Kwenzekeni
Abacwaningi abaxhumene ne-Google Cloud AI Research, iYunivesithi yaseWashington e-St. Louis kanye ne-UNC Chapel Hill bakhiphe i-EnvHarness, ungqimba oluhlelekayo olugoqa izindawo ezingaguquki ze-agent ye-AI ngaphandle kokushintsha izifanisi zazo eziyisisekelo, imisebenzi noma iziqinisekisi ezakhelwe umuntu. I-MarkTechPost ibika ukuthi isistimu isebenzisa izingxenye ezibizwa nge-Stage, Contract and Chain ukuze iguqule izimo zokuqala, izenzo ezivunyelwe, okubonwayo kanye nokwakheka kweziqephu.
I-MarkTechPost ibika ukuthi i-EnvHarness ikhishwe abacwaningi abavela ku-Google Cloud AI Research, i-Washington University e-St. Louis kanye ne-UNC Chapel Hill. Umqondo oyinhloko wephepha ukugoqa indawo ekhona ezingxenyeni ezihlelekayo kunokukhiqiza indawo entsha ngokuphelele. Ama-wrappers asebenza ngemisebenzi evamile yemvelo efana nokusetha kabusha() kanye nesinyathelo(), kushiya isifanisi esingemuva, imisebenzi yebhentshimakhi nesiqinisekisi esakhiwe umuntu singathintwa. Lokhu kuhloselwe ukuvumela ukuqaliswa okukodwa ukuthi kusebenze ezizindeni eziningi kuyilapho kugwema ukuthembela ekhodini yokuqinisekisa esanda kukhiqizwa, okungenzeka ayithembeki. Umthombo uxhumanisa nephepha le-arXiv, inqolobane ye-GitHub kanye nekhasi lephrojekthi, kodwa imiphumela ebikiwe ayizange iqinisekiswe ngokuzimele kulokhu kuhlola.
I-MarkTechPost ichaza izingxenye ezintathu. Isiteji sidlala futhi ukulandelana okungaguquki kwezenzo ngemva kokusetha kabusha(), okuvumela isiqephu ukuthi siqale sisuka kwesinye isimo; isihloko sinikeza ukufihla inkomishi eqondiwe ekhabetheni elivaliwe njengesibonelo esingaphoqa umenzeli ukuthi aseshe kunokuba afinyelele ngokushesha. Inkontileka ifaka amahhuku angavimba izenzo, abhale kabusha izinguquko noma anciphise okubonwayo. I-Chain ibeka indawo yesibili esiqeshini esifanayo ngaphansi kwesabelomali sesinyathelo esabiwe, ngempumelelo edinga ukuthi zombili iziqinisekiso zezingxenye ziphumelele. Umthombo uthi lezi zingxenye zingaqanjwa nokuthi ibhentshimakhi entsha ingaxhunywa ngesixhumi esibonakalayo esifaka ukusetha kabusha, isinyathelo, ukubuka, ukuhlola, get_env_state, save_state kanye nokuthi from_state.
Uhlelo luhlanganisa ne-EnvRigger, iluphu yomklami esekwe ku-LLM. Ngokusho kwe-MarkTechPost, i-EnvRigger ibheka ukukhishwa okuyisisekelo okuhlanu, ixilonga ubuthakathaka benqubomgomo yesistimu, ibhala izingxenye zokugoqa njengekhodi Python futhi izivivinya ekukhishweni okusha okuhlanu. Izindawo zamakhandidethi ezingaxazululeki noma ezingaxazululeka kancane ziyanqatshwa, kuze kufike emizuliswaneni emihlanu yokubuyekeza umsebenzi ngamunye. I-athikili ithi amahhuku akhiqiziwe ahlanganisa ku-subprocess ehlukanisiwe, aguqule ukuguqulwa okubi kube umkhondo orekhodiwe esikhundleni sokugijima okuhlulekile. I-MarkTechPost ibika imiphumela kuyo yonke i-ALFWorld, WebArena, SWE-bench Verified, OfficeQA kanye ne-SpreadsheetBench, okuhlanganisa nokuthuthuka kwamaphuzu angu-9.0 ku-ALFWorld ngaphandle kokusatshalaliswa kokuhlukaniswa kanye nezinyathelo zokusebenzisa ezimbalwa ezingu-9.8% ku-SWE-bench Verified.
Imininingwane yomthombo: marktechpost.com ↗
Kungani kubalulekile
Le ndlela ibhekana nenkinga engokoqobo ekuqeqesheni i-ejenti: izindawo ezimile zingayeka ukunikeza amasignali okufunda awusizo uma umenzeli esewajwayele. Ngokujwayelanisa izindawo ezikhona nobuthakathaka obubonwa ekukhishweni komenzeli uqobo, i-EnvHarness ingenza ukuqeqeshwa nokuhlola kuqondiswe kakhulu kuyilapho ilondoloza ukuqonda okukhona kokuqinisekisa. Izinzuzo ezibikiwe ziyathembisa kodwa zihlala ziyizimangalo ezivela embikweni wesibili mayelana nephepha locwaningo.
Izindawo zokuqeqesha ama-ejenti zivame ukulungiswa: imisebenzi efanayo iziphatha ngendlela efanayo kungakhathaliseki ukuthi i-ejenti ayinaso isipiliyoni noma isivele iyingcweti. I-MarkTechPost ibika ukuthi impendulo evamile iwukukhiqiza izindawo eziningi, kodwa ithi lokhu kudala amapayipi okukhiqiza aqondene nesizinda esithile futhi kudinga izinombolo ezinkulu zokuqinisekisa ezibhalwe nge-LLM ukuthi zihlungwe. I-EnvHarness iqondise ibhodlela ngokulungisa izimo, izenzo kanye nokubona kuyilapho igcina isiqinisekisi sangempela. Uma indlela yokusebenza isebenza njengoba ibikiwe, inikeza indlela yokwenza amabhentshimakhi akhona aphendule kakhudlwana ebuthakathakeni bangempela be-ejenti ngaphandle kokwakha kabusha ibhentshimakhi ngayinye kusukela ekuqaleni.
Imiphumela ebikiwe yebhentshimakhi iphakamisa ukuthi inani livela ekubumbeni kabusha okuqondiwe kunokumane ungeze ikhono endaweni engashintshile. I-MarkTechPost ithi ukusebenza kwe-ALFWorld kukhuphuke kusuka ku-62.4 kuya ku-68.3, ngenzuzo yamaphuzu angu-9.0 ekuhlukaniseni ngaphandle kokusabalalisa. Ku-SWE-bench Verified, izinga elixazululiwe okubikiwe likhuphuke lisuka ku-49.88 laya ku-52.58 ngenkathi izinyathelo ezimaphakathi zehla ukusuka ku-55.01 zaya ku-49.61. I-athikili iphinde ithi amakhono ambiwe ezindaweni ezingalungiswanga enziwe ngaphansi kwesisekelo sokungenakhono ku-SpreadsheetBench ne-WebArena, kuyilapho ukubunjwa kabusha kwenza inqubo yezimayini yaba usizo. Lezi zibalo zingukutholakele okubikiwe, hhayi izilinganiso eziqinisekisiwe ngokuzimela.
Ukubaluleka okungokoqobo kunemibandela. I-MarkTechPost ithi i-EnvHarness ikhishwe ngaphansi kwelayisense ye-Apache-2.0 ku-Python futhi ihlanganisa abashayeli bokukhiqiza kabusha bezindawo eziyisithupha, okungenza ifinyeleleke emaqenjini asevele asebenzisa ama-loop okuhlola ama-ejenti. Indlela ingase ibe usizo ekuqiniseni ukufunda nokuhlola ngoba umthombo ubika intuthuko ngaphansi kokubili kokusungulwa kwamakhono nokuqeqeshwa kwe-GRPO. Ngesikhathi esifanayo, i-athikili ithi uhlelo lunesidingo esinzima: imvelo kufanele ihlelwe kabusha. Lokho akufaki izigaba ezibalulekile zokusetshenziswa komenzeli womhlaba wangempela, okuhlanganisa ama-akhawunti abukhoma namarobhothi aphathekayo, futhi kukhawulela ukuthi imephu yemiphumela yebhentshimakhi iqonde kangakanani ekuziphatheni kokukhiqiza.
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 ebalulekile iwukuthi ingabe abacwaningi abazimele bangakwazi yini ukukhiqiza kabusha ukuthuthukiswa kwebhentshimakhi ebikiwe, ukuthi ithokheni yomklami kanye nokubala kubiza malini iluphu eguquguqukayo, kanye nokuthi indlela idluliswa yini ngaphezu kwezindawo ezilingisiwe ezingasethwa kabusha. I-MarkTechPost ibika ukuthi i-EnvHarness ayiwasekeli ama-akhawunti omsebenzisi abukhoma noma amarobhothi aphathekayo ngoba idinga izindawo ezilungisekayo.
Ukuphindaphinda okuzimele kuyisinyathelo esilandelayo esibaluleke kakhulu. I-MarkTechPost ibika ukuthi i-EnvHarness isebenze kahle kakhulu kunezindawo zangempela kuwo wonke ama-benchmarks ambalwa futhi yehlula i-generator eqondene nesizinda esithile ku-SWE-bench Iqinisekiswe ngamaphoyinti angu-2.46 ngenkathi isebenzisa izinyathelo ezimbalwa ezingu-5.11. Lokho kuqhathanisa kuncike emininingwaneni yokusetshenziswa, isampula yomsebenzi, ukwaziswa, izinguqulo zamamodeli nezinqubo zokuhlola ezingacacisiwe ngokugcwele kumbhalo womthombo. Ukukhiqiza kabusha kufanele kubonise ukuthi izinzuzo ziyaqhubeka yini ngaphansi kwesabelomali esilinganayo sekhompiyutha nakuyo yonke imisebenzi ekhethwe ngokuzimela, kungabi ngaphansi kokusetha kokuhlola kuphela.
Iluphu yomklami eguquguqukayo ingase futhi yethule ukwakheka kwezindleko ezintsha. I-MarkTechPost ibika ukuthi i-EnvRigger ibhala futhi ibuyekeze izisonga ze-Python ngemva kokuhlola ukukhishwa, futhi ihlonze amathokheni omklami njengezindleko zesistimu. Umthombo uthi ukusebenza ezindaweni ezingu-300 kufinyelele ku-54.79, uma kuqhathaniswa no-52.13 wendawo yasekuqaleni kanye no-50.37 kwezakhiwe, ngenxa yokuthi umklami uguqule iqoqo ngalinye ngenqubomgomo yamanje. Iphinde ibike ukuthi isabelo semisebenzi ngaphakathi kwebhendi yezinga lempumelelo okuhlosiwe sikhuphuke sisuka ku-6% saya ku-80%. Ukubika okwengeziwe kufanele kucacise ithokheni, isikhathi sokusebenza kanye nezindleko zobunjiniyela ngemuva kwale miphumela kanye nokuthi izinzuzo ziyazithethelela yini lezo zindleko.
Imingcele yendlela idinga ukunakwa okuseduze. I-MarkTechPost ibika ukwehla okuncane kokuhlukaniswa kwe-ALFWorld ngaphandle kokusatshalaliswa ngaphansi kokuqhathanisa okukodwa kwe-GRPO, nokusebenza okusuka ku-89.6 kuya ku-88.8 naphezu kokwenyuka kokusabalalisa kusuka ku-81.4 kuya ku-87.9. Lowo mphumela uphakamisa ukuthi ukuzivumelanisa nezimo kungase kuthuthukise ukusebenza ekusabalaliseni okuhlosiwe ngaphandle kokuqinisekisa ukwenziwa okuvamile okubanzi. Akwaziwa kusukela emthonjeni ukuthi ama-wrapers angakwazi yini ukulondoloza ukufaneleka kwebhentshimakhi ngaphansi kokusetshenziswa okuphikisayo, noma ngabe iziqinisekisi zangempela zifaka zonke izifundazwe ezisanda kwenziwa kabusha, kanye nendlela isistimu eziphatha ngayo ezindaweni ezinezenzo ezingenakuhlehliswa, abasebenzisi bangaphandle noma imiphumela yomzimba. Asikho isiqinisekiso esizimele sokulungela ukuthunyelwa esitholakala lapha.