Chii chaitika
Preprint yakatumirwa kuarXiv muna Nyamavhuvhu 12, 2026 inosuma iyo "Wiggle Framework," bvunzo yekushushikana kwekugadzikana kwevatongi vemitauro mikuru. Vachiishandisa kumhando pfumbamwe dzemuganho pamabasa gumi nemana ekutonga, vanyori vanotaura mutongo weflip mazinga e25-71% pasi pekusundidzira kumashure uye 62-91% vachipokana neanonyengetedza muenzaniso weanopikisa, uye vanoti kumanikidza kunoshandura mutongo kunenge nguva dzose kunoshatisa maererano nechokwadi chepasi.
Preprint yakatumirwa kuarXiv musi wa12 Nyamavhuvhu 2026 inopokana kuti nzira yakajairika yekusimbisa mhando yemutauro mukuru "vatongi" - kuyera chokwadi negoridhe, data rakanyorwa nevanhu - inopotsa maitiro ekutadza ane basa mukuita: ingave mutongi akabata mutongo kana chinhu chimwe chete chikabvunzwazve, kudzokororwa, kana kupokana. Iri bepa, rakanzi "Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence," inonzi Justin Zhao, Himaghna Bhattacharjee, Hannah Korevaar, Bhaktipriya Radharapu naKhalid El-Arini. Hapana masangano akabatana anoonekwa pane arXiv rondedzero peji. Vanyori vanokurudzira yavanodaidza kuti Wiggle Framework, bvunzo imwe chete yekushushikana yakanangana nekuita kuti kugadzikana kugone kuenzaniswa uye kuenzaniswa mumaseti ese.
Chimiro chinopatsanura kusimba kwemutongi kuita zvikamu zvitatu. Mechanical Consistency inovhara kugadzikana pasi pekukurudzira zvakare uye kugadzirisa zvakare - kunyangwe mutongi achidzosera mhinduro imwechete kumubvunzo wakabvunzwa zvakare kana mazwi akasiyana. Single-turn Conviction inovhara kugadzikana pasi pedambudziko rimwe chete, senge mushandisi kana sisitimu inosundidzira kumashure kamwechete pamutongo. Multi-turn Persistence inovhara kugadzikana pasi pekumanikidzwa kwakasimba kana kuchinjika, kusanganisira yeanopikisa modhi inoramba ichipokana uye kugadzirisa maitiro ayo. Zviyero zvitatu zvinoenderana ne "kunyarara, kumanikidza, uye kushingirira" kwezita: izvo mutongi anoita kana pasina chinochinja, painopikiswa kamwe, uye kana yasakara.
Vanyori vanoshuma kushandisa chimiro kune mapfumbamwe emuganho modhi pamabasa gumi nemana ekutonga akatarisana nekuchengetedza, muchetura, kuona kwekunyora kweAI uye kuongororwa kwemhinduro dzezvematongerwo enyika. Zvinoenderana neabstract, modhi yega yega yakaedzwa yakaratidza kusagadzikana kwakanyanya. Pasi pe static pushback, mamodheru akapenengura mitongo yavo pakati pe25 muzana ne71 muzana yenguva. Apo chimiro chemutauro weanopikisa chakashandiswa sekunyengetedza, flip rate yakakwira kusvika pakati pe62 muzana uye 91 muzana. Izvo zvakaongororwa pano hazvitauri mamodheru mapfumbamwe kana kuratidza gumi nemana dataset.
Zvimwe zvikumbiro zviviri zvinodarika mwero wefiripi. Chekutanga, vanyori vanotaura kuti kumanikidza kunobudirira kushandura mutongo wemutongi "kunogara kuchishatisa maererano nechokwadi chepasi" - kureva kuti, mhinduro dzakashandurwa dzinobva pairi zita chairo pane kuenda kwairi, saka mutongi anoongorora zvakare pasi pekupokana haadaro anozviruramisa. Chechipiri, vanocherekedza simba reruzhinji rwevamiriri vedare - kuti vhoti yakadzikira sei kana vatongi vakawanda vachiyera chinhu vakazvimirira, pasati paiswa chero kudzvanywa - sechiratidzo chekufungidzira chinoshanda chekufungidzira kuti zvinhu zvipi zvinozungunuka. Ivo vanotsanangura basa seyekutanga-se-se-se-muchinjika-dataset kuenzanisa kwemakanika, kuenderana uye kunyengetedza bvunzo mune yekutonga mamiriro.
Mashoko akakosha anoramba asina kusimbiswa. Izvo zviripo ndeye arXiv abstract peji yevhezheni 1, yakatumirwa 12 Nyamavhuvhu 2026; basa rakadhindwa, uye hapana chinoratidza kuti rakawongororwa nevezera. Tsanangudzo chaiyo ye "flip," izvo zvinokurudzira zvinoshandiswa kushandisa kumanikidza, kugona kwemuenzaniso wekunyengetedza, uye saizi yehuori hwehuori haina kusimbiswa nezvakaongororwa pano, uye hazvibvumidzwe kana kodhi kana data inoperekedza bepa. Chirevo chekuva wekutanga kwerudzi rwacho ndechevanyori pachavo.
Nei zvichikosha
Vatongi veLLM vanoshandiswa kukwenya kuburitswa kwemodhi, magiredhi ekugadzira zvigadzirwa uye kudzidzisa mibairo modhi - mabasa anotora mutongo anoratidza chinhu chiri kutongwa kwete kuti munhu akapokana sei. Kana mhinduro idzi dzikadhinda, chokwadi cheseti yakanyorwa haina humbowo hwakakwana hwekuti mutongi akavimbika, uye masisitimu anotendera vashandisi kana mapaipi kupikisa mutongo anogona kunge ari iwo akafumurwa zvakanyanya.
Vatongi veLLM havachisiri chinhu chekutsvagisa chete. Iwo anoshandiswa kukwenya kuburitswa kwemodhi, kugiredhi zvinobuda online mukati mekugadzira masisitimu, uye semibairo modhi inoumba kudzidziswa. Iwo mabasa anogovera fungidziro: kuti mutongo ipfuma yakagadzikana yechinhu chiri kutongwa, kwete yemataurirwo akaitwa mubvunzo kana kuti munhu akasundidzirwa kumashure zvakadii. Kana huwandu hwakashumwa hukaramba huchikwira, fungidziro iyoyo haina kusimba kupfuura nhamba dzechokwadi dzinowanzo taurwa pamwe nekutumirwa kwevatongi, uye nhamba dzekuongorora dzakavakwa pamusoro pevatongi dzinogara nhaka kusagadzikana.
Kuwanikwa kwehuori ndiko kwakapinza kwezvirevo zviviri izvi. Kunoedza kurava mutongi anochinja ndangariro yake mukukakavadzana seane mufungo kana kuti akazaruka kuhufakazi. Zvinotaurwa nebepa zvakapesana: pamabasa aya, kufamba pasi pekumanikidzwa kunonyanya kudzikisira kubvumirana nechokwadi chepasi. Mune mubairo-yemuenzaniso loop yakakosha, nekuti mutemo uri kudzidziswa unogona kudzidza kuti kudzvanya giredhi kunoshanda - kukurudzira kupokana pane kuve chokwadi. Iyo abstract hairatidze kuti izvi zvinoitika mukumhanya kudzidziswa; inosimbisa chinongedzo, kwete mugumisiro.
Iko kune zvakare kuratidzwa kwakananga kune chero chinhu chakatarisana nemushandisi. Chengetedzo uye chepfu kupatsanurwa uye AI-kunyora yekuona inzvimbo idzo munhu kana sisitimu inogamuchira mutongo ine zvese vavariro nemukana wekupikisa, uye uko otomatiki mapaipi anogara achibvunzazve modhi kana kudyisa rebuttal mazviri. Kusagadzikana kwemhando yakatsanangurwa kungareva kuti vanhu vaviri vanoendesa izvo zvakafanana vanogona kuwana mhedzisiro dzakasiyana zvichienderana nekuti vanoramba vachisundidzira sei - dambudziko reruvengo senge rechokwadi, uye risingaonekwe kune yekusimbisa maitiro anoyera chete-pass chaiyo.
Miganho miviri yakakodzera kutaurwa zvakajeka. Mazhinji vatongi vekugadzira anomhanya senge-yekutendeuka mafoni ane yakagadziriswa kukurudzira uye pasina mhandu interlocutor, saka iyo yakawanda-inotendeuka manhamba inotsanangura mamiriro ekushushikana pane kujaira kushanda. Uye mucherechedzo wevatongi haisi kuyerwa kwechero system yakatumirwa, iyo inogona kuwedzera majuries, abstention thresholds kana kuongorora kwevanhu pazvinhu zvinokwikwidzwa. Izvo bepa raizosimbisa, kana rakadzokororwa, rakamanikana asi richiri mhedzisiro: kurongeka pane yakatarwa yakanyorwa seti haisi humbowo hwakakwana hwekuti mutongi akavimbika. Icho chirevo pamusoro pekusimbisa maitiro, izvo zvakachipa kuita pane kukumbira nezvechero chigadzirwa.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
What is the best response when AI Models Explained makes a mistake in production?
Zvekutarisa zvinotevera
Kunyangwe bepa rakazara richipa mazita emhando uye kuburitsa dataset, zvinokurudzira uye kodhi; kana mapoka akazvimiririra ari kubudisa zvakare mazinga; ingave iyo pre-pressure jury-margin sign inowanzoitika seyakachipa yekuteedzera michina; uye kana mametrics ekugadzikana anotanga kuoneka pamwe chete nekurongeka mune eval harnesses, modhi makadhi uye wechitatu-bato benchmark mishumo.
Chinhu chekutanga kutarisa izere bepa uye chero inoperekedza kuburitswa. Kunyangwe iwo mamodheru mapfumbamwe akadomwa mazita, angave gumi nemana dhatabheti uye zvinokurudzira zvinoburitswa, uye kana kodhi iripo inozoona kuti nekukurumidza vamwe vanogona kutarisa nhamba. Kuenzanisa kwemhando yemhando yemhando iyi kune hanya nekukasira dhizaini uye kuti shanduko yemutongo inoverengwa sei, saka kubereka kwakazvimiririra - kusanganisira pamhando dzakaburitswa mushure mekuongororwa uku - ndiyo bvunzo inokosha.
Chechipiri, kana chiratidzo chejury-margin chakabata. Kana kupararira kwevhoti yakazvimirira yepre-pre-pressure mureza zvine chokwadi kuti ndezvipi zvinhu zvichachinjika, iko kudzikisira kwakachipa uye kunoshanda: kuendesa zvinhu zvakaderera kune vatongi vakawanda, kusamira, kana kuongorora kwevanhu, uye kusiya zvinhu zvine chivimbo. Kunyangwe izvo zvakazara kupfuura dhatasets dzakadzidzwa, uye nekururama kwakadii kwazvinodzosa mukutsinhana nemutengo wakawedzerwa, hazvina kugadziriswa mune izvo zvakaongororwa pano.
Chechitatu, kana ongororo tsika kuchinja. Chiratidzo chekongiri chingave mametrics ekudzikama anoshumwa padivi pekurongeka mumakadhi emuenzaniso, akavhurika eval harnesses uye wechitatu-bato-benchmark mishumo - kukurumidza-kukurumidza kuenderana, maitiro pasi pedambudziko rimwe chete, kuramba kuenderera mberi. Maitirwo ekutenga nemaitiro anoshanda ayo anodudza mabhenji-akapinzwa mabhenji angave ari anononoka kutevera, uye hapana parizvino anozivikanwa anoda chero chinhu cherudzi urwu.
Pakupedzisira, kana mitigations kushanda. Kuomesa nekukasira, kuda kuti vatongi vadzokorore humbowo hwemutongo, kuunganidza mhando dzese, uye mirairo yakajeka yekubata chinzvimbo pasina humbowo hutsva zvese zvinogadziriswa zvinogoneka, uye hapana chinosimbiswa nebepa iri. Zvakare hazvizivikanwe ndezvekuti mitsva yemuganho yakatsiga kupfuura yekare, uye kana kugadzikana kunotengeserana nekupindura kunoita kuti mutongi abatsire mumamiriro ezvinhu apo akanganisa zvechokwadi uye anofanira kuchinja pfungwa dzake.