Chii chaitika
Bepa rearXiv rinounza CoVA-SFT, yakarongeka yekudzidzisa dhata rakagadzirirwa kubatsira mamodhi emitauro yakawanda kushandisa zvinoonekwa uchigadzirisa matambudziko ekufunga. Iyo ine 51,900 samples, anopfuura 222,000 multimodal nhanho dzekufunga, mhuri shanu dzekugadzirisa uye gumi nenomwe mabasa akaomarara. Vanyori vanosumawo CoVA-Bench, 1,700-sample yakabatwa-kunze bhenji. Vanoshuma kuti mamodheru akarukwa zvakanaka paCoVA-SFT yakapfuura yakapindirana-cheni-ye-yepfungwa yekutanga nekanopfuura kaviri paavhareji, ichiri kutevera yakasimba-chete cheni-ye-yepfungwa yekutanga.
Iro bepa, rakatumirwa kuarXiv muna Nyamavhuvhu 29, 2026, rinopa CoVA-SFT sekorasi yekudzidzisa yemhando dzemitauro yakawanda. Mhedzisiro yaro yepakati ndeyekuti zvinyorwa-chete-cheni-ye-yekufunga kufunga haina kunaka kuenderana nezvinetso zvekuona nekuti inomanikidza chimiro chekuona kuita prose. Vanyori panzvimbo pacho vanotsanangura nzira iyo mamodheru anopindirana zvinyorwa nezvinooneka paunenge uchigadzirisa mabasa.
Zvinoenderana negwara rebepa, CoVA-SFT ine makumi mashanu neshanu nemazana mapfumbamwe emakumi mapfumbamwe nematatu uye anopfuura mazana maviri nemakumi maviri nemazana mapfumbamwe emazana matanho ekufunga. Mienzaniso inovhara mhuri shanu dzekugadzirisa uye 17 mabasa akaoma. Tsime rinoti corpus inosanganisira akajeka magadzirirwo ekugadzirisa, maagetic renders uye verification loops, ayo anoitirwa kudzidzisa mamodheru maitiro ekuvaka nekuchengetedza yemukati yekuona nzvimbo yekushanda panguva yekufunga.
Vanyori vanosumawo CoVA-Bench, imwe yerimwe ongororo seti ine zviuru zana nemazana manomwe akabatwa-kunze ebvunzo masampula muzvikamu zvakafanana zvebasa. Iyo bhenji inounzwa senzira yekutsigira kudzokororwa kuenzanisa pakati pemaitiro. Kwakabva hakupi mabasa ega ega, kugovaniswa kwemasampuli pakati pawo, kuzivikanwa kwemhando dzakaongororwa kana nzira yakazara yekukoira.
Bepa rinoshuma kuti mamodheru akanyatso kurongedzerwa paCoVA-SFT akakunda ese akapindirana ketani-ye-yepfungwa yekutanga nekanopfuura kaviri paavhareji paCoVA-Bench. Mhedzisiro iyoyo ndeye chirevo chakaitwa nevanyori uye haina kusimbiswa yakazvimirira pano. Iwowo abstract anoti iwo mamodheru achiri kupfupika kune akasimba mameseji-chete-cheni-ye-yekufunga ekutanga, zvichiita kuti mhedzisiro ivandudze pane dzimwe nzira dzakasiyana-siyana pane humbowo hwekuti dambudziko rakakura rekuona-kufunga ragadziriswa.
Nei zvichikosha
Iro basa rinonangana nekupikiswa kwakati mu multimodal AI: modhi inogona kugamuchira zvinoonekwa, asi data rekudzidzisa hariwanzo kuvadzidzisa nzira yekuvaka, kuchengetedza uye simbisa inooneka inomiririra yepakati. Yakakura, inodhinda uye bhenji inogona kupa vaongorori nzira yakafanana yekudzidza dambudziko iri. Iro bepa raro mhedzisiro inokodzera, zvisinei: zvakashumwa zvakawanikwa zvinoyerwa paCoVA-Bench, uye mhando dzakarongedzerwa zvakanaka dzinoramba dziri kumashure kwakasimba mameseji-chete masystem ekufunga.
Multimodal masisitimu anowanzoda kufunga nezve hukama hwepakati, marongero uye zvimwe zvimiro zvakaoma kumiririra zvakatendeka munhevedzano yemashoko. CoVA-SFT inogadzirisa dambudziko iri rekudzidzisa zvakananga nekuita zvinomiririra zvepakati chikamu chemienzaniso. Kana maitiro acho achiwanda, anogona kubatsira mamodheru kubata mabasa uko kuchengetedza chimiro kunokosha sekuziva zvinhu zvega kana kuverenga zvinyorwa.
Huyero hwekorasi yakarongwa yakakosha mukati medambudziko rakamanikana rinotsanangurwa nekwakabva. Zvinopfuura 222,000 nhanho dzekufunga dzinopa vaongorori chinangwa chikuru chekudzidza kuti zvinoonekwa uye kuzvigadzirisa zvinokanganisa sei maitiro emuenzaniso pane kuunganidzwa kudiki kwezviratidziro. CoVA-Bench inowedzera yakamisikidzwa yekuongorora poindi, iyo inogona kuita kuti mhedzisiro ive nyore kuenzanisa pane ese anotevera masisitimu.
Mhedzisiro yakashumwa zvakare ine zvinhu zvishoma. Vanyori vanofananidza neyakaganhurirwa ketani-ye-yepfungwa yekutanga uye vanoshuma zvinopfuura zvakapetwa kaviri avhareji mukana pabhenji yavo, asi vanobvuma kuti mamodheru akakwenenzverwa anoramba ari pazasi akasimba mavara-chete cheni-ye-yekufunga masisitimu. Izvi zvinoratidza kuti iyo inomiririrwa inogona kugadzirisa bhodhoro rimwe chete pasina kuenzanisa kugona kwekufunga kwenzira dzakasimba-dzakavakirwa pamagwaro.
Iko kukosha kunoshanda kunotsamira pane izvo zvisipo kubva pane inotangira peji. Hazvina kujeka kuti , zvirevo, maturusi ekupa kana nzvimbo dzekutarisa dzakadzidziswa dziripo kune veruzhinji, kudhura kwakadii kugadzirisa, kana kuti mienzaniso inoratidza mamiriro ekuona akasangana kunze kwebhenji. Iyo sosi zvakare haipe humbowo nezve kuchengetedzeka, kufukidzwa kwehuwandu hwevanhu, kuwanikwa kana kuita mune yakakwira-stake application.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Muedzo unotevera wakakosha ndewekuti CoVA-SFT inonatsiridza mashandiro kupfuura ayo anosanganisirwa bhenji uye inotamisa mumamodheru, marongero uye chaiwo-enyika ekuona mabasa. Vatsvaguri vanofanirwawo kuongorora kana dhatabheti dzakajeka zvikonzero uye zvishwe zvekusimbisa zvinovandudza kuvimbika kana kunyanya kukurudzira hunhu hwemabhenji. Iyo sosi haigadzirire rezinesi re , ruzhinji kudhawunirodha mamiriro, mutengo wekudzidzira, modhi yezvivakwa, basa-level mhedzisiro kana kuita pane ekunze ongororo.
Kusimbiswa kwekunze kunofanirwa kuve kwekutanga kwekutanga. Mhedzisiro pamataseti akazvimirira ekuona-ekufunga angabatsira kuratidza kana CoVA-SFT ichidzidzisa kugona kuchinjika pane kugadzirisa zvimiro uye mabasa anomiririrwa muCoVA-Bench. Kuenzanisa kunofanirwa kusanganisira iwo mamodheru mamwe chete, mabhajeti uye mamiriro ekukurudzira kuitira kuti mukana wakashumwa ugone kududzirwa zvakaringana.
Vatsvaguri vanofanirwa kutsvaga mutsauko wechikamu chebasa kuseri kweavhareji yakashumwa. Kuvandudza kunopfuura kaviri kwakaunganidzwa kunogona kuvanza mibairo mikuru pane dzimwe mhuri dzemarongerwo uye zvishoma kana kusave nani pane dzimwe. Kwakabva hakutauri kana kuwedzera kwekuita kuchienderana pamabasa ese gumi nemanomwe, uye haatauri maitiro ekukanganisa kana matanho ekuvimba.
Malopu echokwadi anotsanangurwa nevanyori anokodzera kuongororwa. Vanogona kubatsira mamodheru kubata zvikanganiso mune zvepakati zvinoonekwa zvimiro, asi ivo vanokwanisawo kuwedzera mutengo wekufungidzira kana kuburitsa kutaridzika kwekuvimbika pasina kuona mhinduro dzekupedzisira. Ongororo dzenguva yemberi dzinofanira kuyera kurongeka, kuenzanisa, kuverenga uye kutadza kupora zvakasiyana.
Pakupedzisira, munda uchada ruzivo rwakajeka nezve kuwana uye kuberekazve. Sosi yacho inozivisa bepa sekugamuchirwa kuEMNLP 2026 Zvawanikwa, asi haritauri rezinesi re , nzvimbo yekuburitsa, maitiro ekuzivisa kana hardware zvinodiwa. Iwo mameseji ndiwo anozoona kana CoVA-SFT ikava chishandiso chekutsvaga chinoshandiswa zvakanyanya kana kuti inoramba iri nzira yekudzidzisa yakanangana nepepa.