Kwenzekeni
Abacwaningi bethule i-τ^τ-Bench, ibhentshimakhi elenza ukwakhiwa kwe-ejenti yokuphela ukuya ekupheleni umsebenzi wezinhlelo zekhodi ze-AI. Ibhentshimakhi inikeza amarekhodi ebhizinisi lomenzeli kanjiniyela, izidingo zeklayenti, i-API yokukhiqiza, i-codebase ekhona, kanye nemikhawulo yamamodeli nezindleko zokunikezela, bese ihlola i-ejenti yesevisi yekhasimende ewumphumela ngokumelene nabasebenzisi abalingisayo.
Umthombo we-arXiv, othunyelwe ngoSepthemba 4, 2026, uchaza i-τ^τ-Bench njengendawo yokuhlola amasistimu e-AI akha ama-ejenti kunokumane aphendule ukwaziswa kwebhentshimakhi. I-ejenti yonjiniyela ithola amarekhodi ibhizinisi eliwagcinayo, izidingo ezivela kuklayenti, i-API yokukhiqiza okufanele imisebenzi iqhutshwe ngayo, i-codebase ezuzwe njengefa, kanye nemikhawulo ekunikezeni izindleko nokukhetha imodeli. Kufanele isebenzise lezo zinsiza ukuze ilethe i-ejenti yesevisi yekhasimende ephelele.
I-ejenti ewumphumela iyahlolwa ngokuyisebenzisa ngokumelene nabasebenzisi abalingisi ababambekile. Kuyo yonke imisebenzi engama-53 ezizindeni ezine, iphepha libika ukuthi ukucushwa kwalo okuqine kakhulu—Claude Opus 5 esetshenziswa nge-Claude Code—kudlule u-23.9% wezifaniso zokuhlola. Isithenjwa esibhalwe ngochwepheshe sithole u-82.2%. Lena imiphumela ebikwe yiphepha; umthombo awunikezi ukuqinisekiswa okuzimele ekhasini lokubikezela le-arXiv.
Ababhali bahlonza amaphethini okwehluleka abathi afana nezinkinga ezihlangatshezwana nabathuthukisi be-ejenti abangabantu: imibuzo engajulile esikhundleni sokuqonda okujulile kwamarekhodi ebhizinisi, ukuxhumana okuncane neklayenti, kanye nokuhlola okwanele ngezakhiwo ze-ejenti noma izindleko zokuhlinzeka. Babeka uphawu lokulinganisa njengomzamo wokwenza i-ejenti ebambisene yakhe ibe yithagethi elinganisekayo yama-ejenti amakhodi.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Ibhentshimakhi iqondise igebe ekuhlaziyeni kwamanje: ukuthi uhlelo lwe-AI lungakwazi yini ukuletha i-ejenti esebenzisekayo ngaphakathi kwezingqinamba ezingcolile zokusebenzelana kwekhasimende langempela. Igebe lokusebenza elibikiwe phakathi kokucushwa okuvivinywe okuqine kakhulu kanye nesithenjwa sochwepheshe siphakamisa ukuthi ikhono lokubhala ikhodi lilodwa aliqinisekisi ikhono lokuqonda idatha yebhizinisi, ukuxhumana namaklayenti, ukwenza ukukhetha kwezakhiwo, noma ukuphatha izindleko zokusebenza.
Ukuhlola okuningi kuhlukanisa ikhono lemodeli lokukhiqiza ikhodi noma ukuqedela umsebenzi ocaciswe kancane. τ^τ-Ibhentshi esikhundleni salokho ihlola uchungechunge lwezinqumo olunquma ukuthi i-ejenti ingasebenza yini esimweni sokusebenza: ukuhumusha idatha yenhlangano engaphelele, ukuhumusha izidingo zeklayenti zibe ukuziphatha, ukuhlanganisa nesistimu ekhona, ukukhetha amamodeli, nokulinganisa ikhwalithi ngokumelene nezindleko. Lokho kwenza igebe elibikiwe libe wusizo emaqenjini anquma ukuthi kungakanani ukugeleza komsebenzi wokwakha ama-ejenti asakudinga.
Imiphumela iphinde inikeze indlela ebambekayo yokuhlola izimangalo zokuthi ama-ejenti okubhala amakhodi angangena esikhundleni noma enze ngokuzenzakalelayo umsebenzi wokuthuthukisa isoftware kumasistimu e-AI. Ngokusho komthombo, izinhlelo ezihloliwe zivame ukukhiqiza into egijima kodwa ehluleka ukuhlangabezana nezidingo ezijulile zokuthembisana. Ngakho-ke ibhentshimakhi isusa ukunaka kusuka ekukhiqizeni ikhodi iyodwa kuya kukhwalithi yesistimu esetshenzisiwe kanye nokusebenzisana kwayo nabasebenzisi.
Umphumela uhlala uboshiwe. Ikhasi lokubikezela alinikezi imininingwane mayelana nokwakhiwa komsebenzi webhentshimakhi, idizayini yesifanisi, inqubo yokushaya amaphuzu, ukukhetha okuyisisekelo, noma ukungaqiniseki kwezibalo. Futhi ayibonisi ukuthi amaphuzu angu-23.9% abikezela imiphumela yokukhiqiza. Iphepha liyi-arXiv eliphrintiwe ngaphambili, ngakho-ke izimangalo zalo kufanele zithathwe njengemiphumela yocwaningo kusalinde ukucutshungulwa okwengeziwe nokuphindaphindwa.
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
Iphepha aliqinisekisi ukuthi τ^τ-Ibhentshi liqhathaniswa kanjani namanye amabhentshimakhi, noma ngabe abasebenzisi balo abalingisi babikezela ukusebenza namakhasimende angempela, noma ukuthi imiphumela ivele ibe jikelele ngale kwemisebenzi ebikiwe engama-53 nezizinda ezine. Umthombo futhi awukubhalisi ukufinyelela kwebhentshimakhi yomphakathi, izidingo zokusetshenziswa, intengo, noma ukuphindaphinda okuzimele.
Ukuthi ababhali bayayikhipha yini ibhentshimakhi, ukucaciswa komsebenzi, izintambo zokuhlola, kanye nokusetshenziswa kwereferensi kubalulekile ekwenziweni kabusha. Ikhasi lomthombo liqinisekisa iphepha kanye nezixhumanisi ezinguqulweni ze-PDF ne-HTML, kodwa alisho ukuthi ibhentshimakhi ngokwayo ifinyeleleka esidlangalaleni noma ikhomba noma yiziphi izimo zokufinyelela noma intengo.
Ukuhlola okuzayo kufanele kuhlole amamodeli engeziwe, amasistimu e-ejenti yekhodi, izizinda, nezinhlobo zomsebenzi, kuyilapho kucacisa ukuthi abasebenzisi abalingisi bamele kanjani ukuziphatha kwangempela kwamakhasimende. Ukuqhathanisa ne-ejenti ekhona, amakhodi, kanye namabhentshimakhi wobunjiniyela besofthiwe kungasiza ekutholeni ukuthi yimaphi amandla engeziwe τ^τ-Izilinganiso zebhentshi.
Imikhawulo ebikiwe ikhomba ezidingweni zokubuyekezwa okusebenzayo: hlola ukuthi ama-ejenti abuza kanjani amarekhodi enhlangano, adinga ukuxhumana namaklayenti okusobala, uhlole ezinye izakhiwo, futhi uqaphe izindleko zokuphakela ngaphambi kokuthunyelwa. Umthombo awukubiki ukuthunyelwa komhlaba wangempela, imiphumela yamakhasimende, noma izehlakalo zokuphepha, ngakho-ke leyo miphumela ihlala ingaziwa.