Ukuhlanganisa Ukubonisana Nesenzo: I-Synergy of Large Concept Models (LCMs) kanye namamodeli esenzo amakhulu (ama-LAM) kuma-Agentic Systems

Ukuvela kwamamodeli e-AI athuthukisiwe kuye kwaholela emisha endleleni imishini ecubungula ngayo ulwazi, ukuxhumana nabantu, nokwenza imisebenzi kuzilungiselelo zomhlaba wangempela. Izindlela ezimbili zokuphayona ezisafufusa amamodeli amakhulu (LCMs) namamodeli amakhulu wesenzo (ama-LAM). Nakuba bobabili benweba amakhono ayisisekelo ezinhlobo zezilimi ezinkulu (LLMs), izinjongo zabo kanye nezinhlelo zokusebenza ziyehluka.

Ama-LCM asebenzisa izethulo ezingacacile, zokungahloniphi kolimi ezibizwa ngokuthi “imiqondo,” okuzivumela ukuba zicabange ngezinga eliphezulu lokungabonakali. Lokhu kusiza ukuqonda okuguquguqukayo kuzo zonke izilimi nezindlela, ukusekela imisebenzi efana nokucabanga ngomongo omude kanye nokuhlela okunezinyathelo eziningi. Ama-LAM, ngakolunye uhlangothi, aklanyelwe ukwenziwa kwesenzo, ukuhumusha izinhloso zabasebenzisi zibe izinyathelo ezingenzeka kuzo zombili izindawo zedijithali nezomzimba. Lawa mamodeli ahamba phambili ekuhumusheni imiyalo, izinqubo zokuzenzakalela, futhi azivumelanise ngokuguquguqukayo nempendulo yemvelo.

Ama-LCM nama-LAM ahlinzeka ngohlaka olubanzi lokuvala ikhefu phakathi kokuqonda ulimi nesenzo somhlaba wangempela. Ukuhlanganiswa kwabo kunamandla amakhulu wezinhlelo zegrafu ze-ajenti, lapho ama-ejenti ahlakaniphile adinga ukucabanga okuqinile namandla okubulala ukuze asebenze ngempumelelo.

Amamodeli Emiqondo Emikhulu (LCMs): Uhlolojikelele Olujulile

Amamodeli Emiqondo Emikhulu (i-LCMs) ye-FAIR ku-Meta iphakamisa ukucabanga kusuka ekuhlaziyweni okusekelwe kumathokheni kuya kuzinga lomqondo ongacacile, ongenakwaziwa ngolimi, kanye neleveli ye-modality-agnostic. Lawa mamodeli ahlose ukukhiqiza nokucubungula ulwazi ngokuguquguquka okungenakuqhathaniswa kanye nokulinganisa, ukubhekana nemikhawulo ethile yama-LLM avamile. Izakhiwo zabo ezintsha nendlela yokuphatha ulwazi kunikeza amathuba ayingqayizivele wezinhlelo zokusebenza ze-AI ezithuthukisiwe.

I-Abstract and Modality-Agnostic Reasoning

Emgogodleni wama-LCM kukhona ikhono lawo lokusebenza “kumiqondo” kunamathokheni athile olimi. Lokhu kunqamulela kuvumela ama-LCM ukuthi azibandakanye ekucabangeni okweqa imingcele yolimi noma yemikhuba. Esikhundleni sokugxila ebunkingeni bolimi oluthile noma indlela yokufaka, lawa mamodeli acubungula izincazelo nezakhiwo eziyisisekelo, ezivumela ukuthi zikhiqize imiphumela enembile kuzo zonke izimo ezihlukahlukene zolimi nezimo.

Isibonelo, i-LCM eqeqeshelwe idatha yesiNgisi ingakwazi ukwenza kalula amakhono ayo kwezinye izilimi noma izindlela, okuhlanganisa inkulumo nedatha ebonakalayo, ngaphandle kokulungiswa okwengeziwe. Lokhu kukhula kubangelwa isisekelo sakho endaweni yokushumeka ye-SONAR, uhlaka oluyinkimbinkimbi olusekela izilimi ezingaphezu kuka-200 nezindlela eziningi.

Izimpawu Ezisemqoka zama-LCM

1. Isakhiwo Sokulandelana Sokucaca: Ama-LCM asebenzisa ukwakheka kwesigaba esicacile, okuthuthukisa ukufundeka kokuphuma kwefomu ende. Lo mklamo usekela ukukhiqiza okuqukethwe okuhlelekile, okwenza kube lula ukutolika nokuguqula njengoba kudingeka.
2. Ukuphatha Izimo Ezinde: Ngokungafani namamodeli e-transformer endabuko, inkimbinkimbi yawo yokubala ikala ngokuphindwe kane ngobude bokulandelana, ama-LCM athuthukisiwe ukuze aphathe okuqukethwe okubanzi kahle kakhulu. Ngokusebenzisa ukulandelana okufushane kuhlaka lwabo lomqondo, banciphisa imikhawulo yokucubungula futhi bathuthukise amakhono okucabanga efomu elide.
3. I-Zero-Shot Generalization Engenakuqhathaniswa: Ama-LCM ahamba phambili ekwenzeni i-zero-shot generalization, ewavumela ukuthi enze imisebenzi ngezilimi zonkana nezindlela angazange ahlangabezane nazo ngokusobala phakathi nokuqeqeshwa. Isibonelo, ikhono labo lokucubungula izilimi ezisetshenziswa kancane njengesi-Pashto noma isi-Burmese libonisa ukuguquguquka kwazo kanye nokuqina kohlaka lwazo lokucabanga.
4. I-Modularity and Extensity: Ngokuhlukanisa izifaki khodi zomqondo namadekhoda, ama-LCM agwema ukuphazamiseka nokuncintisana okubonwa kuma-LLM anezinhlobo eziningi. Le modularity iqinisekisa ukuthi izingxenye ezihlukene zingenziwa ngokuzimela, zithuthukise ukuguquguquka kwazo kuzinhlelo zokusebenza ezikhethekile.

Izicelo kanye Okujwayelekile

Ama-LCM awusizo emisebenzini edinga ukuqonda okuphelele nokucabanga okuhlelekile, njengokufingqa, ukuhumusha, nokuhlela. Ikhono labo lokuphatha izindlela ezihlukahlukene, okuhlanganisa umbhalo, inkulumo, nedatha ebonakalayo, libenza babe abantu abafanelekile bokuhlanganiswa ezinhlelweni eziyinkimbinkimbi ze-AI. Ngaphezu kwalokho, amandla abo okukhiqiza afakazelwe ngokuhlolwa okubanzi. Isibonelo, ama-LCM enza kahle kakhulu kunamamodeli aqhathanisekayo ekukhiqizeni imiphumela ehambisanayo yemisebenzi yokufingqa ngezilimi eziningi, ikakhulukazi ezilimini ezinezinsizakusebenza eziphansi.

Amamodeli Esenzo Amakhulu (ama-LAM): Uhlolojikelele Olubanzi

I-Microsoft, i-Peking University, i-Eindhoven University of Technology, kanye neNyuvesi yase-Zhejiang basungule amamodeli amakhulu esenzo (ama-LAM) anweba amandla ama-LLM endabuko ukuze avumele ukwenziwa kwesenzo esiqondile endaweni yedijithali nebonakalayo. Lawa mamodeli avala igebe phakathi kokuqonda ulimi nokusebenzelana komhlaba wangempela, okuvumela imiphumela ebonakalayo, egxile emsebenzini.

Ukushintsha kusuka ku-LLM kuya kuma-LAM

Nakuba ama-LLM enza kahle kakhulu ekukhiqizeni umbhalo ofana nomuntu nasekunikezeni imininingwane esekelwe olimini, ngokwemvelo anqunyelwe emiphumeleni yokwenziwa. Abakwazi ukusebenzelana ngokuguqukayo nomhlaba, noma ngabe bazulazula ezindaweni zokuhlangana ezidijithali noma benze imisebenzi engokoqobo. Ama-LAM abhekana nalo mkhawulo ngokwakhela phezu kwamakhono ayisisekelo e-LLM kanye nokuhlanganisa izindlela ezithuthukisiwe zokukhiqiza izinto. Zenzelwe ukuthi:

1. Humusha Izinhloso Zomsebenzisi: Ama-LAM ahlaziya izinhlobo ezihlukahlukene zokufaka—umbhalo, imiyalo yezwi, noma idatha ebonakalayo—ukuze abone izinjongo zabasebenzisi. Ngokungafani nama-LLM, akhiqiza ngokuyinhloko izimpendulo ezisekelwe emibhalweni, ama-LAM ahumusha lezi zinhloso zibe izinyathelo ezingenzeka.
2. Enza Imisebenzi Ezimweni Zomhlaba Wangempela: Ngokusebenzelana nendawo yawo, ama-LAM angakwazi ukwenza ngokuzenzakalelayo imisebenzi efana nokuzulazula kumawebhusayithi, ukuphatha amathuluzi edijithali, noma ukulawula amadivayisi aphathekayo. Leli khono limelela ushintsho olubalulekile ekuhlakanipheni okusebenzisekayo.

Izimpawu Ezisemqoka zama-LAM

1. Isizukulwane Sesenzo: Ama-LAM akhiqiza ukulandelana kwezenzo okunemininingwane, okuqaphela umongo okuhambisana nezidingo zabasebenzisi. Isibonelo, uma iyalwa ukuthenga into ku-inthanethi, i-LAM ingakwazi ukuzulazula ngokuzenzakalelayo iye kuwebhusayithi, iseshe into, futhi iqedele ukuthenga.
2. Ukuzivumelanisa nezimo: Lawa mamodeli angaphinda ahlele futhi alungise izenzo ngokuguquguqukayo ekuphenduleni impendulo yezemvelo, aqinisekise ukuqina nokwethembeka ezimeni eziyinkimbinkimbi.
3. Ubungcweti: Ama-LAM enzelwe imisebenzi eqondene nesizinda esithile. Ukugxila kububanzi bokusebenza obuthile kuzuza ukusebenza kahle nokusebenza okuqhathanisekayo noma okungcono kunama-LLM ajwayelekile. Lobu buchwepheshe bubenza bafanelekele izindawo ezinezisetshenziswa ezifana namadivayisi asemaphethelweni.
4. Ukuhlanganiswa nama-ejenti: Ama-LAM avame ukushumeka ngaphakathi kwezinhlelo zama-ejenti, ezihlinzeka ngamathuluzi adingekayo okuxhumana nendawo. Lawa ma-ejenti aqoqa okubhekwayo, asebenzise amathuluzi, agcine inkumbulo, futhi asebenzise ama-loop wempendulo ukuze asekele ukwenziwa komsebenzi okuphumelelayo.

Izicelo zama-LAM

Ama-LAM asekhombisile ukuthi awusizo emikhakheni eyahlukene. Ku-Automated Digital NavigationAmamodeli afana ne-GPT-V, ahlanganiswe ezinhlelweni ze-ejenti, abonise isithembiso ekwenzeni imisebenzi yokuzulazula kuwebhu. Benza izinqubo ngokuzenzakalelayo njengokusesha ulwazi, ukuqedela okwenziwa ku-inthanethi, noma ukuphatha okuqukethwe kuzo zonke izinkundla eziningi. Futhi, ngoba umsebenzi ozenzakalelayo ezindaweni ze-GUIAma-LAM athuthukisa ukusebenzisana kwekhompuyutha yomuntu ngokwenza imisebenzi yesixhumi esibonakalayo somsebenzisi ngokuzenzakalelayo nokunciphisa umzamo owenziwe ngesandla ekusebenzeni okuphindaphindiwe noma okuyinkimbinkimbi.

Ama-LCM nama-LAM we-Agentic Graph Systems

Amasistimu egrafu e-Ajenti adinga ukucabanga okuyinkimbinkimbi, ukuhlela, namandla okwenza isenzo ukuze asebenze ngempumelelo. Inhlanganisela yama-LCM nama-LAM yakha izakhiwo ezinamandla ezibhekana nalezi zidingo ngokusebenzisa amandla ohlobo lwemodeli ngayinye.

Ama-LCM kuma-Agentic Systems

Ama-LCM aletha uhlaka lomqondo oludlula zonke ekucabangeni nasekucabangeni okungaqondakali. Bangakwazi ukuhlanganisa ulwazi kuzo zonke izimo ezihlukahlukene ngokucubungula ulwazi ngendlela yolimi-agnostic kanye nendlela ye-modality-agnostic. Lokhu kubenza babaluleke kakhulu ekulawuleni izimo zomongo omude, lapho ukuncika kokuqonda nokugcina ukuhambisana kubaluleke kakhulu.

• Ukuhlela kwe-Hierarchical: Amandla e-LCM okusebenza ngezinhlaka ezicacile zezikhundla asiza ekuhlelweni kwezinhlelo nemiphumela. Lokhu kucabanga okulandelanayo kubalulekile ezinhlelweni zegrafu ze-ajenti, ezivame ukubandakanya imisebenzi eyinkimbinkimbi, enezinyathelo eziningi.
• Ukuhlanganiswa kwe-Cross-Modality: Ngokusebenza ekushumekeni kwe-SONAR, ama-LCM asekela okokufaka kwedatha yezindlela eziningi, njengombhalo, inkulumo, nedatha ebonakalayo. Lokhu kuqinisekisa ukuhlanganiswa okungenazihibe kuyo yonke imithombo yolwazi ehlukene kusistimu yegrafu ye-ejenti.

Ama-LAM kuma-Agentic Systems

Ama-LAM, agxile ekwenziweni kwesenzo, ahlinzeka ngesendlalelo sokwenza sezinhlelo ze-ajenti. Bahumusha izinhloso zabasebenzisi futhi bazihumushele ezenzweni ezibambekayo ezisebenzisana nezindawo zedijithali noma ezibonakalayo.

• Ukwenza Umsebenzi: Ama-LAM ahamba phambili ekuhlukaniseni imigomo eyinkimbinkimbi ibe imisebenzi emincane engenziwa. Ukuzivumelanisa nezimo kuthuthukisa leli khono, okubavumela ukuthi bahlele kabusha futhi balungise izenzo ngesikhathi sangempela ngokusekelwe empendulweni.
• Ukusebenzisana Okunamandla: Ama-LAM ahlanganisa nezinhlaka zama-ejenti ukuze ahlanganyele namathuluzi nendawo. Lokhu kusebenzisana kuvumela ukuzulazula kwewebhu, ukulawula uhlelo lokusebenza, nokukhohlisa idivayisi ebonakalayo.

I-Synergy Phakathi kwama-LCM nama-LAM

Ukuhlanganiswa kwama-LCM nama-LAM kusistimu yegrafu ye-ajenti kusebenzisa amandla awo womabili amamodeli. Ama-LCM ahlinzeka ngamakhono okucabanga nawokuhlela adingekayo ukuze kuqondwe izimo eziyinkimbinkimbi, kuyilapho ama-LAM enza lezi zinhlelo kuzilungiselelo zomhlaba wangempela.

• Ukuhlanganiswa Kwegrafu Yolwazi: Amagrafu olwazi asebenza njengohlaka oluhlanganisayo, olwenza womabili amamodeli afinyelele ulwazi oluhlelekile ukuze ahlele futhi afeze kangcono. Lokhu kuthuthukisa ikhono lesistimu lokumodela ubudlelwano, ukugcina inkumbulo, nokukhetha amathuluzi afanelekile.
• Amandla Ahambisanayo: Ngenkathi ama-LCM ekhuluma ngokucabanga okungacacile nokuqonda kwezindlela eziningi, ama-LAM agxila esenzweni somhlaba wangempela. Lokhu kusebenza okuhambisanayo kuqinisekisa ukusebenza okuqinile kuzo zonke izizinda zengqondo nezomzimba, ukuhlangabezana nezidingo zamasistimu e-ajenti ayinkimbinkimbi.

Sengiphetha, ukuhlanganisa ama-LCM nama-LAM kwenza amasistimu ahlanganisa ukucabanga okungacacile nokubulawa okungokoqobo. Ama-LCM ahamba phambili ekucubunguleni imiqondo yezinga eliphezulu, ukuphatha okuqukethwe okude, nokubonisana ngezilimi zonkana nezindlela. Ama-LAM agcwalisa lawa makhono ngokukhiqiza nokwenza izenzo ezifeza izinhloso zabasebenzisi kuzimo zomhlaba wangempela. Kuzinhlelo zegrafu ze-ajenti, ukusebenzisana phakathi kwama-LCM nama-LAM kunikeza indlela ebumbene yokuxazulula imisebenzi eyinkimbinkimbi edinga ukuhlela nokwenza. Ngokusebenzisa amagrafu olwazi, lezi zinhlelo zithola inkumbulo ethuthukisiwe, ukucabanga, namandla okwenza izinqumo, zivula indlela yama-ejenti ahlakaniphe kakhulu nazimele. Nakuba izinselele zisekhona, okuhlanganisa ukuqina, ukuphepha, nokusebenza kahle kwezinsiza, intuthuko eqhubekayo ekwakhiweni kwe-LCM ne-LAM ithembisa ukubhekana nalezi zinkinga.

Imithombo

Futhi, ungakhohlwa ukusilandela Twitter futhi ujoyine yethu Isiteshi socingo futhi I-LinkedIn Grphezulu. Ungakhohlwa ukujoyina wethu 70k+ ML SubReddit.

🚨 [Recommended Read] I-Nebius AI Studio inweba ngamamodeli ombono, amamodeli ezilimi ezintsha, okushumekiwe kanye ne-LoRA ^(Unyusiwe)