Kwenzekeni
Abacwaningi bahlongoza i-Token-Budget Distillation, noma i-TBD, indlela esebenza kahle yepharamitha yokulungisa amamodeli olimi lombono wevidiyo ngaphansi kwebhajethi yethokheni engaguquki. Le ndlela isebenzisa imodeli kathisha enethokheni egcwele ukuze igade imodeli yomfundi ecindezelwe kuyilapho ibuyekeza ama-adaptha e-LoRA kuphela.
Iphepha le-arXiv elithunyelwe ngo-Aug. 28, 2026, lethula i-Token-Budget Distillation yamamodeli olimi lombono wevidiyo. Ababhali bachaza inkinga emaphakathi njengokuhwebelana phakathi kokusebenza kahle kwekhompyutha kanye nokwethembeka kwe-semantic: okokufaka kwevidiyo kukhiqiza amathokheni amaningi abonakalayo, kuyilapho ukucindezela lawo mathokheni kungabangela imodeli eguquliwe ukuthi isuke ekuziphatheni kwesistimu yethokheni yasekuqaleni. Indlela ehlongozwayo iklanyelwe ukusebenza ngaphansi kwebhajethi yethokheni engashintshi kunokumane ilungise kahle okokufaka okucindezelwe. Lokho kwenza uhlaka kusho ukuthi indlela yethulwa njengenqubo yokujwayela imodeli ekhona, isabelomali samathokheni sithathwa njengesithiyo esilolonga ukuqeqeshwa. Ngakho-ke incazelo enikeziwe igcizelela ukuthi uthisha nomfundi baqeqeshwa kanjani ngokuhlobene, kunokuchaza isifaki khodi esisha sevidiyo noma uhlelo olusha oluqondene nomsebenzi othile.
I-TBD imisa umgogodla wemodeli oqeqeshwe kusengaphambili futhi ibuyekeze ama-adaptha e-LoRA kuphela, iphepha elivezwayo njengesu lokuzivumelanisa nepharamitha elisebenza ngempumelelo. Iphinde ihlanganise ukucindezelwa kwethokheni ebonakalayo okusekelwe ku-FlashVID endleleni yevidiyo. Idizayini yokuqeqesha isebenzisa izindlela ezimbili: uthisha onethokheni egcwele uhlinzeka ngokugada, kuyilapho umfundi ocindezelwe efunda kunhloso yomsebenzi kanye namasignali amaningana okukhipha isisu. Lokhu kufaka phakathi i-answer-region KL distillation, i-ground-truth-anchored margin distillation, kanye nokulawula okunokwethenjelwa kolwazi lokugaya. Lezi zingxenye zichazwa njengezingxenye zomklamo wokuqeqesha, ngakho umnikelo obikiwe uyinhlanganisela yokuhlelwa kwesabelomali esingaguquki, indlela yokubona ecindezelweyo, kanye nokugadwa okudluliswa kuthisha kuya kumfundi. Umgogodla usalokhu uyinkomba yokuguquguquka okuchaziwe.
Ababhali babika ukuhlolwa kwevidiyo emithathu ye-VLM backbones: LLaVA-Video, LLaVA-OneVision, kanye ne-Qwen3-VL-8B-Instruct. Bathi le ndlela ihlale isebenza kahle kakhulu kunezisekelo zokuminyanisa kuphela kuwo wonke amabhentshimakhi okuqonda amavidiyo ngaphansi kokubili kokucindezelwa okumaphakathi nokunolaka. Ngesilinganiso esingu-10% sokugcinwa kwethokheni, iphepha libika ukuthi i-TBD igcine u-97.0% wesilinganiso sokunemba semodeli ye-Vanilla ku-LLaVA-Video. Ku-LLaVA-OneVision, ibika isilinganiso samaphuzu angu-58.4 kanye no-100.0% wokunemba okuhlobene esilinganisweni esifanayo sokugcinwa. Leyo miphumela ishiwo njengemiphumela emaphakathi noma ehlobene kusifinyezo sephepha, futhi ihambisana nesilungiselelo sokugcinwa esikhonjwe lapho. I-akhawunti enikeziwe ayingezi ukuhlukaniswa kwebhentshimakhi ngayinye noma imiphumela eyengeziwe yokuhlola ngale kwalokhu kuqhathaniswa okubikiwe.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Amamodeli wevidiyo acubungula izinombolo ezinkulu zamathokheni abonakalayo, angenza ukulungisa kahle nokusho ukuthi kubize kakhulu. Uma imiphumela ebikiwe ibambezeleka ekuhlolweni okubanzi, indlela inganciphisa ukucubungula okubukwayo kuyilapho ilondoloza ikhono lokuqonda levidiyo lemodeli.
Udaba olusebenzayo olusingathwa yiphepha lubalulekile ezinhlelweni zevidiyo ze-AI ngoba okokufaka kwevidiyo kungakhiqiza amathokheni abukwayo kakhulu kunezithombe ezizodwa. Amathokheni engeziwe ngokuvamile andisa umthwalo wokubala wakho kokubili ukucabanga nokuzivumelanisa nezimo. Indlela egcina i-semantics yevidiyo ewusizo kuyilapho icubungula ingxenye encane nje yokumelwa okubonakalayo kwasekuqaleni ingenza ezinye izinhlelo zokusebenza zolimi lwevidiyo zingadingi ukuqalisa noma ukushuna kahle. Ngakho-ke inzuzo engaba khona iboshelwe ekulondolozeni ukusetshenziswa kwe-semantic ngesabelomali esiphansi sethokheni yokubuka. Akukhona, okuvela ezintweni ezinikeziwe kuphela, ukubalwa okuphelele kwezinsiza ezidingekayo kulo lonke ukuqeqeshwa kanye nepayipi lokukhomba.
Idizayini yephepha yothisha nomfundi iqondise ubuthakathaka obuthile bokucindezela okulula. Ukuzijwayeza kokucindezelwa kuphela kungase konge ukubala kodwa kungalahla ulwazi olubalulekile ekuphenduleni imibuzo ngevidiyo. Ngokugcina uthisha onethokheni egcwele ngesikhathi sokuqeqeshwa kanye nokudlulisa izinhlobo ezimbalwa zokugada kumfundi ocindezelwe, i-TBD izama ukufundisa ukumelela okuncane ukulingisa ukuziphatha kwe-semantic kwemodeli engacindezelwe kancane. Lena impendulo eqondiswe kakhulu ekulahlekelweni kwekhwalithi kunokuphatha ukuminyanisa njengesinyathelo sokucubungula kusengaphambili. Lowo mehluko ubalulekile lapho kuchazwa isiphakamiso: isilungiselelo sokuminyanisa siyingxenye yeresiphi yokujwayela ebanzi, futhi indima kathisha ingeyokuqeqesha. Isifinyezo asisho ukuthi uthisha uyadingeka kukho konke ukusetshenziswa kamuva komfundi oshintshiwe.
Imiphumela ebikiwe inamandla amakhulu ngoba ihlanganisa amamodeli amaningi angemuva esikhundleni sesakhiwo esisodwa. Kodwa-ke, umthombo usungula lokhu okutholakele kuphela njengezimangalo ezenziwe ku-preprint. Ayiwahlonzi amabhentshimakhi amane ku-abstract enikeziwe, ilinganise ukuqeqeshwa noma ukusheshisa okucatshangwayo, ibike ukusetshenziswa kwememori noma izindleko zezezimali, ichaza ihadiwe esetshenzisiwe, noma ibonise ukuthi ukunemba okuphelele kushintsha kanjani kuyo yonke imisebenzi. Futhi ayiqinisekisi ukuthi indlela isilungele ukuthunyelwa kokukhiqizwa noma ukuthi izilinganiso zokugcinwa okubikiwe zizodluliselwa kwamanye amamodeli wevidiyo. Ngaleso sizathu, imiphumela ifundwa kangcono njengobufakazi obethulwa ababhali bokusetha okushiwo kokuhlola. Incazelo etholakalayo isekela intshisekelo endleleni, kuyilapho ishiya usayizi nokuvumelana kwanoma iyiphi inzuzo engokoqobo ezosungulwa.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Iphepha liyi-arXiv ephrintiwe ngaphambili, futhi umthombo awunikezi ukuphindaphinda okuzimele, ukubambezeleka okuningiliziwe noma izilinganiso zememori, amagama ebhentshimakhi, ukutholakala kwekhodi, noma ubufakazi bokusetshenziswa. Umsebenzi wokulandelela kufanele uhlole ukuthi izinzuzo ezibikiwe ziyaqhubeka yini kubo bonke ubude bevidiyo, imisebenzi, amazinga okucindezela, nemindeni eyimodeli.
Umbuzo wokuqala wocwaningo lokulandelela ukuthi ingabe ukugcinwa kokunemba okubikiwe kungenziwa kabusha ngokuzimele. Umthombo uchaza imiphumela kuma-backbones amathathu aqanjwe amagama kanye namabhentshimakhi amane, kodwa i-abstract ayinikezi amagama ebhentshimakhi noma imininingwane eyanele yokuhlola ukuze kuhlolwe ukwakheka kwedathasethi, ubunzima bomsebenzi, izivumelwano zokuhlola, noma ukuhluka kwezibalo. Ukubuyekezwa kontanga noma ukuhlola okugcwele kobuchwepheshe kungasiza ekucaciseni ukuthi ukuqhathanisa kuqine kangakanani. Lezi zikhala zithinta ukukhiqizwa kabusha kanye nokutolika. Ngaphandle komongo oshiyiwe, umfundi akakwazi ukusho ku-abstract yedwa ukuthi amasethingi okuhlola akhethiwe amele kangakanani noma ukuthi kungakanani ukungaqiniseki okuzungeze ukuqhathanisa okubikiwe.
Izimangalo zokuphumelela zidinga ukulinganiswa ngokuqondile. Isilinganiso esingu-10% sokugcinwa sibonisa ukuthi mangaki amathokheni abonakalayo asele, kodwa asiqali ngokwaso ukuncishiswa okungama-90% esikhathini sokuchazwa kwewashi lodonga, ukusetshenziswa kwenkumbulo, ukusetshenziswa kwamandla, noma izindleko. Ukuminyanisa okungaphezulu, izindleko zokuqeqeshwa kwemodeli kathisha, ubude bokulandelana, izingxenyekazi zekhompuyutha, usayizi weqoqo, kanye nemininingwane yokusebenzisa kungathinta kakhulu inzuzo ebonakalayo. Lezo zilinganiso azinikezwanga kumbhalo womthombo. Inani lokugcinwa kufanele ngenxa yalokho ligcinwe lihlukile eziphethweni ezisebenza kahle kakhulu. Ukumisa lezo ziphetho kuzodinga izilinganiso ezixhuma isibalo samathokheni ekusetshenzisweni kwensiza ekupheleni ukuya ekupheleni ngaphansi kokusetshenziswa okucacisiwe kanye nomsebenzi.
Kubalulekile futhi ukuhlola indlela ngale kwezilungiselelo ezibikiwe. Ukuhlola okuzayo kungahlola amavidiyo amade nahlukahluka kakhulu, ukucabanga kwesikhashana okuhlaziywe kahle, izehlakalo ezingandile, izilimi eziningi, izilinganiso zokuminyanisa ezihlukene, kanye nemindeni eyimodeli ngaphandle kwemigogodla emithathu esohlwini. Umthombo ulebula umsebenzi we-ACM MM 2026, kodwa awusho ukuthi iphepha lamukelwe, futhi awukhombi ikhodi ekhishiwe, ama-adaptha amamodeli, noma umkhiqizo womphakathi usebenzisa i-TBD. Ukuhlola okunjalo kungasiza futhi izinzuzo ezihlukene ezivela enqubweni yokugaya kusukela emiphumeleni eboshelwe kumgogodla othile noma ukulungiselelwa kokuminyanisa. Kuze kube yileso sikhathi, ubufakazi obunikeziwe buhlala bukhawulelwe kububanzi nesimo esichazwe ukuphrinta kwangaphambili.