Kudzosera Kutuka muLLMs
Kutukwa kunodzoserwa inzira inokatyamadza yekutadza apo modhi yemutauro inodzidza 'A ndiB' isingakwanise kupindura 'B ndiA.' Inoburitsa kuti maLLM anochengeta chokwadi senge-mwe-directional masangano, kwete seruzivo rwesymmetric.
Kudzika Kwakadzika
Yakanyorwa mubepa re2023 naBerglund uye vaanoshanda navo, kutuka kunodzosera kunoratidza kuti kana modhi yakadzidziswa pa 'Amai vaTom Cruise ndiMary Lee Pfeiffer,' inowanzotadza kana yakabvunzwa 'Ndiani mwanakomana waMary Lee Pfeiffer?' kunyange zvazvo mhinduro yacho yakafanana zvine musoro. Mhedzisiro yacho inoenderera kune saizi yemhando uye kunyangwe mushure mekugadzirisa zvakanaka pamazana echokwadi chakadaro. Haisi gaka rekurangarira: modhi yakaona ruzivo, asi mune imwe hurongwa. Nekuti kudzidziswa kunogonesa kufembera kwechiratidzo chinotevera pamusoro peiyo chaiyo yekurongeka kwezwi mu data, iyo nhamba yekubatanidza kubva kuA kusvika kuB haingogadziri chinongedzo kubva kuB kudzokera kuA. Kutsvaga kwakapokana fungidziro yekuti chiyero chega chinoburitsa kuchinjika, kufunga kwemunhu pamusoro pechokwadi.
Technical Insight
Shanduko dzinodzidza nekufanotaura chiratidzo chinotevera chakapihwa mamiriro ekutanga, saka zvigadziriso zvegradient zvinosimbisa mepu inotungamira 'A ipapo B' asi siya 'B ipapo A' isina kubatika kunze kwekunge iyo kurongeka ichionekwawo mukudzidziswa. Nzira mbiri idzi dzinogara munzira dzakasiyana dzehuremu. Vatsvakurudzi vakasimbisa izvi nekuyera log-zvingaitika: mushure mekudzidza chokwadi chemberi, mukana wechirevo chekudzokera shure chakaramba chiri padyo nekutanga, zvichiratidza kuti hapana kudzika kunonzwisisika kwakaitika panguva yekudzidziswa.
Strategic Impact
Kumhanya uye chiyero
Mutauro workflows inogona kufamba nekukurumidza pasina kupira kuenderana.
Svika uye svika
Inopamhidzira kupinda mumitauro yese nemataera ekutaurirana.
Sarudzo dzakajeka
Zvikwata zvinogona kupedza nguva yakawanda pakutonga uku otomatiki ichibata kudzokorora.
Ramangwana Rekudzokera Kutuka muLLMs
Mitigations iri pasi pekudzidza inosanganisira bidirectional data augmentation (kuwedzera zvirevo zvakadzokororwa), zvinangwa zvekudzidzisa zvinofanotaura zviratidzo mumativi ese, uye masisitimu ekudzoreredza anotarisa chokwadi kumusoro anoenderana pane kuvimba nezviremu zvakadzidzwa nemusoro. Zvimwe zvivakwa zvitsva uye reverse-pretraining miedzo inoderedza gaka. Tarisira kuti kutukwa kuderera asi kusapera, sezvo kuchifumura kusawirirana kwakadzama pakati pekudzidza kunotevera-chiratidzo uye symmetric chimiro chehukama hwepasirese.
Real-World Implementation
Chatbot inotaura nemazvo mubereki wemunhu ane mukurumbira asi anotadza kana akakumbirwa kutumidza zita remwana wemubereki ane mukurumbira.
Mumwe muenzaniso unodzokorora 'mutungamiri wechipfumbamwe aiva William Henry Harrison' asi anogumburwa pa 'nhamba ipi mutungamiri aiva William Henry Harrison.'
Mubatsiri wekodhi akadzidza basa-ku-tsananguro mepu haigone kudzoreredza zita rebasa kubva pane tsananguro chete.
A medical QA system yakadzidziswa pa 'Drug X treats Condition Y' inotadza kunyora Mushonga X pawakabvunzwa kuti chii chinobata Condition Y.
Njodzi & Guardrails
Chokwadi chehuroyi chinogona kupinda chinyararire mishumo, kuyerera kwetsigiro, kana tsvakiridzo.
Kunzwa nekukasira kunogona kugadzira mhedzisiro isingaenderane pane zvikumbiro zvakafanana.
Sensitive text data inogona kuburitswa kana zvidhiraivho zvisina kusimba.
Implementation Roadmap
Tsanangura chimiro chekubuda, toni, uye mhando zviyero usati waburitsa.
Mhinduro dzepasi neakavimbika masosi pese pazvine basa.
Chengetedza ongororo yekuongorora yemunhu kune yakakwira-stake zvinobuda.
Tevera maitiro ekutadza uye dzidzisazve kukurudzira kana mafambiro ebasa nguva nenguva.
Ramba Uchiongorora
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
Gaidhi rinotevera
ChatGPT neLLMs
Mibvunzo inowanzo bvunzwa
What is Reversal Curse in LLMs?
Kutukwa kunodzoserwa inzira inokatyamadza yekutadza apo modhi yemutauro inodzidza 'A ndiB' isingakwanise kupindura 'B ndiA.' Inoburitsa kuti maLLM anochengeta chokwadi senge-mwe-directional masangano, kwete seruzivo rwesymmetric.
Chituko chinodzokororwa chinorondedzera chii?
Kutukwa kunodzoserwa ndiko kutadza kutaura kuti 'B ndiA' kubva pakudzidziswa pa'A ndiB,' zvisinei kuti zviviri izvi zvakaenzana.
Sei kutukwa kudzoserwa kuchiitika, nemakanika?
Nekuti kudzidziswa kunogonesa kufanotaura kwechiratidzo chinotevera mune chaiyo yakacherechedzwa kurongeka, iyo reverse mepu haina kumbosimbiswa kunze kwekunge yaonekwawo mukudzidziswa.
Kuwedzera saizi yemuenzaniso zvinogadzirisa kutukwa kunodzosera here?
Chidzidzo chepakutanga chakawana kutukwa kunoitwa mumhando dzakasiyana-siyana dzemasaizi emuenzaniso, zvichiratidza kuti chikero chega hachigadzirise.
Ndekupi kudzikisira kwakanangana nekutukwa kunodzokera kumashure?
Bidirectional data augmentation inofumura modhi kune ese 'A ndiB' uye 'B ndiA,' ichigadzira iyo isipo reverse mubatanidzwa.
Vatsvakurudzi vakasimbisa sei kuti modhi yakanga isina kunyatsodzidza chokwadi chakasiyana?
Chirevo chakadzoserwa mukana wakagara padyo nekutanga mushure mekudzidziswa kwemberi, zvichiratidza kuti hapana kupinduka zvine musoro kwakaitika.