Kwenzekeni
Iphepha lethula i-EDGE, noma i-Experience-Distillation for Guided Exploration, uhlaka lokuqeqesha ama-ejenti angamamodeli olimi ngokufunda okuqinisiwe. Isebenzisa ukuzizwisa okubuyisiwe njengezikafula zesikhashana zokuqeqesha, bese izama ukufaka ngaphakathi ukuziphatha okuwusizo kumodeli ngokwayo.
Umthombo we-arXiv wethula i-EDGE njengempendulo enkingeni ethile ekuqiniseni ukufunda kwama-ejenti angamamodeli olimi: amaphethini okuhlola awusizo emigudwini yokusebenzisana angase alahlwe ngemva kokubuyekezwa kwenqubomgomo. Ababhali bagxila kuma-ejenti abenza imisebenzi eyinkimbinkimbi, ethatha isikhathi eside, lapho indlela ephumelelayo ingaqukatha ulwazi olusebenzisekayo mayelana nokuthi yini okufanele uzame, yini okufanele uyigweme kanye nendlela yokululama ekuhlulekeni. Isiphakamiso sabo esimaphakathi siwukusebenzisa okuhlangenwe nakho komlando ngesikhathi sokuqeqeshwa kuyilapho kunciphisa isidingo sokubonisana nalabo okuhlangenwe nakho kamuva.
I-EDGE ihlukanisa iqembu ngalinye lokukhishwa ngezinhlobo ezimbili zama-trajectories: ezinye zibekwe esimweni esitholiwe futhi ezinye zenziwe ngaphandle kwayo. Leli phepha lithi lesi siqhathaniso silinganisela umnikelo omuhle we-marginal wesipiliyoni futhi sivumela isistimu ukuthi ivume ukuqondiswa okuwusizo ngaphandle kwamasampula engeziwe. Bese ihlakaza ukuziphatha okuwumphumela kunqubomgomo eyisisekelo ngokusebenzisa inhloso ehlehlayo ye-KL esetshenziswa ekusekeleni ngokwamandla kwenqubomgomo. Ngamagama asobala, indlela izama ukukhomba ukuthi yikuphi ukuziphatha okunikezwa ngaphandle okusiza futhi kuqeqeshe lokho kuziphatha kube yimodeli.
Uhlaka luphinde luhlanganise lokho ababhali abakubiza ngokuthi ibhange lesipiliyoni sokuziphendukela kwemvelo. Ngokomthombo, leli bhange lihlanganisa isiqondiso esivela ezindleleni zokuhluleka ezivelayo futhi lisusa okufakiwe okuphelelwa yisikhathi njengoba inqubomgomo ishintsha. Kumabhentshimakhi we-ALFWorld kanye ne-WebShop, ababhali babika intuthuko ngaphezu kwe-GRPO yamaphoyinti esilinganiso sempumelelo angu-8.3 nangu-12.5, ngokulandelanayo, kusetshenziswa amamodeli epharamitha ezigidigidi ezingu-7. Baphinde babike ukuthi ama-ejenti aqeqeshiwe agcine u-96.0% wokusebenza kwawo okuhleliwe lapho ulwazi lwangaphandle lususwa ngesikhathi esinqunyiwe. Umthombo uthi ikhodi iyatholakala nokuthi iphepha lamukelwa engqungqutheleni enkulu ye-EMNLP 2026.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Uma imiphumela ebikiwe ibambelela ngale kwamabhentshimakhi amabili ahloliwe, indlela yokwenza ingenza ama-ejenti we-horizon anganciki ezinhlelweni zokubuyisa futhi akwazi ukusebenzisa kabusha izifundo zemizamo yangaphambili. Lokho kungenza ukusetshenziswa kube lula, nakuba umthombo ungaqinisekisi ukwethembeka komhlaba wangempela noma ukulungela ukukhiqiza.
Ukubaluleka okungokoqobo komsebenzi umzamo wawo wokuxazulula ukungezwani okujwayelekile ekwakhiweni kwe-ejenti. Ukubuyisa kunganikeza umenzeli ulwazi lwangaphambili oluwusizo, kodwa isistimu okufanele iseshe inkumbulo yangaphandle ngokuphindaphindiwe ingase ibangele ukubambezeleka, izindleko zengqalasizinda kanye namanye amaphuzu okuhluleka. Umgomo oshiwo we-EDGE uwukusebenzisa ukubuyisa ngesikhathi sokufunda bese wenza ukuziphatha okuwumphumela kube yingxenye yenqubomgomo. Uma kuphumelele, lokho kungakhiqiza ama-agent aphethe isu lawo lokuhlola elifundiwe ngaphakathi.
Umphumela obikiwe wokugcinwa uhambisana ikakhulukazi nalowo mgomo. Leli phepha lithi ukusebenza kwahlala ku-96.0% wezinga elimisiwe ngemuva kokususwa kokuhlangenwe nakho kwangaphandle ngesikhathi sokunquma. Lokhu kuwubufakazi, ngokusho kwababhali, ukuthi indlela yenza okungaphezu kokuqondisa i-ejenti okwesikhashana: inzuzo enkulu idluliselwe kumodeli. Nokho, isibalo siyisimangalo esivela ekuhlolweni kwephepha, hhayi ukulinganisa okuzimele, futhi umthombo awunikezi amathebula okuhlola, izilinganiso zokungaqiniseki noma imininingwane yokuqhathanisa edingekayo ukuze kuhlolwe ukuqina kwayo.
Imiphumela nayo ibalulekile ngoba iqondise ukuziphatha komenzeli kunemiphumela emile yemodeli yolimi kuphela. I-ALFWorld ne-WebShop ihlanganisa ukusebenzisana kwezinyathelo eziningi, ngakho ukuqhubeka kuzo kungaba usizo kubacwaningi abafunda ukuhlela, ukusetshenziswa kwamathuluzi kanye nokululama emaphutheni. Noma kunjalo, ukuthuthukiswa kwebhentshimakhi akubonisi ngokwakho ukuthi i-ejenti ithembekile kuzilungiselelo ezivulekile. Umthombo awubiki imiphumela yokusebenzisa, izimfuneko zokwengamela komuntu, ukuhlolwa kokuphepha, izindleko, noma ukusebenza ngaphansi kwezimo eziphikisanayo noma ezishintsha ngokushesha.
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
Imibuzo eyinhloko ukuthi ingabe i-EDGE ijwayele ukwenza eminye imisebenzi, amamodeli kanye nezindawo, nokuthi ingabe izinzuzo zayo zihlala ngokwezibalo futhi zibalulekile yini ngaphansi kokuhlolwa okuzimele. Leli phepha liphinde lishiye kuvuliwe ukuthi ibhange layo lesipiliyoni liyanezela kangakanani ikhompyutha nobunjiniyela yinkimbinkimbi.
Inkinga yokuqala okufanele ibukwe ukuphindaphinda. Umthombo ukhomba i-ALFWorld ne-WebShop bese ubika imiphumela esikalini esingu-7B, kodwa awusho ngokufingqiwe ukuthi bangaki ama-run owenziwe, noma ngabe ukuthuthukiswa kubaluleke kangakanani ngokwezibalo, noma ukuthi amabhentshimakhi kanye nemigqaliselo kwalungiselelwa kanjani. Abacwaningi abazimele bazodinga ukuthi baqinisekise izinzuzo ezibikiwe futhi banqume ukuthi zincike ekwazisweni okuthile, izilungiselelo zokubuyisa, amasethi edatha noma amashejuli okuqeqesha.
Indaba yesibili i-. I-EDGE yakhelwe eduze nokufunda kokuqinisa i-ejenti kanye nokuhlola okuqondiswa kokuhlangenwe nakho, kodwa umthombo awuqinisekisi ukuthi indlela efanayo isebenza kuyo yonke imindeni yamamodeli, izikali zepharamitha, indawo noma izinhlobo zomsebenzi. Imiphumela kumabhentshimakhi amabili ingase ingabikezeli ukusebenza kumathuluzi esofthiwe, ukugeleza komsebenzi wocwaningo, izinhlelo zesevisi yamakhasimende noma izindawo ezibonakalayo. Akwaziwa futhi ukuthi ukuzizwisa kwangaphakathi kungenza i-ejenti ingakwazi ukuzivumelanisa nezimo lapho imvelo ishintsha noma uma isu elidala liba yingozi.
Umbuzo wokugcina uphathelene nezindleko kanye nokuphathwa kwebhange lesipiliyoni. Leli phepha lithi ibhange lihlanganisa isiqondiso kusuka ezindleleni zokwehluleka futhi lithene imingenelo ephelelwe yisikhathi, kodwa i-abstract ayibali ukugcinwa, ukubala ukuqeqeshwa, ukuvuselela imvamisa noma izidingo zokunakekelwa. Abacwaningi nabaqambi kufanele bahlole ukuthi ukwehluleka kukhethwa kanjani, ukuthi ukuziphatha okungafuneki kungachithwa kanye nokuwusizo, nokuthi inqubomgomo ewumphumela ingacwaningwa kalula kangakanani. Kuze kube yilapho leyo mibuzo iphendulwa, i-EDGE iqondwa kangcono njengomphumela wocwaningo othembisayo obikwe ababhali bayo, hhayi njengobufakazi bokuthi ama-agent ahlala emkhathizwe alungele ukusetshenziswa ngokungagadiwe. Lena imingcele yobufakazi bamanje kanye nezindawo lapho ucwaningo olwengeziwe luzodingeka ngaphambi kokuba kuthathwe iziphetho ezibanzi mayelana nokuqina, ukudluliswa, ukwethembeka, noma ukulungela.