ZAƁIN MUTANE MILIYAN BAKWAI, ƘARAMAR ƘIRA ƊAYA TA TUNANI

Fiye da ƙarni ɗaya, ilimin halayyar ɗan adam ya samar da ƙa’idoji kan yadda tunani ke aiki. Mun san cewa alaƙa tana ƙara ƙarfi idan abubuwa suka faru tare, cewa koyo yana tafiya ne ta kuskuren hasashe (prediction errors) — tazara tsakanin abin da muke tsammani da abin da ya faru — kuma cewa ƙwaƙwalwa tana tsara bayanai a matakai, daga masu sauƙi zuwa masu zurfi. Amma kowace ƙa’ida an haɓaka ta ita kaɗai, galibi don irin gwaji ɗaya. Nawa daga cikin halayyar ɗan adam suke bayyanawa tare? Ba a taɓa auna wannan ba.

Marcel Binz da Eric Schulz, na Institute for Human-Centered AI a Helmholtz Munich da Max Planck Institute for Biological Cybernetics a Tübingen, sun gina wata ƙira don gano amsar.

Tunani da aka gina da ƙwaƙwalwa

Ƙirarsu, CORE (wato COntextual Retrieval and Encoding), tana karanta cikakken rubutun gwaji — umarni, abin da aka nuna, martani, da amsoshin mahalarcin na baya — kuma tana hasashen wane maɓalli mutumin zai danna na gaba.

Ba ta ƙunshi komai ba sai wasu ƙalilan na sanannun abubuwa:

ana mayar da rubutun zuwa phonemes, wato sautukan magana;
ana rubuta waɗannan cikin ƙwaƙwalwar alaƙa (associative memory), bisa wata tsohuwar ƙa’idar koyo;
abin da ake rubutawa, da ƙarfin rubutun, da abin da ake mantawa, duk sun dogara ne kan mahallin da ake ciki;
ƙwaƙwalwar tana adana kuskuren hasashe, ba bayanan danye ba — abin da ba ta riga ta sani ba kawai;
an jera irin waɗannan ƙwaƙwalwa shida a matakai.

Dukan ƙirar tana da parameter 452,608 kuma an horar da ita cikin kusan awa 24 a kan katin zane-zane (graphics card) guda ɗaya.

Zane: ana mayar da rubutun gwaji zuwa phonemes, a wuce da su ta matakan ƙwaƙwalwa shida, kuma kowane mataki yana rubutawa, kwatantawa da ɗauko bayanai.

Yadda CORE ke aiki: rubutun gwaji (A) yana ciyar da tarin matakan ƙwaƙwalwa (B); kowane mataki yana rubutawa, mantawa da ɗauko alaƙoƙi gwargwadon mahalli (C). — Figure 1, Binz & Schulz (2026), arXiv:2610.13603.

An gwada ta da zaɓi miliyan bakwai

Sun daidaita CORE da Psych-201, babban tarin gwaje-gwajen ilimin halayyar ɗan adam: gwaje-gwaje 137 daga nazari 62, waɗanda suka shafi koyo daga lada, yanke shawara cikin haɗari, koyon rukunai, hukunci na zamantakewa da sauransu — zaɓi 7,036,442 a cikin zama 80,452 don horo, sannan aka ware wasu zaman don gwaji.

A kan zaman da ba ta taɓa gani ba, CORE ta yi hasashen zaɓi fiye da wani daidaitaccen tsarin AI mai girma iri ɗaya (Transformer) a cikin dukan nazari 62.
Ta kuma fi Centaur gaba ɗaya — wani babban samfurin harshe da aka gyara a kan halayyar ɗan adam wanda ke da parameter ninki dubu ɗari (10⁵) fiye — inda ta yi nasara a cikin 47 daga nazari 62.
Cire kowane abu guda ɗaya ya rage mata ƙwarewa. Matakai ne suka fi muhimmanci, sannan mantuwa, sannan phonemes, sannan kuskuren hasashe.

Kamar mahalarci na kwamfuta

Daidaita bayanai bai isa ba: ƙirar tunani ya kamata kuma ta yi hali kamar mutane. Marubutan sun yi amfani da CORE a matsayin mahalarci a cikin gwaje-gwajen asali kuma sun duba ko tasirin da aka wallafa ya sake bayyana. Ta sake samar da tasiri 50 daga 67 (75%).

Ta fi yin kyau a binciken sababbin zaɓuɓɓuka (exploration) (11 daga 11) da koyo daga lada (10 daga 11), kuma ta fi kasawa a harshe da hukunci na zamantakewa (7 daga 12). Marubutan sun bayyana dalilin kasawar: yawancin waɗannan ayyuka sun dogara ne kan ilimin da mutane ke samu a wajen ɗakin gwaji, game da kalmomi, lambobi da duniya — kuma CORE ba ta taɓa ganin komai ba sai rubutun gwaje-gwaje. Tana kuma fama da tsara abin da za a yi gaba, wanda ba ta da wata hanya ta yin sa. Wasu tasirin da ta kasa samarwa sun zama marasa ƙarfi a cikin bayanan ɗan adam ma.

Saitunanta na ciki ma suna kama da na ilimin halayyar ɗan adam. A cikin wasa ɗaya, saurin koyonta yana raguwa yayin da shaida ke taruwa, kamar yadda ka’idojin koyo na Bayes suka yi hasashe. Kuma tsawon ƙwaƙwalwarta yana ƙaruwa a lokacin wasa sannan ya ruguje a iyakar da wasa na gaba — alamar yadda mutane ke yanka abubuwan da suka fuskanta zuwa aukuwa-aukuwa.

Rikicin muradu. Tsarin sarrafa bayanan ya dogara ne kan AI: DeepSeek-V4-Flash ne ya mayar da kowane gwaji zuwa rubutu kuma Claude Fable 5.1, na Anthropic, ya tantance shi; masu kwaikwayon da aka yi amfani da su don gwada tasirin ma wani tsarin AI mai sarrafa kansa ne ya gina su. Claude ne ya rubuta wannan maƙala. Ban da haka, ƙirar Centaur da aka yi amfani da ita a matsayin mai fafatawa da tarin Psych-201 duka aikin baya ne da marubuci na farko ya jagoranta. Marubutan sun bayyana cewa ba su da rikicin muradu.

Zuwa ga ƙirar tunani ta bai ɗaya

Marubutan sun bayyana CORE a matsayin mataki tsakanin tsofaffin ƙirar tunani (cognitive models) da akwatuna baƙaƙe (black boxes): ana iya karanta ƙa’idojinta, amma bayanan da take koyo ba haka suke ba. Ba ta da ƙwaƙwalwa ta dogon lokaci, tun da gwaje-gwajen ba sa wuce awa ɗaya. Burinsu ya fi ƙira ɗaya girma: gwada ka’idojin tunani da dukan tarihin gwaje-gwajen ilimin halayyar ɗan adam a lokaci guda, don fannin ya zama “fiye da jimillar gwaje-gwajensa”.