Byagenze bite
Icapiro ryashyikirijwe arXiv ku ya 24 Kanama ryerekana Credal Large Language Models, cyangwa CLLMs, ikoresha itsinda rya adaptate ya LoRA kugirango ihagararire ibintu byinshi byavuzwe mbere aho gukwirakwizwa rimwe. Abanditsi bakura amanota yikimenyetso hamwe nurwego rwamasomo yo kwiyemeza no kubisuzuma kugirango basubize ibibazo, kalibrasi, guhanura guhitamo, gutahura ibitekerezo, no gutekereza.
Uru rupapuro rusobanura imbogamizi muburyo busanzwe imvugo yerekana imvugo idashidikanywaho: icyitegererezo gisanzwe gitanga igabanywa rimwe ryo guhanura, abanditsi bavuga ko rishobora guhuza ubujiji no kudasobanuka kwukuri. Icyifuzo cyabo CLLM ahubwo ikoresha ensemble ya LoRA adaptate kugirango ikore icyo impapuro zita gushiraho inguzanyo. Mu buryo bufatika, uburyo bugamije kubika ibyo mutumvikanaho cyangwa gukwirakwizwa hagati y’isaranganya rishoboka aho guhosha ibintu byose bidashidikanywaho mu musaruro umwe wa softmax. Inkomoko ntabwo ivuga ko uku guhagararirwa gukora icyitegererezo gikwiye; ivuga ko ishobora gutuma urugero rwicyitegererezo rwiyemeza kurushaho.
Abanditsi batangiza ingamba ebyiri zijyanye. Credal Token Kwiyemeza, cyangwa CTC, ikorera mumwanya wikimenyetso kandi ikomatanya inkunga-yo munsi, ubugari bwinguzanyo, hamwe na entropiya ihuza. Ibisobanuro bivuga ko CTC ishobora kubarwa nta gisekuru cyiyongereye, gishobora kuba ingenzi kuri sisitemu aho gusubiramo inshuro nyinshi byongerera ubukererwe cyangwa ikiguzi. Isezerano rya Semantic Consistency, cyangwa SCC, ryagura igitekerezo kumwanya wa semantique ukoresheje icyitegererezo cyuzuye. Uru rupapuro rusobanura kandi SCC-Gap gupima ibipimo bidahuye hagati yinkunga-urwego rwinkunga ninkunga yo murwego. Aya manota yatanzwe nkibikoresho byo kumenya igihe icyitegererezo cyurwego rwicyizere cyo hejuru gishobora kudahuza nurwego rwibisobanuro byagaragajwe nibisubizo bishoboka.
Isuzuma ririmo Gemma-2-9B, Llama-3.1-8B, na Qwen2.5-7B kuri OpenBookQA, CoQA, TriviaQA, na ARC-Ikibazo. Ukurikije ibisobanuro, CLLM nuburyo bukora neza muburyo bwo gusubiza ibibazo mugihe gikomeza ikosa ryateganijwe. Abanditsi bavuga kandi ko CTC ije mu ijanisha rya 1.5 ku ijana ry’ahantu heza harebwa-hamenyekana munsi yakira ikora ibiranga umurongo mu bice byinshi, nta gisekuru cyiyongera. Kubihitamo byatoranijwe kuri 80%, ibisobanuro byerekana 99.0% byukuri kuri OpenBookQA kuri CLLM hamwe na SCC. Itangira kuvuga ibisubizo bya ARC-Ikibazo kuri CLLM ifite ikizere Csem ariko igabanywa mbere yo gutanga ibisubizo, kugirango ikirego ntigishobora gusuzumwa uhereye kubitangwa.
Ibisobanuro birambuye: arxiv.org ↗
Impamvu ari ngombwa
Urugero rwindimi rushobora gutanga ibisubizo neza hamwe nicyizere kidafite ishingiro. Niba ibisubizo byatangajwe bikomeje, gupima ukutamenya gushidikanya kubishobora kugabanywa byinshi bishobora gufasha sisitemu kumenya ibisubizo bikeneye kugenzurwa, kwifata, cyangwa gusubiramo abantu bidasabye ibisekuruza byiyongera mubibazo byinshi.
Ikibazo nyamukuru ntabwo ari ukumenya niba icyitegererezo cyururimi gishobora gusubiza ikibazo, ahubwo niba gishobora gutandukanya ubumenyi nubudashidikanywaho. Igisubizo cyiza ariko kitari cyo kirashobora guteza akaga kuruta kwangwa byimazeyo mugihe abakoresha bafashe ikizere nkibimenyetso. Uru rupapuro rwasabwe guhagararirwa kwishura rukemura icyo kibazo mugukomeza kutumvikana hagati yubuhanuzi bushingiye. Niba uburyo rusange, bushobora guha abitezimbere ikimenyetso cyicyitegererezo cyo guhitamo igihe cyo gusubiza mu buryo butaziguye, gusaba kugenzurwa, guhuza ikibazo kurindi sisitemu, cyangwa birimo umuntu.
Raporo yatoranijwe yo guhitamo-guhanura irakenewe cyane cyane kubyoherejwe kuko sisitemu yo guhitamo idakeneye gusubiza buri kibazo. Sisitemu ishobora kugumana ukuri kwinshi mugihe ikubiyemo gusa imanza ibona ko ishyigikiwe bihagije irashobora kuba ingirakamaro mugushinga aho amakosa atwara ibiciro bifatika. Raporo 99.0% yerekana neza kuri 80% kuri OpenBookQA irashimishije muri iyo mibare no kuboneza, ariko bigomba kumvikana nkigisubizo cyimpapuro aho kuba ibimenyetso byerekana ko sisitemu yoherejwe yagera kumikorere imwe. Ibisobanuro ntibisobanura umubare wingero, uburyo bwo kugereranya, cyangwa igisobanuro cyibikorwa byo gukwirakwiza.
Nta-yongeyeho-ibisekuruza bisabwa kuri CTC nabyo birashobora kugira akamaro kubishushanyo mbonera. Tekinike nyinshi zidashidikanywaho zishingiye ku gutanga ibintu byinshi byuzuye, bishobora kongera kubara no gutinda. Inkomoko ivuga ko CTC ikomatanya ibintu byinshi bidashidikanywaho nta gisekuru cyiyongereye, mu gihe SCC ikoresha mu buryo bweruye ibyitegererezo byuzuye. Itandukaniro riha impapuro inguni yubuhanga: amanota imwe arashobora kubahendutse kuyashyira mugihe, mugihe ayandi ashobora gufata ubwumvikane buke muburyo butaziguye. Ibisobanuro ntibisobanura ikiguzi cyitsinda rya LoRA ubwaryo, icyakora, igicuruzwa cyose cyagurishijwe ntikiramenyekana.
Uburyo bukoreshwa: Uburyo bukora
Shakisha ikoranabuhanga ryihishe inyuma yiri terambere.
What is a common training objective for an autoregressive language model?
Ibyo kureba
Igisubizo kuri ubu ni umwanditsi-watangajwe mbere yo gusuzuma, ntabwo yigenga yigenga. Ibisobanuro byingenzi bikomeza kutaboneka mubisobanuro byatanzwe, harimo ibisubizo byuzuye bya ARC-Ikibazo, ibyingenzi neza, kubara hejuru, nuburyo uburyo bwimura burenze urugero rwipimishije na datasets.
Ikibazo cya mbere ni ukumenya niba inyungu zavuzwe zirokoka kwigana. Inkomoko ni icapiro rya arXiv ryatanzwe ku ya 24 Kanama 2026, kandi ibikoresho byatanzwe birimo ibisobanuro gusa. Ibisubizo rero nibisabwa n'abanditsi, ntabwo bigenzuwe byigenga. Igenzura ryakagombye gusuzuma imbonerahamwe yuzuye, ibyingenzi, ibisobanuro byizere, itandukaniro ryimibare, kandi niba inyungu zavuzwe zihuye mumibare yose uko ari ine hamwe nimiryango itatu yicyitegererezo.
Interuro ya ARC-Challenge interuro ntabwo yuzuye: ivuga ko CLLM ifite ikizere Csem igera kubisubizo ariko idatanga agaciro. Ayo makuru yabuze agabanya kugereranya nibisabwa byuzuye bya OpenBookQA kandi birinda isuzuma ryuzuye ryimikorere yuburyo bwo gutekereza. Uru rupapuro rutangaza kandi ibisubizo bya -detection mubijyanye no kuba mumanota 1.5 ku ijana ya AUROC nziza cyane mubice byinshi, ariko isoko yatanzwe ntabwo igaragaza amanota yuzuye, uburyo bwiza bwo guhatana, cyangwa nibidasanzwe.
Ibibazo byo kohereza nabyo ni ngombwa. Urupapuro rukoresha itsinda rya adaptori ya LoRA, kandi abstract ntisobanura umubare uhuza adaptate zisabwa, uko zahuguwe, cyangwa igihe cyo kwibuka hamwe nigihe cyo kongeramo. Irasuzuma kandi bitatu gusa byitwa imvugo yindimi hamwe na bine bisubiza ibibazo. Kugeza ubu ntiharamenyekana niba amanota akora kubiganiro birebire, gufungura-kurangiza ibisekuruza, kwinjiza indimi nyinshi, imirimo yihariye, cyangwa moderi hanze yubunini bwageragejwe hamwe nubwubatsi. Kugeza ubwo ibyo bibazo bisubijwe, CLLM ifatwa nkuburyo bwubushakashatsi butanga ikizere cyo gupima ibintu bidashidikanywaho aho kuba uburyo bwemewe bwo gukoresha imigabane myinshi.