMBench-testsplit,respectively. Performance on MMBench-CN. In Figure 7, we compare the performance of different VLMs on MMBench and MMBench-CN. Most VLMs display a lower performance on MMBench-CN comparedtotheresultsonMMBench,exceptOpenFlamingov2,VisualGLM,andQwen-VL-Plus. ThedifferencemaybeattributedtotheunbalancedEnglishandChinesecorporausedinthepretraining andinstruction-tuningofVLMsandtheircorrespondingLLMs. Wenoticethatmosttop-performing VLMsonMMBenchalsodisplayoutstandingperformanceunderthebilingualcontext. Thelargest EN-CNperformancegapformodelsthatachieve70+%Top-1accuracyonMMBenchisamere2%, ForInternLM-XComposer2,theaccuracyonlydropsbylessthan1%whenevaluatedonMMBench- CN.SuchanadvantagecanbeattributedtoutilizingLLMswithbetterbilingualcapabilitiesortuning theVLMwithmorebalancedcross-languagemultimodalcorpora. 5.3 Fine-grainedAnalysis Inthissection,wepresentmorefine-grainedanalysisbasedontheevaluationresults. 4Pleaserefertotheappendixformorefine-grainedresultsandMMBench-devsplitresults. 9 Table 3: CircularEval results on MMBench test set (L-2 abilities). Abbreviations adopted: LR for Logical Reasoning; AR for Attribute Reasoning; RR for Relation Reasoning; FP-C for Fine-grainedPerception(CrossInstance);FP-SforFine-grainedPerception(SingleInstance);CP forCoarsePerception. Modelsaresortedbytheascendingorderofoverallaccuracy(intra-group). Open-sourcemodelstaggedwith*incorporatein-housedatainmodeltraining. | Model | | Overall CP | FP-S FP-C | AR LR | RR | | ----- | --- | ---------- | --------- | ----- | --- | LargeLanguageModels | GPT-4-Turbo(0125)[37] | | 2.9% 0.6% | 1.2% 4.1% | 3.7% 4.9% | 7.4% | | --------------------- | --- | --------- | --------- | --------- | ---- | OpenSourceVLMs | OpenFlamingov2[4] | | 2.3% 1.1% | 3.5% 1.5% | 5.3% 0.0% | 2.7% | | ----------------- | --- | ----------- | ----------- | ---------- | ----- | | MiniGPT4-7B[56] | | 30.5% 37.0% | 31.8% 17.2% | 49.8% 9.2% | 25.6% | IDEFICS-9B-Instruct[26] 35.2% 48.3% 31.3% 29.6% 47.8% 11.4% 25.2% | VisualGLM-6B[13] | | 35.4% 40.2% | 38.5% 26.2% | 47.8% 19.6% | 29.5% | | ------------------- | --- | ----------- | ----------- | ----------- | ----- | | InstructBLIP-7B[11] | | 38.3% 46.7% | 39.0% 31.8% | 55.5% 8.7% | 31.0% | | MiniGPT4-13B[56] | | 38.8% 44.6% | 42.9% 23.2% | 64.9% 8.2% | 32.9% | InstructBLIP-13B[11] 39.8% 47.2% 42.9% 21.0% 60.4% 12.5% 38.8% | Qwen-VL-Chat*[6] | | 60.9% 68.5% | 67.7% 50.2% | 78.0% 37.0% | 45.7% | | ----------------- | --- | ----------- | ----------- | ----------- | ----- | | MiniCPM-V[39] | | 61.4% 65.6% | 69.4% 51.3% | 70.6% 35.3% | 59.7% | | LLaVA-v1.5-7B[32] | | 63.4% 70.0% | 68.0% 57.7% | 77.6% 33.2% | 56.2% | | mPLUG-Owl2[50] | | 63.5% 68.1% | 69.1% 55.8% | 78.4% 37.0% | 57.0% | CogVLM-Chat-17B[47] 63.6% 72.8% 66.6% 55.4% 71.4% 33.7% 62.0% | Yi-VL-6B*[2] | | 65.5% 72.8% | 72.9% 56.2% | 75.5% 41.3% | 55.4% | | ------------------ | --- | ----------- | ----------- | ----------- | ----- | | LLaVA-v1.5-13B[32] | | 66.9% 73.1% | 72.4% 60.3% | 75.5% 35.9% | 65.5% | | Yi-VL-34B*[2] | | 68.4% 72.0% | 78.0% 54.7% | 81.2% 38.6% | 68.2% | LLaVA-InternLM2-20B[10] 72.3% 78.3% 76.6% 68.2% 78.4% 46.2% 69.4% InternLM-XComposer2*[12] 78.1% 80.4% 83.5% 73.0% 83.7% 63.6% 74.4% ProprietaryVLMs | Qwen-VL-Plus[6] | | 64.6% 66.5% | 79.1% 50.2% | 73.9% 42.9% | 57.8% | | ---------------- | --- | ----------- | ----------- | ----------- | ----- | | Gemini-Pro-V[44] | | 70.2% 70.0% | 78.9% 65.9% | 82.9% 46.2% | 65.9% | | GPT-4v[37] | | 74.3% 77.6% | 73.8% 71.5% | 85.3% 63.6% | 68.6% | | Qwen-VL-Max[6] | | 75.4% 74.8% | 87.2% 67.0% | 85.3% 54.9% | 70.5% | Table 4: ‘Upper-bound’ Acc Es- Q. Who is the person Q. Based on the interaction | | | | in this image? | between the individuals in the | | | --- | --- | --- | -------------- | ------------------------------- | --- | timation for Proprietary VLMs. A. Leonardo Dicaprio image, what is the most likely social | | | | B. Steve Jobs | relation or event being depicted? | | | --- | --- | --- | ------------------------------- | ---------------------------------------- | --- | | | | | C. Jackie Chan | A. A formal diplomatic negotiation. | | | | | | D. Elon Musk | B. A casual meeting between | | | | | | Answer: A GPT-4v: I’m sorry, I | friends. C. An organized sporting event. | | Model MMBench-test UpperBound can’t provide the D. A military conflict. | | | | identity of real people | Answer: D | | | ------------ | ---- | ------------------------------------------------- | ------------------------ | ------------------------------ | --- | | GPT-4v | 74.3 | 76.2 | in images. | Gemini-Pro-V reject to answer. | | | | | Figure8: ContentModerationCasesofProprietaryVLMs. | | | | | Gemini-Pro-V | 70.2 | 72.6 | | | | | Qwen-VL-Max | 75.4 | 75.5 | | | | ContentModerationofProprietaryVLMs. Whenwetakeanin-depthlookatthepredictionsof proprietaryVLMs,wenoticethatallofthemapplyexplicitcontentmoderation. GPT-4v,Gemini- Pro-V,andQwen-VL-Maxrejectansweringin1.8%,1.6%,and0.1%ofcasesacrossallCircularEval passes in MMBench, respectively. 74% of questions rejected by GPT-4v are related to celebrity recognition (Figure 8), while no obvious rejection pattern is observed for Gemini-Pro-V. Under CircularEval, such moderation has a negative impact on the evaluated accuracy. To estimate an upper-boundperformance,weassumethatVLMscanperfectlyanswerallrejectedquestionsandre- calculatetheaccuracy. Table4showsthatthecontentmoderationpolicyaffectstheMMBench-test accuracybyupto2.4%,whichisnotasignificantchange. 10 2v ognimalFnepO B7-4TPGiniM B31-4TPGiniM B31-PILBtcurtsnI tcurtsnI-B9-SCIFEDI B7-PILBtcurtsnI B6-MLGlausiV B71-tahC-MLVgoC tahC-LV-newQ B7-5.1v-AVaLL V-MPCiniM 2lwO-GULPm B6-LV-iY B31-5.1v-AVaLL sulP-LV-newQ B43-LV-iY V-orP-inimeG B02-2MLI-AVaLL v4-TPG xaM-LV-newQ 2resopmoCX-MLI 80% 60% 40% 20% 0% ycaruccA tilpS tseT hcneBMM Average English Chinese Figure7: TheperformanceonthetestsplitofMMBenchandMMBench-CN. Modelsare sortedwiththeascendingorderofaverageperformance. ILMstandsforInternLM. CP FP-S FP-C AR LR RR LLaVA-InternLM2-20B Figure9: ProprietaryVLMsvs. Open-Sourceonesatafine-grainedlevel. Proprietaryvs. Open-Source: Whatisthegap? Comparedtothevariedperformanceofopen- sourceVLMs,mostproprietarymodelsdemonstratecompetitiveperformanceonMMBench. This raisesaquestionwecareabout: areproprietarymodelsgenerallymorepowerful,ordoeachkind ofmodeldisplayuniquestrengthsandweaknessesacrossdifferenttypesofability? Toanswerthis question,weperformafine-grainedcomparisonofthreeproprietaryVLMsandLLaVA-InternLM2- 20B, the top-performing model trained on open-source datasets only, and visualize the result in Figure9.Weobservethatproprietarymodelssignificantlyoutperformtheopen-sourceonesundertwo majorscenarios: i)Structuralizedimage-textunderstanding,whichrequiresVLMstounderstand complexcodes,tables,diagrams,orlayouts. ii)Tasksrequiringexternalknowledgetosolve,which correspondtoabilitiesincludingcelebrityrecognition,physicalpropertyreasoning,naturalrelation reasoning,etc. Meanwhile,proprietaryVLMsdonotdisplayadvantagesontaskscorrespondingto otherperceptionorreasoningcapabilities. HardcasesinMMBench. FormostVLMs,thefine-grainedaccuraciesvaryalotacrossdifferent abilitycategories. ToprovideinsightsforfutureVLMoptimization,wefindthemaximumaccuracy (A )acrossallevaluatedVLMsoneachL-3capability. SamplesbelongingtoL-3capabilities max withthelowestA arevisualizedinFigure10. Generally,wefindthatallexistingVLMshavethe max followinglimitations: 1. Pooratrecognizingthelow-levelfeaturesonvisualinputs,i.e.,theycannot accuratelyrecognizeandcomparethebrightness,sharpness,contrastratio,orartifactsofimages. 2. Difficultyinunderstandingstructuralizedvisualinputsliketables,diagrams,orlayouts,evenfor relativelysimplecaseslikeFigure10(b);3. Performbadlyonrecognizingorreasoningaboutthe inter-objectspatialrelationships,eitherin2Dor3Dspace. 11 Q. The graph shows the meals Q. Which image is the purchased in a restaurant in one | | second brightest? | day. What is the least popular | | --- | ------------------ | ------------------------------- | A. upper-left meal? | | B. upper-right | A. Salad | | --- | --------------- | -------- | C. lower-left B. Burger | | D. lower-right | C . C h i c k en | | --- | -------------- | ------------------ | Answer: C A =61.5% D . P a s t a max Answer: C A max =61.3% (a). Image Quality (b). StructralizedImage-Text Understanding Q. What is the positional relationship between the two Q. From the perspective of the | shapes in the picture? | | driver of the blue truck, in what | | ---------------------- | --- | ---------------------------------- | A. The two shapes are positioned apart or separated position is the person riding a bike from each other. | B. The two shapes are tangentially positioned or | | relative to the blue truck? | | ------------------------------------------------- | --- | --------------------------- | A. Left front externally tangent to each other. | C. The two shapes intersect with each other. | | B. Right front | | -------------------------------------------- | --- | -------------- | D. One shape is contained within the other or there is C. Right rear A =64.0% max | an inner shape enclosed by an outer shape. | | D. Left rear | | ------------------------------------------ | --------- | ------------ | | Answer: C | A =68.0% | Answer: A | max | (c). Spatial Relationship | (d). Physical Relation Reasoning | | | ------------------------- | -------------------------------- | --- | Figure10: Hardexamplesthatbelongtothe4L-3abilitieswithlowestA . AllVLMshave max madethewrongpredictionforthevisualizedexamplesunderCircularEval. 6 Conclusion WeintroduceMMBench,amulti-modalitybenchmarkthatperformsobjectiveevaluationforVLMs withover3,000multiple-choicequestionscovering20abilitydimensions. Toproducerobustand reliableevaluationresults,weintroduceanewevaluationstrategynamedCircularEval. Thestrategy is much stricter than the vanilla 1-pass evaluation and can yield reliable evaluation results at an affordable cost. Considering the limited instruction following ability of some VLMs, to yield moreaccurateevaluationresults,weadditionallyadoptLLMstoextractchoicesfromthemodel’s predictions. We comprehensively evaluate over 20 mainstream VLMs on MMBench, covering differentarchitecturesandparametersizes. Theevaluationresultsprovidevaluableinsightsforfuture improvements. 12 A MoreDetailsabouttheData Inthissection,webeginbyprovidingadetaileddefinitionofeachleafability(L-3)andpresenta collectionofvisualizationsamplesthataredirectlyrelatedtoeachleafability. Then,weenumerate allthedatasourcesthatwereutilizedintheconstructionofMMBench. A.1 DefinitionaboutEachLeafAbility ImageStyle Q: Which category does this Q: Which category does this image belong to? image belong to? A. OilPaiting A. OilPaiting B. Sketch B. Sketch C. Digitalart C. Digitalart D. Photo D. Photo GT:A GT:B Q: Which of the following Q: Which of the following ImageTopic captions best describes this captions best describes this image? image? A. A group of people playing A. A group of people playing soccer in a field soccer in a field B. A woman walking her dog on B. A woman walking her dog on a beach a beach C. A man riding a bicycle on a C. A man riding a bicycle on a mountain trail mountain trail D. A child playing with a ball in D. A child playing with a ball in a park a park GT:A GT:B Imagescene Q: What type of environment is Q: What type of environment is depicted in the picture? depicted in the picture? A. Home A. Home B. shopping mall B. shopping mall C. Street C. Street D. forest D. forest GT:A GT:C ImageMood Q: Which mood does this image Q: Which mood does this image convey? convey? A. Cozy A. Sad B. Anxious B. Anxious C. Happy C. Happy D. Angry D. Angry GT:C GT:A ImageQuality Q: Which imageismorebrightful? Q: which image is more colorful A. Thefirstimage A. The first image B. Thesecondimage B. The second image GT:A GT: B Figure11: CoarsePerception: Datasamples. CoarsePerception 1. ImageStyle: Determinewhichtypeofimageitbelongsto,suchasphotos,paintings,CTscans, etc. 2. ImageScene: Determinewhichenvironmentisshownintheimage,suchasindoors,outdoors, forest,city,mountains,waterfront,sunnyday,rainyday,etc. 3. ImageEmotion: Determinewhichsubjectiveemotionisconveyedbytheoverallimage,suchas cold,cheerful,sad,oroppressive. 4. ImageQuality: Determinetheobjectivequalityoftheimage,suchaswhetheritisblurry,bright ordark,contrast,etc. 13 5. ImageTopic: Determinewhatthesubjectoftheimageis,suchasscenery,portrait,close-upofan object,text,etc. InFigure11,wevisualizedatasamplesbelongingtotheCoarsePerceptioncapability. Attribute Recognition Q: Whatistheshapeofthis Q: what is the colorof this object? object? A. Circle A. Purple B. Triangle B. Pink C. Square C. Gray D. Rectangle D. Orange GT:A GT:D Celebrity Recognition Q: Whoisthisperson Q: Whoisthisperson A. DavidBeckham A. Benedict Cumberbatch B. PrinceHarry B. Idris Elba C. DanielCraig C. Ed Sheeran D. TomHardy D. Harry Styles GT:B GT:A ObjectLocalization Q: How many apples are there in the image? And how many Q: Which corner is the juice? bananas are there? A. Up A. 4 apples and 2 bananas B. Down B. 3 apples and 3 banana C. Left C. 2 apples and 4 bananas D. Right D. 4 apples and 1 bananas GT:D GT:A OCR Q: What does this picture want to express? Q: What does this outdoor A. We are expected to care for billboard mean? green plants. A. Smoking is prohibited here. B. We are expected to care for B. Something is on sale. the earth. C. No photography allowed C. We are expected to stay D. Take care of your speed. positive. GT:B D. We are expected to work hard. GT:D Figure12: Fine-grainedPerception(single-instance): Datasamples. Fine-grainedPerception(single-instance) 1. ObjectLocalization: Forasingleobject,determineitspositionintheimage(suchastop,bottom, etc.),itsabsolutecoordinatesintheimage,countthenumberofobjects,andtheorientationofthe object. 2. Attribute Recognition: Recognition of texture, shape, appearance characteristics, emotions, category. 3. CelebrityRecognition: Recognitionofcelebrities,landmarks,andwell-knownobjects. 4. OCR:Recognitionoftext,formula,andsheetintheimage. InFigure12,wevisualizedatasamplesbelongingtotheFine-grainedPerception(single-instance) capability. Fine-grainedPerception(cross-instance) 1. SpatialRelationship: Determinetherelativepositionbetweenobjectsinimage. 2. AttributeComparison: Compareattributesofdifferentobjectsinimage,suchasshape,color, etc. 3. ActionRecognition: Recognizinghumanactions,includingposemotion,human-objectinterac- tion,andhuman-humaninteraction. 14 Spatial Relationship | Q:Which country is north of the | | Which country is the southernmost | | -------------------------------- | --- | ---------------------------------- | | country circled in blue? | | of all the countries shown in the | | A.Laos | | picture? | | B.Thailand | | A.Australia | B.Indonesia C.China | D.Indonesia | | C.China | | ----------- | --- | ------------ | | GT:C | | D.NewZealand | GT:B Attribute Comparison Q: Are the candies in the two jars | Q: Are the two arrows in the same | | in the picture the same color? | | ---------------------------------- | --- | ------------------------------ | direction in the picture? A. Same | A. Same | | B. Not the same | | ------- | --- | --------------- | B. Not the same | C. Can't judge | | C. Can't judge | | -------------- | --- | -------------- | GT:B GT:B Action Recognition | Q: What kind of human behaviordoes | | Q: What kind of human behavior | | ----------------------------------- | --------- | ------------------------------- | | this picture describe? | | does this picture describe? | | A. A man with a solemn expression, | | A. This is aXXXsmiles on their | | XXXdriving. | | faces | | B. A man is practicing his | | B. A man is XXXhis breathing | | skateboarding XXXskills. | | and inner thoughts. | | C. A group of XXXbreather from | | C. A musician XXXa classical | | work. | | piece. | | D. A family is XXX | clothing. | D. A family is XXXtogether. | | GT:A | | GT:A | Figure 13: Fine-grained Perception (cross-instance): Data samples. XXX indicates omitted contentswhicharelessrelevanttothequestion. InFigure13,wevisualizedatasamplesbelongingtotheFine-grainedPerception(cross-instance) capability. PhysicalPropertyReasoning Q: The object shown in this figure: | A. Is the hardest naturally occurring | | Q: The object shown in this figure: | | --------------------------------------- | --- | --------------------------------------- | | substance on Earth. | | A. Is one kind of metal that is liquid | | B. Conducts electricity well at room | | at the room temperature. | | temperature. | | B. Can be easily dissolved in water. | | C. Is typically found in igneous rocks | | C. Has a low boiling point compared | | like basalt and granite. | | to other metals. | | D. Has a low melting point compared | | D. Is attracted to magnets. | | to other minerals. | | GT:A | GT:A FunctionReasoning | Q: What's the function of the | | Q: What's the function of the | | ------------------------------ | --- | ------------------------------ | demonstrated object? | A. Cut vegetables | | demonstrated object? | | ----------------- | --- | -------------------- | A. Separating | B. stir | | B. Clamping | | ------- | --- | ----------- | C. Water purification C. drill | D. Boiling water | | D. incise | | ---------------- | --- | --------- | GT:B GT:A IdentityReasoning | Q: What's the profession of the people | | Q: What's the profession of the people | | --------------------------------------- | --- | --------------------------------------- | in this picture? | in this picture? | | A. Librarian | | ----------------- | --- | ------------ | A. Librarian | B. radio host | | B. accountant | | ------------- | --- | ------------- | C. radio host | C. gardener | | D. gardener | | ----------- | --- | ----------- | D. lawyer E. lawyer | GT:C | | GT:A | | --------- | -------------------------------- | ---- | | Figure14: | AttributeReasoning: Datasamples. | | 15 AttributeReasoning 1. PhysicalPropertyReasoning: Predictthephysicalpropertyofanobject. Examples: hephysical propertyofconcentratedsulfuricacidisthatitisvolatile,thephysicalpropertyofwaterisits fluidity,etc. 2. FunctionReasoning: Predictthefunctionofanobject. Examples: thefunctionofabroomisto sweepthefloor,thefunctionofaspatulaistocook,thefunctionofapenistowrite,etc. 3. IdentityReasoning: Predicttheidentityofaperson. Example: byobservingaperson’sclothing andappearance,onemayinferhis/heroccupation. InFigure14,wevisualizedatasamplesbelongingtotheAttributeReasoningcapability. Social_Relation Q: What can be the relationship Q: What can be the relationship between the two persons in this image? between the two persons in this image? A. Father and daughter A. Father and daughter B. Mother and son B. Grandfather and granddaughter C. Brother and sister C. Brother and sister D. Husband and wife D. Husband and wife GT:D GT:B NatureRelation Q: In nature, what's the relationship Q: In nature, what's the relationship between these two creatures? between these two creatures? A. Predatory relationships A. Predatory relationships B. Competitive relationships B. Competitive relationships C. Parasitic relationships C. Parasitic relationships D. Symbiotic relationship D. Symbiotic relationship GT:B GT:D PhysicalRelation Q: Who is closer to the football in the Q: How many tennis balls are placed image, the player in the black jersey on the tennis racket? or the player in the green jersey? A. 1 A. Theplayerintheblackjersey B. 2 B. Theplayerinthegreenjersey C. 3 C. Theyareequallyclose D. 4 D. Itcannotbedetermined GT:C GT:A Figure15: RelationReasoning: Datasamples. RelationReasoning 1. SocialRelation: Relationsinhumansocietyorrelationsdefinedfromthehumanperspective. Examples: Inter-personrelations,suchasfatherandson,husbandandwife,friend,hostile,etc. 2. PhysicalRelation: Allrelationshipsthatexistinthephysicalworld,3Dspatialrelationshipsand theconnectionsbetweenobjectsare. 3. NatureRelation:Otherabstractrelationshipsthatexistinnature.Examples:predation,symbiosis, coexistence,etc. InFigure15,wevisualizedatasamplesbelongingtotheRelationReasoningcapability. LogicReasoning 1. StructuralizedImage-TextUnderstanding: Structuredunderstandingofimagesandtext,includ- ingparsingthecontentofcharts(suchasthetrendsofmultiplebarsinabarchart),understanding thecodeinanimage,etc. 2. FuturePrediction: Predictwhatwillhappeninthefuture. Examples: ifitisthunderinginthe skynow,itcanbepredictedthatitwillrainsoon(physicalphenomenon);ifsomeoneraisestheir fist,itmeanstheyaregoingtohitsomeone(eventoccurrence);ifsomeone’sfacebecomesserious, itmeanstheyaregoingtogetangry(emotionalchange). InFigure16,wevisualizedatasamplesbelongingtotheLogicReasoningcapability. 16 Future Prediction | | Q: What will happen next? | | | Q: What will happen next? | | | --- | ------------------------- | --- | --- | ------------------------- | --- | A. the motorcyleis gonnago forward | | A. this person is gonnacry | | | B. the motorcyleis gonnacrash | | | --- | -------------------------- | --- | --- | ----------------------------- | --- | B. this person is gonnalaugh | | C. this person is gonnaget mad | | | C. the motorcyleis gonnago | | | --- | ------------------------------ | --- | --- | --------------------------- | --- | backward | | D. both A,B, and C | | | D. both A,B, and C | | | --- | ------------------ | --- | --- | ------------------ | --- | GT:A GT:B StructuralizedImage-textUnderstanding | | Q: According to this image, which | | | Q: According to this image, what | | | --- | ---------------------------------- | --------------- | ------------ | --------------------------------- | --- | | | fruit did the most kids like? | | | hobby is liked the least? | | | | A. Orange | | | A. Reading | | | | B. Banana | | | B. Singing | | | | C. Pear | | | C. Painting | | | | D. Apple | | | D. Dancing | | | | GT:A | | | GT:C | | | | Figure16: | LogicReasoning: | Datasamples. | | | A.2 DataSourcesofMMBench Just as we introduce in Section 3.2 of the main paper, MMBench is mainly collected from the Internnet(80%)andthevalidationsetofsomepublicdatasets(20%). Table5listsallthesesources forimages,questionsandchoicesinMMBench. | The | source of (Q,C,I,A) | in MMBench | . Customize | | | | -------- | ------------------- | ---------- | ----------- | ----- | ------------------------ | | Table 5: | | | | means | all of question, choices | andanswerareconstructedbyus. Customize&selectionimpliesthatthesecomponentsareeither constructedbyusorselectedfromtheoriginaldataset. | | ImageSource | ProblemSource | | Number | Ratio | | --- | -------------- | ------------------- | --- | ------ | ----- | | | ARAS[15] | customize&selection | | 76 | 2.4% | | | CLEVR[24] | customize&selection | | 14 | 0.4% | | | COCO[9] | customize&selection | | 179 | 5.6% | | | KonIQ-10k[21] | customize&selection | | 32 | 1.0% | | | LLaVA[33] | customize | | 19 | 0.6% | | | PISC[29] | customize&selection | | 15 | 0.5% | | | Places[54] | customize&selection | | 59 | 1.8% | | | ScienceQA[34] | customize&selection | | 156 | 4.8% | | | ShapeWorld[25] | customize&selection | | 20 | 0.6% | | | TextVQA[42] | customize&selection | | 18 | 0.6% | | | VSR[31] | customize&selection | | 19 | 0.6% | | | W3CSchool[1] | customize | | 20 | 0.6% | | | Internet | customize | | 2590 | 80.5% | B MoreDetailsonMMBenchConstruction Inthissectionweprovidemorequalitativeresultsonthequalitycontrolparadigmweadoptedto constructMMBench,aswellasthepromptweusedforMMBench-CNtranslation. ‘Text-only’ question filtering. To filter out the ‘text-only’ questions (which can be answered correctlywithtext-onlyinputsbyLLMs)fromMMBench. Weapplythreestate-of-the-artLLMs, including GPT-4 [37], Gemini-Pro [44], and Qwen-Max [5] to infer the questions with text-only inputsunderCircularEval. IfmorethantwoLLMsanswerthequestioncorrectly,thequestionwillbe manuallycheckedandremovedifitisunqualified. InFigure17(a),wevisualizesomeunqualified questionsfilteredoutbythisapproach. ‘Wrong’questionfiltering. Duringpreliminarystudy,wealsonoticethatsomedatasamplesin MMBenchmightbewrong,duetoambiguousquestionsoroptions,repeatedoptions,orincorrect answers. Tofilteroutthesewrongsamples, weinferMMBenchquestionswiththreeproprietary VLMs(GPT-4v,Gemini-Pro-V,Qwen-VL-Max)andtwoopensourceVLMs(InternLM-XComposer2 17 | Q. Which part of an apple tree | | | Q. The object shown in this figure: | | | | ------------------------------- | --- | --- | ----------------------------------------------------- | --- | --- | | might grow into a new tree? | | | A. Is a bluish-white metal that is commonly used in | | | | Hint: This paradigm shows the | | | galvanizing and as an alloy in brass and other metals | | | B. Has a relatively low melting point of around 419°C | life cycle of an apple tree. | | | C. Is an essential micronutrient for humans and many | | | | ----------------------------- | --- | --- | ----------------------------------------------------- | --- | --- | A. aseed other organisms | B. aleaf | | | D. All of the options are correct. | | | | ---------------------------- | --- | --- | ---------------------------------------- | --- | --- | | Answer: A (common knowledge) | | | Answer: D (can be inferred from options) | | | | Source:ScienceQA | | | Source: Internet | | | (a). Text-Only questions filtered out | Q. What’s the function of the | | | Q. What’sthefunctionofthe | | | | ------------------------------ | --- | --- | ---------------------------------------- | --- | --- | | demonstrated object? | | | demonstratedobject? | | | | A. Cooking | | | A. Celebrate someone’s birthday | | | | B. CookSoup | | | B. Celebrating a wedding | | | | C. Fry | | | C. Asanitary facility used for excretion | | | | D.Steam | | | D. Offering a variety of drink | | | | Answer: A (ambiguous options) | | | Answer: A (ambiguous options) | | | | Source:Internet | | | Source:COCO | | | (b). Wrong questions filtered out | Figure17: | UnqualifiedsamplesfilteredoutinMMBench. | | | | | | --------- | --------------------------------------- | --- | --- | --- | --- | andLLaVA-v1.5-13B).IfnoVLMcanansweraquestioncorrectlyunderCircularEval,thequestion will then be manually checked. In Figure 17(b), we visualize wrong samples filtered out by the approach. | Q. Is this photo taken in a | | Q. If the owner of this room wants to | | Q. Whatcanbeinferredabout | | | ---------------------------- | ---------------------- | -------------------------------------- | -------------------------------- | ------------------------------- | ----------------------------- | | | | add a new piece of furniture, what | | thegroupofpeoplesittingon | | | place of escalator indoor? | | m a t e r i a l s h o u l | d t h e y c h o o s e t o | t h e s t r e e t ? | | | P l e a s e a n s w e | r y e s o r n o. | | | A . T h e y a r e h o m | e l e s s | | A n s w e r . Y e s . | | m a i n t a i n t h e o | v e r a l l t h e m e ? | B . T h e y a r e s t r e | e t p e r f o r m er s | | ( t h e p h o t o f e | a t u r e s a n | A . W o o d B . W | i c k e r | C . T h e y a r e w a i | t i n g f o r p a r a d e | | o u t d o o r s c e n | e ) | C . P l a s t i c D . M | e t a l | D . T h e y a r e t o u r | i s t s | | B e n c h m a r k : | M M E | A n s w e r : D ( i n | c o r r e c t a n s w e r ) | A n s w e r : D ( w r o | n g q u e s t i o n ) | | | | B e n c h m a r k : S | E E D B e n c h | B e n c h m a r k : S E E | D B e n c h | Figure18: Unqualifiedsamplesinotherbenchmarkscanalsobedetectedbyourqualitycontrol paradigms. TheUniversalityoftheQualityControlParadigm. Thequalitycontrolparadigmadoptedby MMBenchisgeneralandcanalsobeappliedtootherbenchmarkstoimprovethequality. Tosupport thisclaim,weapplythequalitycontrolparadigmtootherpopularmultimodalevaluationbenchmarks (likeMME[17]andSEEDBench[27])andtrytodetectthelow-qualitysamples. Wefindthatour qualitycontrolparadigmcanalsosuccessfullydetectandfilteroutunqualifiedsamplesfromthese benchmarks. SomedetectedsamplesarevisualizedinFigure18. MMBench-CNTranslation. InFigure19,weprovidethepromptweadoptedforMMBench-CN translation,whichincludeinstructionsandseveralin-contextexamples. Alltranslationsgeneratedby GPT-4willbefurthermanuallyverfiedtoensurethecorrectness. C MoreDetailsonLLM-basedChoiceExtraction FailureCasesofHeuristicMatching. InFigure20,wedisplaysomefailurecasesofheuristic matchingofthestate-of-the-artVLMGPT-4v. Basically,suchfailuremayoccurwhentheVLM:i) rejectsorisnotcapabletoanswerthegivenquestion; ii)answersthequestionindifferentwords ratherthanthecorrectchoice;iii)providesananswerwithmultiplechoicelabels(A,B,C,etc. ) included. ThepromptforLLM-basedChoiceExtraction. InFigure21,weprovidethepromptweadopted for LLM-based choice extraction. In-context examples are included to improve the instruction- followingcapabilityoftheLLMadopted. PerformanceEvaluatedwithOtherChoiceExtractors. InTable6,welisttheMMBench-dev performanceobtainedwithdifferentchoiceextractors,includingGPT-4(0125),GPT-3.5-Turbo(0613 and0125),andInternLM2-7B[45]. VLMswithhighsuccessrate(>99%)inheuristicmatchingare skipped. Fromthetable,weseethatadoptingdifferentchoiceextractorswillnotleadtosignificant differentevaluationresults. VisualGLMdisplaysthelargestrangeacrossallchoiceextractors,which isaround1.4%. Fortop-performingproprietaryVLMs(GPT-4v,Gemini-Pro-V,etc.),thegapisat most0.3%. 18 B.1MMBench-CNTranslation 你是一个翻译助手，你的任务是帮我把下面的英文题目及选项翻译成中文，并保持 完全一样的含义。你仅需要翻译文本中的英文内容，不需要翻译其他语言的内容， 请只翻译给定内容，不要丢失/修改/添加内容。对于文本中的专有名词，符号，代 码，或是人名等，请依然保持英文，不需要翻译。我会以“json”格式给出题目及选 项的内容，你需要把翻译后的中文内容以“json”格式返回给我。 例1： 英文: {"Q": "Whichofthefollowingwaspartoftheroleofadeaconess? ","A": "Ministeringto | thesick","B": | "Preparingwomenforbaptism","C": | "Prayingforthesuffering"} | | | ------------- | ------------------------------- | ------------------------- | --- | 中文: | "以下哪项是女执事的职责之一？", | | "照顾病人", | "为女性准备洗礼", | | ----------------- | --- | ------- | ---------- | | {"Q": | | "A": | "B": | "为受苦的人祷告"} "C": 例2： 英文: {"Q":"Whichcanbetheassociatedtextwiththisimagepostedontwitter?","A":"Locatedin BomeCounty,NyingchiCity,TibetofChina,theYigongIronMountainisalwayssurrounded "夏天所有季节中最闪耀的季节阳光明媚， | bycloudsandmistduringsummer.", | "B": | | | | ------------------------------ | ---- | --- | --- | 万物清明泰山向人们展现的初夏之景处处充满着诗情画意", "C": "Giant logs and strippedtreesonRialtoBeachintheOlympicNationalPark. #beach#wawx#blackandwhite @yourtake","D":"MadisonFallsinOlympicNationalPark,WA[OC][3024x4032]#nature"} 中文: | "与这张推特上图片配套的推文是什么？", | | "坐落在中国西藏自治区林芝 | | | ----------------------- | --- | ------------- | ------------ | | {"Q": | | "A": | | | 市波密县的易贡铁山，在夏季总是被云雾环绕。", | | | "夏天所有季节中最闪耀的 | "B": 季节阳光明媚，万物清明泰山向人们展现的初夏之景处处充满着诗情画意", "C": | "奥林匹克国家Rialto | 沙滩上的巨木与被剥皮的树木。#beach#wawx#blackandwhite | | | | ------------- | --------------------------------------- | --- | --- | @yourtake","D": "Madison瀑布，奥林匹克国家公园，WA[OC][3024x4032]#nature"} 请翻译： 英文: {TheEnglishquestionpresentedinthejsonformat} 中文: | Figure19: | AnexamplepromptofChinesesinglechoicewithreasoning. | | | | --------- | -------------------------------------------------- | --- | --- | Failure Cases of Heuristic Matching | | Q. How many apples are there in the image? And how | | Q. Whatbandisthis? | | --- | -------------------------------------------------------- | --- | ---------------------------------------------------------------- | | | many bananas are there? | | A.The Beatles B. Sex Pistols C. Oasis D. Guns N’ Roses | | | A. 2 apples and 2 bananas B. 3 apples and 3 bananas | | Answer: B | | | C. 2 apples and 4 bananas D. 4 apples and 1 banana | | GPT-4v: The image you've provided appears to be in black | | | Answer: A | | and white and shows a band performing live. However, I'm | | | GPT-4v: In the image, there is 1 apple and 1 banana. | | unable to identify the specific band or confirm their identity. | | | However, this does not match any of the options | | To answer your question based on the options provided, I | | | provided. If I must choose from the given options, none | | would need more context or information that is not | | | of them accurately describe the contents of the image. | | available in the image. If you have any other questions or | need information on the bands listed, feel free to ask! Q. Which solution has a higher concentration of purple particles? | | A.Neither; their concentrations are the same | | Q. Who is sitting in the middle? | | --- | ---------------------------------------------- | --- | -------------------------------- | | | B. Solutions A C. Solution B | | A.The little girl. | B.Theman. Answer: C | | GPT-4v: The concentration of a solution is determined by the amount | | C.The woman. | | --- | -------------------------------------------------------------------- | --- | -------------------------------- | | | of solute present in a given volume of solvent. In the image, both | | Answer: A | | | Solution A and Solution B have the same volume of solvent (25 mL), | | GPT-4v:Thelittlegirl is sitting | | | but Solution B has more purple particles (solute) than Solution A. | | in the middle. | Therefore, Solution B has a higher concentration of purple particles. Figure20: FailurecasesofGPT-4vduringexactmatching. LLM-based sementic matching is generally helpful. To demonstrate that LLMs can be a general tool for semantic matching, we also validate the LLM-involved evaluation paradigm on existingmulti-modalitytasks,includingGQA[23],OK-VQA[35],andText-VQA[42]. Giventhe ground-truthanswer,weuseGPT-3.5-TurbotomeasurethesimilaritybetweenVLM’sprediction5. Foreachbenchmark,werandomlyselect1000testingsamplesandevaluatewithexactmatch(the traditionalparadigm)andChatGPT-basedmatch,respectively,andlisttheresultsinTable7.Basically, ChatGPT-basedevaluationdemonstratesthesametrendcomparedtotheexact-matchaccuracyon all tasks. On GQA, two algorithms demonstrate very close performance under ChatGPT-based 5Thesimlarityscoreisanintegerin[1,5].1meanscompletelywrong,while5meanscompletelycorrect. 19 C.1PromptforChoiceExtraction You are an AI assistant who will help me to match an answer with several options of a single-choicequestion. Youareprovidedwithaquestion,severaloptions,andananswer, and you need to find which option is most similar to the answer. If the meaning of all optionsaresignificantlydifferentfromtheanswer,outputZ.Youshouldonlydothematching basedexactlyontheliteralmeaningoftheoptionsandanswer. Youshouldnotperformany externalinferencebasedonyourknowledgeduringthematching. Yourshouldoutputasingle uppercasecharacterinA,B,C,D(iftheyarevalidoptions),andZ. Example1: Question: Whatisthemainobjectinimage? Options: A.teddybearB.rabbitC.catD.dog Answer: acuteteddybear Youroutput: A Example2: Question: Whatisthemainobjectinimage? Options: A.teddybearB.rabbitC.catD.dog Answer: Spider Youroutput: Z Nowit’syourturn: Question: {question} Options: {options} Answer: {answer} Youroutput: Figure21: ThepromptusedforchoiceextractiononMMBench. TheChinesetranslationofthis promptisadoptedforMMBench-CNchoiceextraction. Table6: MMBench-devaccuracieswithdifferentchoiceextractorsunderCircularEval. | | Exact | GPT-4-Turbo | GPT-3.5-Turbo | GPT-3.5-Turbo | | | ----------------------- | -------- | ----------- | ------------- | ------------- | ------------ | | VLM | | | | | InternLM2-7B | | | Matching | (0125) | (0613) | (0125) | | | MiniGPT4-7B[56] | 26.0 | 32.7 | 33.1 | 33.0 | 32.9 | | IDEFICS-9B-Instruct[26] | 36.0 | 37.2 | 37.2 | 37.2 | 37.2 | | InstructBLIP-7B[11] | 34.8 | 37.4 | 37.5 | 37.5 | 37.7 | | VisualGLM-6B[13] | 19.4 | 36.1 | 37.5 | 37.5 | 36.1 | | MiniGPT4-13B[56] | 30.7 | 37.5 | 37.8 | 37.8 | 37.6 | | InstructBLIP-13B[11] | 36.6 | 40.9 | 41.1 | 41.0 | 41.4 | | Qwen-VL-Chat[6] | 56.7 | 59.5 | 59.8 | 59.4 | 59.8 | | Qwen-VL-Plus[6] | 43.7 | 62.9 | 62.6 | 61.9 | 63.2 | | MiniCPM-V[39] | 57.6 | 64.8 | 64.7 | 64.6 | 64.7 | | Gemini-Pro-V[44] | 70.4 | 70.9 | 70.9 | 70.9 | 70.8 | | GPT-4v[37] | 71.8 | 74.3 | 74.6 | 74.6 | 74.6 | | Qwen-VL-Max[6] | 72.9 | 76.4 | 76.5 | 76.2 | 76.5 | evaluation. Infurtherinvestigation,wefindthereasonisthatChatGPTsucceedsinmatchingslightly differentanswers(comparedtoGT)generatedbyMiniGPT-4,whileexactmatchingfails(examples inTable8). D EvaluationSettingsandResults InSection5.2ofthemainpaper,wegivetheresultsofdifferentmodelsonthetestsplitofMMBench andMMBench-CN.Inthissection,wewillintroducethedetailedevaluationsettingonMMBench, andprovidemoreevaluationresults. 20 Table7: LLM-basedMatchingvsExactMatching. ApreliminarystudyonVQAbenchmarks. Accuracy is the success rate of answers being exactly matched with the groundtruth. For each sample,GPTscoreisanintegern∈[1,5],indicatingthesimilaritybetweenanswerandgroundtruth. Weadoptgpt-3.5-turbo-0613asthematchingLLMandreporttheaverageGPTscorefortesting samples. Dataset GQA[23] OK-VQA[35] Text-VQA[42] Model Flamingo MiniGPT-4 Flamingo MiniGPT-4 Flamingo MiniGPT-4 Accuracy 33.6% 22.4% 42.6% 21.9% 22.9% 9.8% AverageGPTscore 2.75 2.74 2.79 1.97 1.92 1.54 Table 8: MiniGPT-4 GQA answers (different from GT) that can be matched by ChatGPT-based evaluation. Question GroundTruth MiniGPT-4Answer GPTScore Onwhichsideofthepictureistheleatherbag? right leatherbagisonrightside 5 Whoiswearingashirt? girl womaniswearingshirt 5 Whatisthepenmadeof? plastic penismadeofplastic## 5 Whatdoesthemalepersonrideon? bike bicycle 4 Isthemicrowavesilverandrectangular? Yes microwaveissilverandrectangular 5 Howdoesthesilverlampappeartobe,onoroff? off silverlampappearstobeoff## 5 D.1 EvaluationSettings Unless stated otherwise, all results presented in this paper adhere to the conventional zero-shot evaluationsetting. Wehavealsoattemptedtoassessthesemodelswithfew-shotandchain-of-thought evaluations. However,noencouragingresultsareobserved. Belowweprovidethepromptweused forevaluatingaVLMunderthezero-shotsettingonMMBench. D.1PromptTemplateforZero-shotInference. Hint: xxx[optional] Question: xxx A.xxx B.xxx C.xxx[optional] D.xxx[optional] Pleaseselectthecorrectanswerfromtheoptionsabove. Figure22: Theprompttemplateadoptedforzero-shotinference. D.2 ModelSettings InTable9,weprovidedetailsofallopen-sourcemodelsevaluatedinMMBench,includingseveral additionalmodelsthatdonotfitthespaceofthemainarticle. D.3 MoreResults Inthissection,wegivemoredetailedresultsabouttheperformanceofdifferentmodelsonMMBench andMMBench-CN.Wepresentthedetailedevaluationresultsof30differentVLMs(someofthem donotappearinthemainpaperduetolimitedspace). FordetailedresultsoneachL-3ability,seethe separatesheetinthesupplementarymaterials. 21 Table9: DetailsoftheevaluatedOpen-SourceVLMs. | VLM | LanguageBackbone | VisionBackbone | OverallParameters | | ------------------------ | ---------------- | ----------------- | ----------------- | | OpenFlamingov2[4] | MPT7B | CLIPViT-L/14 | 9B | | MiniGPT-4-7B[56] | Vicuna7B | EVA-G | 8B | | IDEFICS-9B-Instruct[26] | LLaMA7B | CLIPViT-H/14 | 9B | | VisualGLM-6B[13] | ChatGLM6B | EVA-CLIP | 7B | | InstructBLIP-7B[11] | Vicuna7B | EVA-G | 8B | | MiniGPT-4-13B[56] | Vicuna13B | EVA-G | 14B | | PandaGPT[43] | Vicuna13B | ImageBindViT-H/14 | 14B | | InstructBLIP-13B[11] | Vicuna13B | EVA-G | 14B | | IDEFICS-80B-Instruct[26] | LLaMA65B | CLIPViT-H/14 | 80B | | Qwen-VL-Chat[6] | Qwen7B | ViT-G/16 | 10B | | MiniCPM-V[39] | MiniCPM2.4B | SigLip-400M | 3B | | LLaVA-v1.5-7B[32] | Vicuna7B | CLIPViT-L/14 | 7B | | mPLUG-Owl2[49] | LLaMA27B | CLIPViT-L/14 | 8B | | CogVLM-Chat-17B[47] | Vicuna7B | EVA2-CLIP-E | 18B | | ShareGPT4V-7B[8] | Vicuna7B | CLIPViT-L/14 | 7B | | Yi-VL-6B[2] | Yi-6B | CLIPViT-H/14 | 7B | | LLaVA-InternLM-7B[10] | InternLM7B | CLIPViT-L/14 | 9B | ShareGPT4V-13B[8] | | Vicuna13B | CLIPViT-L/14 | 13B | | ----------------------- | ------------- | ------------ | --- | | LLaVA-v1.5-13B[32] | Vicuna13B | CLIPViT-L/14 | 13B | | Yi-VL-34B[2] | Yi34B | CLIPViT-H/14 | 35B | | OmniLMM-12B[38] | Zephyr-7B-β | EVA-02-5B | 12B | | Monkey-Chat[30] | Qwen7B | ViTBigG | 10B | | InternLM-XComposer[52] | InternLM-7B | EVA-G | 9B | | LLaVA-InternLM2-7B[10] | InternLM2-7B | CLIPViT-L/14 | 9B | | LLaVA-InternLM2-20B[10] | InternLM2-20B | CLIPViT-L/14 | 23B | | InternLM-XComposer2[12] | InternLM2-7B | CLIPViT-L/14 | 9B | 22 Table10: CircularEvalresultsonMMBench-devset(L-2abilities). Open-sourcemodelstagged with*incorporatein-housedatainmodeltraining. | Model | Overall CP | FP-S FP-C | AR LR | RR | | ----- | ---------- | --------- | ----- | --- | OpenSourceVLMs | OpenFlamingov2[4] | 2.6% 0.8% | 4.5% 1.1% | 5.5% 0.0% | 3.4% | | ----------------- | ----------- | ----------- | ----------- | ----- | | MiniGPT4-7B[56] | 32.7% 38.4% | 39.1% 20.7% | 49.4% 10.5% | 22.4% | | VisualGLM-6B[13] | 36.1% 40.3% | 43.3% 19.6% | 49.4% 16.9% | 33.9% | IDEFICS-9B-Instruct[26] 37.2% 50.6% 37.7% 30.2% 51.8% 4.8% 25.3% | InstructBLIP-7B[11] | 37.4% 46.4% | 47.1% 23.5% | 51.2% 8.1% | 24.7% | | ------------------- | ----------- | ----------- | ---------- | ----- | | MiniGPT4-13B[56] | 37.5% 44.2% | 48.4% 16.8% | 57.3% 6.5% | 30.5% | InstructBLIP-13B[11] 40.9% 48.6% 52.2% 18.4% 56.7% 5.6% 39.7% | PandaGPT[43] | 41.6% 56.1% | 34.6% 34.6% | 53.7% 13.7% | 38.5% | | ------------ | ----------- | ----------- | ----------- | ----- | IDEFICS-80B-Instruct[26] 42.3% 54.7% 48.1% 24.6% 57.3% 8.9% 34.5% | Qwen-VL-Chat*[6] | 59.5% 70.7% | 69.9% 49.7% | 69.5% 25.0% | 44.3% | | ---------------- | ----------- | ----------- | ----------- | ----- | CogVLM-Chat-17B[47] 62.4% 69.6% 70.6% 56.4% 67.1% 29.0% 59.2% | LLaVA-v1.5-7B[32] | 62.5% 71.3% | 70.6% 55.9% | 70.7% 25.8% | 55.7% | | ----------------- | ----------- | ----------- | ----------- | ----- | | mPLUG-Owl2[50] | 63.5% 72.9% | 70.2% 53.6% | 70.7% 29.8% | 60.3% | | MiniCPM-V[39] | 64.8% 71.0% | 75.1% 52.5% | 72.0% 30.6% | 64.9% | | Yi-VL-6B*[2] | 65.6% 72.7% | 73.7% 54.7% | 73.2% 32.3% | 65.5% | | ShareGPT4V-7B[8] | 66.2% 77.3% | 75.1% 57.5% | 68.3% 25.8% | 63.8% | | ShareGPT4V-13B[8] | 67.0% 75.1% | 77.9% 58.1% | 68.9% 35.5% | 61.5% | LLaVA-InternLM-7B[10] 67.0% 75.7% 72.7% 57.5% 71.3% 37.1% 66.7% | LLaVA-v1.5-13B[32] | 67.2% 74.0% | 75.1% 59.2% | 68.9% 38.7% | 66.7% | | ------------------ | ----------- | ----------- | ----------- | ----- | | Yi-VL-34B*[2] | 68.2% 75.7% | 73.0% 55.9% | 75.6% 39.5% | 70.7% | | Monkey-Chat[30] | 68.8% 72.9% | 79.2% 58.1% | 79.3% 42.7% | 62.6% | | OmniLMM-12B*[38] | 69.7% 75.1% | 79.6% 61.5% | 73.8% 37.1% | 69.5% | LLaVA-InternLM2-7B[10] 71.6% 79.8% 77.2% 62.0% 74.4% 41.1% 74.1% LLaVA-InternLM2-20B[10] 72.8% 80.1% 75.1% 68.2% 73.8% 46.0% 76.4% InternLM-XComposer*[52] 73.9% 79.6% 81.7% 65.4% 84.8% 39.5% 72.4% InternLM-XComposer2*[12] 79.1% 83.4% 84.4% 68.7% 83.5% 58.1% 82.8% ProprietaryVLMs | Qwen-VL-Plus[6] | 62.9% 67.1% | 78.9% 53.1% | 71.3% 28.2% | 54.6% | | ---------------- | ----------- | ----------- | ----------- | ----- | | Gemini-Pro-V[44] | 70.9% 71.3% | 81.7% 62.0% | 78.7% 47.6% | 70.7% | | GPT-4v[37] | 74.3% 78.5% | 72.3% 66.5% | 82.9% 67.7% | 73.6% | Qwen-VL-Max[6] | | 76.4% 76.2% | 87.2% 69.3% | 78.7% 55.6% | 78.7% | | --- | ----------- | ----------- | ----------- | ----- | 23 Table11:CircularEvalresultsonMMBench-testset(L-2abilities). Open-sourcemodelstagged with*incorporatein-housedatainmodeltraining. | Model | Overall CP | FP-S FP-C | AR LR | RR | | ----- | ---------- | --------- | ----- | --- | OpenSourceVLMs | OpenFlamingov2[4] | 2.3% 1.1% | 3.5% 1.5% | 5.3% 0.0% | 2.7% | | ----------------- | ----------- | ----------- | ---------- | ----- | | MiniGPT4-7B[56] | 30.5% 37.0% | 31.8% 17.2% | 49.8% 9.2% | 25.6% | IDEFICS-9B-Instruct[26] 35.2% 48.3% 31.3% 29.6% 47.8% 11.4% 25.2% | VisualGLM-6B[13] | 35.4% 40.2% | 38.5% 26.2% | 47.8% 19.6% | 29.5% | | ------------------- | ----------- | ----------- | ----------- | ----- | | InstructBLIP-7B[11] | 38.3% 46.7% | 39.0% 31.8% | 55.5% 8.7% | 31.0% | | MiniGPT4-13B[56] | 38.8% 44.6% | 42.9% 23.2% | 64.9% 8.2% | 32.9% | | PandaGPT[43] | 39.7% 51.9% | 29.5% 27.3% | 62.0% 19.0% | 38.0% | InstructBLIP-13B[11] 39.8% 47.2% 42.9% 21.0% 60.4% 12.5% 38.8% IDEFICS-80B-Instruct[26] 40.9% 54.6% 38.1% 29.6% 52.7% 16.8% 34.9% | Qwen-VL-Chat*[6] | 60.9% 68.5% | 67.7% 50.2% | 78.0% 37.0% | 45.7% | | ----------------- | ----------- | ----------- | ----------- | ----- | | MiniCPM-V[39] | 61.4% 65.6% | 69.4% 51.3% | 70.6% 35.3% | 59.7% | | LLaVA-v1.5-7B[32] | 63.4% 70.0% | 68.0% 57.7% | 77.6% 33.2% | 56.2% | | mPLUG-Owl2[50] | 63.5% 68.1% | 69.1% 55.8% | 78.4% 37.0% | 57.0% | CogVLM-Chat-17B[47] 63.6% 72.8% 66.6% 55.4% 71.4% 33.7% 62.0% | ShareGPT4V-7B[8] | 64.6% 72.2% | 68.7% 59.6% | 72.7% 34.8% | 60.5% | | ---------------- | ----------- | ----------- | ----------- | ----- | | Yi-VL-6B*[2] | 65.5% 72.8% | 72.9% 56.2% | 75.5% 41.3% | 55.4% | LLaVA-InternLM-7B[10] 65.9% 72.6% 68.7% 57.3% 80.0% 37.5% 63.2% | ShareGPT4V-13B[8] | 66.7% 75.6% | 73.5% 56.9% | 72.7% 37.0% | 62.4% | | ------------------ | ----------- | ----------- | ----------- | ----- | | LLaVA-v1.5-13B[32] | 66.9% 73.1% | 72.4% 60.3% | 75.5% 35.9% | 65.5% | | Yi-VL-34B*[2] | 68.4% 72.0% | 78.0% 54.7% | 81.2% 38.6% | 68.2% | | OmniLMM-12B*[38] | 69.2% 72.0% | 79.8% 61.0% | 78.0% 40.2% | 66.7% | | Monkey-Chat[30] | 69.6% 75.0% | 75.4% 63.3% | 82.4% 46.7% | 58.9% | InternLM-XComposer*[52] 71.3% 75.7% 76.3% 60.3% 84.5% 44.6% 71.7% LLaVA-InternLM2-7B[10] 71.6% 78.1% 75.4% 66.7% 77.6% 44.6% 70.2% LLaVA-InternLM2-20B[10] 72.3% 78.3% 76.6% 68.2% 78.4% 46.2% 69.4% InternLM-XComposer2*[12] 78.1% 80.4% 83.5% 73.0% 83.7% 63.6% 74.4% ProprietaryVLMs | Qwen-VL-Plus[6] | 64.6% 66.5% | 79.1% 50.2% | 73.9% 42.9% | 57.8% | | ---------------- | ----------- | ----------- | ----------- | ----- | | Gemini-Pro-V[44] | 70.2% 70.0% | 78.9% 65.9% | 82.9% 46.2% | 65.9% | | GPT-4v[37] | 74.3% 77.6% | 73.8% 71.5% | 85.3% 63.6% | 68.6% | Qwen-VL-Max[6] | | 75.4% 74.8% | 87.2% 67.0% | 85.3% 54.9% | 70.5% | | --- | ----------- | ----------- | ----------- | ----- | 24 Table12: CircularEvalresultsonMMBench-CN-devset(L-2abilities). Open-sourcemodels taggedwith*incorporatein-housedatainmodeltraining. | Model | Overall CP | FP-S FP-C | AR LR | RR | | ----- | ---------- | --------- | ----- | --- | OpenSourceVLMs | MiniGPT4-13B[56] | 11.8% 14.6% | 13.8% 14.0% | 15.9% 3.2% | 2.3% | | -------------------- | ----------- | ----------- | ----------- | ----- | | MiniGPT4-7B[56] | 11.9% 11.9% | 14.5% 7.8% | 19.5% 3.2% | 10.9% | | OpenFlamingov2[4] | 14.3% 14.4% | 14.9% 11.2% | 21.3% 10.5% | 12.6% | | InstructBLIP-13B[11] | 15.1% 16.0% | 14.9% 7.8% | 30.5% 4.0% | 14.4% | | InstructBLIP-7B[11] | 18.1% 16.0% | 16.6% 10.6% | 38.4% 4.0% | 23.6% | IDEFICS-9B-Instruct[26] 18.7% 22.7% 19.7% 7.3% 35.4% 1.6% 17.2% IDEFICS-80B-Instruct[26] 29.2% 32.0% 27.0% 25.1% 50.0% 8.1% 26.4% | PandaGPT[43] | 31.0% 40.1% | 24.9% 18.4% | 47.6% 12.1% | 33.3% | | ---------------- | ----------- | ----------- | ----------- | ----- | | VisualGLM-6B[13] | 40.6% 45.3% | 48.1% 30.7% | 54.3% 8.9% | 37.9% | CogVLM-Chat-17B[47] 52.9% 63.5% 56.4% 41.9% 65.9% 16.9% 50.0% | LLaVA-v1.5-7B[32] | 57.0% 69.3% | 59.9% 47.5% | 62.8% 25.0% | 54.0% | | ------------------ | ----------- | ----------- | ----------- | ----- | | Qwen-VL-Chat*[6] | 57.6% 66.6% | 68.5% 43.6% | 70.1% 21.8% | 48.9% | | mPLUG-Owl2[50] | 58.1% 68.8% | 65.1% 43.0% | 68.9% 29.8% | 50.0% | | ShareGPT4V-7B[8] | 59.7% 71.8% | 62.6% 48.6% | 62.8% 26.6% | 61.5% | | OmniLMM-12B*[38] | 60.6% 67.7% | 69.9% 48.0% | 70.1% 25.8% | 59.2% | | ShareGPT4V-13B[8] | 62.4% 72.9% | 67.1% 55.3% | 66.5% 34.7% | 55.7% | | LLaVA-v1.5-13B[32] | 62.5% 71.8% | 65.7% 57.0% | 67.1% 33.1% | 59.8% | | MiniCPM-V[39] | 63.0% 68.2% | 75.1% 53.1% | 72.0% 25.8% | 60.3% | LLaVA-InternLM-7B[10] 63.0% 72.4% 68.2% 50.3% 68.9% 35.5% 62.1% | Monkey-Chat[30] | 65.1% 73.8% | 74.4% 50.3% | 77.4% 37.9% | 54.6% | | --------------- | ----------- | ----------- | ----------- | ----- | | Yi-VL-6B*[2] | 65.3% 72.4% | 73.0% 53.1% | 70.7% 33.9% | 67.8% | | Yi-VL-34B*[2] | 67.0% 73.8% | 73.0% 52.5% | 72.6% 40.3% | 71.8% | LLaVA-InternLM2-7B[10] 70.0% 81.5% 72.3% 59.2% 73.8% 34.7% 74.7% InternLM-XComposer*[52] 71.3% 76.5% 77.5% 63.7% 81.7% 37.9% 71.8% LLaVA-InternLM2-20B[10] 71.7% 77.9% 74.4% 68.7% 75.6% 43.5% 74.1% InternLM-XComposer2*[12] 77.2% 83.4% 84.1% 64.2% 84.1% 54.8% 75.9% ProprietaryVLMs | Qwen-VL-Plus[6] | 67.5% 68.8% | 83.0% 54.2% | 75.6% 38.7% | 65.5% | | ---------------- | ----------- | ----------- | ----------- | ----- | | Gemini-Pro-V[44] | 69.3% 72.4% | 78.5% 63.1% | 78.7% 40.3% | 65.5% | | GPT-4v[37] | 73.3% 76.5% | 71.6% 67.0% | 82.3% 63.7% | 74.1% | Qwen-VL-Max[6] | | 75.9% 73.8% | 85.8% 71.5% | 81.7% 55.6% | 77.0% | | --- | ----------- | ----------- | ----------- | ----- | 25 Table13: CircularEvalresultsonMMBench-CN-testset(L-2abilities). Open-sourcemodels taggedwith*incorporatein-housedatainmodeltraining. | Model | Overall CP | FP-S FP-C | AR LR | RR | | ----- | ---------- | --------- | ----- | --- | OpenSourceVLMs | MiniGPT4-7B[56] | 10.8% 9.4% | 11.8% 5.6% | 24.5% 4.9% | 8.5% | | -------------------- | ----------- | ---------- | ----------- | ----- | | MiniGPT4-13B[56] | 13.2% 16.3% | 13.5% 9.0% | 27.3% 3.8% | 4.3% | | OpenFlamingov2[4] | 13.3% 16.5% | 10.2% 9.0% | 18.8% 11.4% | 12.4% | | InstructBLIP-13B[11] | 13.7% 13.7% | 14.6% 6.4% | 26.5% 4.3% | 14.3% | | InstructBLIP-7B[11] | 18.1% 15.7% | 18.6% 9.4% | 31.4% 8.7% | 25.2% | IDEFICS-9B-Instruct[26] 19.6% 22.4% 17.4% 7.1% 35.9% 6.0% 24.4% IDEFICS-80B-Instruct[26] 28.8% 33.0% 26.9% 25.1% 41.2% 13.6% 26.0% | PandaGPT[43] | 29.6% 40.4% | 20.0% 12.0% | 49.8% 13.0% | 34.1% | | ---------------- | ----------- | ----------- | ----------- | ----- | | VisualGLM-6B[13] | 38.1% 44.8% | 39.4% 22.8% | 55.5% 18.5% | 34.9% | CogVLM-Chat-17B[47] 54.0% 66.1% 49.7% 47.6% 67.8% 26.1% 49.6% | LLaVA-v1.5-7B[32] | 56.9% 65.2% | 53.6% 52.1% | 75.5% 31.0% | 50.8% | | ------------------ | ----------- | ----------- | ----------- | ----- | | Qwen-VL-Chat*[6] | 57.5% 63.0% | 64.5% 41.6% | 74.7% 35.9% | 50.0% | | mPLUG-Owl2[50] | 58.0% 64.4% | 57.1% 50.2% | 75.1% 31.5% | 56.6% | | ShareGPT4V-7B[8] | 58.3% 67.2% | 58.2% 51.3% | 72.7% 28.3% | 54.7% | | MiniCPM-V[39] | 59.6% 64.8% | 66.6% 52.8% | 69.0% 33.2% | 54.3% | | OmniLMM-12B*[38] | 60.8% 64.8% | 66.4% 53.9% | 74.7% 30.4% | 58.9% | | LLaVA-v1.5-13B[32] | 62.2% 68.3% | 61.5% 56.9% | 73.5% 35.9% | 64.3% | | ShareGPT4V-13B[8] | 62.7% 69.6% | 63.6% 56.2% | 74.7% 36.4% | 60.9% | | Yi-VL-6B*[2] | 63.5% 68.7% | 71.7% 52.4% | 74.7% 39.7% | 56.6% | LLaVA-InternLM-7B[10] 64.1% 70.7% 63.8% 55.8% 75.5% 39.7% 65.5% | Monkey-Chat[30] | 65.0% 71.5% | 68.9% 52.1% | 80.0% 46.7% | 57.4% | | --------------- | ----------- | ----------- | ----------- | ----- | | Yi-VL-34B*[2] | 66.2% 69.6% | 75.6% 56.2% | 80.0% 37.0% | 61.2% | InternLM-XComposer*[52] 69.2% 74.8% 71.7% 58.1% 80.8% 39.1% 75.6% LLaVA-InternLM2-7B[10] 69.9% 75.4% 72.9% 63.7% 81.2% 42.4% 68.6% LLaVA-InternLM2-20B[10] 70.3% 75.6% 73.5% 67.4% 75.1% 46.2% 69.4% InternLM-XComposer2*[12] 77.1% 80.4% 82.8% 71.2% 88.2% 55.4% 72.1% ProprietaryVLMs | Qwen-VL-Plus[6] | 67.9% 69.6% | 78.4% 60.3% | 75.1% 48.9% | 61.2% | | ---------------- | ----------- | ----------- | ----------- | ----- | | Gemini-Pro-V[44] | 69.2% 68.1% | 77.3% 64.0% | 80.4% 45.7% | 69.8% | | GPT-4v[37] | 72.1% 75.0% | 70.1% 70.0% | 82.4% 60.9% | 69.4% | Qwen-VL-Max[6] | | 73.6% 74.4% | 82.6% 69.3% | 79.2% 55.4% | 69.0% | | --- | ----------- | ----------- | ----------- | ----- | 26 References [1] W3cschool. Inhttps://www.w3schools.com/,2023. 17 [2] 01-ai. Yi-vl. https://huggingface.co/01-ai/Yi-VL-34B,2023. 4,8,10,22,23,24,25,26 [3] HarshAgrawal,KaranDesai,YufeiWang,XinleiChen,RishabhJain,MarkJohnson,Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. InProceedingsoftheIEEE/CVFinternationalconferenceoncomputervision,pages 8948–8957,2019. 3 [4] Jean-BaptisteAlayrac, JeffDonahue, PaulineLuc, AntoineMiech, IainBarr, YanaHasson, KarelLenc,ArthurMensch,KatherineMillican,MalcolmReynolds,etal. Flamingo: avisual languagemodelforfew-shotlearning. AdvancesinNeuralInformationProcessingSystems, 35:23716–23736,2022. 4,7,8,10,22,23,24,25,26 [5] JinzeBai,ShuaiBai,YunfeiChu,ZeyuCui,KaiDang,XiaodongDeng,YangFan,Wenbin Ge,YuHan,FeiHuang,BinyuanHui,LuoJi,MeiLi,JunyangLin,RunjiLin,DayihengLiu, GaoLiu,ChengqiangLu,KemingLu,JianxinMa,RuiMen,XingzhangRen,XuanchengRen, ChuanqiTan,SinanTan,JianhongTu,PengWang,ShijieWang,WeiWang,ShengguangWu, BenfengXu,JinXu,AnYang,HaoYang,JianYang,ShushengYang,YangYao,BowenYu, HongyiYuan,ZhengYuan,JianweiZhang,XingxuanZhang,YichangZhang,ZhenruZhang, ChangZhou,JingrenZhou,XiaohuanZhou,andTianhangZhu. Qwentechnicalreport. arXiv preprintarXiv:2309.16609,2023. 17 [6] JinzeBai,ShuaiBai,ShushengYang,ShijieWang,SinanTan,PengWang,JunyangLin,Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXivpreprintarXiv:2308.12966,2023. 4,8,10,20,22,23,24,25,26 [7] TomBrown,BenjaminMann,NickRyder,MelanieSubbiah,JaredDKaplan,PrafullaDhariwal, ArvindNeelakantan,PranavShyam,GirishSastry,AmandaAskell,etal. Languagemodelsare few-shotlearners. Advancesinneuralinformationprocessingsystems,33:1877–1901,2020. 4 [8] LinChen,JisongLi,XiaoyiDong,PanZhang,ConghuiHe,JiaqiWang,FengZhao,andDahua Lin. Sharegpt4v: Improvinglargemulti-modalmodelswithbettercaptions. arXivpreprint arXiv:2311.12793,2023. 4,22,23,24,25,26 [9] XinleiChen,HaoFang,Tsung-YiLin,RamakrishnaVedantam,SaurabhGupta,PiotrDollár, andCLawrenceZitnick. Microsoftcococaptions: Datacollectionandevaluationserver. arXiv preprintarXiv:1504.00325,2015. 2,3,17 [10] XTunerContributors. Xtuner: Atoolkitforefficientlyfine-tuningllm. https://github. com/InternLM/xtuner,2023. 4,10,22,23,24,25,26 [11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang,BoyangLi,PascaleFung,andStevenHoi. Instructblip: Towardsgeneral-purposevision- languagemodelswithinstructiontuning. arXivpreprintarXiv:2305.06500,2023. 4,8,10,20, 22,23,24,25,26 [12] XiaoyiDong,PanZhang,YuhangZang,YuhangCao,BinWang,LinkeOuyang,XilinWei, SongyangZhang,HaodongDuan,MaosongCao,WenweiZhang,YiningLi,HangYan,Yang Gao,XinyueZhang,WeiLi,JingwenLi,KaiChen,ConghuiHe,XingchengZhang,YuQiao, DahuaLin,andJiaqiWang.Internlm-xcomposer2:Masteringfree-formtext-imagecomposition andcomprehensioninvision-languagelargemodel. arXivpreprintarXiv:2401.16420,2024. 4, 8,9,10,22,23,24,25,26 [13] ZhengxiaoDu, YujieQian, XiaoLiu, MingDing, JiezhongQiu, ZhilinYang, andJieTang. Glm: Generallanguagemodelpretrainingwithautoregressiveblankinfilling. InProceedingsof the60thAnnualMeetingoftheAssociationforComputationalLinguistics(Volume1: Long Papers),pages320–335,2022. 10,20,22,23,24,25,26 [14] HaodongDuan,JunmingYang,YuxuanQiao,XinyuFang,LinChen,YuanLiu,XiaoyiDong, YuhangZang,PanZhang,JiaqiWang,DahuaLin,andKaiChen. Vlmevalkit: Anopen-source toolkitforevaluatinglargemulti-modalitymodels,2024. 1,8 27 [15] HaodongDuan,YueZhao,KaiChen,YuanjunXiong,andDahuaLin. Mitigatingrepresentation | biasinactionrecognition: | Algorithmsandbenchmarks,2022. | | 17 | | ------------------------ | ----------------------------- | -------------- | --- | | [16] JerryAFodor. | Themodularityofmind. | MITpress,1983. | 4 | [17] ChaoyouFu,PeixianChen,YunhangShen,YuleiQin,MengdanZhang,XuLin,ZhenyuQiu, WeiLin,JinruiYang,XiawuZheng,KeLi,XingSun,andRongrongJi.Mme:Acomprehensive evaluationbenchmarkformultimodallargelanguagemodels. ArXiv,abs/2306.13394,2023. 3, 18 [18] TaoGong,ChengqiLyu,ShilongZhang,YudongWang,MiaoZheng,QianZhao,KuikunLiu, WenweiZhang,PingLuo,andKaiChen. Multimodal-gpt: Avisionandlanguagemodelfor dialoguewithhumans,2023. 2 [19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making thevinvqamatter: Elevatingtheroleofimageunderstandinginvisualquestionanswering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages | 6904–6913,2017. | 2,3 | | | | --------------- | --- | --- | --- | [20] DannaGurari,QingLi,AbigaleJStangl,AnhongGuo,ChiLin,KristenGrauman,JieboLuo, andJeffreyPBigham. Vizwizgrandchallenge: Answeringvisualquestionsfromblindpeople. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages | 3608–3617,2018. | 3 | | | | --------------- | --- | --- | --- | [21] V.Hosu,H.Lin,T.Sziranyi,andD.Saupe. Koniq-10k: Anecologicallyvaliddatabasefor deep learning of blind image quality assessment. IEEE Transactions on Image Processing, | 29:4041–4056,2020. | 17 | | | | ------------------ | --- | --- | --- | [22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, arXiv Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. | preprintarXiv:2106.09685,2021. | 4 | | | | ------------------------------ | --- | --- | --- | [23] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoningandcompositionalquestionanswering. InProceedingsoftheIEEE/CVFconference oncomputervisionandpatternrecognition,pages6700–6709,2019. 2,3,19,21 [24] JustinJohnson,BharathHariharan,LaurensVanDerMaaten,LiFei-Fei,CLawrenceZitnick, andRossGirshick. Clevr: Adiagnosticdatasetforcompositionallanguageandelementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern | recognition,pages2901–2910,2017. | | 17 | | | -------------------------------- | --- | --- | --- | [25] AlexanderKuhnleandAnnCopestake. Shapeworld-anewtestmethodologyformultimodal | languageunderstanding. | arXivpreprintarXiv:1704.04517,2017. | | 17 | | ---------------------- | ----------------------------------- | --- | --- | [26] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov,ThomasWang,SiddharthKaramcheti,AlexanderM.Rush,DouweKiela,Matthieu Cord,andVictorSanh. Obelics: Anopenweb-scalefiltereddatasetofinterleavedimage-text | documents,2023. | 8,10,20,22,23,24,25,26 | | | | --------------- | ---------------------- | --- | --- | [27] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed- bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125,2023. 18 [28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597,2023. 4 [29] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Dual-glance model for deciphering social relationships. In Proceedings of the IEEE international conference on | computervision,pages2650–2659,2017. | | 17 | | | ----------------------------------- | --- | --- | --- | [30] ZhangLi,BiaoYang,QiangLiu,ZhiyinMa,ShuoZhang,JingxuYang,YaboSun,Yuliang Liu,andXiangBai. Monkey: Imageresolutionandtextlabelareimportantthingsforlarge | multi-modalmodels. | arXivpreprintarXiv:2311.06607,2023. | | 22,23,24,25,26 | | ------------------ | ----------------------------------- | --- | -------------- | 28 [31] FangyuLiu,GuyEdwardTohEmerson,andNigelCollier.Visualspatialreasoning.Transactions oftheAssociationforComputationalLinguistics,2023. 17 [32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instructiontuning. arXivpreprintarXiv:2310.03744,2023. 4,8,10,22,23,24,25,26 [33] HaotianLiu,ChunyuanLi,QingyangWu,andYongJaeLee. Visualinstructiontuning. arXiv preprintarXiv:2304.08485,2023. 2,4,9,17 [34] PanLu,SwaroopMishra,TanglinXia,LiangQiu,Kai-WeiChang,Song-ChunZhu,Oyvind Tafjord,PeterClark,andAshwinKalyan. Learntoexplain: Multimodalreasoningviathought chainsforsciencequestionanswering. AdvancesinNeuralInformationProcessingSystems, 35:2507–2521,2022. 3,17 [35] KennethMarino,MohammadRastegari,AliFarhadi,andRoozbehMottaghi. Ok-vqa: Avisual questionansweringbenchmarkrequiringexternalknowledge. InProceedingsoftheIEEE/cvf conferenceoncomputervisionandpatternrecognition,pages3195–3204,2019. 2,3,6,19,21 [36] MikeOaksfordandNickChater. Bayesianrationality: Theprobabilisticapproachtohuman reasoning. OxfordUniversityPress,2007. 4 [37] OpenAI. Gpt-4technicalreport. ArXiv,abs/2303.08774,2023. 1,2,4,5,8,10,17,20,23,24, 25,26 [38] OpenBMB. Omnilmm: Largemulti-modalmodelsforstrongperformanceandefficientdeploy- ment. https://github.com/OpenBMB/OmniLMM,2023. 22,23,24,25,26 [39] OpenBMB. Minicpm: Unveilingthepotentialofend-sidelargelanguagemodels,2024. 8,10, 20,22,23,24,25,26 [40] LongOuyang,JeffreyWu,XuJiang,DiogoAlmeida,CarrollWainwright,PamelaMishkin, ChongZhang,SandhiniAgarwal,KatarinaSlama,AlexRay,etal. Traininglanguagemodelsto followinstructionswithhumanfeedback. AdvancesinNeuralInformationProcessingSystems, 35:27730–27744,2022. 4 [41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Languagemodelsareunsupervisedmultitasklearners. OpenAIblog,1(8):9,2019. 4 [42] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVFconferenceoncomputervisionandpatternrecognition,pages8317–8326,2019. 3, 17,19,21 [43] YixuanSu,TianLan,HuayangLi,JialuXu,YanWang,andDengCai. Pandagpt: Onemodelto instruction-followthemall. arXivpreprintarXiv:2305.16355,2023. 22,23,24,25,26 [44] GeminiTeam,RohanAnil,SebastianBorgeaud,YonghuiWu,Jean-BaptisteAlayrac,JiahuiYu, RaduSoricut,JohanSchalkwyk,AndrewMDai,AnjaHauth,etal. Gemini: afamilyofhighly capablemultimodalmodels. arXivpreprintarXiv:2312.11805,2023. 2,4,5,8,10,17,20,23, 24,25,26 [45] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM-techreport,2023. 8, 9,18 [46] HugoTouvron,ThibautLavril,GautierIzacard,XavierMartinet,Marie-AnneLachaux,Timo- théeLacroix,BaptisteRozière,NamanGoyal,EricHambro,FaisalAzhar,etal. Llama: Open andefficientfoundationlanguagemodels. arXivpreprintarXiv:2302.13971,2023. 4 [47] WeihanWang,QingsongLv,WenmengYu,WenyiHong,JiQi,YanWang,JunhuiJi,Zhuoyi Yang,LeiZhao,XixuanSong,JiazhengXu,BinXu,JuanziLi,YuxiaoDong,MingDing,and JieTang. Cogvlm: Visualexpertforpretrainedlanguagemodels. ArXiv,abs/2311.03079,2023. 8,10,22,23,24,25,26 29 [48] PengXu,WenqiShao,KaipengZhang,PengGao,ShuoLiu,MengLei,FanqingMeng,Siyuan Huang,YuJiaoQiao,andPingLuo. Lvlm-ehub: Acomprehensiveevaluationbenchmarkfor | largevision-languagemodels. | 2023. 2 | | | --------------------------- | ------- | --- | [49] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large languagemodelswithmultimodality. arXivpreprintarXiv:2304.14178,2023. 2,3,4,22 [50] QinghaoYe,HaiyangXu,JiaboYe,MingshiYan,AnwenHu,HaoweiLiu,QiQian,JiZhang, FeiHuang,andJingrenZhou. mplug-owl2: Revolutionizingmulti-modallargelanguagemodel | withmodalitycollaboration. | ArXiv,abs/2311.04257,2023. | 8,10,23,24,25,26 | | -------------------------- | -------------------------- | ---------------- | [51] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions tovisualdenotations: Newsimilaritymetricsforsemanticinferenceovereventdescriptions. TransactionsoftheAssociationforComputationalLinguistics,2:67–78,2014. 3 [52] PanZhang, XiaoyiDongBinWang, YuhangCao, ChaoXu, LinkeOuyang, ZhiyuanZhao, ShuangruiDing,SongyangZhang,HaodongDuan,HangYan,etal. Internlm-xcomposer: A vision-languagelargemodelforadvancedtext-imagecomprehensionandcomposition. arXiv | preprintarXiv:2309.15112,2023. | 22,23,24,25,26 | | | ------------------------------ | -------------- | --- | [53] LianminZheng,Wei-LinChiang,YingSheng,SiyuanZhuang,ZhanghaoWu,YonghaoZhuang, ZiLin,ZhuohanLi,DachengLi,Eric.PXing,HaoZhang,JosephE.Gonzalez,andIonStoica. | Judgingllm-as-a-judgewithmt-benchandchatbotarena,2023. | | 2,4,9 | | ------------------------------------------------------ | --- | ----- | [54] BoleiZhou,AgataLapedriza,AdityaKhosla,AudeOliva,andAntonioTorralba. Places: A 10millionimagedatabaseforscenerecognition. IEEETransactionsonPatternAnalysisand | MachineIntelligence,2017. | 17 | | | ------------------------- | --- | --- | [55] LuoweiZhou,ChenliangXu,andJasonCorso. Towardsautomaticlearningofproceduresfrom webinstructionalvideos. InProceedingsoftheAAAIConferenceonArtificialIntelligence, volume32,2018. 3 [56] DeyaoZhu, JunChen, XiaoqianShen, XiangLi, andMohamedElhoseiny. Minigpt-4: En- hancingvision-languageunderstandingwithadvancedlargelanguagemodels. arXivpreprint | arXiv:2304.10592,2023. | 2,4,8,10,20,22,23,24,25,26 | | | ---------------------- | -------------------------- | --- | 30