Komawa Labarai
Bidi'aAI Understanding takaitaccen bayani

Preprint yana ba da shawarar ƙididdiga na wahala mai ci gaba don rage ƙimar dacewar kai LLM

Wani sabon bugu na arXiv yana ba da shawarar daidaiton kai mai sassauƙa, hanyar da ke ƙididdige yadda rashin tabbas na ƙirar harshe game da tambaya da daidaita adadin hanyoyin tunani. Marubutan sun ba da rahoton tanadin alamar har zuwa 76% yayin da suke kiyaye daidaito kwatankwacin daidaitattun daidaiton kai a cikin daban-daban…

5 min readRead the primary source
Primary-source image accompanying Preprint proposes continuous difficulty estimates to reduce LLM self-consistency costs
Takardun tushe na farkoAn rubuta tushen tushe
Mawallafi
arxiv.org
Tushen hanyar haɗin gwiwa
arxiv.orghttps://arxiv.org/abs/2608.24590
Nau'in tushe
Takardun farko - sanarwar hukuma, takarda, yin rajista, ko shafi na farko da muka karanta kai tsaye.
MaganaFahimtar wannan a cikin daƙiƙa 60

Fara a nan

Mabuɗin sharuddan

Babban Samfurin Harshe (LLM)
Samfurin harshe da aka horar akan babban haɗin gwiwar rubutu don samarwa da tantance rubutu.
Rabewa
Aiki inda samfurin ke sanya shigarwa zuwa ɗaya ko fiye da ƙayyadaddun ƙayyadaddun bayanai.
Alamar alama
Daidaitaccen gwaji ko saitin bayanai da aka yi amfani da shi don aunawa da kwatanta aikin ƙira.
Gwada kankaAI Model An Bayyana Tambayoyi

Me ya faru

Masu bincike suna ba da shawarar daidaita kai mai sauƙi, ko FSC, hanyar yanke hukunci don manyan nau'ikan harshe waɗanda ke amfani da ci gaba da ƙididdige wahalar matsala don yanke shawarar hanyoyin tunani nawa don samarwa. Takardar ta ce FSC na iya ajiyewa har zuwa kashi 76% na alamun yayin da ake kiyaye daidaito kwatankwacin daidaiton kai.

Takardar arXiv, da aka ƙaddamar a kan Agusta 25, 2026, ta yi bayani game da amfani da alamar dacewar kai, dabarun yanke hukunci don manyan nau'ikan harshe. A cikin tsarin da marubutan suka bayyana, samfurin yana haifar da hanyoyi daban-daban na tunani don tambaya ɗaya kuma ya zaɓi amsar da ta bayyana mafi daidaito a cikin waɗannan hanyoyin. Takardar ta gabatar da wannan a matsayin mai amfani ga rikitattun matsalolin tunani, amma ta gano maimaita tsararraki a matsayin babban iyakancewar lissafi. Wannan tsararru yana kiyaye takardar mayar da hankali kan ƙididdige lokaci-lokaci da kuma adadin hanyoyin da aka samar da ake amfani da su don kowace tambaya.

Marubutan sun ce hanyoyin da suka dace a baya sukan ware albarkatu ta amfani da ƴan ƙaramin ƙayyadaddun rukunan wahala. Babban gardamarsu ita ce, wahalar tunani ta bambanta ci gaba, don haka rabe-rabe na binary ko mai ƙima na iya ware hanyoyi da yawa ga wasu tambayoyi kuma kaɗan ga wasu. An ƙera FSC don amsa wannan bambance-bambancen ta hanyar canza kasafin kuɗin samfurinta cikin sassauƙa daga shigarwa ɗaya zuwa na gaba.

FSC tana amfani da binciken da aka riga aka horar don yin hasashen abin da ake fitarwa na tambayar shigarwa. Takardar ta yi la'akari da cewa entropy ya annabta a matsayin mai nuna rashin tabbas na ƙirar kuma yana amfani da shi don ƙayyade yawan hanyoyin tunani don samfurin. A zahiri, hanyar da aka gabatar tana da niyyar kashe ƙarin ƙoƙarin yin samfur akan tambayoyin inda ƙirar ta bayyana ƙarancin tabbas da ƙarancin ƙoƙarin inda rashin tabbas ɗin sa ya ragu.

Dangane da taƙaitaccen bayanin, gwaje-gwaje a cikin samfura daban-daban da maƙasudai sun gano cewa FSC tana kiyaye daidaito daidai da daidaitattun daidaito yayin da ake samun tanadin alamar har zuwa 76%. Madogararsa baya bayar da sunayen samfura ko maƙasudai, cikakkun alkaluman daidaito, rarraba tanadi, ko yanayin da aka kiyaye matsakaicin tanadi. Waɗannan cikakkun bayanai ne na kayan aiki don tantance sakamakon.

Bayanan tushe: arxiv.org ↗

Me ya sa yake da mahimmanci

Daidaitaccen daidaito na iya inganta amsoshi ta hanyar yin la'akari da hanyoyi masu yawa, amma yana amfani da ƙarin alamu. Idan sakamakon da aka ruwaito ya zama cikakke fiye da gwaje-gwajen takarda, ware ƙarin ƙididdiga kawai lokacin da samfurin ya bayyana rashin tabbas zai iya sa wannan dabarar tunani ta fi dacewa.

Takardar ta yi niyya ga takamaiman rashin aiki a hanyar da aka tattauna sosai don inganta tunanin-samfurin harshe: samar da hanyoyi da dama don kowace matsala ba tare da la’akari da wahalar shigar da bayanai ba. Hanyar da za ta iya bambanta tsakanin tambayoyin da ke buƙatar samfuri mai yawa da tambayoyin da ke buƙatar ƙasa na iya rage tsararraki mara amfani tare da kiyaye fa'idar kwatanta hanyoyi da yawa.

Sakamakon da aka ruwaito yana da yuwuwar yin tasiri saboda ya shafi adadin da aka samar da rubutu da ake buƙata don dabarun tunani, maimakon ƙaramin canji zuwa makin ƙima. Ƙididdiga na da'awar tanadi na har zuwa 76%, kodayake wannan adadi wani sakamako ne na gwaji da marubuci ya bayar da rahotonsa daga abin da ya dace kuma bai kamata a kula da shi azaman garantin aiki mai zaman kansa ba.

FSC kuma tana kwatanta zaɓin ƙira mai faɗi a cikin tsarin ƙirar harshe: ko don amfani da kasafin ƙima iri ɗaya don kowace buƙatu ko daidaita ƙididdiga zuwa ƙiyasin rashin tabbas. Idan hanyar tana aiki da dogaro, rarraba daidaitacce zai iya taimakawa masu ƙirƙira tsarin daidaita ingancin amsa da ingancin lissafi. Madogararsa baya kafa tasiri akan jinkiri, tsadar kuɗi, amfani da makamashi, ko wadatar mai amfani, don haka waɗannan abubuwan suna kasancewa cikin sharadi maimakon nunawa.

Shawarar na iya zama mai dacewa musamman inda hanyoyin tunani ke da tsawo ko kuma inda aka riga aka samar da samfura da yawa. A lokaci guda kuma, ƙiyasin rashin tabbas zai iya zama ajizi. Samfurin na iya kasancewa da tabbaci akan tambaya mai wahala ko rashin tabbas akan mai sauƙi, kuma ƙayyadaddun bayanan baya nuna yadda FSC ke aiki a waɗannan lokuta. Da'awar kwatankwacin daidaiton da'awar saboda haka yana buƙatar fassarar cikin takamaiman gwaje-gwajen da aka ruwaito a cikin cikakkiyar takarda.

Interactive Mechanism

Ingantacciyar hanyar sadarwa: Yadda A zahiri yake Aiki

Bincika fasahar da ke bayan wannan ci gaban ta hanyar mu'amala.

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.
Duba ra'ayi na hulɗa+10 Points
AI Models Explained Quiz

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

Abin kallo na gaba

Madogaran sabon bugu ne da aka ƙaddamar, kuma taƙaitaccen bayaninsa baya gano samfuran da aka gwada, maƙasudi, tsarin horar da bincike, ko cikakken sakamakon kwatanta. Za a buƙaci kwafi mai zaman kanta da cikakken bincike na takarda don tantance yadda aka riƙa riƙon tanadin da aka ruwaito da kuma ko daidaito ya tsaya tsayin daka a cikin ayyuka.

fifiko na farko shine bincika cikakken saitin gwaji. Ƙididdigar ƙididdiga ba ta gano ƙirar harshe, nau'ikan ɗawainiya, alamomi, saitunan samfuri, ko saiti na asali da aka yi amfani da su a cikin kwatancen daidaitattun daidaiton kai. Idan ba tare da waɗannan cikakkun bayanai ba, ba zai yiwu a fayyace ko sakamakon ya shafi gabaɗaya ko ya mai da hankali a cikin takamaiman ayyukan tunani ba.

Binciken da aka riga aka horar wani muhimmin abin da ba a sani ba ne. Majiyar ta ce tana yin hasashen fitarwar fitarwa, amma bai bayyana yadda aka horar da shi ba, menene bayanan da aka yi amfani da shi, nawa ƙarin ƙididdigewa yake buƙata, ko kuma ya haɗa da ƙira da ayyuka a waje da yanayin horo. Waɗancan abubuwan za su ƙayyade ko tanadin alamar daga samfurin daidaitawa ya zarce farashin samar da kimanta wahala.

Ya kamata kwafi mai zaman kanta ya gwada fiye da mafi kyawun rahoton ceto. Sakamakon biyo baya mai fa'ida zai haɗa da matsakaicin ajiyar alama, daidaito a ƙayyadaddun kasafin kuɗi, aiki akan tambayoyi tare da siginonin rashin tabbas na yaudara, da sakamako a cikin iyalai daban-daban na ƙira da nau'ikan ma'auni. Ma'anar kalmar "har zuwa 76%" tana kwatanta iyakar, ba lallai ba ne sakamakon da aka saba.

A ƙarshe, masu karatu ya kamata su kalli shaida daga turawa ko ƙarin kimantawa. Madogarar ta kafa sabon tsarin bincike da aka ƙaddamar kuma tana ba da rahoton sakamakon gwaji, amma ba ta kafa samuwan samarwa, ɗauka ta mai bada samfuri, ko haɓakawa a cikin aikace-aikacen ainihin duniya ba. Har sai an amsa waɗannan tambayoyin, FSC an fi fahimtar FSC azaman dabara mai yuwuwar ingantacciyar fa'ida wacce ta rage ga tabbatarwa.

Jagorori masu alaƙa & tambayoyin tambayoyi

AI Model ya bayyanaAI horoMasu canjiGwada abin da kuka sani - gwada gwajin AI kyautaNemo kalmar AI a cikin ƙamus ɗin muBi samfurin AI na sakin tracker
An sami wannan yana da amfani?