학술논문

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context
Document Type
Working Paper
Author
Gemini TeamGeorgiev, PetkoLei, Ving IanBurnell, RyanBai, LibinGulati, AnmolTanzer, GarrettVincent, DamienPan, ZhufengWang, ShiboMariooryad, SorooshDing, YifanGeng, XinyangAlcober, FredFrostig, RoyOmernick, MarkWalker, LexiPaduraru, CosminSorokin, ChristinaTacchetti, AndreaGaffney, ColinDaruki, SamiraSercinoglu, OlcanGleicher, ZachLove, JulietteVoigtlaender, PaulJain, RohanSurita, GabrielaMohamed, KareemBlevins, RoryAhn, JunwhanZhu, TaoKawintiranon, KornraphopFirat, OrhanGu, YimingZhang, YujingRahtz, MatthewFaruqui, ManaalClay, NatalieGilmer, JustinCo-Reyes, JDPenchev, IvoZhu, RuiMorioka, NobuyukiHui, KevinHaridasan, KrishnaCampos, VictorMahdieh, MahdisGuo, MandyHassan, SamerKilgour, KevinVezer, ArpiCheng, Heng-Tzede Liedekerke, RaoulGoyal, SiddharthBarham, PaulStrouse, DJNoury, SebAdler, JonasSundararajan, MukundVikram, SharadLepikhin, DmitryPaganini, MichelaGarcia, XavierYang, FanValter, DashaTrebacz, MajaVodrahalli, KiranAsawaroengchai, ChulayuthRing, RomanKalb, NorbertSoares, Livio BaldiniBrahma, SiddharthaSteiner, DavidYu, TianheMentzer, FabianHe, AntoineGonzalez, LucasXu, BiboKaufman, Raphael LopezShafey, Laurent ElOh, JunhyukHennigan, TomDriessche, George van denOdoom, SethLucic, MarioRoelofs, BeccaLall, SidMarathe, AmitChan, BettyOntanon, SantiagoHe, LuhengTeplyashin, DenisLai, JonathanCrone, PhilDamoc, BogdanHo, LewisRiedel, SebastianLenc, KarelYeh, Chih-KuanChowdhery, AakankshaXu, YangKazemi, MehranAmid, EhsanPetrushkina, AnastasiaSwersky, KevinKhodaei, AliChen, GowoonLarkin, ChrisPinto, MarioYan, GengBadia, Adria PuigdomenechPatil, PiyushHansen, StevenOrr, DaveArnold, Sebastien M. R.Grimstad, JordanDai, AndrewDouglas, SholtoSinha, RishikaYadav, VikasChen, XiGribovskaya, ElenaAustin, JacobZhao, JeffreyPatel, KaushalKomarek, PaulAustin, SophiaBorgeaud, SebastianFriso, LindaGoyal, AbhimanyuCaine, BenCao, KrisChung, Da-WoonLamm, MatthewBarth-Maron, GabeKagohara, ThaisOlszewska, KateChen, MiaShivakumar, KaushikAgarwal, RishabhGodhia, HarshalRajwar, RaviSnaider, JavierDotiwalla, XerxesLiu, YuanBarua, AdityaUngureanu, VictorZhang, YuanBatsaikhan, Bat-OrgilWirth, MateoQin, JamesDanihelka, IvoDoshi, TulseeChadwick, MartinChen, JilinJain, SanilLe, QuocKar, ArjunGurumurthy, MadhuLi, ChengSang, RuoxinLiu, FangyuLamprou, LamprosMunoz, RichLintz, NathanMehta, HarshHoward, HeidiReynolds, MalcolmAroyo, LoraWang, QuanBlanco, LorenzoCassirer, AlbinGriffith, JordanDas, DipanjanLee, StephanSygnowski, JakubFisher, ZachBesley, JamesPowell, RichardAhmed, ZafaraliPaulus, DominikReitter, DavidBorsos, ZalanJoshi, RishabhPope, AedanHand, StevenSelo, VittorioJain, VihanSethi, NikhilGoel, MeghaMakino, TakakiMay, RhysYang, ZhenSchalkwyk, JohanButterfield, ChristinaHauth, AnjaGoldin, AlexHawkins, WillSenter, EvanBrin, SergeyWoodman, OliverRitter, MarvinNoland, EricGiang, MinhBolina, VijayLee, LisaBlyth, TimMackinnon, IanReid, MachelSarvana, ObaidSilver, DavidChen, AlexanderWang, LilyMaggiore, LorenChang, OscarAttaluri, NithyaThornton, GregoryChiu, Chung-ChengBunyan, OskarLevine, NirChung, TimothyEltyshev, EvgeniiSi, XianceLillicrap, TimothyBrady, DemetraAggarwal, VaibhavWu, BoxiXu, YuanzhongMcIlroy, RossBadola, KartikeyaSandhu, ParamjitMoreira, EricaStokowiec, WojciechHemsley, RossLi, DongTudor, AlexShyam, PranavRahimtoroghi, ElaheHaykal, SalemSprechmann, PabloZhou, XiangMincu, DianaLi, YujiaAddanki, RaviKrishna, KalpeshWu, XiaoFrechette, AlexandreEyal, MatanDafoe, AllanLacey, DaveWhang, JayAvrahami, ThiZhang, YeTaropa, EmanuelLin, HanzhaoToyama, DanielRutherford, ElizaSano, MotokiChoe, HyunJeongTomala, AlexSafranek-Shrader, ChalenceKassner, NoraPajarskas, MantasHarvey, MattSechrist, SeanFortunato, MeireLyu, ChristinaElsayed, GamaleldinKuang, ChenkaiLottes, JamesChu, EricJia, ChaoChen, Chih-WeiHumphreys, PeterBaumli, KateTao, ConnieSamuel, RajkumarSantos, Cicero Nogueira dosAndreassen, AndersRakićević, NemanjaGrewe, DominikKumar, AviralWinkler, StephanieCaton, JonathanBrock, AndrewDalmia, SidSheahan, HannahBarr, IainMiao, YingjieNatsev, PaulDevlin, JacobBehbahani, FeryalProst, FlavienSun, YanhuaMyaskovsky, ArtiomPillai, Thanumalayan SankaranarayanaHurt, DanLazaridou, AngelikiXiong, XiZheng, CePardo, FabioLi, XiaoweiHorgan, DanStanton, JoeAmbar, MoranXia, FeiLince, AlejandroWang, MingqiuMustafa, BasilWebson, AlbertLee, HyoAnil, RohanWicke, MartinDozat, TimothySinha, AbhishekPiqueras, EnriqueDabir, ElaheUpadhyay, ShyamBoral, AnudhyanHendricks, Lisa AnneFry, CoreyDjolonga, JosipSu, YiWalker, JakeLabanowski, JaneHuang, RonnyMisra, VedantChen, JeremySkerry-Ryan, RJSingh, AviRijhwani, ShrutiYu, DianCastro-Ros, AlexChangpinyo, BeerDatta, RominaBagri, SumitHrafnkelsson, Arnar MarMaggioni, MarcelloZheng, DanielSulsky, YuryHou, ShaoboPaine, Tom LeYang, AntoineRiesa, JasonRogozinska, DominikaMarcus, DrorBadawy, Dalia ElZhang, QiaoWang, LuyuMiller, HelenGreer, JeremySjos, Lars LoweNova, AzadeZen, HeigaChaabouni, RahmaRosca, MihaelaJiang, JiepuChen, CharlieLiu, RuiboSainath, TaraKrikun, MaximPolozov, AlexLespiau, Jean-BaptisteNewlan, JoshCankara, ZeyncepKwak, SooXu, YunhanChen, PhilCoenen, AndyMeyer, ClemensTsihlas, KaterinaMa, AdaGottweis, JurajXing, JinweiGu, ChenjieMiao, JinFrank, ChristianCankara, ZeynepGanapathy, SanjayDasgupta, IshitaHughes-Fitt, StephChen, HengReid, DavidRong, KeranFan, Hongminvan Amersfoort, JoostZhuang, VincentCohen, AaronGu, Shixiang ShaneMohananey, AnhadIlic, AnastasijaTobin, TaylorWieting, JohnBortsova, AnnaThacker, PhoebeWang, EmmaCaveness, EmilyChiu, JustinSezener, ErenKaskasoli, AlexBaker, StevenMillican, KatieElhawaty, MohamedAisopos, KostasLebsack, CarlByrd, NathanDai, HanjunJia, WenhaoWiethoff, MatthewDavoodi, ElnazWeston, AlbertYagati, LakshmanAhuja, ArunGao, IsabelPundak, GolanZhang, SusanAzzam, MichaelSim, Khe ChaiCaelles, SergiKeeling, JamesSharma, AbhanshuSwing, AndyLi, YaGuangLiu, ChenxiBostock, Carrie GrimesBansal, YaminiNado, ZacharyAnand, AnkeshLipschultz, JoshKarmarkar, AbhijitProleev, LevIttycheriah, AbeYeganeh, Soheil HassasPolovets, GeorgeFaust, AleksandraSun, JiaoRrustemi, AlbanLi, PenShivanna, RakeshLiu, JeremiahWelty, ChrisLebron, FedericoBaddepudi, AnirudhKrause, SebastianParisotto, EmilioSoricut, RaduXu, ZhengBloxwich, DawnJohnson, MelvinNeyshabur, BehnamMao-Jones, JustinWang, RenshenRamasesh, VinayAbbas, ZaheerGuez, ArthurSegal, ConstantNguyen, Duc DungSvensson, JamesHou, LeYork, SarahMilan, KieranBridgers, SophieGworek, WiktorTagliasacchi, MarcoLee-Thorp, JamesChang, MichaelGuseynov, AlexeyHartman, Ale JakseKwong, MichaelZhao, RuizheKashem, SheleemCole, ElizabethMiech, AntoineTanburn, RichardPhuong, MaryPavetic, FilipCevey, SebastienComanescu, RamonaIves, RichardYang, SherryDu, CosmoLi, BoZhang, ZizhaoIinuma, MarikoHu, Clara HuiyiRoy, AurkoBijwadia, ShaanZhu, ZhenkaiMartins, DaniloSaputro, RachelGergely, AnitaZheng, StevenJia, DaweiAntonoglou, IoannisSadovsky, AdamGu, ShaneBi, YingyingAndreev, AlekSamangooei, SinaKhan, MinaKocisky, TomasFilos, AngelosKumar, ChintuBishop, ColtonYu, AdamsHodkinson, SarahMittal, SidShah, PremalMoufarek, AlexandreCheng, YongBloniarz, AdamLee, JaehoonPejman, PedramMichel, PaulSpencer, StephenFeinberg, VladimirXiong, XuehanSavinov, NikolaySmith, CharlotteShakeri, SiamakTran, DustinChesus, MaryBohnet, BerndTucker, Georgevon Glehn, TamaraMuir, CarrieMao, YiranKazawa, HidetoSlone, AmbroseSoparkar, KedarShrivastava, DishaCobon-Kerr, JamesSharman, MichaelPavagadhi, JayAraya, CarlosMisiunas, KarolisGhelani, NimeshLaskin, MichaelBarker, DavidLi, QiujiaBriukhov, AntonHoulsby, NeilGlaese, MiaLakshminarayanan, BalajiSchucher, NathanTang, YunhaoCollins, EliLim, HyeontaekFeng, FangxiaoyuRecasens, AdriaLai, GuangdaMagni, AlbertoDe Cao, NicolaSiddhant, AdityaAshwood, ZoeOrbay, JordiDehghani, MostafaBrennan, JennyHe, YifanXu, KelvinGao, YangSaroufim, CarlMolloy, JamesWu, XinyiArnold, SebChang, SolomonSchrittwieser, JulianBuchatskaya, ElenaRadpour, SoroushPolacek, MartinGiordano, SkyeBapna, AnkurTokumine, SimonHellendoorn, VincentSottiaux, ThibaultCogan, SarahSeveryn, AliakseiSaleh, MohammadThakoor, ShantanuShefey, LaurentQiao, SiyuanGaba, MeenuChang, Shuo-yiinSwanson, CraigZhang, BiaoLee, BenjaminRubenstein, Paul KishanSong, GanKwiatkowski, TomKoop, AnnaKannan, AjayKao, DavidSchuh, ParkerStjerngren, AxelGhiasi, GolnazGibson, GenaVilnis, LukeYuan, YeFerreira, Felipe TiengoKamath, AishwaryaKlimenko, TedFranko, KenXiao, KefanBhattacharya, IndroPatel, MiteyanWang, RuiMorris, AlexStrudel, RobinSharma, VivekChoy, PeterHashemi, Sayed HadiLandon, JessicaFinkelstein, MaraJhakra, PriyaFrye, JustinBarnes, MeganMauger, MatthewDaun, DennisBaatarsukh, KhuslenTung, MatthewFarhan, WaelMichalewski, HenrykViola, FabioQuitry, Felix de ChaumontLan, Charline LeHudson, TomWang, QingzeFischer, FelixZheng, IvyWhite, ElspethDragan, AncaAlayrac, Jean-baptisteNi, EricPritzel, AlexanderIwanicki, AdamIsard, MichaelBulanova, AnnaZilka, LukasDyer, EthanSachan, DevendraSrinivasan, SrivatsanMuckenhirn, HannahCai, HonglongMandhane, AmolTariq, MukarramRae, Jack W.Wang, GaryAyoub, KareemFitzGerald, NicholasZhao, YaoHan, WoohyunAlberti, ChrisGarrette, DanKrishnakumar, KashyapGimenez, MaiLevskaya, AnselmSohn, DanielMatak, JosipIturrate, InakiChang, Michael B.Xiang, JackieCao, YuanRanka, NishantBrown, GeoffHutter, AdrianMirrokni, VahabChen, NanxinYao, KaishengEgyed, ZoltanGalilee, FrancoisLiechty, TylerKallakuri, PraveenPalmer, EvanGhemawat, SanjayLiu, JasmineTao, DavidThornton, ChloeGreen, TimJasarevic, MimiLin, SharonCotruta, VictorTan, Yi-XuanFiedel, NoahYu, HongkunChi, EdNeitz, AlexanderHeitkaemper, JensSinha, AnuZhou, DennySun, YiKaed, CharbelHulse, BriceMishra, SwaroopGeorgaki, MariaKudugunta, SnehaFarabet, ClementShafran, IzhakVlasic, DanielTsitsulin, AntonAnanthanarayanan, RajagopalCarin, AlenSu, GuolongSun, PeiV, ShashankCarvajal, GabrielBroder, JosefComsa, IuliaRepina, AlenaWong, WilliamChen, Warren WeilunHawkins, PeterFilonov, EgorLoher, LuciaHirnschall, ChristophWang, WeiyiYe, JingchenBurns, AndreaCate, HardieWright, Diana GagePiccinini, FedericoZhang, LeiLin, Chu-ChengGog, IonelKulizhskaya, YanaSreevatsa, AshwinSong, ShuangCobo, Luis C.Iyer, AnandTekur, ChetanGarrido, GuillermoXiao, ZhuyunKemp, RupertZheng, Huaixiu StevenLi, HuiAgarwal, AnanthNgani, ChristelGoshvadi, KatiSantamaria-Fernandez, RebecaFica, WojciechChen, XinyunGorgolewski, ChrisSun, SeanGarg, RoopalYe, XinyuEslami, S. M. AliHua, NanSimon, JonJoshi, PratikKim, YelinTenney, IanPotluri, SahityaThiet, Lam NguyenYuan, QuanLuisier, FlorianChronopoulou, AlexandraScellato, SalvatoreSrinivasan, PraveenChen, MinminKoverkathu, VinodDalibard, ValentinXu, YamingSaeta, BrennanAnderson, KeithSellam, ThibaultFernando, NickHuot, FantineJung, JunehyukVaradarajan, ManiQuinn, MichaelRaul, AmitLe, MaigoHabalov, RuslanClark, JonJalan, KomalBullard, KaleshaSinghal, AchintyaLuong, ThangWang, BoyuRajayogam, SujeevanEisenschlos, JulianJia, JohnsonFinchelstein, DanielYakubovich, AlexBalle, DanielFink, MichaelAgarwal, SameerLi, JingDvijotham, DjPal, ShaliniKang, KaiKonzelmann, JaclynBeattie, JenniferDousse, OlivierWu, DianeCrocker, RemiElkind, ChenJonnalagadda, Siddhartha ReddyLee, JongHoltmann-Rice, DanKallarackal, KrystalLiu, RosanneVnukov, DenisVats, NeeraInvernizzi, LucaJafari, MohsenZhou, HuanjieTaylor, LillyPrendki, JenniferWu, MarcusEccles, TomLiu, TianqiKopparapu, KavyaBeaufays, FrancoiseAngermueller, ChristofMarzoca, AndreeaSarcar, ShouryaDib, HilalStanway, JeffPerbet, FrankTrdin, NejcSterneck, RachelKhorlin, AndreyLi, DinghuaWu, XihuiGoenka, SonamMadras, DavidGoldshtein, SashaGierke, WilliZhou, TongLiu, YaxinLiang, YannieWhite, AnaisLi, YunjieSingh, ShreyaBahargam, SanazEpstein, MarkBasu, SujoyLao, LiOzturel, AdnanCrous, CarlZhai, AlexLu, HanTung, ZoraGaur, NeerajWalton, AlannaDixon, LucasZhang, MingGloberson, AmirUy, GrantBolt, AndrewWiles, OliviaNasr, MiladShumailov, IliaSelvi, MarcoPiccinno, FrancescoAguilar, RicardoMcCarthy, SaraKhalman, MishaShukla, MrinalGalic, VladoCarpenter, JohnVillela, KevinZhang, HaibinRichardson, HarryMartens, JamesBosnjak, MatkoBelle, Shreyas RammohanSeibert, JeffAlnahlawi, MahmoudMcWilliams, BrianSingh, SankalpLouis, AnnieDing, WenPopovici, DanSimicich, LeninKnight, LauraMehta, PulkitGupta, NisheshShi, ChongyangFatehi, SaaberMitrovic, JovanaGrills, AlexPagadora, JosephPetrova, DessieEisenbud, DanielleZhang, ZhishuaiYates, DamionMittal, BhavishyaTripuraneni, NileshAssael, YannisBrovelli, ThomasJain, PrateekVelimirovic, MihajloAkbulut, CanferMu, JiaqiMacherey, WolfgangKumar, RavinXu, JunQureshi, HaroonComanici, GheorgheWiesner, JeremyGong, ZhitaoRuddock, AntonBauer, MatthiasFelt, NickGP, AnirudhArnab, AnuragZelle, DustinRothfuss, JonasRosgen, BillShenoy, AshishSeybold, BryanLi, XinjianMudigonda, JayaramErdogan, GokerXia, JiaweiSimsa, JiriMichi, AndreaYao, YiYew, ChristopherKan, StevenCaswell, IsaacRadebaugh, CareyElisseeff, AndreValenzuela, PedroMcKinney, KayPaterson, KimCui, AlbertLatorre-Chimoto, EriKim, SolomonZeng, WilliamDurden, KenPonnapalli, PriyaSosea, TiberiuChoquette-Choo, Christopher A.Manyika, JamesRobenek, BronaVashisht, HarshaPereira, SebastienLam, HoiVelic, MarkoOwusu-Afriyie, DeneseLee, KatherineBolukbasi, TolgaParrish, AliciaLu, ShawnPark, JaneVenkatraman, BalajiTalbert, AliceRosique, LambertCheng, YuchungSozanschi, AndreiPaszke, AdamKumar, PraveenAustin, JessicaLi, LuSalama, KhalidKim, WooyeolDukkipati, NanditaBaryshnikov, AnthonyKaplanis, ChristosSheng, XiangHaiChervonyi, YuriUnlu, CaglarCasas, Diego de LasAskham, HarryTunyasuvunakool, KathrynGimeno, FelixPoder, SiimKwak, ChesterMiecnikowski, MattDimitriev, AlekParisi, AaronLiu, DangyiTsai, TomyShevlane, TobyKouridi, ChristinaGarmon, DrewGoedeckemeyer, AdrianBrown, Adam R.Vijayakumar, AnithaElqursh, AliJazayeri, SadeghHuang, JinCarthy, Sara McHoover, JayKim, LucyKumar, SandeepChen, WeiBiles, CourtneyBingham, GarrettRosen, EvanWang, LisaTan, QijunEngel, DavidPongetti, Francescode Cesare, DarioHwang, DongseongYu, LilyPullman, JenniferNarayanan, SriniLevin, KyleGopal, SiddharthLi, MeganAharoni, AsafTrinh, TrieuLo, JessicaCasagrande, NormanVij, RoopaliMatthey, LoicRamadhana, BramandiaMatthews, AustinCarey, CJJohnson, MatthewGoranova, KremenaShah, RohinAshraf, ShereenDasgupta, KingshukLarsen, RasmusWang, YichengVuyyuru, Manish ReddyJiang, ChongIjazi, JoanaOsawa, KazukiSmith, CelineBoppana, Ramya SreeBilal, TaylanKoizumi, YumaXu, YingAltun, YaseminShabat, NirBariach, BenKorchemniy, AlexChoo, KiamRonneberger, OlafIwuanyanwu, ChimezieZhao, ShubinSoergel, DavidHsieh, Cho-JuiCai, IreneIqbal, ShariqSundermeyer, MartinChen, ZheBursztein, ElieMalaviya, ChaitanyaBiadsy, FadiShroff, PrakashDhillon, InderjitLatkar, TejasiDyer, ChrisForbes, HannahNicosia, MassimoNikolaev, VitalyGreene, SomerGeorgiev, MarinWang, PidongMartin, NinaSedghi, HanieZhang, JohnBanzal, PraseemFritz, DougRao, VikramWang, XuezhiZhang, JiagengPatraucean, VioricaDu, DayouMordatch, IgorJurin, IvanLiu, LewisDubey, AyushMohan, AbhiNowakowski, JanekIon, Vlad-DoruWei, NanTojo, ReikoRaad, Maria AbiHudson, Drew A.Keshava, VaishakhAgrawal, ShubhamRamirez, KevinWu, ZhichunNguyen, HoangLiu, JiSewak, MadhaviPetrini, BryceChoi, DongHyunPhilips, IvanWang, ZiyueBica, IoanaGarg, AnkushWilkiewicz, JarekAgrawal, PriyankaGuo, DanhaoXue, EmilyShaik, NaseerLeach, AndrewKhan, Sadh MNMWiesinger, JuliaJerome, SammyChakladar, AbhishekWang, Alek WenjiaoOrnduff, TinaAbu, FolakeGhaffarkhah, AlirezaWainwright, MarcusCortes, MarioLiu, FrederickMaynez, JoshuaPetrov, SlavWu, YonghuiHassabis, DemisKavukcuoglu, KorayDean, JeffreyVinyals, Oriol
Source
Subject
Computer Science - Computation and Language
Computer Science - Artificial Intelligence
Language
Abstract
In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.