Subira ku makuru
Guhanga udushyaAI Understanding ibisobanuro

Impapuro zerekana uburyo bwigihe gito bwo kumenya salusiyo kurwego rwikimenyetso

Icapiro rya arXiv risobanura icyuma cya salusiyo ihuza imibare yinyandiko, ibimenyetso byerekana hamwe nururimi-ntangarugero bitunguranye bikurikiranye aho gucira ibimenyetso byigenga. Moderi yayo ya BiGRU yageze kuri AUC ya 0.840 kuri RAGTruth, nkuko impapuro zibitangaza.

6 min readRead the primary source
Source-provided image accompanying Paper reports a temporal method for detecting hallucinations at the token level
Inyandiko y'ibanzeInkomoko yanditse
Umwanditsi
arxiv.org
Ihuza ry'inkomoko
arxiv.orghttps://arxiv.org/abs/2608.18115
Ubwoko bw'inkomoko
Inyandiko y'ibanze - itangazo ryemewe, impapuro, dosiye, cyangwa urupapuro rwambere-dusoma mu buryo butaziguye.
ImirongoSobanukirwa ibi mumasegonda 60

Tangira hano

Amagambo y'ingenzi

Token
Igice kinini cyinyandiko yatunganijwe nicyitegererezo cyururimi, nkijambo igice cyangwa ikimenyetso.
Gushiraho Isuzuma
Dataset ifashwe ikoreshwa mu gupima ubuziranenge bw'icyitegererezo nyuma y'amahugurwa.
Hallucination
Iyo icyitegererezo gitanga amakuru meza ariko yibinyoma cyangwa adashyigikiwe.
IsuzumeModeri ya AI Yasobanuwe Ikibazo

Byagenze bite

Igicapo gishya cya arXiv gitanga igitekerezo cyo gufata salusiyo nk'ibice bigenda bikurikirana, aho kuba nk'amakosa yo mu rwego rwa wenyine. Uburyo bukomatanya ibintu 33 bivuye mumyandiko yatanzwe, imvugo-karemano-yerekana-indimi hamwe nururimi-icyitegererezo gitunguranye, hanyuma ikabitunganya hamwe nurugero rukurikirana. Uru rupapuro rutangaza ko GRU ifite ibyerekezo bibiri yageze kuri AUC ya 0.840 kuri RAGTruth ku mbuto 10, kunguka amanota 11 hejuru y’ibikorwa byigenga byo kwisubiza inyuma.

Inkomoko ni inyandiko ya arXiv ya “Temporal Multi-Signal Fusion for -Level Detection,” yanditswe na Igor Itkin ikanashyikirizwa ku ya 30 Kamena 2026. Ikadiri rero yerekana nkikimenyetso cyikurikiranya: buri kimenyetso cyakira amanota aturutse kumurongo wibintu, mugihe icyitegererezo gishobora gukoresha amakuru avuye mumyanya ituranye kugirango hamenyekane niba ikinyuranyo kidashyigikiwe cyangwa atari cyo.

Icyifuzo kiranga icyerekezo gifite ibipimo 33. Ukurikije ibivugwa, ibyo biranga guhuza imibare yinyandiko, imvugo-karemano-yerekana-imvugo-ntangarugero. Detector ntabwo ikoresha moderi ibyara ibikorwa byimbere cyangwa ibindi byimbere. Uru rupapuro rugerageza ibice bibiri byerekanwe inshuro nyinshi kuri ibyo biranga kandi rutangaza agace kari munsi yabakiriye ikora umurongo wa 0.840 kuri dataset ya RAGTruth, ukoresheje imbuto 10. Iragereranya ibisubizo hamwe na base yigenga ya logistique-regression kandi ivuga raporo yiterambere ryamanota 11, hamwe na p-agaciro ka 0.002 uhereye kubizamini bya Wilcoxon byashyizweho umukono.

Uru rupapuro ruvuga kandi ko igeragezwa ryangirika rigamije gutandukanya agaciro k'urutonde rw'agateganyo n'agaciro k'icyitegererezo gikomeye. Igisobanuro cyacyo nuko inyungu nyinshi zituruka kumiterere yigihe gito aho kuba ubushobozi bwikitegererezo: imyanya yizeye irashobora gutanga ibimenyetso kumyanya idasobanutse mubituranyi bitagaragara. Ikintu kimwe kigera kuri 0.845 cyerekana igisenge kivugwa murwego rusubirwamo, leta-umwanya hamwe nubwubatsi bwubaka, harimo Mamba na moderi ishingiye kubitekerezo. Abanditsi bakoresha icyo gisenge gisubirwamo kugirango bemeze ko ikintu kigabanya ari ikintu cyatoranijwe cyashyizweho, ntabwo cyihariye cyubatswe.

Inkomoko ikomeza ivuga ko detector ishobora gukora ku nyandiko yakozwe na moderi ifunze-isoko kuko isoma inyandiko yatanzwe n'ibimenyetso byo hanze gusa. Iratangaza kandi ko detector ikomeje gukora ku nyandiko ziva mu ngero z’ururimi zidahari mu makuru y’amahugurwa, hamwe n’igihombo kiri munsi ya 4% muri AUC. Ibi nibisabwa na progaramu imwe. Inkomoko yatanzwe ntabwo itanga amazina yicyitegererezo yasuzumwe, ibisobanuro byubwubatsi bwa dataset, protocole-igerageza protocole, uburyo bwo kurenga, gutondeka, kwibutsa, kalibrasi, ubukererwe cyangwa ingero zamakosa. Ntabwo kandi ishyiraho ibyemezo byigenga byemejwe.

Ibisobanuro birambuye: arxiv.org

Impamvu ari ngombwa

Inzira yashizweho kugirango ikore itabonye uburyo bwimiterere yimbere yimbere, ishobora gutuma ikoreshwa kuri sisitemu ifunze-isoko. Niba ibisubizo byatangajwe bifite hanze yisuzuma ryimpapuro, birashobora gufasha kumenya umwanya ukemangwa mubisubizo byongerewe ibisubizo hamwe no gushyigikira gusubiramo cyangwa kugenzura ibikorwa. Igisubizo gikomeza kuba intangiriro: isoko ni arXiv icapiro, kandi inyandiko yatanzwe ntabwo ishyiraho kwigana cyangwa gukora umusaruro.

Ubusobanuro bufatika nicyitegererezo cyatanzwe cyo gushakisha. Uburyo bwinshi bwo kugenzura-moderi buterwa no guhagararirwa imbere, ibimenyetso byerekana cyangwa andi makuru aboneka kubakoresha icyitegererezo gusa. Detector ishingiye kumyandiko yatanzwe hamwe nibimenyetso byo hanze birashobora, mubisanzwe, gushyirwa hafi yicyitegererezo imbere kidashoboka. Ibyo bituma igitekerezo kijyanye nimiryango ikoresha imiterere yururimi yakiriwe, nubwo isoko iterekana ko yinjijwe muri serivisi nzima cyangwa ikageragezwa munsi yimodoka.

Urutonde rushingiye kumurongo rukemura kandi imbogamizi nyayo mugereranya impapuro: ikimenyetso gisa nkibisanzwe mu bwigunge gishobora kuba igice cyigihe kirekire kidashyigikiwe. Gukoresha ibimenyetso bituranye bishobora kwemerera sisitemu gushyira ibendera igice kugirango bisubirwemo aho kwerekana urutonde rurerure rwamanota ya amanota. Mubikorwa byo kugarura-byongerewe akazi, ikimenyetso nkicyo gishobora gukoreshwa mugusaba ibimenyetso byinyongera, inzira igisubizo kumuntu, cyangwa guhagarika umwanya utazwi neza. Nibishobora gukoreshwa bivuye muburyo, ntabwo byoherejwe ninkomoko.

Raporo ya AUC ivugwa iragaragara cyane mubigeragezo byavuzwe kuko byerekana ko gutumiza amakuru bishobora kuba ngombwa kuruta gusimbuza umurongo fatizo hamwe na detector nini. Igisenge cyambukiranya inyubako nacyo ni ingirakamaro nkubushakashatsi bwakozwe: niba ibyiciro bitandukanye byicyitegererezo bihurira hafi yamanota amwe, kongera ubushobozi byonyine ntibishobora gukemura amakosa asigaye. Urupapuro rwisobanura rwonyine rugamije kunoza ibimenyetso byihishe inyuma, nkubwiza bwa entailment cyangwa ibintu bitunguranye, aho gutekereza ko ubundi buryo bukurikirana buzakemura ikibazo.

Imipaka nayo ni ngombwa. AUC ya 0.840 ivuga muri make imikorere yerekana urutonde; ntivuga umubare wa salusiyo wafatwa ku gipimo cyo kumenyesha ibikorwa, ni bangahe ibice nyabyo byashyirwaho ibendera nabi, cyangwa niba amanota ya detector yahinduwe nkibishoboka. RAGTruth ntishobora guhagararira buri gice cyibisobanuro cyangwa imiterere yo gusubiza. Kuberako ibikoresho byemewe hano ni progaramu imwe ya arXiv, ibisubizo ntabwo byigenga byashizweho ninyandiko yatanzwe. Basomyi bagomba gufata ibyabonetse nkibimenyetso byo gukomeza kwipimisha, ntabwo ari gihamya ko gutahura kwa salusiyo byakemuwe.

Interactive Mechanism

Uburyo bukoreshwa: Uburyo bukora

Shakisha ikoranabuhanga ryihishe inyuma yiri terambere.

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.
Kugenzura Ibitekerezo Byagenzuwe+10 Points
AI Models Explained Quiz

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

Ibyo kureba

Ibizamini byingenzi bizakurikiraho ni kwigana imibare yinyongera, gusuzuma muri domaine nimiryango yicyitegererezo, no gupima ibyiza, ibinyoma nubukererwe. Urupapuro rwerekana ntirugaragaza imiterere yasuzumwe, urwego rukora rwa detector, urwego rwarwo cyangwa rwibutsa, cyangwa uburyo imikorere ihinduka mugihe ibimenyetso bituzuye, bidasobanutse cyangwa bihanganye. Ibyo bisobanuro bizagaragaza niba uburyo ari uburinzi bwingirakamaro cyangwa ibisubizo byibanze.

Kwigana bigomba gutangirana nimpapuro zuzuye protocole hanyuma ikagura. Abashakashatsi bagomba gutanga ibisubizo kuri datasets nyinshi za salusiyo, bakandika imiterere yindimi zikoreshwa mugukora no gusuzuma inyandiko, hamwe na domaine yikizamini aho amakosa atwara ingaruka zitandukanye. Kugereranya bigomba kuba bikubiyemo amagambo yoroheje kandi asobanura, ibyiciro bya kalibatifike hamwe nuburyo bukoresha amakuru yimbere-imbere mugihe ayo makuru aboneka. Ikibazo cyingenzi nukumenya niba inyungu zavuzwe zirokoka impinduka mumibare, generator hamwe nuburyo bwo gutangaza amakuru.

Ibipimo by'imikorere bizagira akamaro nkukusanya AUC. Isuzuma ry'ejo hazaza rigomba gutangaza neza no kwibutsa ku mbuga zihariye zo kumenyesha, kugabanura umurongo, kugereranya urwego hamwe n'ibirango by'abantu, hamwe n'igihe cyo kubara cyo gutanga buri cyemezo. Ikimenyetso cyerekana uturere twagutse ariko ntigishobora gutandukanya ukutamenya gushidikanya kutavuzwe nabi bishobora kuba bigoye gukoresha. Inkomoko ntivuga niba uburyo bugamije gutahura neza, gusubiramo ibisekuruza cyangwa byombi, bityo ubukererwe bwoherejwe hamwe nuburyo sisitemu ishobora gutabara bikomeza kutamenyekana.

Igeragezwa rikomeye rigomba gusuzuma ibimenyetso bituzuye byo gushaka, inkomoko zivuguruzanya, ibisobanuro bisobanuwe, ibisubizo birebire hamwe ninyandiko zakozwe nkana. Kubera ko detector ikoresha ibimenyetso byo hanze, ubwiza nubwigenge bwibyo bimenyetso bishobora kugenzura imikorere. Ururimi rwururimi rushobora kandi guhindura imiterere yarwo, kalibrasi cyangwa ikosa ryakozwe nyuma yo koherezwa, birashobora guca intege disiketi yatojwe kumusaruro ushaje. Uru rupapuro ruvuga ko igihombo kiri munsi ya 4% ya AUC kurugero rwururimi rutagaragara, ariko ibisobanuro byatanzwe ntibisobanura gutandukanya icyitegererezo cyangwa kwerekana niba ibyo bisubizo bigera kuri domaine zitagaragara no guhindura sisitemu yo gushaka.

Hanyuma, impapuro zisubiramo igisenge gikwiye gusuzumwa. Irashobora kwerekana imipaka nyayo mumiryango itatu yerekana ibimenyetso, cyangwa irashobora kwerekana imibare, ibirango cyangwa igishushanyo mbonera. Urupapuro rwuzuye rugomba gusobanura uburyo urwego rwa salusiyo rwanditseho, uko ibimenyetso byatoranijwe, kandi niba amakuru yose avuye mu isuzuma ashobora gutemba mubiranga. Gushyira mubikorwa byigenga hamwe nibizamini kubisubizo nyabyo kubakoresha-ibisubizo byatanga ibimenyetso bikomeye kuruta ibyubatswe byubatswe. Kugeza icyo gihe, ingaruka zitazwi cyane ntabwo ari ukumenya niba detector ishobora gutondekanya ingero mubipimo byatangajwe, ariko niba ishobora kunonosora ibyemezo mugihe cyagenwe aho imvugo yicyitegererezo idashyigikiwe itera ingaruka mbi.

Ibijyanye nuyobora & ibibazo

Moderi ya AI YasobanuweChatGPT na LLMsAmahugurwa ya AIImyitwarire ya AIGerageza ibyo uzi - gerageza ikibazo cya AI kubuntuReba ijambo AI mumagambo yacu
Basanze ari ingirakamaro?