Chii chaitika
Vatsvagiri vanomwe vakaisa preprint pa arXiv vachipokana kuti kusimbaradza kudzidza mushure mekudzidziswa kunodonhedza kusiyana kwemhinduro dzemuenzaniso wemutauro, uye kuti nzira dzekushanduka-shanduka - kugadzirisa zvakananga munzvimbo yehuremu kuburikidza nekukanganisika kwakangoerekana kwaitika - chengetedza kusiyana uye kusimudza pass@k. Iro bepa rakatumirwa musi wa13 Nyamavhuvhu 2026 uye harina kuongororwa nevezera. Chinyorwa chinoonekwa neruzhinji hachina mazita ekuenzanisa, saizi dzemhando, kana nhamba dzakayerwa.
Preprint yakanzi "Beyond the Best Guess: Kuvandudza LLM Solution Coverage neEvolution Strategies" yakaendeswa kuarXiv musi wa13 Nyamavhuvhu 2026 uye yakanyorwa pasi pehungwaru hwekugadzira (cs.AI) ine yechipiri runyorwa muneural uye evolutionary computing (cs.NE). Inotendwa kune vanyori vanomwe - Conor F. Hayes, Elliot Meyerson, Kajetan Schweighofer, Roberto Dailey, Babak Hodjat, Risto Miikkulainen naXin Qiu. Iyo arXiv rekodhi haina kunyora masangano, uye chirevo ichi hachitauri basa kune chero sangano.
Kuumbwa kwebepa kunotanga kubva pamashandisirwo emamodheru emitauro anowanzo shandiswa muzvirongwa zvekuwanikwa zvakaita semasvomhu nesainzi: dambudziko rinounzwa, uye mhinduro imwe yemuenzaniso inotorwa semhinduro yakarongwa. Vanyori vanopokana kuti iyi "yakanyanya kufungidzira" modhi inosiya kukosha patafura, nekuti yekuwedzera-yenguva yekuyedza komputa inogona kushandiswa kugadzira akawanda evamiriri mhinduro pane imwe. Hutongi ihwohwo hunowanzoyerwa nepass@k, metric inoverenga dambudziko sekugadziriswa kana imwe chete ye k sampled yekuedza ndeyechokwadi.
Chirevo chepakati ndechekuongororwa kwemugumisiro wepadivi mukuita kwazvino. Mushure mekudzidzisa modhi nekudzidza kwekusimbisa, vanyori vanonyora, inopfupisa iyo modhi inobuda kugovera kutenderedza-mubairo wepamusoro-uye kuti kupfupika kunoita kuti kuvharwa kwemhinduro kudonha. Mune mamwe mazwi, RL inogona kuita kuti mhinduro yekutanga ive nani ichiita seti yemhinduro dzakasiyana diki, izvo zvinonyatso ranga pass@k hurumende ine vanyori vane hanya nayo.
Seimwe nzira, bepa rinokurudzira nzira dzekushanduka-shanduka: nzira yehuwandu-yakavakirwa, gradient-isina-yemashure-yekudzidzira iyo inokwenenzvera yakananga munzvimbo inorema nekushandisa zvisina tsarukano kuvhiringidza kuenzanisira paramita uye kusarudza pane zvabuda, pane kudzosera chiratidzo chemubairo. Iyo abstract inotaura kuti ES inowana zvinogara yakakwira pass@k pane RL, inogadzira yakafararira yekugovanisa ine yakakura mhinduro kufukidzwa, uye kuti kufukidzwa uku kunoshandura kuita zvirinani mhedzisiro yemabhenji akajairwa.
Izvo zvisingapiwi neabstract ndihwo humbowo hwakawanda. Iyo haina kudoma mhuri dzemhando kana maverengero eparamendi, hapana chaiwo mabhenji, hapana makoshero e k, hapana baseline RL algorithm, uye hapana manhamba mhinduro - mazwi "anogara akakwira" uye "zvirinani mhedzisiro" ndiwo ega maratidziro anopihwa. Iko kutumirwa ndeye 449 KB uye zvinyorwa zvizere zviripo sePDF uye yekuedza HTML, saka izvo zvinyorwa zvinogona kunge zviri mubepa; havapo kubva pane zvinyorwa zvakaongororwa pano. Iyo vhezheni-imwe preprint isina chiratidzo chekuongorora kwevezera, hapana inooneka kodhi kana data link, uye hapana yakazvimirira kudzokorora.
Nei zvichikosha
Munzvimbo umo mhinduro yemumiriri inogona kutariswa - humbowo, kodhi, fungidziro dzesainzi - sirin'i inobatsira haisi yekuti fungidziro yekutanga yemuenzaniso ndeyechokwadi asi kuti mhinduro chaiyo inoonekwa chero kupi mubatch yekuedza. Kana RL post-training ikadzikama dziva iro, iyo indasitiri inotungamira yekumisikidza inogona kunge iri kutengesa kure neiyo chaiyo pfuma inotsamira pakutsvaga nekutsvaga kweajenti.
Musiyano unodhonzwa nebepa - pakati pemhinduro imwe chete yakanakisa uye hupamhi hwezvainogona kuburitsa mumasampure mazhinji - haisi yedzidzo. Chikamu chiri kukura cheyakanyanya-kukosha AI basa rinomhanya mukugadzira-ipapo-kusimbisa loop: kurudzira zvirongwa zvakawanda zvevamiriri uye mhanyisa bvunzo sutu, taura matanho mazhinji ehumbowo uye utarise iwo nemakanika, kurudzira fungidziro dzakawanda uye udzionere. Mune ese aya marongero, muongorori anosarudza kuti ndeupi mumiriri ari mugwara, saka chinosunga ndechekuti mumiriri akakodzera akambogadzirwa here. Kuvhara ndiyo sirin'i, uye iko kurongeka pakuedza kwekutanga ingori imwe chete pasi payo.
Izvo zvinoita kuti kuongororwa kuve kwakakoshera kune yakakura post-kudzidziswa resipi. Kusimbisa kudzidza kubva kune mibairo masaini magadzirirwo mazhinji emazuva ano emuganho modhi mushure mekudzidzira, uye kurodza kugovera kwekubuda kuri padyo nenzvimbo yechiitwa. Kana kurodza uku kuchidhura kuvharika, saka pombi yakajairwa inogona kunge ichinatsiridza zvibodzwa zvakayerwa pa k=1 uku ichidzikisira kuita kwechinyararire muhurumende umo modhi dziri kuramba dzichiiswa. Kunetseka kwekuti RL inoderedza kusiyana kwemuenzaniso kwakasimudzwa kare mumabhuku ekutsvagisa; mupiro unotaurwa pano ndeimwe nzira yechokwadi kwete kuongororwa kutsva.
Mushonga wakarongwa unotakura kutengeserana kwawo kunozivikanwa. Mazano ekushanduka-shanduka anodzivirira magradients zvachose, izvo zvinodzivirira kumwe kusagadzikana kweRL kunyatso-tuning, asi ivo vagara vachibhadhara izvo nemuenzaniso wakanaka: kufungidzira inobatsira yekuvandudza nzira kubva kune zvakangoitika uremu zvinokanganisa zvinowanzoda akawanda kumberi kupfuura nepakati pehuwandu, iyo inofananidzwa zvakanaka asi inodya compute. Kana masvomhu iwaya anoshanda pamwero wemienzaniso yemitauro yemazuva ano ndiwo mubvunzo ungada kupindurwa nemuverengi, uye chidimbu hachitaridzi mutengo zvachose.
Kune veruzhinji uye kune varapi, hapana chinochinja nhasi. Ichi chirevo chekutsvagisa nezve nzira yekudzidziswa, kwete chigadzirwa, kuburitswa kwemuenzaniso, kana kuchengetedza kutsvaga. Kukosha kwayo kwaizosvika zvisina kunanga - kuburikidza nemamodheru akarongedzerwa kuongorora kwete kuita, kana kuburikidza nevanopa vachifumura mapfundo anotengeserana chokwadi chemhinduro yekutanga nehupamhi. Iko hakuna humbowo mune sosi yekuti chero system yakatumirwa inoshandisa iyi nzira, uye hapana mutengesi akatsanangurwa achiitora.
Izvo zvakakoshawo kutaura miganhu yezvinogona kugumiswa kubva kune preprint abstract. Chirevo chekuti ES inorova RL pa pass@k ishumo yevanyori yezviyedzo zvavo, kwete chokwadi chakazvimiririra. Kuenzanisa pakati pemaitiro ekugadzirisa kunonyanya kutariswa nekugadzirisa kushanda nesimba, komputa bhajeti, uye sarudzo yekutanga, uye mhedzisiro inomiririra imwe modhi sikero kana imwe bhenji mhuri kazhinji haina kuzara.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which question best defines a clear goal for using AI Training?
Zvekutarisa zvinotevera
Iwo anotakura zvinhu ari muPDF yakazara uye kwete abstract: ndeapi mamodheru akadzidziswa, ndeapi mabhenji akashandiswa, ndeapi maitiro e k, uye zvakanyanya kana ES neRL zvakaenzaniswa pamakomputa akaenzana. Zvakare zvakakosha kutariswa ndezvekuti kodhi yakaburitswa here, kana mhedzisiro yacho ichiburitsa kunze kwemasvomhu, uye kana chero muganhu werabhu unotora nzira.
Chinhu chekutanga chekutarisa ndeye yakazara bepa kuyedza kuseta, uye kunyanya kana iyo ES neRL inomhanya yaive compute-inofananidzwa. Nzira yemahara-yemahara iyo inoshandisa zvakanyanya kudzidziswa komputa pane yayo RL yekutanga inogona kutaridzika zvirinani nekuda kwezvikonzero zvisingaenderane nemagadzirirwo anopiwa nevanyori. Vaverengi vanofanirwa kutsvaga saizi yemhando yakadzidziswa, iyo RL algorithm inoshandiswa sekuenzanisa, huwandu hwekuvhiringidza pachizvarwa, uye kukosha kwe k uko kwaiyerwa pass@k - kuvhara macurves anowanzo kuyambuka, uye nzira inokunda pahombe k inogona kurasikirwa nek = 1.
Chechipiri, tarisa nzvimbo. Iyo abstract's kongiri yekuda nezve yakadzika mabhenji inogumira ku "standard math benchmarks," domain ine yakachipa otomatiki verification uye inorema yekutanga optimization. Kunyangwe mukana wekuvharidzira uchiendeswa kukugadzira kodhi, kugadzirwa kwesainzi, kana mabasa akazaruka eajenti - uko kuonesa kuri kuita ruzha uye iko kurongeka kusiri kwebhinari - hakuna kusimbiswa nezvinhu zviripo.
Chechitatu, tarisa zvigadzirwa uye kudzokorora. Iyo arXiv rondedzero inoratidza hapana yakabatana kodhi kana kuburitswa kwedata. Kuitwa kwakaburitswa, kana kudhirowa kweboka risingadyidzane nevanyori, kunobvisa izvi kubva pakudaidzira kune mhedzisiro. Kusipo izvozvo, chimiro chakakodzera kufarira pane kuvimba, uye chero chiratidzo chekugamuchira chinofanira kuuya nenhamba dzayo.
Chechina, mubvunzo weverifier unosungira kuti imarii chero yekuvhara yakakosha. Yepamusoro pass@k inongoshandura kuita inobuda inobatsira kana chimwe chinhu chichigona kuratidza mumiriri akakodzera pakati pevazhinji. Mumasvomhu uye software, macheki aripo; mumashandisirwo mazhinji ezvokutengeserana havaite, uye kusarudza mhinduro chaiyo kubva padziva rakakura rinova dambudziko rayo risati ragadziriswa. Kutevera basa rekubatanidza ES-akadzidziswa mamodheru ane nzira dzekusarudza angave echisikigo nhanho inotevera.
Pakupedzisira, tarisa chinyorwa chinobudiswa. Iyo bepa ndeye v1 preprint isina nzvimbo yakataurwa kana chimiro chekuongorora. Kudzokororwa, kutumira kwemusangano, kana kutsoropodzwa neruzhinji kubva kune vamwe vaongorori vanoshanda paRL post-training siyana zvese zvinorodza kuti zvakakomba sei chirevo chepakati chinofanira kutorwa.