I-Benchmarking Factual Robustness of LLMs nge-Multi-conversation Persuasion

Idokhumenti yomthombo oyinhlokoUmthombo urekhodiwe Sep 16, 2026 05:20 UTC

Umshicileli: arxiv.org
Isixhumanisi somthombo: arxiv.orghttps://arxiv.org/abs/2609.16777
Uhlobo lomthombo: Idokhumenti eyisisekelo — isimemezelo esisemthethweni, iphepha, ukugcwalisa, noma ikhasi lomuntu wokuqala esilifunda ngokuqondile.

UmongoQonda lokhu ngemizuzwana engama-60

Qala lapha

Imigomo ebalulekile

Ukuqina: Ikhono lemodeli lokugcina ukusebenza ngaphansi komsindo, amashifu, noma okokufaka okuphikisayo.
Inkumbulo (Inkumbulo yomenzeli): Ingqikithi egciniwe umenzeli we-AI usebenzisa ezinyathelweni zonke noma izikhathi ukuze athuthukise ukuqhubeka.
Isethi yedatha: Iqoqo lezibonelo ezihlelekile noma ezingahlelekile ezisetshenziselwa ukuqeqeshwa, ukuqinisekiswa, noma ukuhlola.

Kwenzekeni

Abacwaningi bethule uhlaka lwe-SAST-IR ukuze bahlole ukuqina kwamamodeli Olimi Olukhulu ngokumelene nokuhlaselwa kokuncenga. Uhlaka lulingisa ukulungiselelwa kwesitha esibi kakhulu ngokuphoqelela ukusula inkumbulo kumodeli eqondiwe kuyilapho igcina umlando womhlaseli. Ukuhlolwa kwesethi yedatha yangokwezifiso ye-CounterFact-Strict kuveze imiphumela ethusayo, ngamasu alula okuhlasela athola izinga lempumelelo elingu-96%.

Uhlaka lulingisa ukulungiselelwa kwesitha esibi kakhulu ngokuphoqelela ukusula inkumbulo kumodeli eqondiwe kuyilapho igcina umlando womhlaseli.

Ukuhlolwa kwesethi yedatha yangokwezifiso ye-CounterFact-Strict kuveze imiphumela ethusayo, ngamasu alula okuhlasela athola izinga lempumelelo elingu-96%.

Imininingwane yomthombo: arxiv.org ↗

Kungani kubalulekile

Ukuqina Kwamamodeli Olimi Olukhulu ngokumelene nokuhlaselwa kokuncenga kuwukukhathazeka okubalulekile kwezokuphepha. Uhlaka lwe-SAST-IR luhlinzeka ngethuluzi elisha lokuhlola ukuqina kwalawa mamodeli nokuhlonza ubungozi obungaba khona.

Uhlaka lwe-SAST-IR luhlinzeka ngethuluzi elisha lokuhlola ukuqina kwamamodeli Olimi Olukhulu ngokumelene nokuhlaselwa ngokuncenga.

Uhlaka lulingisa isilungiselelo esiphikisayo esibi kakhulu, silwenze ithuluzi elibalulekile lokuhlonza ubungozi obungaba khona kulawa mamodeli.

Imiphumela yokuhlolwa kudathasethi ye-CounterFact-Strict yangokwezifiso iyethusa, ngamasu alula okuhlasela azuza izinga lempumelelo elingu-96%.

Uhlaka lwe-SAST-IR lungasetshenziswa ukuhlonza ubungozi obungase bube Kumamodeli Olimi Olukhulu futhi kuthuthukiswe amasu okuvikela aqinile.

Uhlaka lungabuye lusetshenziselwe ukuhlola ukusebenza kahle kwamasu okuvikela ahlukene kanye nokuhlonza izindlela eziphumelela kakhulu.

Interactive Mechanism

I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela

Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.

Document Size:128K tokens

Needle Placement Depth (Location in document):50% into text

Attention Context Buffer Map:

Target Fact (50%)

Equivalent Pages~320Standard book pages

Retrieval Accuracy99.9%Needle recall score

RAM / KV Cache5.1 GBMemory overhead

Prompt CachingActive~80% discount on reuse

Core takeaway: Million-token context windows allow querying whole codebases or legal archives in one prompt. However, KV cache memory scales with context length, making prompt caching crucial for real-time production.

I-Interactive Concept Check+10 Points

What is AI? Quiz

A route planner searches possible journeys using explicit rules. What does this illustrate about AI?

Ongakubuka ngokulandelayo

Ukuthuthukiswa kwamasu okuzivikela aqinile ngokumelene nokuhlaselwa kokuncenga.

Ukwakhiwa kwamasu okuzivikela aqinile ekuhlaselweni okuncenga kubalulekile ukuze kuqinisekiswe ukuphepha nokwethembeka kwamamodeli olimi amakhulu.

Uhlaka lwe-SAST-IR luhlinzeka ngethuluzi elisha lokuhlola ukuqina kwalawa mamodeli nokuhlonza ubungozi obungaba khona.

Uhlaka lungasetshenziswa ukuhlonza ubuthakathaka obungaba khona Kumamodeli Olimi Olukhulu kanye nokuthuthukisa amasu okuzivikela aqine kakhulu.

Uhlaka lwe-SAST-IR lungasetshenziswa futhi ukuhlola ukusebenza kahle kwamasu ahlukene okuvikela kanye nokuhlonza izindlela ezisebenza kahle kakhulu.

Ukwakhiwa kwamasu okuzivikela aqine ngokwengeziwe ngokumelene nokuhlaselwa kokuncenga kuzodinga ukubambisana kwabacwaningi, onjiniyela, nochwepheshe bemboni.

Imihlahlandlela ehlobene nemibuzo

Yini i-AI?ChatGPT ne-LLMs Ukuziphatha kwe-AI Ama-AI Agents Hlola okwaziyo — zama imibuzo ye-AI yamahhala Bheka igama le-AI kuhlu lwethu lwamagama Landela i-tracker yokukhishwa kwemodeli ye-AI

Uthole lokhu kuwusizo?