Ababhali:  (1) Zhan Ling, UC San Diego kunye negalelo elilinganayo;  (2) Yunhao Fang, UC San Diego kunye negalelo elilinganayo;  (3) Xuanlin Li, UC San Diego;  (4) Zhiao Huang, UC San Diego;  (5) uMingu Lee, uphando lwe-Qualcomm AI kunye noPhando lwe-Qualcomm AI  (6) URoland Memisevic, uPhando lwe-Qualcomm AI;  (7) Hao Su, UC San Diego.  Itheyibhile yoQhagamshelwano   Isishwankathelo kunye nentshayelelo   Umsebenzi onxulumeneyo   INkuthazo kunye nokuqulunqwa kweNgxaki   I-Deductively Verrifiable Chain-of-ingQingqa yokuqiqa   Iimvavanyo   Ukulinganiselwa   Ukuqukumbela, Imibulelo kunye neeReferensi   Ukuqinisekiswa kweDeductive kunye neeModeli zeVicuna   B Iingxoxo ezingakumbi malunga nokuPhuculwa koQinisekiso oluChanekayo xa kuthelekiswa noPhuculo lokuChaneka kokuGqibela   C Iinkcukacha ezingakumbi malunga nokutsalwa kwempendulo   D Iingcebiso   E Imizekeliso yoQinisekiso eMininzi  2 Umsebenzi onxulumeneyo    Iimodeli zolwimi ezinkulu zamvanje (LLMs) [3, 8, 57, 47, 38, 18, 9, 37] zibonise isakhono esimangalisayo sokusombulula imisebenzi entsonkothileyo yokuqiqa. Esikhundleni sokuvumela ii-LLMs zivelise ngokuthe ngqo iimpendulo zokugqibela njengemveliso, umsebenzi wangaphambili ubonise ukuba ngokukhuthaza inyathelo ngenyathelo lokuqiqa ngokukhuthaza okufanelekileyo, njenge-Chain-of-Thought (CoT) ekhuthazayo [50] kunye nabanye abaninzi [21, 59], 58, 44, 48, 60, 25, 54], iiLLMs zibonisa ukusebenza okungcono kakhulu kwimisebenzi eyahlukeneyo yokuqiqa. Ukuqhubela phambili ukuphucula inkqubo yokuqiqa ngenyathelo, ezinye izifundo zamva nje ziye zaphanda izisombululi zangaphandle ezixhasayo ezifana neetoliki zeprogram [39, 5, 27], uqeqesho kunye nokubiza iimodyuli zangaphandle zokuqiqa [11], okanye ukwenza uphando olucacileyo ukuvelisa amanyathelo okunciphisa. [2, 46]. Ngokunxuseneyo nale misebenzi, asithembeli kwiimodyuli zangaphandle kunye ne-algorithms, kwaye sixhasa ngokuthe ngqo isakhono sokufunda kumxholo wee-LLMs ukuvelisa iingcamango ezichanekileyo nezingqongqo. Ukuqiqa ngemizekelo yolwimi olukhulu.    Ukusebenzisa imifuziselo yolwimi ukuvavanya imizekelo yezizukulwana ibiyinto ekudala ikho [22, 36, 40, 4]. Njengoko ii-LLMs zibonisa amandla anomtsalane kuyo yonke imisebenzi eyahlukeneyo, iba ngumbono wendalo ukusebenzisa iiLLM njengovavanyo kunye nezixhobo zokuqinisekisa. Umzekelo, [10, 11, 33] finetune LLMs ukuqinisekisa izisombululo kunye namanyathelo aphakathi. Ii-LLM ezihambelana ne-RLHF [32, 31, 48] nazo ziye zaqeshwa ukuthelekisa izizukulwana ezahlukeneyo zemodeli. Ukongeza, imisebenzi yamva nje efana ne- [43, 52, 28, 6] iphakamisa uyilo olukhawulezayo ukuvumela ii-LLMs ukuba ziqinisekise, zisulungekise, kwaye zizilungise ngaphandle kwesidingo sokulungiswa. Nangona kunjalo, le misebenzi ayigxininisi kubungqongqo kunye nokuthembeka kweenkqubo zokuqiqa ezixhuzulayo kuwo onke amanyathelo okuqiqa. Kulo msebenzi, siphakamisa ifomathi yokuqiqa esekelwe kulwimi lwendalo evumela ii-LLMs ukuba ziqinisekise ngokwazo zonke inyathelo eliphakathi lenkqubo yokuqiqa, ngaloo ndlela iphucula ubungqongqo kunye nokuthembeka kokuqiqa.  Iimodeli zolwimi ezinkulu njengabaqinisekisi.  Ukongezelela, ngelixa eminye imisebenzi yamva nje [12, 53, 15, 34] ineendlela ezicetywayo zokuqinisekisa amanyathelo omntu ngamnye kwinkqubo yokuqiqa, indlela yethu iyahlula kule misebenzi kule mibono ilandelayo:   Indlela yethu yokuphucula i-in-context yokufunda ukufezekisa. ungqinisiso lwengqiqo, ngaphandle kwemfuneko yokulungiswa kwemodeli yolwimi.   Indlela yethu yokuqinisekisa i-LLM esekelwe kwiNkqubo yeNdalo ayichongi nje kuphela amanyathelo angasebenziyo okuqiqa, kodwa inika neengcaciso ezicacileyo zokuba kutheni ingasebenzi, ichaza iimpazamo ezithile zokuqiqa ezibandakanyekayo.   Indlela yethu yokuqiqa esekelwe kwiNkqubo yeNdalo kunye nokuqinisekisa iyahambelana nemisebenzi yokuqiqa engaphakathi kumxholo apho amanyathelo okuqiqa angenabo ubungqina obuquka ubume obuqukayo. Ngokomzekelo, indlela yethu yokujonga iyahambelana nomsebenzi weeNcwadi zokuGqibela, apho i-LLM iyalelwe ukuba ikhuphe ukudibanisa koonobumba bokugqibela bawo onke amagama ngokulandelelana njengempendulo yokugqibela.   Indlela yethu yeNkqubo yeNdalo ivumela ukusetyenziswa kolwazi oluqhelekileyo olungadweliswanga ngokucacileyo kwiindawo. Ngokomzekelo, khawucinge ngale ngxaki: “UMrin utya ama-apile ama-4 ngosuku. Utya ama-apile amangaphi ngoNovemba?” Nangona "uNovemba uneentsuku ze-30" engadweliswanga ngokucacileyo kwindawo, iNkqubo yeNdalo ivumela ukusetyenziswa kolwazi oluqhelekileyo olunjalo ngaphakathi kwinqanaba lokuqiqa. Inkqubo yethu yoqinisekiso olungaphakathi kumxholo ikwanakho ukuphatha ezi ndawo zifihlakeleyo (umz., ukuba iziphumo zeLLM “uNovemba uneentsuku ezingama-29” kwinyathelo lokuqiqa, iya kumakishwa njengengekho mthethweni). (1) (2) (3) (4)  Eli phepha   phantsi CC BY 4.0 DEED ilayisenisi. liyafumaneka arxiv

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

cosmological.TECH

Read My Stories

From Big Bang's singularity to galaxies' cosmic dance the universe unfolds its majestic tapestry of space and time.

Cosmological's blog

Le audio iveliswe ngolwimi lokuqala lwebali!

Ukusombulula iNgxaki ye-AI ye-Hallucination kunye neeNkqubo zeNdalo zokuZiqinisekisa

About Author

IZIMVO

ZIJONGE IIMPAWU

ELI NQAKU LINIKEZELWE KU

Related Stories

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps