TAKARDUN ZAMANIN AI SUNA CEWA "RATHER THAN" SAU BAKWAI FIYE DA DA

Manyan samfuran harshe (large language models) yanzu suna taimakawa wajen rubuta babban kaso na takardun kimiyya, kuma suna barin alamu: kalmomin da suka fi so, wani salon magana. Wata ɗabi’a ta fara ɓata wa masu bincike a sarrafa harshen ɗan adam (NLP) rai: jumlolin da aka gina a kan “X rather than Y” (X maimakon Y), waɗanda ke bayyana wani iƙirari wani ɓangare ta abin da ba shi ba. A cewar takardar, masu bitar sabon zagayen takardun da aka tura zuwa taron NLP sun yi ƙorafi a kai, kuma masu shirya taro yanzu suna tattauna rubutu “da ya cika da ɗabi’un AI”.

Olga Zamaraeva, Adrián Gude, Roi Santos-Ríos da Carlos Gómez-Rodríguez, a Jami’ar A Coruña a Spain, sun tashi don auna shi — da wani take da ke bayyana manufar da kansa.

Sau bakwai fiye

Sun kwatanta takardu 4,744 daga tarukan Association for Computational Linguistics na 2019, kafin mataimakan rubutu na AI su zama ruwan dare, da takardun NLP 1,821 da aka saka a arXiv a 2026 a tsarin taro iri ɗaya.

Yawan bayyana a kowace kalma 1,000 a takardun 2019 (shuɗi) da 2026 (lemu): “rather than” da “X, not Y” sun yi tashin gwauron zabo, yayin da “instead of” da “as opposed to” suka ragu. — Figure 1, Zamaraeva et al. (2026), arXiv:2610.10092.

• “Rather than” ya tashi daga amfani 0.134 zuwa 0.993 a kowace kalma 1,000, kusan sau bakwai fiye.
• “X, not Y” mai alaƙa da shi ya ninka sau huɗu; a halin yanzu “instead of” da “as opposed to” sun zama ƙasa da yawa.
• Kashi 94% na takardun 2026 suna ɗauke da “rather than”, idan aka kwatanta da kashi 36% a 2019, kuma suna amfani da shi kusan sau takwas a matsakaici. Wata takarda ta 2026 ta yi amfani da shi sau 69, kusan sau ɗaya a kowace kalma 170; babu takardar 2019 da ta wuce 12.

Tawagar ta kuma sa samfura su rubuta cikakkun takardu daga take da taƙaitaccen bayanin ainihin takardu. GPT-5.6-sol da GPT-6-sol sun yi amfani da kalmar a kowace takarda, fiye da marubutan 2026. Claude Haiku 4.5, Sonnet 5.5 da Opus 5.5 na Anthropic sun yi amfani da ita kusan a kowace takarda su ma, da ɗan ƙasa da haka. Wani samfuri na buɗaɗɗen tushe, Tülu 3 8B, da ƙyar ya yi amfani da ita a kowane mataki na horonsa. Gyara takardu bai cire ɗabi’ar ba: sababbin nau’ikan arXiv na takardun 2026 guda suna ɗauke da ɗan ƙari.

Masu ɓata rai, amma waɗanne?

Yawan amfani kaɗai ba ya tabbatar da komai: kalmar tana da amfani na halal da yawa. Don haka biyu daga cikin marubutan, gogaggun masu bita, sun yi wa amfani 1,000 da aka nuna musu ba tare da sanin asalinsu ba alama, a matsayin “mai ɓata rai” ko “na halal”.

• Kusan babu amfani na 2019 da ya ɓata musu rai (kashi 1.6% ga ɗaya, kashi 0% ga ɗayan).
• Kusan amfani ɗaya cikin goma na 2026 ya ɓata musu rai (kashi 11.5% da 8.3%).
• Da ƙyar suke yarda kan waɗanne amfani ne masu ɓata rai. Amma tun da takardar 2026 tana ɗauke da su da yawa, marubutan sun kiyasta cewa kusan rabin takardun 2026 suna ɗauke da aƙalla amfani ɗaya da ke ɓata wa kowannensu rai, kuma kusan kashi biyu cikin uku suna da ɗaya da ke ɓata wa aƙalla ɗaya daga cikinsu rai — idan aka kwatanta da kusan kashi 1% a 2019.

Abin da ke bambanta amfani masu ɓata rai galibi shi ne yadda suke ɗaukar zaɓin da aka ƙi. Sukan yabi wani zaɓi kuma su rage darajar ɗayan, kuma wasu mutane sun fi ɗaukar zaɓin da aka ƙi a matsayin straw man, wato matsayin da babu wani ƙwararren mai bincike da zai kare shi. Masu karatu kuma sun ɗauki kalmar kanta a matsayin alamar rubutun AI, duk da cewa ba za su iya bambanta sassan da GPT ta rubuta da na ɗan adam da aka rubuta a 2019 ba.

Sakamakon Claude, wanda mai alama ɗaya kawai ya yi wa alama zuwa yanzu, gauraye ne. Daftarin farko na Claude ba su ɓata rai fiye da takardun 2019 ba kuma sun yi ƙasa da na GPT-6-sol. Gyare-gyarensa sun ɓata rai daidai da na GPT-6-sol, kusan ninki biyu na daftarinsa na farko. Bambancin ya fito ne daga “musantawa” (disavowals), inda marubuta ke ƙin wani iƙirari mai ƙarfi game da aikinsu (“a hypothesis rather than a tested mechanism”, wato hasashe maimakon hanyar da aka gwada), waɗanda suka ƙaru lokacin da samfuran suka gyara takardunsu.

Masu tantancewa suna ba da lada

Daga ina ɗabi’ar ta fito? Ana gyara chatbot a kan nau’in amsoshi biyu-biyu, ɗaya daga cikinsu wani mutum ko alƙalin AI ya fi so. A cikin tarin bayanan jama’a huɗu na irin waɗannan zaɓuɓɓuka, amsoshin da ke ɗauke da “rather than” sun fi samun damar a zaɓe su a cikin uku, musamman idan alƙalan mutane ne. “Samfuran lada” (reward models) huɗu na buɗaɗɗen tushe sun ba wa jumla maki mafi girma tare da sashenta na “rather than” fiye da ba tare da shi ba, ko da an maye gurbinsa da wani sashe na tsaka-tsaki mai tsawo iri ɗaya. Kuma gaya wa samfuri ya zama “mai gaskiya” ya sa ya ƙara amfani da kalmar kaɗan.

Marubutan sun yi hasashen cewa wannan ginin wani tasiri ne na gefe na irin wannan horo: a cikin amsar chatbot guda, musantawa a hankali tana kama da gaskiya; idan aka maimaita ta a duk takarda, da duk wallafe-wallafe, tana karantuwa kamar kare wani ra’ayi. Hana kalmar watakila zai kawai mayar da aiki guda zuwa wani wuri, kamar yadda GPT-6-sol ke yi lokacin da yake musanya “rather than” da “X, not Y” yayin gyara.

Salon yana yaɗuwa

Binciken yana da iyakoki bayyanannu: masu alama biyu waɗanda su ma marubuta ne, kuma kwatancen Claude ɗaya daga cikinsu ne kawai ya yi masa alama zuwa yanzu; neman tsari da zai iya rasa wasu amfani; da bayanan zaɓi da ke nuna alaƙa kawai. Amma ƙarshensa ya wuce kalma ɗaya. Takardun da mutane suka sa hannu sun karɓi wannan ginin, don haka salon wallafe-wallafen kimiyya yana karkata zuwa na samfuran — waɗanda su kuma, za su koya daga gare shi.

Rikicin moriya. Marubutan sun bayyana cewa sun yi amfani da Claude Code (Claude Sonnet 5, Sonnet 5.5 da Opus 5.5) wajen rubuta code ɗinsu, gudanar da nazarce-nazarce da rubuta daftarin rubutu a duk takardar, har da sassan Methods da Results, kafin su gyara shi da hannu — suna cire, tsakanin wasu abubuwa, “rather than” goma da aka saka lokacin samarwa. Samfuran Claude suna cikin tsarin da aka yi nazari a kai kuma sun yi aiki a matsayin masu tantancewa na atomatik. Claude ne kuma ya rubuta wannan labarin.