Dzokera kuNhau
InnovationAI Understanding muchidimbu

LURE bepa rinokurudzira kutsvaga-kunzvenga-kutamba kudzidzisa LLM kufunga pasina data rebasa

Bepa idzva rearXiv rinopa LURE, zero-data-self-play nzira umo modhi yemutauro mumwe inoseta kuoma kwebasa uku imwe ichigadzirisa zvinonetsa kufunga. Vanyori vanoshuma kwakasimba kunze-kwe-kugovera zero-pfuti kurongeka kupfuura yakadzidziswa mabhenji mumabhenji mapfumbamwe akabatwa, asi abstract haaiti…

5 min readRead the primary source
Primary-source image accompanying LURE paper proposes pursuit-evasion self-play to train LLM reasoning without task data
Primary-source documentKwakanyorwa
Muparidzi
arxiv.org
Source link
arxiv.orghttps://arxiv.org/abs/2608.21871
Source type
Gwaro rekutanga - chiziviso chepamutemo, bepa, faira, kana peji rebato rekutanga ratinoverenga zvakananga.
ContextNzwisisa izvi mumasekonzi makumi matanhatu

Tanga pano

Matemu akakosha

Mutauro Mukuru (LLM)
Mutauro wemodhi yakadzidziswa pane yakakura text corpora kugadzira nekuongorora zvinyorwa.
Kuzvitamba
Seti yekudzidzisa apo modhi inovandudzwa nekugadzira data kuburikidza nekudyidzana kana makwikwi nemakopi ayo.
Kusimbisa Kudzidza
Kudzidziswa nemasaini masaini apo mumiririri anodzidza zviito zvinowedzera kudzoka kwenguva refu.
Zviedze iwe pachakoAI Models Inotsanangurwa Mibvunzo

Chii chaitika

Bepa rearXiv rinounza LURE, nzira yekusimbisa-yekudzidza inovavarirwa kuvandudza magadzirirwo emutauro wakakura pasina kuvimba nekuunganidzwa kwemabasa akakura akachekwa nevanhu. Iyo inoronga kudzidziswa semutambo wekutsvaga-kunzvenga: anonzvenga anoisa mabasa pamwero wekuomerwa, nepo murongi-anodzingirira anoedza kuzvigadzirisa kuburikidza nekudyidzana kunogoneka.

Iro bepa, rakatumirwa kuarXiv muna Nyamavhuvhu 22, 2026, rinotsanangura LURE senzira ye zero-data yekuzvitamba mumutauro wakakura womuenzaniso kufunga. Vanyori vanoti kudzidza kwakajairwa kusimbaradzwa nemibairo inovimbika kunowanzo tora kuwana kuunganidzwa kwakakura kwemabasa akacheneswa nevanhu. Chinangwa chavo chakataurwa ndechekubvisa kutsamira uku nekuva nemamodheru anogadzira kana chinzvimbo chebasa panguva yekudzidziswa pane kuvimba zvizere nedziva rebasa rakagadzirirwa. Sosi inosimbisa izvi sedambudziko rekutsvaga kwebepa uye chikumbiro; hairatidze kuti LURE inzira yakatumirwa kana nzira yekugadzira-yakagadzirira yekudzidzisa.

LURE inokamura maitiro ekudzidzisa kuita mabasa maviri. Mudzivisi weLLM anoisa mabasa padivi pekuomerwa kwezvakatipoteredza, nepo anoronga-anoteedzera anoedza kugadzirisa iwo mabasa kuburikidza nekudyidzana kune mhedzisiro inogona kusimbiswa. Iro bepa rinoti anonzvenga anogashira mubairo wekutora-muganho uri wepamusoro kana mugadziri aitora pahafu chaiyo yekuburitswa kwayo. Mukugadzirisa kwevanyori, mubairo iwoyo unoshandura "kusabatika" kuisa basa kuita chimwe chinhu chakadzidzwa pane kumanikidzirwa nechikumbaridzo chakasarudzwa nemaoko. Iwo abstract haatsanangure iwo chaiwo nharaunda, mafomati ebasa, kana verifier kuita.

Iyo nzira zvakare inoshandura maitiro ekugadzirisa modhi anogamuchira kiredhiti yekudzidziswa. Panzvimbo pekuvimba chete pamubairo wepakati pekupedzisira, LURE inoshandisa inonzi nebepa kubatwa-yakamisikidzwa dense process credit. Chidimbu chinoti monotone verifier kufambira mberi ndeyeboka-yakajairwa pamwe chete neterminal kubatwa, pasi peyakatenderedzwa-yakamisikidzwa KL inomanikidza kudzikamisa co-evolution pakati pekuita-basa uye kugadzirisa basa. Vanyori vanorondedzera bvunzo munzvimbo nhatu dzinogoneka dzekufunga uye mhuri nhatu dzemusana, vachienzanisa yakabatana uye nyanzvi marongero. Kwainobva haazivise mazita emusana, mabhajeti ekudzidzisa, ablations, kana chaiwo magadzirirwo ekutanga.

Kwakabva mashoko: arxiv.org ↗

Nei zvichikosha

Kana mhedzisiro yakashumwa ikakamira, LURE inogona kupa vaongorori nzira yekugadzira uye kusarudza anobatsira ekufunga matambudziko pasina kuunganidza akakura akanyorwa dhataset. Nzira yaro inonangana nematambudziko maviri anoramba achidzidziswa kamwechete: kusarudza mabasa akaoma asi anodzidzika, uye kupa chikwereti kumatanho epakati pane kungopindura mhinduro dzekupedzisira.

Izvo zvakakosha zvechikumbiro ndechekuedza kwayo kuderedza kutsamira pazviitwa zvinogadzirwa nevanhu. Kugadzira mahombe, emhando yepamusoro kuunganidzwa kwemabasa kunodhura, uye kuunganidzwa kwakagadziriswa kunogona kuita kuti zviome kuchengeta matambudziko ekudzidzisa padanho rakakodzera. LURE's evader-vateveri setup yakagadzirirwa kugadzirisa kuoma sezvo solver inovandudza. Kana iyo michina ikashanda nekuvimbika, inogona kuita kuti kudzidzira kutamba wega kuve kuchinjika uye nekugona kuderedza huwandu hwebasa remanyorero rinodiwa kune mamwe matunhu ekufunga. Icho chinhu chinogona kuitika chekugadzira, kwete mhedzisiro yakazvimirira.

Iro bepa zvakare rinotarisa kupihwa kwechikwereti, nyaya yepakati mune akawanda-nhanho kudzidzisa kufunga. Mugadziri anogona kusvika pamhedzisiro yekupedzisira mushure mekutevedzana ine sarudzo dzinobatsira uye dzisingabatsire, nepo mubairo wekupedzisira wega unopa ruzivo rudiki pamusoro pematanho ane basa. Nekubatanidza yepakati verifier kufambira mberi kune yekupedzisira kubatwa chiitiko, LURE ine chinangwa chekupa denser yekudzidza chiratidzo pasina kurasa kukosha kweiyo yakasimbiswa mhedzisiro. Iyo abstract mishumo yekuti musanganiswa uyu wakapfuura mabhesiro epamusoro muzviedzo zvevanyori, asi hazviratidze kana kuvandudzwa kwacho kwakauya zvakanyanya kubva pakusarudzwa kwebasa, denser credit, KL kudzikamisa temu, kana kudyidzana kwavo.

Mhedzisiro yakasimba yakashumwa ndeyokuti modhi yakabatana yakawana huchenjeri hwepamusoro-soro kunze kwekugovera zero-pfuti kupfuura ese akadzidziswa mabhenjimaki ekumisikidza anotora mhuri nhatu dzebasa. Chirevo ichocho, kana chichigadzirwazve, chinoratidza kuti nzira yacho inogona kuvandudza kufambisa kwete kungokwirisa nharaunda dzinoshandiswa panguva yekudzidziswa. Zvakadaro, "kuunganidzwa kwakasimba" hakuna kukwana kuona kukosha kunoshanda: iyo abstract haipe zvibodzwa, nguva dzekuvimba, pa-benchmark zvabuda, kuenzanisa neakakura kana akawanda macomputer-akasimba masisitimu, kana ruzivo rwekuti mabasa akaitwa akasarudzwa sei. Basa iri saka rinonzwisiswa zvakanyanya semhedzisiro yekutsvagisa inoda kumwe kuongororwa, kwete sehumbowo hwekuti zero-data wega-kutamba yakagadzirisa kudzidziswa kwekufunga-chinangwa.

Interactive Mechanism

Interactive Mechanism: Iyo Inonyatsoshanda

Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.

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.
Interactive Concept Check+10 Points
AI Models Explained Quiz

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

Zvekutarisa zvinotevera

Iro bepa pfumbamwe-mapeji arXiv preprint, uye sosi rinopa chete chidimbu. Mibvunzo yakakosha inoramba iri pamusoro pehukuru hwenhamba yezvakaburitswa zvakashumwa, mutengo wekombuta, kuvaka basa, saizi dzemodhi, maumbirwo ebhenji, kuberekana, uye kuti nzira yacho inofambiswa kupfuura nharaunda nhatu uye mhuri nhatu dzebasa dzakaedzwa.

Danho rekutanga rekuongorora ndiro bepa rakazara uye matafura aro ekuedza. Vaverengi vanofanira kutsvaga kuzivikanwa uye hukuru hwemhuri nhatu dzemusana, tsanangudzo chaidzo dzenzvimbo nhatu idzi, nhamba nemhando yemabasa ekudzidzisa, nzira dzekutanga, uye komputa yakashandiswa. Iwo mameseji ndiwo anozoona kana zvakawanikwa zveLURE zvichiratidza rakawanda rinobatsira rekudzidzisa resipi kana mhedzisiro yakasungirirwa kune imwe bhenji dhizaini. Kwakabva kwabva kunotsinhira kuti bepa rine matafura mashanu nenhamba shanu, asi haripe zvirimo.

Reproducibility imwe nyaya yakavhurika. Iyo yakapihwa sosi haitaure kana kodhi, basa jenareta, verifier maitirwo, macheki, kana data rekudzidzisa riripo. Nekuti iyo nzira inotsamira pane-inoshanduka-shanduka evader uye anoteedzera, sarudzo diki mukuyeresa mubairo, kuburitsa sampling, normalization, kana kudzikamisa zvinogona kukanganisa mhedzisiro. Kudzokororwa kwakazvimiririra kumhuri dzakasiyana dzemhando uye nharaunda dzinogoneka kwaizobatsira kuona kuti hunhu hwakataurwa hwakasimba here kana hunoenderana nekuitwa kwevanyori.

Chekupedzisira, chiyero chemaitiro chinoda kuyedzwa kupfuura zvakataurwa nebepa mabhenji mabhenji uye mhuri nhatu dzebasa. Mamiriro ekuona echokwadi anobatsira nekuti anopa mhinduro yechinangwa, asi akawanda-chaiyo-yepasirese mabasa ekufunga anoshaya otomatiki verifier kana kuti ane maitiro asina kujeka ebudiriro. Izvo zvinoramba zvisingazivikanwe kana LURE inogona kugadzira anobatsira kuomerwa kosi mune izvo zvigadziriso, kuti kutariswa kwevanhu kwaizoramba kuchidikanwa here, uye kana muvharidzi angadzidzira kushandisa kusasimba mune inosimbisa pane kuburitsa matambudziko akakosha. Iyo abstract zvakare inosiya isingazivikanwe iyo nzira yekudzidzira mutengo, latency, kukundikana modhi, uye kuita pamabasa akareba kana mashoma akarongwa.

Related guides & Quizzes

AI Models InotsanangurwaKudzidziswa kweAITransformersAI AgentsEdza zvaunoziva - edza yemahara AI quizTarisa kumusoro izwi reAI mune yedu glossaryTevedza iyo AI modhi yekuburitsa tracker
Wakawana izvi zvinobatsira?