Kwenzekeni
Abacwaningi bahlongoza Ukufunda Okuphambene Nokuqaphela Ubude Kwabenzeli be-GUI, noma i-LACL-GUI, uhlaka lokufunda oluqinisayo lwama-ejenti we-GUI ehlukene. Indlela yengeza amasignali ekhwalithi yeleveli yokulandela ekugadweni okusekelwe kumphumela, okuhloswe ngayo ukwenza ukuziphatha okuphumelelayo kufinyeze kakhudlwana futhi kunikeze ukuhlukana okuhle phakathi kwemizamo ehlulekile. Iphepha libika ukuthuthukiswa kokusebenza okungaguquki kunezindlela zangaphambilini kumabhentshimakhi womenzeli we-GUI.
Umthombo oyinhloko yi-arXiv ephrintiwe ngaphambili eyathunyelwa ngo-Agasti 22, 2026, enesihloko esithi "Ngaphandle Kwempumelelo Nokwehluleka: Ukufunda Okuphambene Nobude Kokuqaphela Kwama-GUI Agents." Ithinta i-AI ngokuqondile: ama-ejenti amamodeli wolimi amakhulu asebenzisa i-multimodal asebenza ngokusebenzisa ukuxhumana okucacile komsebenzisi. Ababhali bahlela ukufunda okuqiniswayo njengendlela yokuqeqesha evelele yalezi zinhlelo futhi bagxile ekutheni isignali yokuqeqesha yakhiwa kanjani lapho i-ejenti izama umsebenzi. Kuleso silungiselelo, ukugxila kwephepha akusona isixhumi esibonakalayo esisha noma isici esibheke kumsebenzisi. Kuyindlela ehlongozwayo yokubumba ukufunda kumarekhodi emisebenzi ezanyiwe.
Leli phepha lithi izindlela ezisetshenziswa kabanzi ezifana ne-Group Relative Policy Optimization zingaba nenkinga yalokho elikubiza ngokuthi ukungaqondi kahle kwe-gradient, kukhiqizwe ukusebenza kahle noma okungazinzile. Ibeka umsebenzi wayo phakathi kwemizamo yakamuva yokwenza kabusha ukufunda okuqiniswayo ngemiklomelo eqinisekiswa njengezinjongo eziphikisanayo noma ezisekelwe ngokwezigaba. Ngokusho kwe-abstract, lezo zindlela zingathuthukisa ukuzinza ngokugwema ukuziphatha okunezinkinga ze-gradient, kodwa ngokuvamile ziqondisa kuphela umphumela wokugcina we-trajectory. Umehluko ubalulekile ngoba ilebula yokugcina efanayo ingafihla umehluko endleleni umenzeli afinyelele ngayo kuyo. Ngakho-ke i-LACL-GUI iphatha i-trajectory njengengxenye yesignali yokuqeqesha.
I-LACL-GUI yethulwa njengohlaka oluphikisanayo lokufunda-ngemivuzo eqinisekisayo olungeza ulwazi mayelana nekhwalithi yokulandelana kwesenzo esigcwele. Ngaphakathi kwezindlela eziphumelelayo, isungula izintandokazi ezihloselwe ukuvumela ukubulawa okufushane. Ngaphakathi kwemizila ehlulekile, ihlukanisa imizamo ngokwehlukana kwayo nemigudu eyimpumelelo. Ukuhlolwa kwemibiko ye-abstract kumabhentshimakhi we-ejenti ye-GUI futhi ithi indlela ikhiqize amasiginali okufunda asebenza kahle futhi yathuthukisa ukusebenza ngokuqhubekayo kunezindlela zangaphambilini. Ayiwahlonzi amabhentshimakhi, ukulungiselelwa kwemodeli, ukubalwa kwemisebenzi, imiphumela yezinombolo, noma ukuhlolwa kwezibalo. Lokhu kwenza ukwakheka komgudu kube phakathi kwenhloso eshiwo yendlela. Umphumela obikiwe umayelana nokuziphatha kokufunda kumabhentshimakhi, nobufakazi obuthile bokuhlola obushiywe emininingwaneni yephepha.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Ama-ejenti e-GUI aklanyelwe ukuqedela imisebenzi ezindaweni zedijithali, ngakho ukusebenza kahle kokuqeqeshwa nokuthembeka kuthinta ngokuqondile ukuthi angasiza yini ngaphandle kokuboniswa. Umnikelo omaphakathi wephepha uyindlela yokukhipha ulwazi olwengeziwe kokubili emizamweni ephumelelayo nengaphumeleli esikhundleni sokuphatha umphumela ngamunye njengempumelelo noma ukwehluleka nje. Ngenxa yokuthi umthombo awunikezi amagama ebhentshimakhi, amaphuzu, isisekelo, noma ubufakazi bokusetshenziswa, isikali esisebenzayo sokuthuthukiswa okubikiwe sihlala singacacile.
Ucwaningo lubhekana nobuthakathaka obuphathekayo kuma-ejenti okuqeqesha okumele enze izenzo eziningi zokusebenzelana. Umphumela kanambambili ungabonisa ukuthi umzamo uphumelele noma wehlulekile, kodwa awubonisi ngokwawo ukuthi umzamo ophumelelayo usebenzise izinyathelo ezingadingekile noma ukuthi ukwehluleka okukodwa bekuseduze kakhulu nokuqedwa kunomunye. Indlela ehlongozwayo ithatha lokho kuhluka njengokugada okuwusizo, okungase kunikeze isithuthukisi ulwazi olwengeziwe kusuka kumzila ngamunye orekhodiwe. Ngokulandelayo isiphakamiso sincike ekutheni lezo zintandokazi zichazwa futhi zisetshenziswe kanjani ngesikhathi sokuthuthukisa. Lezo zinketho zedizayini ziwumongo obalulekile wokuhumusha ukuthuthukiswa kokusebenza okubikiwe.
Konjiniyela babenzeli be-GUI, lowo mbono ungaba nendaba ngoba imisebenzi yedijithali ngokuvamile ihilela ukulandelana kunokuqagela okukodwa. Indlela yokuqeqesha ekhuthaza izindlela eziphumelelayo ezimfishane ingase inciphise ukusebenzisana okungadingekile, kuyilapho ukwelashwa okulinganiselwe kokwehluleka kungasiza i-ejenti ukuthi ifunde kokuseduze esikhundleni sokukuhlanganisa ngemizamo engalungile. Lena imithelela yokwakheka kwendlela, hhayi okutholwe ngokuzimele okukhonjiswa umthombo ngaphandle kwesimangalo sebhentshimakhi sababhali. Lokho kwenza uzimele kuphinde kushiye kuvuliwe ukuthi ingakanani inzuzo etholakalayo uma imisebenzi ihluka ngobunzima noma lapho ukuziphatha kwesixhumi esibonakalayo kushintsha. Umthombo awuyixazululi leyo mibuzo.
Umphumela uqondwa kangcono njengentuthuko yocwaningo kunobufakazi bomkhiqizo osulungele ukuphakelwa. I-abstract ithi i-LACL-GUI ithuthukisa ngokungaguquki ukusebenza kunezindlela zangaphambilini, kodwa ayinikezeli osayizi bomphumela, imiphumela eqondene nomsebenzi othile, izidingo zekhompuyutha, noma imininingwane yokuqhathanisa. Futhi ayibonisi ukuthi le ndlela ithuthukisa ukuphepha, ukwenza okuvamile, ukufinyeleleka, noma ukwethembeka kuzixhumi ezibonakalayo zomhlaba wangempela. Leyo mikhawulo yenza umsebenzi ube usizo ekuqondeni isiqondiso sokuqeqeshwa kuyilapho ishiya umthelela wawo womphakathi ungaqinisekile. Ngokwezinto ezibonakalayo, umbono uxhuma ukusebenza kahle nekhwalithi yokugada. Ayinqumi ngokwayo ukuthi i-ejenti kufanele ilinganisele kanjani isivinini, ukuqapha, kanye nokululama. Ukuhlola okucophelela kuyodinga ukwehlukanisa umnikelo wendlela emandleni emodeli eyisisekelo nemisebenzi ekhethiwe. Lokho kuhlukaniswa akuchazwanga esifingqini esitholakalayo.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
Ongakubuka ngokulandelayo
Ukulandelela okubalulekile ukuthi izinzuzo ezibikiwe ze-LACL-GUI zibamba ezindaweni ezihlukene ze-GUI, amamodeli we-multimodal, ubude bemisebenzi, nokuhlola okuzimele. Abafundi kufanele futhi babheke ubufakazi mayelana nezindleko zokuqeqesha, ukuziphatha okucatshangelwayo, ukukhiqizwa kabusha, nokuthi ingabe ama-trajectories ayimpumelelo amafushane ahlala elungile ngaphansi kokusebenzelana okushintshile noma imisebenzi eyinkimbinkimbi. Umthombo awuqinisekisi ukuthi indlela ikhishwe esidlangalaleni, isilungele ukukhiqizwa, noma iphakeme ngaphandle kokuhlolwa okuchazwe ephepheni.
Ukuphindaphinda okuzimele kufanele kuhlole ukuthi inzuzo ebikiwe ivela ekugadweni kokwazi ubude ngokwayo noma kwezinye izinketho ekusethweni koqeqesho. Ukuqhathanisa okuwusizo kungahlukanisa okuthandwayo kwe-trajectory yempumelelo, isignali yekhwalithi yokwehluleka, kanye nenjongo eyisisekelo ehlukile. Umthombo awusho ukuthi ukukhishwa okunjalo kufakiwe yini, ngakho umnikelo wengxenye ngayinye uhlala umbuzo ovulekile. Ukuhlola okunjalo kuzocacisa ukuthi ingabe indlela ishintsha kuphela amandla okuthuthukisa noma iphinde ikhiqize ukuziphatha okuhlala kuthembekile ngaphandle kwesilungiselelo sokuhlola. Umthombo njengamanje ushiya lowo mehluko ungakaxazululwa.
Ukwenziwa ngokujwayelekile kungenye into ebalulekile engaziwa. Ama-ejenti e-GUI angase ahlangabezane nezakhiwo ezihlukene, izivumelwano zokusebenzisana, ama-horizons omsebenzi, nezinguquko zesixhumi esibonakalayo. Umthombo uthi kuphela ukuhlola kwenziwe kumabhentshimakhi e-ejenti ye-GUI; ayisunguli ukusebenza ezindaweni ezingabonakali, ukugeleza komsebenzi okuhlala isikhathi eside, okokufaka okubonakalayo okunomsindo, noma imisebenzi lapho indlela emfushane ingeyona indlela ephephe kakhulu noma ethembeke kakhulu. Ngakho-ke ukuhlola okuzayo kufanele kulinganisele ukunemba kanye nenani lesenzo, ukululama emaphutheni, nokuqina koshintsho.
Ukutholakala kwephepha nesimo sokusetshenziswa nakho kudinga ukuqinisekiswa. Umthombo ukhomba ukuthunyelwa kwe-arXiv nezixhumanisi zephepha, kodwa awusho ukuthi ikhodi, idatha yokuqeqesha, izindawo zokuhlola, noma i-ejenti itholakala esidlangalaleni. Kanjalo ayinikezi ulwazi mayelana nelayisensi, ihadiwe, ubude besikhathi sokuqeqeshwa, amacala okuhluleka, noma ukugadwa komuntu. Kuze kube yilapho kubikwa leyo mininingwane, isiphetho esiqine kakhulu esisekelwe ukuthi ababhali bahlongoza futhi balinganisele indlela yokuqeqesha engase ibe usizo, hhayi ukuthi ama-ejenti e-GUI ayisebenzisayo alungele ukuthunyelwa okubanzi. Lawo ma-artifact angekho azokwenza kube lula ukuhlola indlela nokukhiqiza kabusha ukuqhathanisa okubikiwe. Ukungabikho kwabo emthonjeni kunciphisa lokho okungaphethwa mayelana nokutholwa okungokoqobo.