Chii chaitika
Vatsvagiri Oguz Serdar naCuneyt Mertayak vakatumira preprint inotsanangura SteerBench-Work, bhenji rekutanga-kuita "enderera kana kubata" sarudzo munzvimbo yebasa AI vamiririri. Vanoshuma kuti pane makumi matatu emhando mamiriro, mamodheru akabatwa zvisirizvo, ane humbowo-akacheneswa basa pa28.1% yemikana uye zvisizvo akabvumira basa risina kuchengeteka pa1.
Vatsvagiri vaviri, Oguz Serdar naCuneyt Mertayak, vakatumira preprint ku arXiv musi wa12 Nyamavhuvhu 2026 vachiunza SteerBench-Work, bhenji rakavakirwa padanho rimwe mukati mekufamba kwenguva refu kwemumiririri weAI: kuita chiito kana kuibata kuti iongororwe nevanhu kana mutemo. Vanyori vanodaidza iyi danho rekutungamira uye voiisa pane yavanodaidza muganho wechiito - iyo poindi mumiriri asati atumira email, kubatanidza chikumbiro chekudhonza, kana waya mubhadharo. Chiyereso hachina zvibodzwa kana mumiririri achikwanisa kupedza basa. Inoita zvibodzwa kuti mumiririri akayambuka here kana kuti akabata muganho nemazvo.
Kuburitswa kwakatsanangurwa muabstract kwakanyorwa kuti v2026-05 uye ine 106 scenarios inotora mashandiro evagadziri, basa revatengi, mari, zvemutemo, zvekurapa, HR, uye chengetedzo. Mamiriro acho akatsigirwa muzviitiko zveveruzhinji. Imwe neimwe inopetwa neizvo vanyori vanodana girazi-rakadzoserwa girazi - vhezheni yemamiriro akafanana umo humbowo huripo hunonongedza imwe nzira - padivi pezvinodzora zvinodzora. Mabhii akapatsanurwa zvakaenzana pakati pekuenderera mberi nekubata, izvo zvinonzi nevanyori ndezvemaune: inopa nzira mbiri dzekukanganisa padyo nenhamba imwechete yemikana yekuitika, saka modhi haigone kukora zvakanaka nekungotadza kungwarira. Muchinhu chimwe nechimwe, muenzaniso unoratidzwa chiito chakarongwa uye uchapupu huripo, uye inodzorera sarudzo yegedhi.
Pakati pemamiriro emhando makumi matatu, vanyori vanotaura kuti kukundikana kunomhanya zvakazara munzira imwe. Mamodheru akabatwa zvisirizvo akatenderwa, humbowo-akacheneswa basa pa28.1% yemikana, uku achibvumira zvisirizvo basa risina kuchengeteka pa1. Chikamu chakaomesesa, maererano neabstract, ndeizvo vanyori vanodana kuti njodzi-yakagadziriswa inoita: zviitiko apo njodzi chaiyo inokonzeresa yakadzingwa asi yakasainwa kana humbowo hwakarongwa hwatozvibvisa, uye mhinduro chaiyo ndeyekuenderera. Kuita zvakare kwakasiyana zvakanyanya pakati pezviitiko zvakakurumbira uye magirazi avo - 98.5% pazviitiko pachazvo zvichipesana ne63.8% pane humbowo-akadzoserwa shanduro.
Vanyori vanotora mumwe mutsauko pakati pekugona kuzere uye kutungamira calibration. Mienzaniso yepamusoro-soro, ivo vanonyora, kazhinji-kuramba-kuramba pamuganhu wekuzvipira, uye kufunga kwekuwedzera kunogona kugadzirisa gedhi risina simba uku richisiya rave-calibrated one flat. Iwo abstract anonongedzera kune yeruzhinji leaderboard, asi iro tsime peji rinopa kero sechibatiso kwete chekushanda chinongedzo.
Zvinhu zvakati wandei hazvina kusimbiswa nezvinhu zviripo pano. Iyo abstract haitauri mamodheru akaedzwa, haitsanangure izvo zvakakosha seyakasiyana "yemhando mamiriro" pakati pemakumi matatu, uye haipe pa-domain kana per-modhi yekuputsika. Izvo hazvitsanangure kuti ndiani akanyora pasi-chokwadi mavara kana kuti kusawirirana kwakagadziriswa sei, uye vhezheni v2026-05 haitsanangurwe maererano nezuva rekutumira Nyamavhuvhu. Iyi vhezheni-imwe preprint isina chiratidzo chekuongorora kwevezera, uye hapana kana nhamba yakaitwa yakazvimirira.
Kwakabva mashoko: arxiv.org β
Nei zvichikosha
Yakawanda yekuongorora kuchengetedza mumiririri inoyera kuti mamodheru anovharira zviito zvinokuvadza. Iri basa rinoyera nzira dzese dzekukanganisa uye rinowana kutadza kukuru kuri kuvharika, izvo zvinotakura mutengo chaiwo wekushandisa uye zvinogona kusundira vanhu murubber-stamping. Nhamba idzi dzinobva kune isina kuongororwa preprint uye hadzina kusimbiswa dzakazvimirira.
AI vamiririri vari kuwedzera kuisirwa kuita zviito pane kungoburitsa zvinyorwa, uye poindi apo zano rinova mhedzisiro isingadzoreki ndipo panoitika njodzi. Basa rakawanda rakaburitswa rekuchengetedza agent rinobvunza mubvunzo mumwechete - iyo modhi yakaramba chinhu chinokuvadza here? - uye sisitimu inoramba zvese zvinokwana pamubvunzo iwoyo asi isina basa. Nekuenzanisa enderera mberi uye kubata mavara uye kushuma ese ari maviri mareti ekukanganisa, iyi bhenji inoyera tradeoff panzvimbo yedivi rimwe racho.
Kana iyo asymmetry yakashumwa ikawanda, dambudziko rinoshanda mukutumirwa kwevamiriri vebhizinesi rinogona kunge riri pedyo nekuvharira-kuvharira pane kuita kutiza. Kubata kwenhema hakusi kwemahara. Imwe neimwe nzira inodzokera kumunhu, izvo zvinokanganisa nyaya yekuita otomatiki basa zvachose, uye inowedzera latency kune maitiro akadai sekugadzirisa tikiti kana ongororo yekodhi uko kunonoka pachayo kune mutengo. Kune njodzi yakavanzika zvakare: mutsara wekudzokorora uzere nekukwira kusingakoshi kunodzidzisa vanoongorora kuti vabvumire nekukurumidza, izvo zvinonetesa hutarisiro hwevanhu hunowanikwa kuti hupe. Iyo yakadzika mhedzisiro mhedzisiro inonzwisisika kubva pakuseta, kwete chimwe chinhu chinoyerwa.
Mukana uri pakati pe98.5% pazviitiko zvine mukurumbira uye 63.8% pazvionioni zvavo zvakadzoserwa umboo ndiyo mhedzisiro inonyanya kutariswa. Inopindirana nemhando dzinoziva chimiro chekukundikana kwakanyorwa zvakanaka uye kupindura kukuziva ikoko pane kuverenga uchapupu huri pamberi pavo. Kana kuverenga ikoko kuchibata, zvinonetsa kuti vatengi vanofanira kududzira sei kuchengetedzwa kwemumiririri zvibodzwa kazhinji: modhi inogona kutaridzika yakavimbika pamamiriro ezvinhu akafanana neakaburitswa ngano dzeyambiro asi uchibata zvakafanana mamiriro ezvimiro zvisina kunaka kana chokwadi charongwazve. Vanyori '"kugadziriswa-kugadziriswa kuita" chikamu bvunzo chaizvo izvi - humbowo huripo hunobvisa njodzi yakakonzerwa, uye modhi yacho inofanirwa kuigadzirisa.
Izvo zvakawanikwa zvakare zvine chekuita nemanyorerwo emunhu-in-the-loop zvinodiwa. Mitemo inoraira kuongororwa kwezvikamu zvechiito zvinofungidzira kuti gedhi ndiro rakachengeteka default. Iri basa rinoratidza kuti gedhi rine mwero waro wekukanganisa, munzira yekuvharira basa rakacheneswa, uye kuti mwero unofanirwa kuyerwa kwete kufungidzirwa kuti uri pedyo ne zero.
Kuyera pamusoro peizvi zvese: bhenji inoongorora mutongo mumwe chete usati waitwa wakapihwa curated umboo pakiti, kwete mumiririri anoshandisa chaiyo chishandiso loop ine ruzivo rusina kukwana rwainofanira kuzviunganidza. Real deployments zvakare inogara kuseri kwemvumo masisitimu uye mamiriro esangano ayo mamiriro haagone kumiririra zvizere. Kunyangwe 106 akavakwa mamiriro, zvisinei akanyatso kusimbiswa, kufanotaura maitiro mukugadzira mubvunzo wakavhurika, uye zvinorehwa pano ndezvavanyori pachavo.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
What is the most accurate way to describe what AI Agents can do today?
Zvekutarisa zvinotevera
Kunyangwe bepa rakazara, dhatabheti, uye bhodhi rekutungamira rinoratidza mhando dzakaedzwa; kana mapoka akazvimirira anobereka mukaha wekuramba; kana humbowo-yakadzoserwa girazi mhedzisiro inosimudza sechiratidzo chekufananidza-pateni; uye kana vatengi uye regulators vanotanga kubata enhema akabata seakayerwa kukundikana pane yakachengeteka default.
Chinhu chakanyanya kukurumidza kutsvaga ndechekufumura. Iro bepa rakazara, chero dhatabheti rakaburitswa kana kodhi, uye bhodhi rekutungamira mareferensi angaratidza kuti ndeapi mamodheru akaedzwa, maitiro makumi matatu akaitwa, uye kuti 28.1% nhamba inogovera sei mumamodheru nemadomasi. Nhamba yakaunganidzwa inotora 30 mamiriro inogona kuvanza kusiyana kwakakura, uye mubvunzo unoshanda kune chero munhu anosarudza modhi ndeapi masisitimu anogara kumagumo eiyo renji.
Kudzokorora kwakazvimirira kunonyanya kukosha kupfuura nhamba yemusoro. Tarisa kune mamwe mapoka anomhanyisa mamiriro, kuti uongororwe pasi-echokwadi zvinyorwa pamatambudziko akaomarara e "njodzi-yakagadziriswa" nyaya, uye kuti chiitiko-chinopesana-girazi gap chinopona here kuyedzwa nevanhu vasina kuvaka bhenji. Iyo gaka ndiyo inonyanya kukosha yepepa, uye zvakare ndiyo inonyanya kutsamira pamagadzirirwo akaitwa magirazi.
Zvakare zvakakosha kuteedzera ndezvekuti iyo imwe-nhanho yekumisikidza inofambiswa. Mutevedzeri wakakumbirwa kugedhi chiitiko chimwe chakarongwa nehumbowo hwakapihwa uri mune imwe nzvimbo kubva kune mumiririri wepakati-basa iro rinofanirwa kusarudza humbowo hwekuunganidza usati waita. Basa rekutevera rinoisa zvimiro zvakafanana mukati meajenti akazara zvishwe zvaizoyedza kana iyo calibration yakayerwa pano inofanotaura maitiro chaiwo.
Padivi rekugamuchira, mubvunzo ndewekuti mareti ekunyepa anopinda mukutenga uye makadhi ehurongwa pamwe chete nemitengo yekuramba, uye kuti vanogadzira modhi vanotanga here kugadzirisa padanho rekuita kwete kungwarira chete. Kutaura kwevanyori kuti kufunga kwakawedzerwa kunobatsira magedhi asina kusimba asi asiri atove akakwenenzverwa, kana akasimbiswa, anosundira kupokana nefungidziro yekuti yakawanda inference-nguva komputa inovandudza kutonga kwakakodzera.
Pakupedzisira, tarisa shanduro. yakanyorwa v2026-05 uye yakamisikidzwa muzviitiko zveveruzhinji inotarisana nedambudziko rekusvibiswa sezvo mamiriro ayo ari kutenderera uye mamodheru emangwana anodzidzisa pazviri. Vanyori vanozorodza sei seti - uye kana kuita kwegirazi kuchivandudza kuburikidza nekuwana kufunga kwechokwadi kana kuburikidza nekufumurwa - kucharatidza kuti nhamba dzinoreva kureba zvadzinoreva nhasi.