AI TALATIN, KUSKURE ƊAYA
Hikimar taron jama’a (wisdom of crowds), tunanin da takardar ta danganta da Galton a 1907, tana cewa: a ɗauki matsakaicin ƙiyasai masu yawa masu zaman kansu, kurakuran za su share juna. Idan kurakuran ba su da karkata (unbiased) kuma masu zaman kansu ne, kuskuren matsakaicin yana raguwa kamar ɗaya raba da tushen murabba’i (square root) na adadin ƙiyasan. Tsare-tsare da yawa da ke haɗa samfuran harshe na AI (AI language models) da dama — ta hanyar ɗaukar samfur da jefa ƙuri’a (sampling and voting), ta muhawara, ko ta kwamitocin samfura daban-daban — sun dogara da wannan hujja: babbar ƙungiya za ta iya maye gurbin wakili (agent) mafi ƙwarewa.
Carolina Fortuna da Blaž Bertalanič, a Cibiyar Jožef Stefan (Jožef Stefan Institute) a Ljubljana, sun nuna cewa ƙarfin hujjar bai wuce na ginshiƙanta (premises) ba, kuma sun tashi don gwada su.
Tambayoyin Fermi a matsayin wurin gwaji
Tambayoyin Fermi suna neman adadin da sai dai a yi ƙiyasinsa, kamar abubuwa nawa ne za su iya shiga cikin wani akwati. Amsoshinsu adadi ne da suka mamaye matakan girma masu yawa, don haka ana iya auna kuskure a cikin matakan girma (orders of magnitude) — ninki goma yana ƙidayuwa a matsayin ɗaya — kuma kurakurai masu alama (sign) saɓani za su iya share juna da gaske.
Masu binciken sun tattara amsoshi miliyan 2.51 daga samfura 15 masu buɗaɗɗun nauyi (open-weight models), daga parameters biliyan 3 zuwa 72, kan tambayoyi 529 na ma’aunin gwajin realFP (realFP benchmark), sannan suka tabbatar da sakamakon a kan wasu rukunan tambayoyi biyu. Kowane samfuri ya amsa kowace tambaya kusan sau 350, kuma an ɗauki matsakaicin ƙungiyoyin amsoshi 2 zuwa 30 a kan ma’aunin logarithm (logarithmic scale).
Su kaɗai, samfuran ba su da ƙwarewa a wannan wasa. Amsa ɗaya tana samun daidai matakin girma a kan 12 zuwa 34% na tambayoyin, kuma tana kuskure da fiye da ninki ɗari a kan 30 zuwa 63%.
Kuskuren da kowane kwafi ke rabawa
Muhimmin mataki shi ne raba kuskuren kowace amsa zuwa sassa biyu. Ɗaya shi ne kuskuren da samfuri ke yi a kan wata tambaya ta musamman, wanda dukkan amsoshinsa gare ta suke rabawa. Ɗayan kuma shi ne sashen bazata (random) da ke sauyawa daga wata amsa zuwa ta gaba. Na biyun ne kaɗai zai iya sharewa idan aka ɗauki matsakaicin amsoshi.
Ashe sashe na farko ne ya fi rinjaye. A kan realFP, shi ke ɗaukar tsaka-tsakin (median) 0.906 na bambancin kuskure (error variance). Idan aka san tambayar, babu daidaito (correlation) tsakanin amsoshi biyu daga ƙungiya ɗaya, kamar yadda ɗaukar samfur mai zaman kansa ke nufi; amma kuskurensu ɗaya ne.
Takardar ta mayar da wannan zuwa adadi ɗaya, da wata sananniyar dabara da ake kira tasirin tsarin Kish (Kish’s design effect): ainihin adadin amsoshi masu zaman kansu a cikin ƙungiyar N shi ne N / (1 + (N − 1)ρ), inda ρ shi ne wannan kaso da ake rabawa. Da ρ = 0.906, ƙungiyar talatin tana daidai da wakilai (agents) 1.10 masu zaman kansu.

Hagu: kuskuren ƙungiya a kan girman ƙungiya N ga samfura 15. Dama: abu ɗaya idan aka kwatanta da wakili guda (layuka masu launi, kusa da 1), a kan abin da amsoshi masu zaman kansu marasa karkata za su bayar (baƙaƙen layuka masu karya-karya da masu karya-da-ɗigo) da kuma ƙimar 1/√N (layin ɗigo-ɗigo). — Hoto na 2, Fortuna & Bertalanič (2026), arXiv:2611.16315.
A aikace, ƙungiyoyin talatin sun rage kuskuren wakili guda da tsaka-tsakin 1.7%, kuma 5.4% a mafi yawa. Amsoshi masu zaman kansu waɗanda su ma ba su ɗauke da wani kuskure da ake rabawa a kan kowace tambaya da sun rage shi da 74 zuwa 79%.
Komai yana karkata zuwa miliyan ɗaya
Wani sashe na kuskuren da ake rabawa yana da siffar da ake iya gane ta. Kowane samfuri yana ƙiyasta ƙananan adadi fiye da ainihinsu, manya kuma ƙasa da ainihinsu, yana matse ƙiyasansa zuwa kusan miliyan ɗaya (10⁶). Wannan yanayi yana bayyana a cikin dukkan samfura 15, amma yana bayyana tsaka-tsakin 17% kawai na kuskuren da ake rabawa. Samfuran kuma suna kuskure a kan tambayoyi iri ɗaya ta hanya ɗaya: kurakuran kowace tambaya na samfura biyu daban-daban har yanzu suna da daidaito (correlation) na 0.60 zuwa 0.69 bayan an rage tasirin lamura masu tsauri (extreme cases) kuma an cire karkatar zuwa miliyan ɗaya.

Tsaka-tsakin kuskuren samfura 15 a kan kowace tambaya, a kan ainihin amsa (ninkin goma, powers of ten): ana ƙiyasta ƙananan adadi fiye da ainihinsu, manya kuma ƙasa da ainihinsu. — Hoto na 1 (hagu), Fortuna & Bertalanič (2026), arXiv:2611.16315.
Tattaunawa ba ta taimaka ba
To idan wakilan suka tattauna fa? A zagaye uku na muhawara, kowane wakili ya ga amsoshin abokan ƙungiyarsa sannan ya yi gyara. Yaɗuwar amsoshi a cikin ƙungiyoyin biyar ta ruguje da 93% — zuwa zagaye na uku, ga samfura 12 cikin 15, wakilai uku cikin biyar sun ba da adadi iri ɗaya a kan aƙalla rabin tambayoyin. Amma kuskuren da ake rabawa bai motsa ba, kuma jimillar ribar ita ce 2.1%, da kuɗin lissafin kwamfuta (computing cost) sau goma sha biyar. Wakilan da kawai suka gyara amsarsu ta kansu, ba tare da ganin takwarorinsu ba, sun yi aƙalla daidai da haka.
Abin da ya ɗan taimaka
Ƙara zafin ɗaukar samfur (sampling temperature) ya sa amsoshi suka ƙara bambanta ba tare da inganta ƙungiyoyi da fiye da kusan 1% ba. Haɗa samfura daban-daban ya fi taimakawa: samfura biyar masu kusan daidaito iri ɗaya sun fi yin kyau da 6.6% a tare fiye da a cikin ƙungiyoyinsu na kansu. Duk da haka babu wani haɗin samfura 15 ɗin da ya inganta sosai (significantly) a kan ƙungiyar da aka gina daga samfuri guda mafi daidaito, gpt-oss-20b. Wani gyara da aka koya a kan tambayoyin da aka san amsoshinsu ya cire karkatar samfuran gaba ɗaya (overall bias), amma ba kuskuren tambaya-bayan-tambaya ba.
Marubutan sun jaddada iyakokin aikin: buɗaɗɗun samfura masu matsakaicin girma, ba tare da yanayin “tunani” (thinking modes) nasu ba, a kan waɗannan ma’aunan gwaji da umarni (prompts); ba a gwada samfuran sahun gaba (frontier models) da samfuran tunani mai zurfi (reasoning models) ba. Matsayarsu darasi ce ta zane (design). Ƙara wakilai, ko ƙara saɓani, ba ya gyara kuskuren da kowane kwafi na samfuri ke rabawa. Marubutan sun ba da shawarar a kai wa wannan kuskuren da ake rabawa hari kai tsaye, misali ta hanyar ba wakilai bayanai na musamman ga kowace tambaya — hanyar da har yanzu ba su gwada ba.
