Kwenzekeni
Ithimba labacwaningi lihlongoze i-Consilience, uhlaka lwe-orchestration lwesikhathi sokucabanga lwezinhlelo eziningi zemodeli yezilimi zama-agent amakhulu. Ilandelela ukungaqiniseki, ukungaboni ngaso linye, ukuzuza kobufakazi, ukuphelelwa umsebenzi kanye nokuvumelana ngaphambi kwesikhathi, bese ikhetha ukuthi ama-ejenti kufanele aphikise, acacise, afune ubufakazi noma ahambise umzila ingxoxo, kanye nokuthi iyiphi i-ejenti okufanele ikhulume. Iphepha libika imiphumela kuwo wonke amamodeli olimi angu-12 anesisindo esivaliwe futhi esivaliwe emisebenzini yesitayela se-HiddenBench.
Iphrinta ye-arXiv, ethunyelwe ngo-Aug. 20, 2026, igxile ekucabangeni okufihliwe kwama-ejenti amaningi. Kulesi silungiselelo, i-ejenti yemodeli yolimi ngayinye ibona ingxenye yobufakazi obudingekayo ukuze kuthathwe isinqumo esifanele. Ababhali baphikisana ngokuthi izindlela zokuxhumanisa ezivamile-amashejuli ahleliwe, ukushintshaniswa kwe-round-robin kanye nenkulumo-mpikiswano engahleliwe-akunikezi isiqinisekiso sokuthi isenzo esithile sokuxoxa sifanelekile. Ukuvumelana kwethulwa njengendlela yokulawula lokho kushintshana ngesikhathi, kunokuba njengemodeli yolimi olusha noma inqubo entsha yokuqeqesha.
Uhlaka lugcina isimo sezingxoxo ezihlangene. Ngokusho kwe-abstract, leso simo sifinyeza ukungaqiniseki, ukungaboni ngaso linye, ukuzuza kobufakazi, ukungadingeki kanye nokuvumelana ngaphambi kwesikhathi. Esikhathini ngasinye, isistimu isebenzisa lezo zimpawu ukuze ikhethe kokubili ukungenelela nesipika. Ukungenelela okukhona ukuphonsela inselelo isikhundla esikhona, ukucacisa ulwazi, ukufuna ubufakazi obengeziwe noma ukuhambisa ingxoxo. Lokhu kwenza ukukhetha kokuxhumana kube yinkinga yokulawula esobala: isistimu inquma ukuthi ingxoxo kufanele iqhubeke nini, hlobo luni lokushintshisana oludingekayo nokuthi iyiphi i-ejenti esesimweni esihle kakhulu sokuyihlinzeka.
Isimangalo sobuchwepheshe esimaphakathi sephepha siyindlela yokulinganisa ehlakaniphile ehambisanayo. Ababhali bathi inikeza isiqinisekiso samahhala sokusabalalisa, esinesampula-esikhawulelwe: okunemibandela engxoxweni efinyelela umzuliswano othile, ukuzisola okuyisinyathelo esisodwa sesenzo esihlongozwayo somlawuli kuboshwe umkhawulo olinganisiwe onamathuba asemaceleni okungenani angu-1 ukususa i-alpha. Indlela yokwamukela isuke isetshenziselwa ukuphoqelela isiqinisekiso esifanayo sesenzo esenziwe ngempela, esikhundleni seziphakamiso ezithathwa njengezingamukeleki ngokomthetho wokulinganisa. Umthombo wethula lokhu njengesendlalelo sesitifiketi sezinqumo zokuxhumana, hhayi njengesiqinisekiso sokuthi yonke impendulo yokugcina ilungile.
Ababhali babika ukuhlolwa kwemisebenzi yephrofayela efihliwe yesitayela se-HiddenBench ehlanganisa amamodeli ezilimi angu-12 anesisindo esivulekile nesivaliwe. I-abstract ithi Consilience ithuthukisa ukunemba kwesinqumo nokusebenza kahle kokuxhumana uma kuqhathaniswa nezivumelwano zezingxoxo ezingaguquki nezingahlelekile, futhi ngezinye izikhathi zeqa isisekelo solwazi olugcwele lapho yonke i-ejenti ibona bonke ubufakazi. Umthombo awunikezi usayizi wezinombolo walokho kuthuthukiswa kokuqukethwe okunikezwe lapha. Ihlonza ukuthunyelwa njengokuphrinta kusengaphambili okungamakhasi angama-39 okunezibalo ezimbili, amathebula ayisishiyagalolunye, isithasiselo esiqukethe ubufakazi Besiphakamiso 1, imiphumela eyandisiwe, ukukhishwa kanye nezifanekiso eziphelele ezisheshayo. Umthombo awuqinisekisi ukuthi umsebenzi ubuyekezwe kontanga.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Umsebenzi ubhekana nenkinga engokoqobo kuma-agent amaningi we-AI: ukwengeza ama-ejenti amaningi noma ulwazi olwengeziwe akukhiqizi izinqumo ezingcono ngokuzenzakalelayo. Umnikelo we-Consilience oceliwe uyindlela yokulinganisa nokwamukela okuhloswe ngayo ukwenza ukukhetha kokuxhumana kuthembeke kakhulu futhi kunikeze isibopho esilinganisekayo ekuzisoleni kwesinyathelo esisodwa somlawuli.
Amasistimu we-AI enama-agent amaningi avame ukuhlukanisa ulwazi phakathi kwezibonelo ezimbalwa futhi athembele engxoxweni ukuze ayihlanganise. Lelo lungiselelo lidala inkinga yokuxhumanisa ehlukile emandleni anoma iyiphi imodeli ngayinye. I-ejenti ingase idinge ukucela ubufakazi obungekho, inselele isiphetho esibuthakathaka noma iyeke ukuphinda ulwazi oseluvele lwabiwe. Ukuvumelana kuqondise leso singqimba ngokuqondile, okwenza ukukhetha kwesenzo sokuxhumana kube yingxenye yenqubo yesinqumo yesistimu.
Ukuqhathanisa okubikiwe nesisekelo solwazi olugcwele kuyaphawuleka ngoba kukhomba ekuhwebeni okungase kube khona ukusebenza kahle. Uma umphumela wephepha uba jikelele, ukuxhumana okulawulwa ngokucophelela kwesinye isikhathi kunganxephezela ama-ejenti angakwazi ukufinyelela kuwo wonke ucezu lobufakazi. Lokho kungabaluleka kumasistimu aklanywe eduze kolwazi olusabalalisiwe noma olubucayi, lapho ukwabelana ngabo bonke ubufakazi ne-ejenti ngayinye kungase kubize noma kungafuneki. Umthombo, nokho, awukhombisi ukuvikeleka kobumfihlo noma uqinisekise ukuthi i-Consilience ivimbela ulwazi olubucayi ukuthi lungadalulwa.
Isimangalo sokulinganisa singase futhi sinikeze abaklami besistimu indlela ecacile yokuhlola ukuziphatha kwe-orchestration. Ishejuli yengxoxo engaguquki ingaba lula ukuyisebenzisa kodwa ingase ichithe amajika ekuhwebeni okungafuneki noma okuqondiswe kabi. Isilawuli esiqinisekisiwe singanikeza umkhawulo osemthethweni wokwenqaba izenzo ezithile ezihlongozwayo, okuvumela ama-opharetha ukuthi ahlukanise isinqumo sokuxhumana esamukelwe kwesingagunyaziwe. Isiqinisekiso esichazwe ku-abstract siphathelene nokuzisola okuyisinyathelo esisodwa ngaphansi kwezimo zokulinganisa ezishiwo; akufanele ifundwe njengesiqinisekiso sokwethenjelwa esijwayelekile sohlelo oluphelele lwama-agent amaningi.
Inani elingokoqobo lihlala lingelesikhashana. Ubufakazi obubikiwe buvela emisebenzini yesitayela se-HiddenBench kanye nesethi yamamodeli ezilimi angu-12, ezingenazo izibalo emthonjeni onikeziwe ezibonisa imiphumela yesisekelo esiphelele, izindleko ze-orchestration, ukubambezeleka, amacala okuhluleka noma ukusebenza emthwalweni wokusebenza. Futhi akucaci ukuthi indlela ibucayi kangakanani ekukhethweni kwe-alpha, idatha yokulinganisa, isimo sesifinyezo noma imiksi yemodeli eyisisekelo. Leyo mikhawulo yenza lokhu kube isiqondiso socwaningo esingase sibe usizo, kunobufakazi bokuthi ukuxhumanisa kwe-AI yama-agent amaningi kuyaxazululwa.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
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 evulekile ukuthi ingabe izinzuzo ezibikiwe zibambelela ngale kwezilungiselelo zebhentshimakhi zephepha, kungakanani ukubala nokubambezeleka kwe-orchestration eyengezayo, nokuthi ukulinganisa kuziphatha kanjani lapho imisebenzi, amamodeli noma ukusatshalaliswa kobufakazi kushintsha. Umthombo awunikezi izinzuzo zokunemba kwezinombolo, izibalo ezisebenza kahle noma imiphumela yomhlaba wangempela yokusebenzisa.
Okubalulekile kokuqala irekhodi eligcwele lokuhlola. Imibiko engabonakali ithuthukisa ukunemba nokusebenza kahle kokuxhumana kodwa ayisho ubukhulu bomphumela, izibalo zemisebenzi, izikhawu zokuzithemba noma incazelo enembile yokusebenza kahle. Leyo mininingwane izonquma ukuthi ingabe umphumela uwukuthuthukela okuphusile noma umphumela omncane wokulinganisa. Ukukhishwa kufanele kubonise ukuthi yikuphi ukuguquguquka kwesimo nokungenelela okunomthelela omkhulu kumphumela.
Abacwaningi nabakhi bohlelo kufanele bahlole izimo zokusebenza zesiqinisekiso. Isibopho esishiwo sinemibandela ekufinyeleleni umzuliswano wezingxoxo futhi sichazwa njengesiqinisekiso samathuba amancane. Akucaci emthonjeni onikeziwe ukuthi idatha yokulinganisa iqoqwa kanjani, ukuthi ama-threshold ashintsha kanjani kuyo yonke imizuliswano noma ukuthi kwenzekani lapho ukusatshalaliswa komsebenzi obukhoma kuhluka ekusabalazweni kokulinganisa. Ukuqina ngaphansi kwezizinda ezintsha, ubufakazi obuphikisayo kanye nokushintsha kwamamodeli abantu kungumbuzo obalulekile ongaphenduliwe.
Izindleko zokuqalisa zingolunye udaba oluvulelekile. Ukuvumelana kwengeza ukulandelela isimo, ukukhethwa kwesenzo, ukukhethwa kwesipikha nokwamukelwa kwesiphakamiso ohlelweni oseluvele lwengxoxo. Isistimu ingathola ukunemba kuyilapho ilahlekelwa inani elingokoqobo uma lezo zilawuli zidinga amakholi wamamodeli amaningi, amathokheni noma isikhathi. Ukubika okuzayo kufanele kulinganise lezo zindleko ngokuqhathanisa nokongiwa kokuncishiswa kokuphinda kuqhathaniswe indlela nemithetho elula yomzila noma yokumisa.
Okokugcina, umsebenzi udinga ukuhlolwa ngaphandle kwezilungiselelo zesitayela sebhentshimakhi. Umthombo awubiki ukuthunyelwa, ukugadwa komuntu, ukuphathwa kwedatha ebucayi noma imiphumela ezinqumweni eziphakeme. Ukuphindaphinda okuzimele kuzosiza ukuthola ukuthi okutholakele kuncike ezifanekisweni ezisheshayo zababhali, ukukhethwa kwemodeli noma ukwakhiwa kwebhentshimakhi. Kuze kube yileso sikhathi, leli phepha liqondwa kangcono njengesiphakamiso esisemthethweni esinemiphumela ethembisayo ebikiwe kanye nokungaqiniseki okuzwakalayo mayelana nokwenza okuvamile.