Chii chaitika
Iyo itsva arXiv preprint inopa Evidence-State Reliability (ESR), nhanho yekuongorora yeakawanda-nhanho-yemutauro-modhi yemapaipi. ESR inobvunza kana humbowo hwepakati hunoramba hwakakwana, hwakavakirwa, hunoenderana nemukati uye huchishandika kune chinhanho chinotevera, pane kungotarisa kana izvo zvinobuda zvichitevera fomati inotarisirwa.
Iro bepa rinoongorora mapaipi e-multi-stage LLM ari pasi pekudzora kuderedzwa. Inofananidza humbowo hwakachena nehutatu hwakashandurwa mamiriro: humbowo hwakamanikidzwa-kurasikirwa, chikamu chekudonha uye humbowo hunopesana. Iyo pombi inosanganisira sarudzo, yekuongorora uye nhanho dzekuwedzera, uye vanyori vanoongorora kurongeka kweparser kutendeseka zvakasiyana kubva kana humbowo hunoramba hwakakodzera basa rakapihwa nhanho.
Vanyori vanoshuma 720 yakarongwa uye yakadzoserwa mafoni, ayo makumi manomwe nenhatu mitsara yekuuraya yakachengetwa. Ongororo yakashandiswa GLM-5.2 uye makumi matanhatu egasi rehutsanana makesi. Pakati pezvipfumbamwe zvakaenzanirana kuenzanisa pakati pemamiriro akashatiswa uye akachena, yega yega nhanho-yekubudirira fungidziro yaive isina kunaka, uye yega yega 95% bootstrap nguva yakaramba iri pazasi zero, maererano neabstract.
Kutendeseka kweParser kwakaenda kwakapesana: fungidziro dzemapoinzi mapfumbamwe dzaive dzakanaka, kunyangwe nguva dzekuenzanisa kutatu-kudonha kwaisanganisira zero. Mune mamwe mazwi, zvinobuda mupombi zvinogona kuramba—kana kuita sezvingangoita kuramba—zvichienderana kunyangwe humbowo hwekubudirira huchiwedzera.
Bepa racho zvakare rinopatsanura kuziva humbowo hwakadzikiswa kubva pakuwana kubva kwairi. Pakati pe-parser-inoshanda yakaderedzwa yekuongorora yakabuda, kuderedzwa kwekuona kwakashumwa se1.0 mune yega yega yakasvibiswa mamiriro, asi manyepo-yekuvimbisa mazinga anga achiri asiri zero. Pakati pe-parser-chaiyo yakaderedzwa yekukwira kwezvakabuda, kudzoreredza kwaive 0.0 mumamiriro ese akashatiswa. Vanyori vanotsanangura izvi seyakasungirirwa kuvimbika-layer divergence muyakaongororwa gadziriso.
Zvakatorwa pamwechete, dhizaini inochengeta mibvunzo miviri yakasiyana panguva yese yekuongorora: kana chinobuda chinogona kugamuchirwa muchimiro chinotarisirwa chechimiro, uye kana humbowo hunotsigira icho chabuda huchiri kushandiswa padanho rakapihwa. Mienzaniso yakashumwa inobata mutsauko iwoyo pasi pemamiriro akataurwa. Ivo saka vanotsanangura maitiro akaita matanho muongororo iyi, vachisiya nzvimbo yakakura yekusiyana yakavhurika kune kumwe kuyedzwa.
Nei zvichikosha
Chidzidzo ichi chinosimbisa maitiro ekutadza kunoshanda kweAI masisitimu anodarika macheki ezvimiro achivimba nehumbowo husina kukwana, hwakamanikidzwa kana kupokana. Musiyano iwoyo une basa pese apo imwe modhi nhanho inopa ruzivo kune imwe, kusanganisira sarudzo, kuongorora uye kukwira kwekufamba kwemabasa.
Mazhinji maAI masisitimu anoshandisa fomati uye schema cheki seyekutanga kuchengetedza. Iwo macheki anogona kuona kuti chinobuda chinokwanisa kushandiswa-semuenzaniso, kuti chine minda inotarisirwa-pasina kuratidza kuti humbowo huripo hwakakwana, hunoenderana kana hwakakodzera sarudzo inotevera. ESR inotarisirwa kuyera iyo yechipiri pfuma.
Musiyano uyu unonyanya kukosha mumapaipi apo modhi yekutanga inopfupisa kana kurongedza ruzivo uye gare gare nhanho yekuongorora, sarudza kana kukwira zvichienderana nemhedzisiro yepakati. Kana humbowo hwakashatiswa hukashandurwa kuita chitarisiko chakachena, software yepasi inogona kuzvigamuchira vasingazive kuti ruzivo rwunodiwa pakuita basa racho rwakaneteswa.
Mhedzisiro yekukwira kwebepa inonyanya kukosha semuganho pane izvo zvinoonekwa zvinogona kuitwa. Iyo abstract mishumo yekuti iyo yakaongororwa nhanho yekukwira haina kupora mune chero mamiriro akashatiswa, zvisinei neparser-chaiyo mhedzisiro. Izvo hazviratidze kuti ese LLM escalation masisitimu anotadza, asi inoratidza nei kuona dambudziko uye kudzoreredza humbowo hwakavimbika zvakapatsanurwa engineering zvinodiwa.
Zvakawanikwa zvinogona kubatsira masangano kugadzira ongororo dzinoedza kupfuura mafomati anobuda. Sisitimu inogona kuda matanho akaparadzana ehumbowo huzere, kudzika, kuenderana kwemukati, budiriro yebasa uye maitiro ekudzoreredza, padivi pezvakajairwa schema kana parser cheki. Iro bepa haritauri kuti ESR ndeye general indasitiri mwero kana kuti inovandudza-chaiyo-yenyika mhedzisiro; inopa uye inoshandisa hurongwa mune imwe yakashumwa ongororo.
Mhedzisiro yakakura saka ndeye pamusoro peicho cheki yekuvimbika inofanirwa kukumbirwa kuyerwa. Kuenderana kwechimiro kunogona kuramba kuchibatsira sechivakwa cheinjiniya, asi haipindure mubvunzo wehumbowo wega. Hurongwa hwechidzidzo hunoisa zvivakwa izvi padivi peumwe kuitira kuti pombi iongororwe kune ese ari maviri fomu inoshandisika uye inoshandisika tsigiro, pasina kubata chero chiyero senhoroondo yakazara yekuvimbika.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
What is a common training objective for an autoregressive language model?
Zvekutarisa zvinotevera
Mhedzisiro yacho inoda kuyedzwa kupfuura iyo yepepa imwe modhi yekumisikidzwa, dhizaini yepombi uye sanitized makesi. Basa remangwana rinofanirwa kuongorora kana ESR uye kusiyana kwakashumwa kunobata mhando dzese, mabasa, mhando dzehumbowo uye yakakura kana yakazvimiririra yakadzokororwa ongororo.
Kudzokorora ndiwo mubvunzo wepakati wakavhurika. Vanyori vanodzika mhedziso yavo kune yakaongororwa modhi dhizaini, dhizaini yepombi, yakasarudzwa sanitized makesi, scoring maitiro uye single scaled run. Iyo abstract hairatidze kuti mibairo yaizochinja sei nemamwe mamodheru emitauro, dhataseti hombe, marekodhi ebhizinesi mhenyu kana mamwe mapaipi ekuvaka.
Iro bepa rinoti kodhi uye zvigadziriso zvekushandisa zviripo, asi sosi rakapihwa haritsanangure zvirimo, ruzivo rwekuita kana kuti vaongorori vakazvimiririra vakaburitsa mhedzisiro. Izvo zvinhu uye zvekunze zvinodzokororwa zvichave zvakakosha pakutarisa maitiro ekukoira, iyo bootstrap kuongororwa uye zvinorehwa nematanho ekubudirira akashumwa.
Ongororo dzenguva yemberi dzinofanira kuyedza kana mutsauko wakafanana uchionekwa munzvimbo dzinoshanda nekutadza kweuchapupu hwakasiyana. Ikozvino kunobva mazita kumanikidzana, kurasikirwa kwechidimbu uye kukakavara, asi haipe ruzivo rwakakwana muchidimbu kuti uone kuti ndedzipi mhando dzehumbowo dzainyanya kukuvadza, ingave yekudzikisira kuomarara kwakasiyana zvakarongeka kana kuti nyaya dzakasarudzwa sei.
Vaverengi vanofanirawo kuona kuti ESR inofananidzwa sei nekuvimbika kuripo, kuenzanisa, kudzoreredza-pasi uye kusavimbika matanho. Kwakabva kunotsigira mhedziso yakamanikana yekuti parser chokwadi uye humbowo-inotarisisa nhanho budiriro yakapatsanurwa muchiyedzo ichi. Iyo haitsigire zvikumbiro pamusoro pehuwandu hwekutadza kwakawanda, njodzi yekugadzira kana kuvimbika kwemapaipi eLLM kazhinji.
Miganhu chikamu chekududzirwa kwemhedzisiro. Zviyero zvakashumwa zvinoratidza zvakaitika mukati mezvakatsanangurwa zvigadziriso uye hazvigadzirise kana patani imwe chete ichizoenderera kune imwe nzvimbo. Zvishandiso zvekugadzira zvakare, kuongorora kwakazvimirira uye kuenzanisa nematanho ane hukama kunogona kujekesa kuti huremu hwakadii hwekuisa pachimiro uye pamusiyano wakaonekwa muchidzidzo chakapihwa.