Me ya faru
Wani sabon tsarin bincike mai suna LogicTrack an gabatar da shi don magance matsalar manyan nau'ikan harshe (LLMs) suna samar da ingantattun amsoshi na ƙarshe ta hanyar sarƙoƙin tunani mara kyau. LogicTrack yana aiki azaman tsarin neuro-alama wanda ke daidaita matakan mutum kai tsaye a cikin tsarin Tsarin Tsara-Thought (CoT) zuwa wakilcin alama. Ana tabbatar da waɗannan wakilcin ta amfani da ƙa'idodin ka'idar atomatik don tabbatar da daidaiton ma'ana. Tsarin yana gabatar da tsarin ladabtarwa na tushen mai warwarewa (SBR), wanda ke ba da maki mai hikima ta mataki don jagorantar binciken bishiyar baya yayin tunani. Bugu da ƙari, masu binciken sun yi amfani da LogicTrack don samar da bayanan daidaitawa mai kyau (SFT), ƙyale samfura su koyi yin nazari a mataki-mataki azaman iyawar ciki.
LogicTrack yana magance yanayin 'akwatin baƙar fata' na tunani na Sarkar-Thought ta hanyar gabatar da alamar neuro-alama. Maimakon dogara kawai ga ƙirar ƙila ta cikin rarraba don tantance mataki na gaba, tsarin yana canza kowane matakin tunani zuwa harshen alama na yau da kullun.
Tsarin yana amfani da masu tabbatar da ka'idar ka'ida ta atomatik don bincika ingancin waɗannan matakan alamar. Idan an gano matakin ba shi da ma'ana a hankali, tsarin Rarraba-Based Backtracking Reward (SBR) yana haifar da binciken bishiyar baya, yana tilasta ƙirar don gano wasu hanyoyin tunani waɗanda suka dace daidai da ma'ana.
Bayan tantancewa-lokaci-lokaci, masu binciken sun yi amfani da tsarin don ƙirƙirar bayanan 'hanyoyin baya.' Ta hanyar daidaita ƙirar ƙira akan waɗannan bayanan, sun ba wa samfuran damar yin nau'i na tantance kansu, inda ƙirar ke koya don ba da fifiko ga hanyoyin tunani mai ma'ana ba tare da buƙatar ƙa'idodin ƙa'idar waje ba a kowane mataki na gaba.
Me ya sa yake da mahimmanci
Dogaro da sakamako na tushen sakamako a cikin horon LLM na yanzu galibi yana rufe matakan ''haushe' ko a zahiri mara inganci matakan tunani, wanda ke haifar da babban haɗari a cikin manyan yankuna kamar magani, doka, ko injiniyan injiniya inda tsari ke da mahimmanci kamar sakamakon. Ta hanyar haɗa tabbaci na yau da kullun a cikin yanayin tunani, LogicTrack yana ba da tsari don aiwatar da ƙaƙƙarfan hankali. Wannan juyi daga fitowar mai yuwuwa zalla zuwa ingantaccen dabaru na alama yana haɓaka amincin tsarin AI. Ƙarfin shigar da wannan tsarin tantancewa ta hanyar daidaitawa mai kyau yana nuna hanya zuwa ga ƙira waɗanda suka fi dogaro da gaske kuma basu da kusanci ga kurakurai masu ma'ana, koda lokacin aiki a waje da yanayin tabbatarwa na yau da kullun. An nuna tasirin tsarin a cikin ma'auni takwas na dalilai guda takwas da LLM daban-daban guda bakwai, yana nuna fa'ida don inganta amincin ƙirar.
Siffofin horo na LLM na yanzu suna ba da fifiko ga amsa ta ƙarshe, wanda zai iya haifar da 'madaidaitan' amsoshin da aka samo daga dabarar da ba daidai ba. Wannan yana da matsala a cikin manyan mahalli inda tsarin tunani dole ne ya zama abin dubawa da tabbatarwa.
LogicTrack yana gadar rata tsakanin hanyoyin sadarwa mai yiwuwa na jijiya da ƙayyadaddun dabaru na alama. Ta hanyar aiwatar da daidaiton ma'ana, yana rage yuwuwar samfura su isa kan madaidaicin ƙarshe ta matakai mara kyau ko marasa ma'ana.
Nasarar tsarin a cikin LLM guda bakwai daban-daban yana nuna cewa hanyar ƙirar ƙira ce-agnostic, tana ba da daidaitacciyar hanya don haɓaka ingancin sarƙoƙin tunani a cikin gine-gine daban-daban.
Ingantacciyar hanyar sadarwa: Yadda A zahiri yake Aiki
Bincika fasahar da ke bayan wannan ci gaban ta hanyar mu'amala.
Which component of an AI application is the machine-learning model itself?
Abin kallo na gaba
Babban abin da ba a san shi ba shine babban lissafin lissafin da ke da alaƙa da gudanar da abubuwan tabbatar da ka'idojin sarrafa kai tsaye yayin ƙaddamarwa, wanda zai iya iyakance tura lokacin aiki a aikace-aikacen latency. Ci gaba na gaba zai yiwu ya mayar da hankali kan inganta tsarin tsarawa ta atomatik don aiwatar da ƙarin hadaddun, ayyukan tunani marasa lissafi inda wakilcin alama ke da wahala a halin yanzu. Ya rage a ga yadda wannan tsarin ya kai girma, mafi girman ƙira, da kuma ko nasarorin da aka samu a cikin daidaiton tunani yana fassara zuwa dogaro na gaske a cikin mahallin da ba na asali ba. Masu amfani yakamata su saka idanu ko an haɗa wannan hanyar a cikin bututun horar da ƙirar kasuwanci ko kuma idan ta kasance da farko kayan aikin mataki-bincike don ƙwararrun ayyuka na tabbatarwa.
Binciken bai ƙayyadad da tasirin latency na tafiyar da ka'idojin tabbatarwa ba yayin ƙaddamarwa. Ɗaukar aiki na zahiri zai dogara ne akan ko za'a iya rage girman wannan sama don yanayin samarwa.
Matsakaicin 'tsari-kai-tsaye' yana da iyakancewa mai mahimmanci. Duk da yake tasiri ga ma'auni na lissafi da ma'ana, ba a san yadda ingantaccen tsarin zai iya tsara tunani a cikin ma'auni ko ma'ana ba.
Samuwar lambar da ƙayyadaddun ƙayyadaddun ƙa'idodin ƙa'idar da aka yi amfani da su ba a dalla-dalla a cikin sanarwar, barin samun damar wannan tsarin don tabbatarwa mai zaman kansa ko aiwatarwa a halin yanzu ba a san shi ba.