CACAR TANTANCEWAR TAKWARORIN MASANA

Kimiyya tana tafiya ne kan tantancewar takwarorin masana (peer review): kafin a wallafa wani sakamako, wasu ƙwararru suna yi masa hukunci. A manyan tarurrukan koyon na’ura (machine learning), ’yan masu tantancewa kaɗan suna ba kowace takardar da aka tura maki kuma a karɓe ta ko a ƙi ta. Bayan wannan al’ada akwai wata tambaya mara daɗi. Ka tura takarda ɗaya zuwa wasu masu tantancewa, shin hukuncin zai kasance iri ɗaya?

Gwaje-gwaje biyu masu tsada

Amsoshi kai tsaye kawai sun fito ne daga gwaje-gwaje biyu inda wani taro ya tantance wasu takardu sau biyu, da kwamitoci biyu masu zaman kansu. Takardar ta tuno da sakamakonsu:

• NeurIPS 2014: cikin takardu 166 da aka tantance sau biyu, kwamitocin biyu sun saɓa kan 43 — kashi 25.9. Kusan rabin takardun da wani kwamiti ya karɓa, ɗayan da ya ƙi su.
• NeurIPS 2021: kan takardu 882, kwamitocin sun saɓa kan kashi 23.0.

Irin waɗannan gwaje-gwaje ba kasafai ake yin su ba saboda suna da tsada: suna nufin ninka tantancewar ɗaruruwan takardu. Feilian Huang, wani mai bincike mai zaman kansa, ya tambaya ko za a iya kiyasta wannan lamba daga bayanan da aka bayyana a fili kawai.

Kwaikwayon kwamiti na biyu

Binciken yana amfani da tantancewar da aka bayyana a fili na taron ICLR daga 2017 zuwa 2025: takardu 36,113 da tantancewa 134,912. Wani samfurin ƙididdiga yana raba kowane maki zuwa sassa biyu: ingancin takardar da ke ɓoye, da “hayaniya” (noise) da kowane mai tantancewa ke ƙarawa. Samfuri na biyu yana sake samar da yadda maki ke zama hukunci. Daga nan kwamfutar tana tunanin, ga kowace takarda, kwamitoci biyu masu zaman kansu na masu tantancewa biyu, uku ko huɗu, sau dubu, kuma tana ƙirga sau nawa suke saɓawa.

Kafin a amince da sakamakon, marubucin ya duba shi sau biyu. Idan aka kwatanta da gwajin NeurIPS 2021, da masu tantancewa uku a kowace takarda kamar yadda aka yi a can, kwaikwayon ya ba da rashin jituwa kashi 23.3 maimakon kashi 23.0 da aka auna. Kuma kan takardun ICLR masu tantancewa aƙalla huɗu, raba ainihin masu tantancewa zuwa biyu-biyu bazata kawai yana ba da amsa ɗaya da ta samfurin, cikin maki ɗaya, a 2018 da kowace shekara daga 2021 zuwa 2025.

Kason bambancin maki da ke fitowa daga takardar kanta, shekara bayan shekara: kusan rabi, sauran hayaniyar masu tantancewa ce. — Hoto na 2, Huang (2026), arXiv:2610.06591.

Kusan takarda ɗaya cikin huɗu

Kason bambancin maki da ke fitowa daga takardar kanta yana tsakanin 0.38 zuwa 0.58 ya danganta da shekara — sauran hayaniyar masu tantancewa ce, daidai da kiyasin 2014 na kusan rabi. Sakamakon:

• Da masu tantancewa biyu a kowane kwamiti, hukuncin yana juyawa ga kashi 22.8 (2017) zuwa kashi 30.3 (2025) na takardu. Da masu tantancewa huɗu, kashi 17.7 zuwa 24.2.
• Kashi 30 zuwa 50 na takardun da aka karɓa da wani kwamiti mai zaman kansa ya ƙi su — har zuwa rabi a 2021.
• A tsawon shekaru tara da takardun da aka tura suka ƙaru kusan ninki 24 (489 a 2017, 11,663 a 2025), binciken bai samu wani yanayi mai ƙarfi ba: hayaniyar ba ta ƙaru ko ta ragu a fili. Marubucin ya jaddada cewa da bayanan shekaru tara kawai, wannan shaida ce kan rashin wani yanayi mai ƙarfi, ba hujjar daidaito ba.

Kwaikwayon rashin jituwa tsakanin kwamitoci biyu, idan aka kwatanta da ainihin gwaje-gwajen NeurIPS biyu. — Hoto na 3, Huang (2026), arXiv:2610.06591.

Ma’aunin kansa yana ƙara sa’a

Shekaru biyu sun fita daban. A 2020, ICLR ta yi amfani da wani ma’aunin maki mai kauri na matakai huɗu (1, 3, 6, 8): kashi 23.7 na takardu sun samu maki iri ɗaya daga dukkan masu tantance su, idan aka kwatanta da kashi 4.7 zuwa 13.1 a sauran shekaru. Don gwada tasirin, marubucin ya ɗauki makin 2021, waɗanda aka bayar daga cikin 10, ya sake rubuta su a kan ma’aunin matakai huɗu. Sakamako: kusan kashi 8 na siginar ya ɓace, rashin jituwa tsakanin kwamitoci ya ƙaru da kusan maki 7, kuma kason takardun da aka karɓa da za su juya ya ƙaru da kusan maki 11. Ma’auni mai kauri ba wai kawai yana sake sanya sunaye ga ra’ayoyi iri ɗaya ba ne; yana ƙara sa’a.

2021 ta bambanta: ma’auni mai laushi, amma siginar mafi rauni a cikin shekaru taran, kuma kashi 30.4 na takardun da aka karɓa suna zaune ne ɗan sama da iyakar karɓa. Takardu da yawa kusa da layin, hayaniya mai yawa: juyawa cikin sauƙi.

Kuma chatbots fa?

Binciken ya shirya neman canji a halayen tantancewa bayan 2023, lokacin da samfuran harshe (language models) suka yaɗu. Bayanan da aka bayyana a fili ba su ƙunshi rubutun tantancewa ko yarda da kai na masu tantancewa bayan 2022 ba, kuma gwajin ƙididdigar da ake da shi kusan ba shi da ƙarfi. Don haka marubucin bai yanke wata matsaya ta kowace hanya ba — kuma ya gabatar da wannan kamewa a matsayin da gangan.

Binciken cacar

Hanyar tana buƙatar maki da hukunci ne kawai, waɗanda tarurruka da dama tuni suke wallafawa. Kowanne daga cikinsu zai iya lissafa yawan “cacar” nasa ba tare da ninka ko da tantancewa ɗaya ba. Hukuncin marubucin kan shekaru tara na ICLR: tantancewa ba ta rushewa kuma ba ta inganta — tana ci gaba da hayaniya a kusan mataki ɗaya. Binciken ya ƙunshi ICLR kawai, kuma marubucinsa mai bincike ne guda ɗaya mai zaman kansa.