Kwenzekeni
Iphepha elithunyelwe ku-arXiv ngomhlaka-22 Agasti liphakamisa i-ToSCA, uhlaka lokufundisa olunamazinga amabili okuqinisa ama-ejenti ezingxoxo. Ibeka isisekelo sokwenziwa kwempendulo yethokheni ngethokheni kuhlelo lokukhuluma lwezinga lokukhuluma, ukuhlanganisa ukuhlela okusezingeni eliphezulu nokukhiqizwa kolimi okusezingeni eliphansi.
Umthombo irekhodi le-arXiv elithi “ToSCA: Leveraging Hierarchical on Temporal and Strategic Abstractions of Conversation Agents,” ngababhali abayisishiyagalombili. Ithi iphepha lathunyelwa ngo-Agasti 22, 2026, futhi lamukelwa ku-EMNLP 2026 Findings. Isihloko esiqondile sephepha ukuqeqeshwa nokuhlolwa kwama-ejenti ezingxoxo e-AI, esikhundleni sengxoxo evamile yokufunda okuqiniswayo noma amamodeli olimi. Ngamanye amazwi, irekhodi lichaza ukwakheka kwe-ejenti ethile kanye nesifundo, kanye nezikhundla ezakha umbono wokuhlela wephepha.
Isistimu ehlongozwayo isebenzisa amazinga amabili esikhashana. Ezingeni eliphezulu, i-ejenti ikhetha isu lombhalo osobala ezingeni lokukhuluma. Ezingeni eliphansi, umenzeli unquma ithokheni yokuphendula ngethokheni kuyilapho ebeka leso sizukulwane kuhlelo olukhethiwe. Ababhali bafaka lo mklamo njengebhuloho phakathi kwezindlela zokufunda zokuqinisa ezisebenza kuphela kumathokheni nezindlela ezisebenza kuphela kumazwi aphelele. Ngakho-ke umehluko uphakathi kokukhetha isu elihlosiwe lenkulumo kanye nokuqaphela lokho kukhetha ngokusebenzisa amathokheni empendulo ngayinye.
Iphepha limodela umsebenzi njengenqubo yesinqumo sika-Markov enamazinga amabili. Ngokusho kwe-abstract, abacwaningi basebenzisa i-Q-network ejulile, noma i-DQN, yokuhlaziya izinga eliphezulu kanye nokwenza ngcono inqubomgomo ye-proximal, noma i-PPO, kumdlali ogxeka umlingisi wezinga eliphansi. Umthombo uchaza lokhu kuhlukaniswa njengokugqugquzelwa ukutholwa kokuvela kwethiyori kanye nokucatshangelwa kokusebenza kahle, kodwa awunikezi ngokuphuma noma imininingwane yokusetshenziswa kokubalulekile okunikeziwe.
Ukuze kubhekwane nemiklomelo eyingcosana, ababhali bethula indlela yomklomelo we-dual-granularity. Ihlanganisa isikolo sokwaneliseka kwezinga lezwi kanye nokukhuthazwa okungaphakathi kweleveli yethokheni kanye nenhlawulo ye-K-L. Imibiko engabonakali ihlola ezingxoxweni zansuku zonke nezingxoxo ezisekela imizwa, lapho i-ToSCA isebenze kahle kakhulu kunesethi yezisekelo ezingashiwongo ekunqumeni amasu kanye nekhwalithi yokuphendula. Umthombo uphinde uthi ukuqaliswa kuyatholakala, nakuba irekhodi elinikeziwe lingabandakanyi isixhumanisi sendawo.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Umsebenzi ubhekana nenselelo eyinhloko ku-AI yengxoxo: ukuxhumanisa imigomo yokuxhumana ebanzi namagama ngamanye akhiqizwa yi-ejenti. Uma kuqinisekiswa ngale kokuhlolwa okubikiwe, lokhu kuhlukaniswa kunganikeza indlela ehleleke kakhudlwana yokuqeqesha ama-ejenti ezingxoxo ezinezinyathelo eziningi futhi kwenze ukukhetha kwabo kwamasu kube lula ukuhlola.
Umehluko ohlongozwayo phakathi kwesu namagama ubonisa inkinga engokoqobo ezinhlelweni zezingxoxo. Impendulo ingaba eshelelayo ngenkathi iphishekela umgomo ongalungile wokusebenzelana, noma ingakhetha umgomo ophusile kodwa iwuveze kabi. Ngokwenza isu libe yisenzo esisobala esimaphakathi, i-ToSCA izama ukuhlukanisa la maphuzu amabili okuhluleka ngesikhathi sokuqeqeshwa nokuhlola.
Leso sakhiwo singaba usizo kumasistimu okulindeleke ukuthi aqhubekisele phambili izingxoxo izikhathi eziningi. Isu lezinga eliphezulu lingase limele umgomo wokusebenzelana, kuyilapho ukukhiqizwa kwezinga lamathokheni kusingatha ukukhetha kolimi lwasendaweni oludingekayo ukuze kuvezwe. Umthombo awuqinisekisi ukuthi i-ToSCA isebenza ezingxoxweni ezinde, kodwa idizayini ye-hierarchical ihambisana nemizamo yokwenza ama-ejenti ezingxoxo abe ngamabomu kakhulu futhi anganciki ezinqumweni ezihlukile zethokheni elandelayo.
Idizayini yomvuzo nayo ibalulekile. Ukufunda okuqinisiwe kwesizukulwane solimi kungathola impendulo kuphela ngemva kwempendulo ephelele, okwenza kube nzima ukukhomba ukuthi yiziphi izinqumo ezisize noma ezilimaza umphumela. Indlela yephepha ekabili-granularity izama ukunikeza impendulo kuwo womabili amaleveli ezwi nethokheni. I-abstract ayibonisi ukuthi ingabe lokhu kukhiqiza ukuqeqeshwa okuzinzile, izindleko eziphansi, noma ukuziphatha okungcono ngaphansi kokwaziswa okunzima noma okuphikisayo.
Imiphumela ebikiwe iyakhuthaza kodwa iboshiwe. Ziphathelene nokuhlolwa ezingxoxweni zansuku zonke nezeseka imizwelo futhi zethulwa ababhali bephepha. Umthombo onikeziwe awunikezi imiphumela yezinombolo, izikhawu zokuzethemba, osayizi bedathasethi, iphrothokholi yokuhlola komuntu, noma ukuphindaphinda okuzimele. Ngakho-ke isekela ukubika indlela kanye nokuqhathaniswa okufunwa ngababhali, kodwa hhayi isiphetho esibanzi sokuthi ukufunda kokuqinisa ukuhlelwa kwezigaba ngokuvamile kuthuthukisa i-AI yengxoxo.
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
Imiphumela isalokhu iyizicelo ezivela ephepheni elilodwa futhi kufanele ihlolwe ngokuzimela. Okungaziwa okubalulekile kufaka phakathi amasethi edatha anembile, isisekelo, izilinganiso zokuhlola, osayizi bomphumela, izindleko zokubala, ukusebenza kuzo zonke izilimi nezizinda, nokuthi ingabe izinzuzo ziyaqhubeka yini ezingxoxweni zomhlaba wangempela noma izilungiselelo ezizwelayo zokuphepha.
Inkinga yokuqala okufanele ibukwe ukuphindaphinda. Umthombo uthi ukusetshenziswa kuyatholakala, kodwa umbhalo onikeziwe we-arXiv awuyikhombi inqolobane noma uchaze ilayisense yayo, okuncikile, idatha yokuqeqeshwa, noma izimfuneko zezingxenyekazi zekhompuyutha. Abacwaningi abazimele bazodinga leyo mininingwane ukuze banqume ukuthi izinzuzo ezibikiwe zingaphinda zenziwe kabusha nokuthi ingabe indlela iyasebenza ngaphandle kokumiswa kwababhali.
Idizayini yokuhlola izobaluleka. Amagama angabonakali nsuku zonke nezingxoxo ezisekela imizwa kodwa awawahlonzi amasethi edatha, izilimi, inani lokujika, inani labantu ababambiqhaza, noma incazelo "yekhwalithi yokuphendula." Futhi ayisho ukuthi ukunqunywa kwamasu kwalinganiswa kanjani noma ukuthi abahloli babazi ukuthi yiluphi uhlelo olukhiqize impendulo. Lokho okweqiwe kuvula ithuba lokuthi ukusebenza kuyehluka kakhulu ngomsebenzi, isizinda, noma indlela yokuhlola.
Ukuphepha nokwethembeka kudinga ukucutshungulwa ngokukhethekile ezilungiselelweni ezisekela imizwa. Isu elithuthukisa umphumela wokwaneliseka lingase lingathuthukisi ukunemba kweqiniso, ukuphatha izinhlekelele, ukuvikelwa kobumfihlo, noma ukukhushulwa okufanele osizweni lwabantu. Umthombo awenzi izimangalo zokuphepha futhi awubiki ukuhlolwa kwezicelo eziyingozi, abasebenzisi abasengozini, amashifu okusabalalisa, noma ukwehluleka okubangelwa isu lezinga eliphezulu elingalungile.
Umsebenzi owengeziwe kufanele uhlole ukuthi ingabe isendlalelo sesu esicacile sithuthukisa ukwengamela kanye nokusebenza. Ubufakazi obuwusizo buzobandakanya ukuchithwa kwe-DQN, i-PPO, izingxenye zomvuzo, kanye nenhlawulo ye-K-L; ukuqhathanisa nezinhlelo zesimanje eziqinile; izilinganiso ze-latency nokubala; nokuhlola ngezingxoxo ezinde, zezilimi eziningi, nezomhlaba wangempela. Kuze kube yilapho ubufakazi obunjalo butholakala, i-ToSCA iqondwa kangcono njengesiphakamiso socwaningo esinezinzuzo zokuhlola ezibikiwe, hhayi impumelelo ebonisiwe yokukhiqiza.