Buyela Ezindabeni
UkuphephaAI Understanding ukwaziswa

Iphepha lixwayisa ngokuthi abenzeli be-LLM bangabuyisa ulwazi olususiwe ngamathuluzi

Iphepha elisha le-arXiv libonisa igebe ekungafundini kwe-LLM: umenzeli angase ayeke ukukhumbula ulwazi ezisindweni zalo kodwa alithole ngokusesha iwebhu, ukubuyisa noma amathuluzi esizindalwazi. Ababhali bahlongoza indlela enezigaba ezimbili zokunciphisa zombili izinhlobo zokuthola kabusha kuyilapho kugcinwa ukusetshenziswa kwamathuluzi okusemthethweni.

5 min readRead the primary source
Primary-source image accompanying Paper warns that LLM agents can recover deleted knowledge through tools
Idokhumenti yomthombo oyinhlokoUmthombo urekhodiwe
Umshicileli
arxiv.org
Isixhumanisi somthombo
arxiv.orghttps://arxiv.org/abs/2608.21544
Uhlobo lomthombo
Idokhumenti eyisisekelo — isimemezelo esisemthethweni, iphepha, ukugcwalisa, noma ikhasi lomuntu wokuqala esilifunda ngokuqondile.
UmongoQonda lokhu ngemizuzwana engama-60

Qala lapha

Imigomo ebalulekile

Imodeli Yolimi Olukhulu (LLM)
Imodeli yolimi eqeqeshwe ku-massive text corpora ukuze ikhiqize futhi ihlaziye umbhalo.
Ukuqinisa Ukufunda
Ukuqeqeshwa ngamasignali omklomelo lapho umenzeli efunda izenzo ezandisa imbuyiselo yesikhathi eside.
Inkumbulo (Inkumbulo yomenzeli)
Ingqikithi egciniwe umenzeli we-AI usebenzisa ezinyathelweni zonke noma izikhathi ukuze athuthukise ukuqhubeka.
ZihloleImibuzo ye-AI Agents

Kwenzekeni

Iphepha le-arXiv lethula i-Agentic Tool Unlearning, uhlaka oludizayinelwe ama-ejenti angamamodeli olimi angabiza amathuluzi angaphandle. Ababhali baphikisa ngokuthi ukungafundi okuvamile kungase kucindezele imodeli ukukhumbula okuqondile kolwazi ngaphandle kokuvimbela i-ejenti ekubuyiseleni ulwazi olufanayo ngokusesha iwebhu, ukubuyisa noma ukubheka isizindalwazi.

Iphepha lichaza imodi yokwehluleka eliyibiza ngokuthi ukutholwa kwethuluzi. Kumodeli yolimi evamile, ukungafundi ngokuvamile kuhlolwa ngokuhlola ukuthi imodeli isengakwazi yini ukukhumbula ngokuqondile ithagethi eqokiwe kumapharamitha wayo. Ababhali baphikisa ngokuthi lokhu kuhlola akuphelele ku-ejenti impendulo yayo ingancika kumakholi amathuluzi kanye nokubonwa kwangaphandle. Umenzeli onjalo angase ehluleke ukusho okuqondiwe ekhanda kodwa athole ulwazi olufanayo ngomthombo wangaphandle. Umehluko ubalulekile ngoba impendulo ebonakalayo ingahlala ikhona ngisho nalapho impendulo yangaphakathi yemodeli isishintshile. Kuleso silungiselelo, ukuhlola imodeli ngaphandle kokuhlola izenzo zayo kungase kugeje umzila lapho ithagethi itholwa khona.

Indlela ehlongozwayo, i-Agentic Tool Unlearning, inezigaba ezimbili. Okokuqala, isistimu isebenzisa ulwazi lwe-parametric unlearning okuhloswe ngalo ukucindezela ukukhumbula okuqondile. Okwesibili, isebenzisa ukufunda okuqiniswayo kwezinga le-trajectory ezindaweni ezilingisayo ezithuthukisiwe zamathuluzi. Ngokombono wephepha, lesi sigaba sijezisa kokubili ukuziphatha kwamathuluzi okufuna okuqondiwe kanye nokuvuza empendulweni yokugcina. Umgomo uwukunciphisa ukululama ngezenzo ze-ejenti, hhayi nje ngokushintsha izisindo zemodeli. Lokhu kwenza ukulandelana kwezinqumo ze-ejenti kube yingxenye yenkinga yokungafundi, okuhlanganisa ukukhetha kokufuna ulwazi kanye nendlela impahla ebuyisiwe ehlanganiswa ngayo empendulweni.

Ababhali babika izivivinyo ku-RWKU kanye ne-MUSE yokuyeka ukufunda kumabhentshimakhi kuzo zonke izinhlobo zezakhiwo zemodeli yolimi. Bathi le ndlela ifinyelela ukulingana okungcono phakathi kokukhohlwa okuhlosiwe kanye nokugcina insiza evamile yolwazi okufanele luhlale lukhona. Umthombo onikeziwe awunikezi imiphumela yezinombolo, amagama emodeli, izindleko zokuqeqesha, ukuqhathaniswa kwesisekelo noma imininingwane yamathuluzi alingisa, ngakho amandla kanye nobubanzi bokuthuthuka okubikiwe akukwazi ukuhlolwa ngokusuka ku-abstract kuphela. Lokho okweqiwe kushiya umphumela uqondwe kangcono njengesiphakamiso socwaningo ngokuhlolwa okubikiwe, kunokuba kube ubufakazi bokuthi indlela iqinisekisiwe kuwo wonke amasistimu asetshenzisiwe.

Imininingwane yomthombo: arxiv.org ↗

Kungani kubalulekile

Iphepha ligqamisa inkinga esebenzayo yokuphepha kanye nemfihlo kumasistimu ahlanganisa amamodeli olimi namathuluzi angaphandle. Ukususa ulwazi kumapharamitha emodeli kungase kunganele uma umenzeli esakwazi ukuthola noma akhe kabusha okuqondiwe ngamathuluzi azungezile.

Okutholakele kubalulekile ngoba ukufinyelela kwamathuluzi kushintsha ukuthi kusho ukuthini ukuthi uhlelo lwe-AI lukhohlwe okuthile. Imodeli ingase ingasakwazi ukufaka ikhodi noma ikhiqize kabusha ucezu lolwazi ngokuqondile, kodwa i-ejenti ephelele isengakwazi ukuyikhiqiza ngokusesha, ukubuyisa noma ukubuza imininingwane egciniwe exhunyiwe. Kumasistimu aphatha ulwazi lomuntu siqu, lobunikazi noma ngenye indlela ekhawulelwe, iyunithi efanelekile yokuhlola ingase ibe ukuhamba komsebenzi kwe-ejenti okugcwele kunokuba imodeli yodwa. Lo mbono obanzi ulandela ulwazi kuyo yonke indlela engaveza impendulo, esikhundleni sokuphatha amapharamitha emodeli njengokuphela komthombo ongaba khona.

Umehluko uphinde ube nomthelela wokuthi izinhlangano zitolika kanjani izimangalo zokususwa noma zokususwa. Uma isicelo sihloselwe ukuvimbela i-ejenti ekukhiqizeni ithagethi ethile, ukulungisa amapharamitha wemodeli kukodwa kungase kushiye enye indlela ivuliwe. Iphepha aliqinisekisi ukuthi noma iyiphi isistimu esetshenzisiwe ehlulekayo njengamanje, kodwa inikeza uhlaka lokuhlola oluphathekayo lokuhlola ukuthi amathuluzi e-ejenti ayayibukela phansi yini inqubo yokungafundi. Lolo hlaka lungasiza ukuhlukanisa ushintsho kulokho imodeli ekukhumbulayo kusukela kushintsho kulokho isistimu ehlanganisiwe esakwazi ukuyithola. Futhi igcina ububanzi besimangalo sokususwa esiboshelwe ekuziphatheni abasebenzisi abangakubona.

Kukhona ukuhwebelana okunzima kwezobunjiniyela kunhloso ehlongozwayo. Ukusetshenziswa kwethuluzi ngokuvamile kuyadingeka ukuze i-ejenti iphendule imibuzo mayelana nolwazi okufanele ilugcine. Ukujezisa ukufuna amathuluzi kakhulu kungalimaza ukuziphatha okuwusizo, kuyilapho ukujezisa okuncane kakhulu kungashiya ithagethi eselikhohliwe ikwazi ukuphinda itholakale. Inhloso yaleli phepha yokulondoloza ulwazi olugciniwe yenza lokhu kuhwebelana kube sobala, kodwa umthombo awubonisi ukuthi le ndlela isingatha kahle kangakanani izicelo ezingacacile, imibuzo engaqondile noma amathuluzi aqukethe ulwazi olugqagqene. Ngakho-ke indlela ewusizo kufanele ibambezele umzila ongafuneki ngenkathi iqhubeka nokweseka ukusetshenziswa kwethuluzi okuhlala kusemthethweni, ibhalansi engenakucatshangelwa ngokungaqondakali kuphela.

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 Agents Quiz

An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?

Ongakubuka ngokulandelayo

Umbuzo obalulekile ukuthi ngabe indlela ebikiwe ijwayelekile ngale kwezindawo ezilingisayo zephepha namabhentshimakhi. Imininingwane eyengeziwe iyadingeka olwazini oluqondiwe, ukulungiselelwa kwamathuluzi, amamodeli ezakhiwo, ukuhwebelana okulinganiselwe kanye nokuthi indlela isebenza ngokwethembeka yini ngaphandle kokuphazamisa ukusetshenziswa kwethuluzi elisemthethweni.

Iphepha eligcwele kufanele licacise ukuthi yini ebaluleke njengokukhohlwa ngempumelelo. Imininingwane ebalulekile ihlanganisa ukuthi ingabe ukuhlola kuhlola kuphela ukukhiqizwa okuhlosiwe okuqondile noma ukuphimisela, izimpendulo ezingaqondile kanye nokwakhiwa kabusha kwezinyathelo eziningi. Kufanele futhi ibonise indlela indlela ehlukanisa ngayo ukufuna okuhlosiwe okunqatshelwe ekutholeni okusemthethweni kolwazi olugciniwe oluhlobene, njengoba lowo mehluko uzonquma ukuthi le ndlela iyasebenza yini. Idizayini yokuhlola izoba nendaba kakhulu njengomphumela wesihloko: ukuhlola okuncane kungase kubonise ukuthi impendulo ethile ivinjiwe ngaphandle kokubonisa ukuthi ulwazi oluyisisekelo alukwazi ukufinyelelwa ngenye indlela. Izincazelo ezicacile zingenza ibhalansi ebikiwe phakathi kokukhohlwa nokusetshenziswayo kube lula ukuyitolika.

Ukuphindaphinda kuzoba okubalulekile ngoba ukuhlolwa okubikiwe kusebenzisa i-RWKU ne-MUSE kanye nezimo ezilingisayo ezithuthukiswe ngamathuluzi. Abafundi kufanele babheke ukuhlolwa okunezinhlobo ezahlukene zamathuluzi, amasistimu okubuyisa, imininingo egciniwe kanye nemindeni eyimodeli, kanye nokuhlola okwenziwa ngaphandle kokuhlelwa kokuqeqeshwa kwababhali. I-abstract ayisho ukuthi ikhodi, indawo noma amamodeli aqeqeshiwe ayatholakala, ngakho ukukhiqizwa kabusha akwaziwa. Ukuhlola okuzimele kungasiza futhi ukunquma ukuthi ingabe indlela incike endleleni ethile amathuluzi alingisa aveza ngayo ulwazi noma ukuthi izithiyo zakhona zihlala zisebenza yini lapho isistimu ezungezile ilungiswa ngendlela ehlukile. Ngaphandle kwalokho kuqhathanisa, ukujwayela kwendlela kusalokhu kungumbuzo ovulekile.

Ukuhlola okuzayo kufanele kulinganise kokubili ukuvikeleka nokusebenza phezu kwama-trajectories ama-ejenti amade. Isistimu evimba ukuvuza okuqondile ezivivinyweni ezimfushane ingase iziphathe ngendlela ehlukile lapho ingahlela, iphinde izame, ishayele amathuluzi amaningana noma ihlanganise ukuqaphela ingxenye. Umthombo uphinde ushiye kuvuliwe ukuthi ingabe i-Agentic Tool Unlearning ingabhekana yini nolwazi olukopishelwe kuzinkomba zamathuluzi noma kusizindalwazi ngokwazo, nokuthi ingabe lungasetshenziswa kuma-ejenti asetshenzisiwe ngaphandle kokuqeqesha kabusha amasistimu azungezile. Le mibuzo ixhuma inqubo yezinga lemodeli endaweni ebanzi lapho kungenzeka khona ukululama. Ziphinde zibonise ukuthi kungani ukuboniswa okuphumelelayo kuzodinga ukuhlola kokubili lokho umenzeli anqaba ukukudalula nokuthi yiluphi ulwazi oluwusizo eqhubeka nokuluthola.

Imihlahlandlela ehlobene nemibuzo

Ama-AI AgentsChatGPT ne-LLMsAmamodeli e-AI AchaziweUkuziphatha kwe-AIHlola okwaziyo — zama imibuzo ye-AI yamahhalaBheka igama le-AI kuhlu lwethu lwamagamaLandela isilandeleli sokulawula i-AI
Uthole lokhu kuwusizo?