Buyela Ezindabeni
UkuqambaAI Understanding ukwaziswa

Iphepha libika indlela yesikhashana yokuthola imibono engekho ezingeni lethokheni

Iphrinta ye-arXiv ichaza umtshina wokuphupha ohlanganisa izibalo zombhalo, amasiginali ahlanganisayo kanye nokumangala kwemodeli yolimi kuwo wonke ukulandelana esikhundleni sokwahlulela amathokheni ngokuzimela. Imodeli yayo ye-BiGRU ifinyelele ku-AUC engu-0.840 ku-RAGTruth, ngokusho kwephepha.

6 min readRead the primary source
Source-provided image accompanying Paper reports a temporal method for detecting hallucinations at the token level
Idokhumenti yomthombo oyinhlokoUmthombo urekhodiwe
Umshicileli
arxiv.org
Isixhumanisi somthombo
arxiv.orghttps://arxiv.org/abs/2608.18115
Uhlobo lomthombo
Idokhumenti eyisisekelo — isimemezelo esisemthethweni, iphepha, ukugcwalisa, noma ikhasi lomuntu wokuqala esilifunda ngokuqondile.
UmongoQonda lokhu ngemizuzwana engama-60

Qala lapha

Imigomo ebalulekile

Uphawu
Isiqephu sombhalo esicutshungulwe amamodeli olimi, njengesiqephu segama noma uphawu.
Isethi Yokuhlola
Idathasethi ebanjiwe esetshenziselwa ukukala ikhwalithi yemodeli ngemva kokuqeqeshwa.
I-hallucination
Uma imodeli ikhiqiza ulwazi olushelelayo kodwa olungamanga noma olungasekelwe.
ZihloleImibuzo Ecacisiwe yamamodeli e-AI

Kwenzekeni

Iphrinta entsha ye-arXiv iphakamisa ukuphatha imibono engemihle njengobubanzi obuvela ngokulandelana, kunokuba kube njengamaphutha eleveli yamathokheni ahlukanisiwe. Indlela ihlanganisa izici ezingu-33 ezivela embhalweni okhiqiziwe, inkomba yolimi lwemvelo kanye nokumangala kwemodeli yolimi, bese izicubungula ngemodeli yokulandelana. Iphepha libika ukuthi i-GRU eqondiswa kabili ifinyelele ku-AUC engu-0.840 ku-RAGTruth ezimbewuni ezingu-10, inzuzo enamaphuzu angu-11 phezu kwesisekelo esizimele sokuhlehla kwezinto.

Umthombo uyirekhodi le-arXiv le-"Temporal Multi-Signal Fusion for -Level Detection," ebhalwe ngu-Igor Itkin futhi yathunyelwa ngoJuni 30, 2026. Isisekelo esiyinhloko saleli phepha ukuthi ukubona izinto ezingekho emthethweni kuvame ukunwetshwa okwesikhashana esikhundleni seqoqo lamathokheni amabi azimele. Ngakho-ke ifaka ukutholwa kozimele njengokulebula kokulandelana: ithokheni ngayinye ithola isikolo kusukela kuchungechunge lwezici, kuyilapho imodeli ingasebenzisa ulwazi oluvela ezindaweni eziseduze ukuze inqume ukuthi ingabe i-span ayisekelwe noma ayilungile.

Ukusakaza kwesici okuhlongozwayo kunobukhulu obungu-33. Ngokuvumelana ne-abstract, lezo zici zezibalo zombhalo we-fuse, i-natural-language-inference-inference kanye nokumangalisa kwemodeli yolimi. Umtshina awusebenzisi ukwenza kusebenze kwangaphakathi kwemodeli ekhiqizayo noma ezinye izinto zangaphakathi zobunikazi. Iphepha lihlola iyunithi ephindaphindayo enesango eliphindwe kabili phezu kwalezo zici futhi libika indawo engaphansi kwejika lesici sokusebenza komamukeli elingu-0.840 kudathasethi ye-RAGTruth, kusetshenziswa imbewu engu-10. Iqhathanisa lowo mphumela nesisekelo esizimele sokuhlehla kwezinto futhi ibike ukuthuthukiswa kwamaphoyinti angu-11, ngevelu engu-p engu-0.002 kusukela kuhlolo lwezinga elisayiniwe le-Wilcoxon.

Iphepha liphinde libike ukuhlolwa kokubola okulawulwayo okuhloselwe ukuhlukanisa inani lokuhleleka kwesikhashana kunani lemodeli enamandla kakhulu. Ukuhumusha kwakho ukuthi inzuzo enkulu ivela ekwakhiweni kwesikhashana esikhundleni samandla okuyimodeli: izikhundla ezizethembayo zingadlulisa ubufakazi ezikhundleni ezingomakhelwane ezingacacile ngaphakathi kwenkathi ekhohliwe. Uphahla olufanayo lokusebenza olucishe lube ngu-0.845 lubikwe kuzo zonke izakhiwo eziphindaphindwayo, zendawo yesifunda kanye nokunaka, okuhlanganisa iMamba namamodeli asekelwe ukunaka. Ababhali basebenzisa lolo phahla oluphindaphindayo ukuphikisa ngokuthi isici esikhawulelayo yisethi yesici esikhethiwe, hhayi ukwakheka okulandelanayo okuthile.

Umthombo uqhube wathi umtshina ungasebenza embhalweni okhiqizwe amamodeli emithombo evaliwe ngoba ufunda umbhalo okhiqiziwe kanye namasiginali angaphandle kuphela. Iphinde ibike ukuthi umtshina uyaqhubeka nokusebenza ngombhalo ovela kumamodeli olimi angekho kudatha yawo yokuqeqeshwa, ngokulahlekelwa okungaphansi kuka-4% ku-AUC. Lezi izimangalo ezenziwe ngokuphrinta kwangaphambili okukodwa. Umthombo onikeziwe awunikezi amagama emodeli ahloliwe, imininingwane yokwakhiwa kwedathasethi, iphrothokholi yokuhlola isitimela, inqubo yomkhawulo, ukunemba, ukukhumbula, ukulinganisa, ukubambezeleka noma izibonelo zamaphutha. Futhi ayisunguli ukuqinisekiswa okuzimele kokutholwe.

Imininingwane yomthombo: arxiv.org

Kungani kubalulekile

Le ndlela iklanyelwe ukusebenza ngaphandle kokufinyelela ezifundeni zangaphakathi zemodeli, okungenza isebenze kumasistimu wemithombo evaliwe. Uma umphumela obikiwe ubambe ngaphandle kwesilungiselelo sokuhlola sephepha, kungasiza ukukhomba izikhathi ezingabazekayo ezimpendulweni ezithuthukisiwe zokubuyiswa kanye nokubuyekezwa kosekelo noma ukugeleza komsebenzi wokuqinisekisa. Umphumela uhlala ungowokuqala: umthombo uyi-arXiv ephrintiwe ngaphambili, futhi irekhodi elinikeziwe alisunguli ukuphindaphinda okuzimele noma ukusebenza kokukhiqiza.

Ukubaluleka okungokoqobo imodeli yokufinyelela ehlongozwayo yomtshina. Amasu amaningi okuqapha amamodeli ancike ekwethulweni kwangaphakathi, emathubeni amathokheni noma olunye ulwazi olutholakala ku-opharetha oyimodeli kuphela. Umtshina osuselwe kumbhalo okhiqiziwe kanye namasiginali angaphandle, empeleni, ungabekwa eduze kwemodeli okungeneki kuyo okwangaphakathi. Lokho kwenza umbono uhlobane ezinhlanganweni ezisebenzisa amamodeli olimi olusingathiwe, nakuba umthombo ungabonisi ukuthi luhlanganiswe nesevisi ebukhoma noma luhlolwe ngaphansi kwethrafikhi yokukhiqiza.

Uhlaka olususelwa ekulandeleni luphinde lubhekane nomkhawulo wangempela ekuqhathanisweni kwephepha: ithokheni elibukeka lijwayelekile uma lihlukaniswa lingase libe yingxenye yesimangalo eside esingasekelwe. Ukusebenzisa ubufakazi obungomakhelwane kungavumela isistimu ukuthi ihlabe umkhosi indinyana ukuze ibuyekezwe esikhundleni sokwethula uhlu olude lwezikolo zamathokheni anqanyuliwe. Ekugelezeni komsebenzi okuthuthukisiwe kokubuyisa, isignali enjalo ingase isetshenziselwe ukucela ubufakazi obengeziwe, ukuhambisa impendulo eya kumuntu, noma ukucindezela isikhathi esingaqinisekile kakhulu. Lezi izinhlelo zokusebenza ezingaba khona eziqondiswe endleleni, hhayi ukusetshenziswa okuboniswa umthombo.

Ukuthuthukiswa kwe-AUC okubikiwe kuyaphawuleka phakathi kokuhlolwa okushiwo ngoba kuphakamisa ukuthi ulwazi loku-oda lungase lubaluleke ngaphezu kokumane umiselele isisekelo somugqa ngomtshina omkhulu. Uphahla lwe-cross-architecture luwusizo futhi njengokutholwa kocwaningo: uma amamodeli amamodeli ahlukene ehlangana eduze kwamaphuzu afanayo, ukungeza umthamo kukodwa kungase kungabheki amaphutha asele. Incazelo yephepha uqobo ikhomba ekuthuthukiseni amasiginali angaphansi, njengekhwalithi yezinto ezifakiwe noma ezimangazayo, kunokuthatha ngokuthi imodeli ehlukile yokulandelana izoxazulula inkinga.

Imikhawulo ibaluleke ngokulinganayo. I-AUC ka-0.840 ifingqa ukusebenza kwezinga phakathi kwemikhawulo; ayisho ukuthi mangaki ama-allucinations angabanjwa ngezinga lesexwayiso sokusebenza, mangaki amaphaseji alungile angamakwa ngokungalungile, noma ukuthi amaphuzu omtshina alinganiswa njengamathuba. I-RAGTruth ingase ingameli yonke indawo yesihloko noma ifomethi yempendulo. Ngenxa yokuthi okokusebenza okugunyaziwe lapha kuwukuphrinta ngaphambili kwe-arXiv eyodwa, umphumela awusungulwa ngokuzimela irekhodi elinikeziwe. Abafundi kufanele baphathe okutholakele njengobufakazi bokuhlolwa okwengeziwe, hhayi njengobufakazi bokuthi ukutholwa kokubona izinto ezingekho kuxazululiwe.

Interactive Mechanism

I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela

Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.

Agent Lifecycle Stage:
1
User Intent & Planning: "Audit customer refund request #4092 and settle payment."
2
Tool Calling: Emits structured JSON call crm_get_transaction(id='4092').
3
Guardrail & Verification:🛡️ Paused: High-value action requires human operator sign-off.
4
Final Settlement: Refund recorded, email receipt dispatched, and audit log stored.
Core takeaway: An AI agent is not just a language model—it is a closed loop of planning, tool invocation, and environment feedback. Production systems require self-healing retries and strict human approval guardrails.
I-Interactive Concept Check+10 Points
AI Models Explained Quiz

In AI, what are a model's "parameters"?

Ongakubuka ngokulandelayo

Ukuhlola okulandelayo okubalulekile ukuphindaphinda kumasethi edatha engeziwe, ukuhlolwa kuzo zonke izizinda nemindeni eyimodeli, kanye nokukalwa kokuphozithiyo okungamanga, ukulinganiswa nokubambezeleka. I-abstract yephepha ayicacisi amamodeli ahloliwe, umkhawulo wokusebenza komtshina, ukunemba kwawo noma ukukhumbula, noma ukuthi ukusebenza kushintsha kanjani lapho ubufakazi bungaphelele, buyindida noma buphikisana. Leyo mininingwane izonquma ukuthi ingabe indlela iyisivikelo esiwusizo noma ngokuyinhloko iwumphumela wokulinganiswa.

Ukuphindaphinda kufanele kuqale ngephrothokholi eqondile yephepha bese kuyandisa. Abacwaningi kufanele babike imiphumela kumadathasethi amaningi wokukhohlisa, babhale phansi amamodeli olimi asetshenziselwa ukukhiqiza nokuhlola umbhalo, kanye nezizinda zokuhlola lapho amaphutha ethwala imiphumela ehlukene. Ukuqhathanisa kufanele kufake izisekelo ezilula zesichazamazwi kanye nezihlanganisayo, izihlukanisi ezilinganiselwe nezindlela ezisebenzisa ulwazi lwemodeli-yangaphakathi uma lolo lwazi lutholakala. Umbuzo obalulekile ukuthi ingabe inzuzo ebikiwe iyasinda yini ekushintsheni kwedatha, amajeneretha nezinqubo zezichasiselo.

Izilinganiso zokusebenza zizobaluleka njenge-AUC ehlanganisiwe. Ukuhlola okuzayo kufanele kushicilele ukunemba nokukhumbula emikhawulweni ethile yesexwayiso, amajika okulinganisa, ukugqagqana kwezinga le-span namalebula abantu, kanye nesikhathi nezindleko zekhompuyutha zokukhiqiza isinqumo ngasinye. Umtshina ohlonza izifunda ezibanzi kodwa ongakwazi ukuhlukanisa ukungaqiniseki okungenabungozi esitatimendeni esingamanga esiwumphumela singase sibe nzima ukusisebenzisa. Umthombo awusho ukuthi ingabe indlela ihloselwe ukutholwa kokusakaza-bukhoma, ukubuyekezwa kwesizukulwane sangemuva noma kokubili, ngakho ukubambezeleka kokuphakwa kanye nephuzu lapho isistimu ingangenela khona kuhlala kungaziwa.

Ukuhlolwa kokuqina kufanele kuhlole ubufakazi obungaphelele bokubuyiswa, imithombo ephikisanayo, izimangalo ezishiwo ngamagama, izimpendulo ezinde kanye nombhalo oklanywe ngamabomu. Njengoba umtshina usebenzisa amasignali angaphandle, ikhwalithi nokuzimela kwalawo masignali kungase kulawule ukusebenza. Imodeli yolimi ingase futhi iguqule isitayela sayo, ukulinganisa noma amaphethini amaphutha ngemva kokusetshenziswa, okungenzeka yenze buthaka umtshina oqeqeshwe emikhiqizweni emidala. Iphepha libika ukulahlekelwa okungaphansi kuka-4% kwe-AUC kumamodeli olimi angabonakali, kodwa i-abstract enikeziwe ayichazi ukuhlukaniswa kwemodeli noma ibonisa ukuthi lowo mphumela udlulela ezizindeni ezingabonakali kanye nokushintsha amasistimu okubuyisa.

Okokugcina, isilingi sokusebenza esiqhubekayo sephepha sidinga ukucutshungulwa. Ingase ibonise umkhawulo wangempela emindenini emithathu yamasignali, noma ingase ibonise idathasethi, amalebula noma umklamo wokuhlola. Iphepha eligcwele kufanele licacise ukuthi izikhala ze- zilebulwe kanjani, ukuthi ubufakazi bukhethwa kanjani, nokuthi noma yiluphi ulwazi oluvela kusethi yokuhlola lungangena ezicini. Ukusetshenziswa okuzimele nokuhlola okulindelwe ezimpendulweni zangempela ezibhekene nabasebenzisi kuzonikeza ubufakazi obunamandla kunokufakwa esikhundleni okwengeziwe kwezakhiwo. Kuze kube yileso sikhathi, okungaziwa okuwumphumela omkhulu akukhona ukuthi umtshina ungalinganisa izibonelo kubhentshimakhi ebikiwe yini, kodwa ukuthi ungathuthukisa yini izinqumo ngokwethembeka kuzilungiselelo lapho izitatimende zemodeli ezingasekelwe zidala ukulimala okungokoqobo.

Imihlahlandlela ehlobene nemibuzo

Amamodeli e-AI AchaziweChatGPT ne-LLMsUkuqeqeshwa kwe-AIUkuziphatha kwe-AIHlola okwaziyo — zama imibuzo ye-AI yamahhalaBheka igama le-AI kuhlu lwethu lwamagama
Uthole lokhu kuwusizo?