Byagenze bite
Abashakashatsi berekanye RecurSE, uburyo bwo kunonosora buri gihe sisitemu ya LLM-nkumucamanza nta kugenzura zahabu yo hanze mugihembo-cyo kwiga-gushimangira. Uburyo bukoresha umucamanza utozwa, umugenzuzi uhuza hamwe na monitor yo kwemeza kugirango abuze kandi asuzume inzira.
Uru rupapuro rugaragaza RecurSE, ngufi yo Kwisubiramo Kwisubiramo, nkuburyo bugarukira bwo kwisubiramo kwisubiramo kuri sisitemu ya LLM-nkumucamanza. Sisitemu isuzuma inyandiko ifunguye, nkibisubizo byabakandida, irwanya rubrics kandi irashobora no kuyobora kuyobora nyuma y amahugurwa. Abashakashatsi bashimangira ikibazo nyamukuru nko kunoza umucamanza ubwayo badashingiye ku birango bya zahabu byo hanze, ibihembo byerekana ibihembo cyangwa gutandukanya umwarimu ukomeye. Icyifuzo cyabo cyo guhugura ahubwo gikoresha icyitegererezo cyo gusuzuma ubushobozi bwo gutanga ibimenyetso byo kwiga neza. Inkomoko isobanura ibi nkuburyo bufunze, ariko ntibivuga ko icyitegererezo cyateye imbere hatabayeho imiterere cyangwa isuzuma; uburyo bukubiyemo rubrics, umugenzuzi nuburyo bwo kwemeza.
Sisitemu yatanzwe ifite passes ebyiri. Muri Pass 1, umucamanza uhugura asuzuma ibisubizo byabakandida munsi ya rubrics kandi agatanga umwanzuro. Muri Pass 2, kugenzura politiki-kopi igenzura igenzura ibitekerezo byumucamanza kuri meta-rubrics kandi itanga ibihembo bya scalar. Abashakashatsi bagaragaza uburyo bwananiranye umucamanza ashobora gukoporora cyangwa gukoresha ibimenyetso byaciwe kugirango atange ibihembo byayo. Kugira ngo ibyo bibazo bikemuke, RecurSE ikoresha interineti igaragara: amanota ya scalar ya cheque yatandukanijwe nibimenyetso byumucamanza. Ibi bigamije kuvanaho kwangirika kwa token-gukoporora shortcut no gukora ibihembo byabyaye umusaruro kurushaho. Uru rupapuro ruvuga kandi ko imyigire yisubiramo idafite inanga yo hanze iba ifite imipaka. Itangiza Pairwise Advantage Validity, cyangwa PAV, nkigenzura ryemeza rifatanije gukurikirana ukuri kwabacamanza nubudahemuka. Intego yavuzwe ni ukumenya idirishya ryiza-rihagarara, aho kubireka bikomeza.
Ibisobanuro byerekana ubushakashatsi burimo Qwen3.5-9B, Gemma-4-E4B-it na Qwen3.6-27B. Ivuga ko ubwo buryo bwatanze umusaruro uhoraho muri rusange ku buvuzi bwakozwe, mu buryo bubiri, mu ncamake no ku bipimo by’umwuga, kandi ko ubushakashatsi bwo gukuraho bwasanze abacamanza-bagenzuzi bahuje ubwihindurize bakomeye kurusha abagenzuzi bakonje, meta-bacamanza bo hanze, kwishyira ukizana no gutesha agaciro abarimu. Inkomoko ntabwo itanga ingano yumubare wibyo byungutse mumyandiko yatanzwe. Abashakashatsi bakomeza bavuga ko guhitamo gukosorwa n'umucamanza wabo byateje imbere politiki yo hasi. Ibi byagura uburyo burenze gusuzuma ibisubizo: ibisubizo byumucamanza bitangwa nkamakuru yingirakamaro mumahugurwa yo guhuza indi politiki. Uru rupapuro rufata umwanzuro ni uko imbibi zisubirwamo zo kwiteza imbere ku bacamanza ba LLM ari ingirakamaro mugihe agaciro k'ibihembo byatanzwe ubwako kaciwe kandi kagakurikiranwa. Nkibisobanuro bya arXiv aho kuba urungano rwasuzumwe n’urungano, ibisabwa ni ibisubizo byatanzwe n’abanditsi, ibyifuzo byabo, ntabwo byagaragaye ku bwigenge.
Ibisobanuro birambuye: arxiv.org ↗
Impamvu ari ngombwa
Abacamanza ba LLM barushijeho gukoreshwa mugusuzuma ibisubizo byafunguye no kuyobora icyitegererezo nyuma y amahugurwa, ariko kunoza abo bacamanza birashobora gusaba ibisobanuro byabantu bihenze cyangwa icyitegererezo cyabarimu gikomeye. RecurSE itanga inzira ishoboka yo kugenzura-kugenzura, mugihe kandi yerekana impamvu kwiteza imbere bisaba kurinda ibihembo bizenguruka cyangwa biyobya.
Akamaro k'umurimo kari mu giciro no ku gipimo cyo gusuzuma icyitegererezo. Abasesengura abantu barashobora kubahenze kandi batinda, mugihe abarimu bakomeye bigezweho cyangwa ibihembo bitandukanye byahawe ibihembo byongera ibikorwa remezo no kwishingikiriza. Niba icyitegererezo gishobora gutanga ibimenyetso byingirakamaro byokwiga kunoza isuzuma, abitezimbere barashobora kwagura isuzuma cyangwa nyuma yimyitozo yakazi hamwe na annotasiyo yo hanze. Inkomoko ishyigikira gusa ibisobanuro bigufi bivuga ko abanditsi bagerageje uburyo na raporo yunguka muburyo bwatoranijwe. Ntabwo byerekana ko ubugenzuzi bwo hanze bushobora kuvaho muri rusange cyangwa ko RecurSE ihendutse muri buri gikorwa. Uru rupapuro kandi rukemura ikibazo cyo kwizerwa hagati mugusuzuma AI: sisitemu irashobora kugaragara nkiterambere kuko yize guhaza abayisuzuma aho guca imanza nziza.
Gutandukanya gutandukana hagati yigihembo cya scalar nigihembo cyumucamanza cyateganijwe kugabanya inzira imwe yihariye. Ikurikiranabikorwa rya PAV rigamije kumenya igihe imyitozo isubiramo ikomeje kunoza isuzuma ryemewe aho gutembera mu kwikomeza. Ibi bituma akazi kadakorwa neza gusa ahubwo no kwizerwa ryisuzuma ryikora rikoreshwa mugutezimbere icyitegererezo. Ibipimo ngenderwaho byavuzwe bikubiyemo ubuvuzi, byombi, incamake n'imirimo y'umwuga aho kuba ikizamini kimwe. Isuzuma ryakozwe, nkuko byasobanuwe muri abstract, rigamije gukemura niba uburyo rusange burenze ingero zikoreshwa mugihe cy'amahugurwa. Biracyaza, inkomoko ntabwo itanga ibisobanuro birambuye hano kubyerekeye iyubakwa ryamakuru, ibipimo ngenderwaho byo gusuzuma, indangagaciro zifatizo, ingano yikizamini cyangwa ubunini nubwinshi bwiterambere. Hatariho ibisobanuro birambuye, abasomyi ntibashobora kumenya ingano nini, ikomeye cyangwa mubikorwa byingenzi byavuzwe.
Ikirego cyo guhuza ibice gishobora kuba ikibazo mugihe umucamanza watoranijwe agabanya imirimo ikora mugutanga amakuru yamahugurwa. Nyamara umucamanza utezimbere guhuza rubrici yacyo ntabwo ahita umucamanza ufata ibyifuzo byabantu, amahame ya domaine cyangwa ibisabwa byumutekano. Inkomoko ivuga ko guhitamo byombi byongereye politiki yo guhuza politiki yo hasi, ariko ntigaragaza neza ibyo bakunda bisobanura guhuza, uko guhuza byapimwe cyangwa niba abasesengura abantu bemeje impinduka zavuyemo. Iri tandukanyirizo ni ingenzi kubantu bose batekereza uburyo muburyo bworoshye cyangwa bukomeye. Muri rusange, RecurSE itanga urugero rufatika rwimipaka yo kwiteza imbere. Igishushanyo cyacyo gitekereza ko ibihembo byakozwe ubwabyo bishobora kugirwa akamaro binyuze mu gutandukanya imiterere no gukurikirana; ntibikuraho gukenera gusobanura rubrics, meta-rubrics cyangwa intego yo kwemeza. Umusanzu rero urumvikana neza nkurwego rwateganijwe rwo guhugura no kugenzura icyiciro runaka cyabasesengura LLM. Agaciro rusange kazaterwa nimba nyuma yimirimo yemeza ko umutekano wacyo ukomeza kuba ingirakamaro mugihe icyitegererezo, imirimo, rubrics cyangwa ibyifuzo bihinduka.
Uburyo bukoreshwa: Uburyo bukora
Shakisha ikoranabuhanga ryihishe inyuma yiri terambere.
Which component of an AI application is the machine-learning model itself?
Ibyo kureba
Inkomoko ivuga ibyungutse muburyo butandukanye hamwe nubwoko bwibipimo ariko ntibitanga ingano yingaruka, ingano ya dataset, ubusobanuro bwibarurishamibare cyangwa ibisobanuro byigenga byo kwigana mubisobanuro. Ubundi bugenzuzi bugomba gusuzuma niba uburyo bukomeza kwizerwa hanze yimiterere yatangajwe kandi niba monitor yayo ihagarara ishobora kumenya kwibeshya-kwiteza imbere.
Ikibazo cya mbere ni ukumenya niba ibyatangajwe byavuzwe bikomeza kwigana. Inkomoko yatanzwe igaragaza imiryango itatu yicyitegererezo cyangwa verisiyo hamwe nibyiciro byinshi, ariko ntisobanura umubare wimirimo, imbuto zidasanzwe, imyitozo cyangwa ibizamini byibarurishamibare. Isuzuma ry'ejo hazaza rigomba gutanga amakuru arambuye kandi ukagereranya RecurSE ukurikije ibice bimwe munsi ya compte ihuye, amakuru hamwe nuburyo bwo guhugura. Ibyo byasobanura niba inyungu ziva mubishushanyo bisubirwamo ubwabyo, kongera ibitekerezo, guhitamo ibipimo byihariye cyangwa ibindi bitandukanye muburyo bwo kugerageza.
Ikibazo cya kabiri ni ukumenya niba PAV ishobora kumenya neza ingingo aho kwiteza imbere bigomba guhagarara. Ibisobanuro bisobanura PAV nk'umugenzuzi wo kwemeza utabogamye hamwe ukurikirana ukuri k'umucamanza n'ubudahemuka bw'abagenzuzi, ariko ntibisobanura iyubakwa rya monitor cyangwa ngo ritange ibimenyetso byerekana ko ryakozwe. Abashakashatsi n’abakoresha bagomba gusuzuma imanza aho umucamanza n’umugenzuzi bahuriza ku mpamvu zitari zo, aho bombi bazungura ahantu hamwe cyangwa aho imikorere igenda neza kuri rubrici mu gihe itesha agaciro ireme ry’umuntu. Ibizamini nkibi byerekana niba itegeko ryateganijwe ryo guhagarika ryerekana iterambere nyaryo aho guhuza imbere.
Uburyo bushingiye kuri rubrics na meta-rubrics nabwo bukwiye kwitabwaho. Inkomoko ivuga ko umucamanza asuzuma ibisubizo hakurikijwe amategeko ya rubrics kandi igenzura rikagenzura ibitekerezo biterwa na meta-rubrics, ariko ntirisobanura uburyo aya ruburasi yanditswe, avugururwa cyangwa arinzwe bidasobanutse. Amategeko adasobanutse neza arashobora guha repursive loop intego ihamye ariko itifuzwa. Akazi kazoza kagomba gusuzuma ibipimo bivuguruzanya, amabwiriza atuzuye, ibisubizo birwanya hamwe na domaine aho gukosora bigoye kugabanuka kugeza kuri rubricike yanditse, harimo nubuvuzi bwatanzwe nubuvuzi.
Hasi yo gukoresha imikoreshereze yumucamanza-yatunganijwe ni akandi gace ko kugenzura. Inkomoko ivuga ko politiki yatunganijwe neza, ariko kohereza bifite akamaro bisaba ibimenyetso byerekana ireme ryiza, kutemeranya nabashinzwe gusuzuma no kunanirwa kwatangijwe na gahunda yo gutunganya. Kugeza ubu ntiharamenyekana niba inzira yimurwa mu bunini bw'icyitegererezo, indimi, indangarugero n'intego zo gusuzuma, cyangwa niba ubudahemuka bw'umugenzuzi bugwa iyo umucamanza ahuye n'imirimo itamenyerewe. Ibyo bitazwi bigabanya uburyo bwagutse imyanzuro ya abstract igomba gukoreshwa.
Hanyuma, impapuro ni icapiro rya arXiv ryatanzwe ku ya 25 Kanama 2026, kandi isoko yatanzwe ikubiyemo gusa urupapuro rudasobanutse na bibliografiya. Impapuro zuzuye zirashobora kubura ibisobanuro byubushakashatsi, ariko ntibishobora gufatwa hano. Basomyi bagomba gushakisha uburyo bwuzuye, kode cyangwa amakuru yasohotse, ibisubizo byo gukuraho, gusesengura amakosa hamwe nubushakashatsi bwigenga. Kugeza icyo gihe, RecurSE ni gihamya yicyerekezo cyihariye cyubushakashatsi na raporo yabanditsi yerekana ibisubizo bitanga icyizere, ntabwo ari gihamya ko abacamanza ba LLM bashobora kwiteza imbere neza cyangwa bizewe muri rusange.