Buyela Ezindabeni
UkuqambaAI Understanding ukwaziswa

I-FaithSieve isebenzisa i-Lean ukuthuthukisa ukutholwa kwe-AI kwamaphutha okuqala ebufakazini bezibalo

Ukuphrinta kusengaphambili okusha kwethula i-FaithSieve, uhlaka olusizwa nge-Lean oluhlukanisa ubufakazi bezibalo obukhiqizwe yi-AI baba amayunithi okucabanga endawo futhi lubike ukunemba okuphezulu kokwenziwa kwendawo kwephutha lokuqala kunokwahlulela kwemodeli eqondile.

5 min readRead the primary source
Primary-source image accompanying FaithSieve uses Lean to improve AI detection of first errors in math proofs
Idokhumenti yomthombo oyinhlokoUmthombo urekhodiwe
Umshicileli
arxiv.org
Isixhumanisi somthombo
arxiv.orghttps://arxiv.org/abs/2608.26310
Uhlobo lomthombo
Idokhumenti eyisisekelo — isimemezelo esisemthethweni, iphepha, ukugcwalisa, noma ikhasi lomuntu wokuqala esilifunda ngokuqondile.
UmongoQonda lokhu ngemizuzwana engama-60

Qala lapha

Imigomo ebalulekile

Ukujwayela
Imodeli isebenza kahle kangakanani kudatha entsha, engabonakali ngaphandle kwesethi yokuqeqeshwa.
Ukulinganisa
Ukuthi amaphuzu okuzethemba kwemodeli afanelana kahle kangakanani namathuba okulutha.
Isichasiselo
Amalebula engeziwe ngabantu noma imethadatha esetshenziselwa ukuqeqesha noma ukuhlola amamodeli okufunda omshini.
ZihloleImibuzo Ecacisiwe yamamodeli e-AI

Kwenzekeni

Abacwaningi bethula i-FaithSieve, uhlaka lokuhlola ubufakazi bezibalo bolimi lwemvelo olukhiqizwa amamodeli ezilimi ezinkulu. Ichitha izinyathelo zobufakazi obubanzi ibe amayunithi okucabanga endawo, ikhiphe izibopho zobufakazi ezithayiphiwe futhi isebenzisa ukuhlola okusemthethweni okusekelwe ku-Lean lapho amaphuzu okuqondanisa kwe-semantic kukhombisa ukuthi isitatimende esisemthethweni silondoloza umongo wesimangalo sangempela, izinto kanye nefomu elinengqondo. Kumadathasethi amabili aqinisekiswe uchwepheshe, ababhali babika ukuthi i-FaithSieve ithuthukise ukwenziwa kwasendaweni okuyiphutha lokuqala ngaphezu kwesisekelo sokwahlulela okuqondile.

Ukuphrinta ngaphambilini, okuhanjiswe ku-arXiv ngo-Agasti 26, kwethula i-FaithSieve njengempendulo enkingeni ethile yokuhlola: amamodeli olimi amakhulu angakwazi ukukhiqiza ubufakazi obude, obunezinyathelo eziningi zezibalo, kodwa abahloli bangase bathwale kanzima ukukhomba iphuzu lokuqala lapho ukucabanga kungasebenzi khona. Ababhali baphikisa ngokuthi izindlela ezikhona zivame ukuncika ekwahluleleni kolimi lwemvelo okusekelwe kumodeli, okungase kugeje izikhala zasendaweni noma ngabe zabela ukuhlola okuhle okubanzi. Ngakho-ke inhloso eshiwo yincane kunokwahlulela bonke ubufakazi njengobulungile noma obungalungile: ukwenza indawo yokwehluleka kokuqala kubonakale kakhulu ngaphakathi kobufakazi.

I-FaithSieve yephula izinyathelo zobufakazi obumahhadlazelayo ibe amayunithi amancane okucabanga futhi iguqule lawo mayunithi abe yizibopho zobufakazi ezithayiphiwe. Ibese isebenzisa i-ejenti yokuhlola esemthethweni ene-Lean, isistimu yokuqinisekisa i-theorem, ukuze ihlole izibopho eziwumphumela. Uhlaka aluphathi isheke le-Lean eliyimpumelelo njengokwanele ngokwalo. Esikhundleni salokho, ukuqinisekiswa okusemthethweni kufaka phakathi amaphuzu wokuqondanisa kwe-semantic okuhloswe ngawo ukunquma ukuthi isitatimende esisemthethweni siyawugcina yini umongo wobufakazi bangempela, izinto zezibalo kanye nesimo esinengqondo. Lokho ku-oda kugcina umehluko phakathi kokuhlola isibopho esisemthethweni kanye nokunquma ukuthi isibopho siwukumela okwanele kweyunithi engakahleleki.

Ababhali babika imiphumela kumadathasethi amabili aqinisekisiwe ochwepheshe: I-ProofLoc-Olympiad iqukethe izinkinga ezingu-350, kuyilapho i-ProofLoc-University iqukethe izinkinga ezingu-200 ezihlanganisa izizinda eziyisithupha ezithuthukile. Ngomgogodla we-GPT-5.4, i-FaithSieve izuze ukunemba okunembile okungu-81.43% ekutholeni iphutha lokuqala kusethi ye-Olympiad, uma kuqhathaniswa no-72.29% wokwahlulela okuqondile. Ngokwesilinganiso senyuvesi, ifinyelele ku-84.5%, uma iqhathaniswa no-75.0% wejaji eliqondile. Lezi izimangalo zokuphrinta ngaphambilini, hhayi imiphumela esungulwe ngokuzimela. Ukuqhathanisa kwethulwa njengokuhlolwa kokwenziwa kwendawo, kunokuba njengobufakazi bokuthi bonke ubufakazi obuyisisekelo noma ukwenziwa kube semthethweni kulungile.

Imininingwane yomthombo: arxiv.org ↗

Kungani kubalulekile

Umsebenzi ubhekana nobuthakathaka obungokoqobo ekuhloleni ukucabanga kwe-AI: ubufakazi bungabukeka bukholisa kuyilapho buqukethe iphutha lendawo langaphambi kwesikhathi, futhi isisho esisemthethweni kwesinye isikhathi singaqinisekisa isitatimende esingameleli ngokwethembeka impikiswano yasekuqaleni. Umhloli onembe kakhudlwana angasiza abacwaningi, othisha nabathuthukisi bakhombe lapho ukucabanga kwezibalo kwehluleka kunokuthembela kuphela esahlulelweni esisodwa sokulunga. Imiphumela ebikiwe iyathembisa kodwa iphuma ekuphrinteni kwangaphambilini okusha futhi ayisunguli ukusebenza kukho konke ukubhala kwezibalo noma amasistimu e-AI asetshenzisiwe.

Ukubaluleka okumaphakathi kuyi-methodological. Ukuhlola ukuthi ubufakazi obuphelele bungenziwa ngokusemthethweni akufani nokunquma ukuthi isinyathelo ngasinye esingakahleleki sivumelekile yini. Isaga esisemthethweni singaqinisekisa okuqondiwe okubanzi ngokweqile, noma isitatimende esisemthethweni singase sisuke kulokho okushiwo ubufakazi bokuqala. Inhlanganisela kaFaithSieve yokubola kwendawo kanye nesango le-semantic yakhelwe ukunciphisa lezo zindlela ezimbili zokwehluleka. Inzuzo ehlosiwe iwukuxhumana okucophelela kakhudlwana phakathi kombhalo wobufakazi kanye nokuhlola okusemthethweni okusetshenziselwa ukukuhlola.

Konjiniyela bezinhlelo ze-AI zezibalo, ukwenziwa kwasendaweni kwephutha lokuqala kuyenzeka kakhulu kunomphumela owodwa wokudlula noma wokufeyila. Ukwazi ukuthi ingxabano iqala kuphi kungasekela ukuqeqeshwa kabusha okuhlosiwe, ukulungisa amaphutha okungcono kanye nokuqhathanisa okufundisayo phakathi kwamamodeli. Kungase futhi kwenze ukuhlola kube lula ukucwaningwa, ngoba izibopho ezisemthethweni zinganikeza ubufakazi obuxhumene namayunithi okucabanga angawodwana kunokuncika ngokuphelele esahlulelweni sephrozi somhloli. Lokhu kungenza kube lula ukuhlola ukungezwani komuntu ngamunye, kuyilapho kushiya ukwahlulela okubanzi mayelana nobufakazi obugcwele.

Igebe elibikiwe mayelana nokwahlulela okuqondile likhulu ngokwanele ukuba libonakale phakathi kwalezi zivivinyo: Amaphesenti angu-9.14 amaphuzu ku-benchmark ye-Olympiad kanye namaphuzu angu-9.5 ku-benchmark yenyuvesi. Noma kunjalo, ubufakazi bukhawulelwe ekuhlolweni kwephepha uqobo. Umthombo awunikezi ulwazi olwanele ukuze uhlole ikhwalithi yesichasiselo, ukungaqiniseki kwezibalo, izimo zokuhluleka, isikhathi sokusebenza, izindleko, ukuzwela kumgogodla we-GPT-5.4 noma ukusebenza ngokumelene nochwepheshe bezibalo abangabantu. Lokho kweqiwa kubalulekile ngoba kuthinta indlela imiphumela yezinombolo okufanele ihunyushwe ngayo nokuthi abanye bangaphinda bakhiqize kanjani ukuqhathanisa.

Interactive Mechanism

I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela

Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
I-Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Ongakubuka ngokulandelayo

Imibuzo ebalulekile iwukuthi ingabe i-FaithSieve yenza ngokujwayelekile ngale kwamabhentshimakhi amabili, ukuthi isango lalo lokuqondaniswa kwe-semantic likwenqaba kaningi kangakanani ukwenza okusemthethweni okudukisayo, kanye nokuthi inqubo yokuhlola idinga ulwazi olungakanani lomuntu noma ukubala. Ukucutshungulwa okwengeziwe kufanele kuhlole ukwakhiwa kwebhentshimakhi, izigaba zamaphutha, izilungiselelo zemodeli nejaji, nokuqhathanisa nezinye izindlela zokuhlola ubufakazi. Umthombo awuqinisekisi ukuthi i-FaithSieve ifakazela ukuthi izimpikiswano ezingakahleleki zilungile, ingena esikhundleni sosolwazi bezibalo noma ivimbela zonke izinhlobo zokucabanga ezihunyushwa ingekho.

Inkinga yokuqala yokulandelela iwukuba jikelele. Izilinganiso zichazwa njengeziqinisekiswe uchwepheshe futhi zihlanganisa izinkinga ze-Olympiad kanye nezizinda eziyisithupha ezithuthukisiwe zezinga lenyuvesi, kodwa umthombo awuwuhlonzi ukuxuba kwesizinda esigcwele, izitayela zolimi, amafomethi obufakazi noma ukusatshalaliswa kwamaphutha. Imiphumela ingase ishintshele ebufakazini obungakhiwe kancane, izinkambu zezibalo ezihlukene, amamodeli abuthaka noma anamandla, noma izimpikiswano ezincike ezincazelweni ezingamelelwe kalula ohlelweni oluhlelekile. Ngakho-ke incazelo etholakalayo isekela umphumela webhentshimakhi, kodwa hhayi isiphetho sokuthi ukuziphatha okufanayo kuzoba khona kuzo zonke izilungiselelo.

Abacwaningi kufanele futhi bahlole isigaba sokuqondanisa kwe-semantic. Inhloso yawo ibalulekile ngoba ubufakazi obusemthethweni buwusizo kuphela uma buhambisana nesimangalo esingakahleleki esihlolwayo. Imibiko engacacile yokuthi i-FaithSieve isebenzisa isikolo sokuqondanisa, kodwa ayisho ukulinganiswa kwesikolo, ukukhethwa komkhawulo, izinga lamaphutha noma ukuthi izinqumo zokuqondanisa zenziwa abantu, amamodeli noma enye inqubo. Leyo mininingwane izonquma ukuthi uhlaka lugwema ngokuthembekile kangakanani ukuqinisekisa isitatimende esingalungile. Ngaphandle kwalezo zilinganiso, isango liyingxenye ebalulekile echazwe ukusebenza kwayo okuhlala kuwumbuzo ovulekile.

Okokugcina, ukuthunyelwa okusebenzayo kuzoncika ngaphezu kokunemba okuqondile. Umsebenzi wesikhathi esizayo kufanele ubike ukuthi singakanani isikhathi nokubala okudingekayo ukuze kukhishwe ubufakazi, kukhiqizwe izibopho ezisemthethweni kanye nokuhlola ama-Lean, kanye nokuthi uhlelo lusingatha kanjani ukuhlelwa okusemthethweni okuhlulekile kanye nephrozi engaqondakali. Umthombo futhi ushiya kungaziwa ukuthi i-FaithSieve ingakwazi yini ukuhlola ubufakazi ekugelezeni komsebenzi wangempela kwezemfundo noma kocwaningo, ukuthi ubufakazi bayo buyaqondakala yini kwabangebona ochwepheshe nokuthi ingabe ukuthuthukiswa kuyaqhubeka yini lapho ababhali besilinganiso, amamodeli nabahloli beshintshwa. Le mibuzo iphathelene nokusebenziseka kohlaka ekusebenzeni futhi ayiguquli ukuqhathanisa kwebhentshimakhi okubikiwe.

Imihlahlandlela ehlobene nemibuzo

Amamodeli e-AI AchaziweUkuqeqeshwa kwe-AIAma-TransformersHlola okwaziyo — zama imibuzo ye-AI yamahhalaBheka igama le-AI kuhlu lwethu lwamagamaLandela i-tracker yokukhishwa kwemodeli ye-AI
Uthole lokhu kuwusizo?