Juya La'anar a cikin LLMs
La'anar juyawa ita ce yanayin gazawa mai ban mamaki inda samfurin harshe wanda ya koyi 'A shine B' ba zai iya amsawa 'B shine A' da aminci ba. Yana bayyana cewa LLMs suna adana bayanai azaman ƙungiyoyin jagora ɗaya, ba azaman ilimin ƙima ba.
Zurfafa nutsewa
An rubuta a cikin takarda na 2023 da Berglund da abokan aiki suka yi, la'anar da aka yi amfani da ita ta nuna cewa idan an horar da samfurin a kan 'mahaifiyar Tom Cruise ita ce Mary Lee Pfeiffer,' sau da yawa yakan kasa lokacin da aka tambaye shi 'Wane ne ɗan Mary Lee Pfeiffer?' duk da cewa amsar ita ce a hankali. Tasirin yana ci gaba a cikin girman samfuri har ma bayan daidaitawa akan ɗaruruwan abubuwan gaskiya. Ba ƙwaƙwalwar ƙwaƙwalwar ajiya ba ce: samfurin ya ga bayanin, amma a cikin tsari ɗaya kawai. Saboda horo yana inganta hasashe na gaba akan ainihin tsari na kalma a cikin bayanan, hanyar haɗin ƙididdiga daga A zuwa B ba ta haifar da hanyar haɗi kai tsaye daga B zuwa A. Binciken ƙalubalen zato wanda ma'auni kaɗai ke haifar da sassauƙa, tunani irin na ɗan adam akan gaskiya.
Fahimtar Fasaha
Masu canza canji suna koya ta hanyar tsinkayar alama ta gaba da aka bayar a baya, don haka sabuntawar gradient yana ƙarfafa taswirar shugabanci 'A sannan B' amma suna barin 'B sannan A' ba a taɓa su ba sai dai idan wannan tsari shima ya bayyana a cikin horo. Hanyoyi biyu suna rayuwa ne a cikin hanyoyin nauyi daban. Masu bincike sun tabbatar da hakan ta hanyar auna yuwuwar log: bayan koyon gaskiyar gaba, yuwuwar bayanin ta tsaya kusa da tushe, ba tare da nuna wata juyi na ma'ana ba da ta faru yayin horo.
Dabarun Tasiri
Gudu da sikelin
Gudun aikin harshe na iya tafiya da sauri ba tare da sadaukar da daidaito ba.
Shiga ku isa
Yana faɗaɗa damar shiga cikin harsuna da salon sadarwa.
Shawarwari masu haske
Ƙungiyoyi za su iya ciyar da ƙarin lokaci akan hukunci yayin da aiki da kai ke sarrafa maimaitawa.
Makomar La'anar Juyawa a cikin LLMs
Ragewar da ke ƙarƙashin binciken sun haɗa da haɓaka bayanai na bidirectional (ƙara juzu'i na jujjuya), makasudin horarwa waɗanda ke hasashen alamu a bangarorin biyu, da tsarin dawo da bayanai waɗanda ke kallon gaskiya daidai gwargwado maimakon dogaro da ma'aunin da aka haddace. Wasu sabbin gine-ginen gine-gine da gwaje-gwajen horo na baya-bayan nan suna rage gibin. Yi tsammanin la'anar za ta ragu amma ba za ta ɓace ba, saboda tana fallasa rashin daidaituwa mai zurfi tsakanin koyo na gaba da tsarin daidaitaccen tsarin dangantakar duniya.
Aiwatar da Gaskiyar Duniya
A chatbot daidai yana faɗin iyayen sananne amma ya kasa lokacin da aka tambaye shi sunan shahararren ɗan wannan iyayen.
Wani abin ƙira yana karanta 'shugaban tara shine William Henry Harrison' duk da haka ya yi tuntuɓe akan 'wanne shugaba ne William Henry Harrison'.
Mataimakin coding wanda ya koyi taswirar aiki-zuwa-bayani ba zai iya dawo da sunan aikin daga bayanin shi kaɗai ba.
Tsarin QA na likita wanda aka horar akan 'Drug X yana magance Yanayin Y' ya kasa lissafin Drug X lokacin da aka tambaye shi abin da ke kula da Yanayin Y.
Hatsari & Tsare-tsare
Abubuwan da aka ruɗe suna iya shigar da rahotanni cikin nutsuwa, kwararar tallafi, ko abubuwan bincike.
Hankali na gaggawa na iya ƙirƙirar sakamako mara daidaituwa a cikin buƙatun iri ɗaya.
Za a iya fallasa bayanan rubutu mai ma'ana idan ikon samun dama yana da rauni.
Taswirar Hanya
Ƙayyade tsarin fitarwa, sautin, da ma'auni masu inganci kafin fitowa.
Amsa a ƙasa tare da amintattun tushe a duk lokacin da daidaito ya shafi mahimmanci.
Ajiye wurin binciken ɗan adam don abubuwan da ake samu masu girma.
Bibiyar tsarin gazawar kuma sake horar da tsokaci ko tafiyar aiki akai-akai.
Ci gaba da Bincike
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Reversal Curse in LLMs quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Jagora na gaba
ChatGPT da LLM
Tambayoyin da ake yawan yi
What is Reversal Curse in LLMs?
La'anar juyawa ita ce yanayin gazawa mai ban mamaki inda samfurin harshe wanda ya koyi 'A shine B' ba zai iya amsawa 'B shine A' da aminci ba. Yana bayyana cewa LLMs suna adana bayanai azaman ƙungiyoyin jagora ɗaya, ba azaman ilimin ƙima ba.
Menene la'anar juyawa ta bayyana?
La'anar juyawa ita ce kasawar 'B shine A' daga horarwa akan 'A shine B' duk da cewa biyun sun yi daidai.
Me yasa la'anar juyawa ta faru, ta hanyar injiniyoyi?
Saboda horo yana inganta hasashen alama na gaba a daidai tsarin da aka lura, ba a taɓa ƙarfafa taswirar baya ba sai idan kuma ya bayyana a cikin horo.
Shin haɓaka girman samfurin yana gyara la'anar juyawa?
Binciken na asali ya gano la'anar da aka gudanar a cikin nau'ikan nau'ikan nau'ikan nau'ikan nau'ikan, yana nuna ma'auni kaɗai ba ya warware shi.
Wanne ragewa ne kai tsaye ke nufi da la'anar juyawa?
Ƙirƙirar bayanan bidirectional yana fallasa samfurin ga duka 'A shine B' da 'B shine A,' ƙirƙirar ƙungiyar da ta ɓace.
Ta yaya masu bincike suka tabbatar da samfurin bai koyi gaskiyar a fakaice ba?
Yiwuwar bayanin da aka juyar ya tsaya kusa da tushe bayan horo na gaba, yana nuna babu wani juyi na ma'ana da ya faru.