Amaluphu Wokuklama, Hhayi Ukwaziswa | Mayelana neSayensi Yedatha

“Asisayibhali imiyalo. Sidizayina izihibe.” – othile e-Anthropic ngoJuni 2026

i-agent loop, ukuzigxeka akwenzanga ngcono kunokungenzi lutho. Isiqinisekisi esinqunyiwe, esinehange emthonjeni sinciphise izinga lokubona izinto ezingekho cishe phakathi.

Ulayini usuka emasontweni ambalwa edlule futhi usuvele uzizwa uyiqiniso. Siyekile ukushuna ukwaziswa okukodwa okuhle futhi saqala ukwakha amasistimu azamayo, ahlole umsebenzi wawo, futhi athuthuke ngezinyathelo ezimbalwa. Imodeli engabuyekeza ibaluleke kakhulu kunemodeli ephendula kanye bese ima. Kulokhu, umugqa ulungile.

Okushiya ngaphandle umthethosivivinywa. Iluphu inzima kakhulu ukuyiqinisekisa kunocingo olulodwa: ngekholi eyodwa uhlola okukhiphayo okukodwa, kodwa ku-loop zonke izinyathelo zingakhukhuleka, futhi izindlela ezingahamba kahle ziphindaphindeke ngokuphindaphinda ngakunye. Ingxenye enzima iyeka ukuba isizukulwane. Iba ukuqinisekiswa. Noma, uma uthanda: ukwazi ukuthi iluphu iyithola kahle yini. Futhi indlela ezenzakalelayo yokuqinisekisa – vumela imodeli ihlole umsebenzi wayo – iphenduka isixhumanisi esibuthakathaka kakhulu kuchungechunge.

Ngakho-ke lokhu akukona ukuxabana no “idizayini loops, hhayi imiyalo.” Iwukubamba okufihlayo, kukalwa: ukuhlola okungiqinisekise, ngezinombolo kanye nendlela, ukuze ukwazi ukuzihlolela ngokwakho.

Indawo yokuqinisekisa ikhula ngesinyathelo ngasinye

Ucingo olulodwa lunendawo eyodwa lapho lungalungi khona: impendulo. Iluphu yezinyathelo ezintathu inohlaka lokuqala, ukugxeka okusalungiswa, ukubuyekezwa, ukuhluzwa kwesibuyekezo, kanye nesinqumo sokumisa. Ngayinye yalezo iwumphumela wemodeli, futhi ngayinye ingaba nephutha ngokuzethemba. Awuzange ususe inkinga yokuqinisekisa ngokungeza iluphu. Uyiphindaphinde.

Iluphu izenzo ngezinqumo zayo. Uma isheke lithi “kuhle,” iluphu iyama bese ihamba ngomkhumbi. Uma isheke lingalungile, iluphu ithumela iphutha – futhi okubi nakakhulu, ingase iqhubeke nokupholisha lelo phutha ngokuphindaphindiwe kuze kube yilapho ifundeka ngokugculisayo. Iluphu ithembekile kuphela njengento eqinisekisa ngayo.

Isixhumanisi esibuthakathaka: imodeli ehlela umsebenzi wayo

Isiqinisekisi esivame kakhulu imodeli ngokwayo. Ngemva kokubhala, uyabuza: “Ingabe lokhu kulungile?” Ishibhile, ayidingi ingqalasizinda eyengeziwe, futhi izwakala njengokuzindla.

Inkinga ukuthi imodeli ilungiselela ini. Uma i-LLM ikala okukhiphayo, iklomelisa lokho umsindo kwesokudla. Impendulo yokuzethemba, eshelelayo, nengalungile izwakala ilungile nayo. Ngakho-ke ukuzigxeka kuvame ukudlulela kukho kanye ukwehluleka ofuna ukukubamba, futhi kwesinye isikhathi kuzikhulumela ngaphandle kwempendulo efanele. Alikho iqiniso langaphandle ku-loop – ukusatshalaliswa okufanayo kuphela okukhiqize iphutha, manje eliceliwe ukulithola.

Bengifuna ukukukala.

Isheke elihlukile: i-deterministic kanye ne-source-anchored

Okunye okokuqinisekisa okungabuzi nhlobo umbono wemodeli. Kufanele sicabangele izici ezimbili ezifanele:

• Umthombo-ihange. Isheke likala ukuthi impendulo isekelwe emthonjeni wangempela, hhayi ukuthi ifundeka kahle yini. Uma impendulo iqhela kumthombo, isiqinisekisi siyayihlaba umkhosi – ngaphandle kokuthi iphrozi izwakala iqiniseka kangakanani.

• Ukunquma. Okokufaka okufanayo, isinqumo esifanayo, ngaso sonke isikhathi. Ungayihlola, uyiloge, futhi uyithembe kuyo yonke imisebenzi.

Ijaji lesitokofela elishintsha umqondo walo akusona isisekelo iluphu engamela kuso.

Isiqinisekisi engisisebenzisile i-geometric. Ishumeka umbuzo, impendulo yekhandidethi, kanye nomthombo ku-vector hypersphere futhi ifunde i- ama-engeli phakathi kwabo. Impendulo enesisekelo ihlezi eduze nomthombo wayo; okhohliwe uyakhukhuleka abheke embuzweni asuke emthonjeni. I-Semantic Grounding Index (SGI) iyisilinganiso sama-engeli amabili anjalo; i-comanion score (i-DGI) isilinganiso sokusatshalaliswa komhlaba esilinganiswa ngamapheya asekelwe phansi. Zombili ziyijiyomethri emsulwa ngaphezu kwesishumeki esigxilile, ngakho-ke zinqunywa ngokwakhiwa. Ukuqaliswa kuwumthombo ovulekile (Groundlens); iphuzu lalesi sihloko akusona izibalo kodwa kwenzekani uma ufaka isheke elinjalo ngaphakathi kweluphu.

Okokuqala, ingabe i-geometry iyakubandlulula ngisho nokubona izinto ezingekho? Kubhentshimakhi ye-HaluEval QA, amaphuzu asekelwe ezimpendulweni ezikhohliwe:

Ithebula 1: Ukutholwa ku-n = amapheya wezimpendulo angama-300; izikhawu zokuzethemba ze-bootstrap.

Isignali ehlanganisiwe ihlukanisa ngokuhlanzekile nezimpendulo ezihlosiwe. Yilokho umbandela. Manje umbuzo ukuthi ingabe isheke leli elinembile, elibekwe ngaphakathi kweluphu, empeleni lenza izimpendulo zokugcina zeluphu zibe ngcono kunokuzigxeka.

Ukuhlolwa

Idizayini ihlukanisa okuhlukile okukodwa: ukuthi iluphu iqinisekisa ini (Umfanekiso 1).

Ijeneretha iphendula imibuzo eyiqiniso incwadi evaliwe – kusuka enkumbulweni yayo, engenamthombo phambi kwayo – ngakho-ke ibona izinto ezingekho njalo futhi isiqinisekisi sinokuthile okufanele sikulungise. Umbuzo ngamunye uhamba ngezingalo ezine, futhi a unompempe oyimodeli amamaki kuyo yonke impendulo yokugcina, ngakho-ke ayikho imodeli ezihlulela yona ekutholeni amaphuzu:

• Ireferensi yencwadi evuliwe — ijeneretha imane inikezwe umthombo. Alikho isheke. Lona usilingi.
• Okukodwa (incwadi evaliwe) – impendulo eyodwa, akukho isheke. Iphansi leli.
• Ukuzigxeka – incwadi evaliwe; imodeli yahlulela impendulo yayo futhi ibuyekeze ize yaneliseke (kuze kube ukuphindaphinda kathathu).
• Umthombo-ihange – incwadi evaliwe; isiqinisekisi sejiyomethri sithola impendulo, futhi efulegini sijova umthombo futhi sicele ukubhala kabusha okusekelwe phansi (okuphindaphinda kathathu).

Ukusetha, ukukhiqiza kabusha: i-generator Claude Opus 4.8; unompempe i-GPT-5.5 (izilinganiso ezihlukile); ibhentshi HaluEval QA; isifaki khodi all-MiniLM-L6-v2; temperature=0 (uma ikhona); seed=0; ama-loop threshold alinganiswe kuhlelo olusalungiswa lwemodeli yokuqeqeshwa kwezincwadi ezivaliwe; $n=120$ izinto ngokusebenzisa izihibe.

I-asymmetry eyodwa yenziwe ngamabomu. Futhi yilo lonke iphuzu: ingalo enehange yomthombo inokufinyelela emthonjeni weqiniso ngesiqinisekisi sayo, futhi ingalo yokuzigxeka ayinakho.

I-hypothesis ngaphansi kokuhlolwa ayikona ukuthi “i-geometry idlula ukuzigxeka ngolwazi olufanayo.” “Isiqiniseko esinesisekelo somthombo esishintsha ijeneretha ekhohlisayo yencwadi evaliwe ibe ngesisekelo, kuyilapho ukuzigxeka ngokwakho kungenakukwazi.” Incwadi evulekile kanye nezingalo zodwa zibophe okungenzeka phezulu nangaphansi.

Imiphumela

Ukufundwa okubili, nezikhawu zokuzethemba zinquma kokubili.

Ukuzigxeka akusizanga. Ku-43.3%, uma kukhona, kubi kakhulu kunesitezi esingu-40.0%, kanye nesikhawu saso. [34.8%, 52.3%] kweqa kwephansi [31.7%, 48.9%] cishe ngokuphelele. Iziphindaphindo ezengeziwe azithenganga lutho. Imodeli ezihlola yona yachitha isikhathi esiningi ibala lapho iqale khona – futhi ukukhukhuleka okuncane kuya phezulu kuhambisana nokuzigxeka okuvame ukuketula izimpendulo ezilungile, okuyimodi yokwehluleka ongayibikezela lapho lingekho iqiniso langaphandle ku-loop.

Ukuqinisekisa okusekelwe kumthombo cishe kunciphise inani lephutha. Isuse phansi isuka ku-40.0% yaya ku-19.2%, ukuncishiswa kwesihlobo ngo-52%, ngenani elifanayo lokuphindaphinda iluphu yokuzigxeka esetshenzisiwe. Lokhu akukho emsindweni: isikhawu esisetshenziselwe phezulu siphuma ku-27.1%, ngaphansi lapho isikhawu sesitezi siqala ku-31.7%. Okubili akuhambelani. Ukuthuthukiswa kuwuphawu lwangempela, hhayi ukugijima kwenhlanhla.

Ukuma komphumela yindaba. Ijeneretha efanayo, isabelomali seluphu efanayo, ukukhubazeka kwencwadi evaliwe efanayo. Okuwukuphela kwento eyashintsha kwaba lokho iluphu ethembele – ukwahlulela kwayo, noma ukulinganisa okunqunyiwe ngokumelene nomthombo. Omunye walabo wanyakazisa inaliti ngesigamu. Omunye akazange awunyakaze nhlobo.

Awukwazi ukuqamba amanga ku-loop

I-intuition ilula. I-ejenti ifunda empendulweni yayo, ngakho awukwazi ukuqamba amanga kuyo. Isheke elivuza impendulo eqinisekile engalungile lenza lokho kanye — ukuphakela iluphu isignali yomvuzo ehlotshaniswa nokuqephuza kuneqiniso. Iluphu ilungiselela kahle isignali enikezwayo futhi ipholisha iphrozi. Ukuhlola okusekelwe kumthombo kunikeza iluphu umvuzo ohlotshaniswa nesisekelo esikhundleni salokho, futhi iluphu ithuthukisa lokho.

Asisho ukuthi i-geometry iyalazi iqiniso. Isiqinisekisi sikala ukuthi impendulo isebenzisana nomthombo wayo yini, hhayi ukuthi umthombo ulungile yini nokuthi impendulo iyiqiniso ngomqondo othile ophelele. Kubhentshimakhi eyakhelwe ukuhlola ubuqiniso esikhundleni sokubeka phansi, isignali efanayo iseduze kwethuba. Ukubeka phansi kanye neqiniso kuyizinhloso ezihlukene, futhi le ndlela ikhuluma kuphela okokuqala. Ukuwina kuyinto enesizotha: isiqiniseko esinehange lomthombo singcono isisekelo seluphu kunokuzigxeka, hhayi i-oracle.

Ukulinganiselwa

Ngeke ngishicilele umphumela mayelana nokuqinisekisa ngaphandle kokusho ukuthi kuma kuphi.

• I-asymmetry ingokoqobo futhi ihlosiwe. Ingalo eqinile ingafinyelela emthonjeni; ingalo yokuzigxeka ayikwazi. Okutholakele kumayelana nokunikeza iluphu ihange langaphandle, elinqumayo, hhayi mayelana nokuzihlola kwejiyomethri ngendlela ephumelelayo kulwazi olulinganayo.
• Ukubeka phansi akulona iqiniso. I-SGI ikala ukubandakanyeka komthombo. Kubhentshimakhi yeqiniso isignali efanayo cishe yithuba (AUROC ≈ 0.48). Uma imodi yakho yokwehluleka ingumthombo ongalungile kunempendulo engenasisekelo, lokhu akusizi.
• Ijeneretha eyodwa, ibhentshimakhi eyodwa, isifaki khodi esisodwa. Umphumela oqinile uthi u-Claude Opus 4.8 ku-HaluEval QA enemodeli eyodwa yokushumeka umusho. Angizange ngibonise ukuthi ibamba kuwo wonke amajeneretha nezizinda; ukuqaliswa kwangaphambi kwesikhathi ngejeneretha ehlukile nokucushwa akuzange kubonise ukuzuza okufanayo, yingakho impela ukuphindaphinda kwe-cross-generator kuyisinyathelo esilandelayo kunombhalo waphansi.
• Incwadi evaliwe iyisilungiselelo se-headroom. Ukuphoqa imodeli ukuthi iphendule ngenkumbulo kukhulisa izinga lephutha eliyisisekelo ukuze isiqinisekisi sibe nendawo yokusebenza. Epayipini elivamile le-RAG lapho umthombo usuvele ungaphakathi komongo, izinombolo eziphelele zizoba zincane – nakuba lokho futhi kuwumbuso lapho isheke eliyisisekelo lishibhile ukwengeza.
• Izilinganiso zephoyinti lembewu eyodwa. Izikhawu nguWilson; ukulinganisa imbewu kuzoziqinisa nakakhulu.

Ozokuthatha

“Izihibe zokuklama, hhayi ukwaziswa” kulungile. Kodwa iluphu iphephe kuphela njengento eqinisekisa ngokumelene nayo, futhi okuzenzakalelayo okulula – ukwahlulela kwemodeli ngokwayo – ingxenye engase yehluleke kakhulu. Kulesi sivivinyo ukuziqinisekisa akuphumelelanga ukwenza lutho, kuyilapho isheke eliqinisekile, elisekelwe emthonjeni lisika isilinganiso samaphutha ngohhafu kusabelomali esifanayo.

Uma wakha amalophu e-ejenti, umnyakazo osebenzayo uwukhomba isiqinisekisi seluphu kokuthile okungaphandle kombono wemodeli: isheke elinqumayo, elihlolekayo ngokumelene nomthombo wangempela. Uthola izihibe ezisebenza kahle kakhulu futhi ezinokwethenjelwa, futhi uthola isinqumo ongasifaka futhi ukhiqize kabusha esikhundleni se-vibe.

Isiqinisekisi esisetshenziswe lapha siwumthombo ovulekile, futhi incwadi yokubhalela egcwele ikhiqiza kabusha yonke inombolo engenhla (ijeneretha, okhiye onompempe, kanye neyodwa PROVIDER shintsha): github.com/groundlens-dev/groundlens. Ukungavumelani kwamukelekile — kuwumgomo, ngakho ungazihlolela wena.

Izithenjwa

• U-Huang, J., Chen, X., Mishra, S., Zheng, HS, Yu, A., Song, X., & Zhou, D. (2024, May). Amamodeli olimi amakhulu awakwazi ukuzilungisa ngokwawo okwamanje. Ku Inkomfa yamazwe ngamazwe mayelana nezinkulumo zokufunda (Umqulu. 2024, amakhasi 32808-32824).
• Kamoi, R., Zhang, Y., Zhang, N., Han, J., & Zhang, R. (2024). Ngabe ama-llms angawalungisa nini amaphutha awo? inhlolovo ebucayi yokuzilungisa kwe-llms. Ukwenziwa kwe-Association for Computational Linguistics, 121417-1440.
• U-Marín, J. (2025). Inkomba yesisekelo se-Semantic: Imingcele yejometri ekuzibandakanyeni komongo ezinhlelweni ze-RAG. I-arXiv preprint arXiv:2512.13771.
• U-Chen, KY, Su, FY, & Chiang, JH (2026). Inkohliso Yokuzilungisa: Ama-LLM Alungisa Abanye Kodwa Hhayi Bona. I-arXiv preprint arXiv:2606.05976.
• U-Marín, J. (2026). I-Geometric Taxonomy of Hallucinations in LLMs. I-arXiv preprint arXiv:2602.13224.