Subira ku makuru
Guhanga udushyaAI Understanding ibisobanuro

Igiti cyo kwibuka cyibiro byerekana raporo yunguka kubakozi barebare ba LLM

Icapiro rishya ryerekana sisitemu yo kwibuka yibice igena ibice bigize amateka yimikorere ya AI igomba gukomeza gukora. Kuri GAIA-Umwandiko, abanditsi batangaza amakuru yukuri, gukoresha vuba-token gukoresha no gukomeza kutagira amakuru yuburozi kuruta kwibuka kumurongo.

5 min readRead the primary source
Primary-source image accompanying Weighted Memory Tree reports gains for long-horizon LLM agents
Inyandiko y'ibanzeInkomoko yanditse
Umwanditsi
arxiv.org
Ihuza ry'inkomoko
arxiv.orghttps://arxiv.org/abs/2608.20631
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

Ururimi runini (LLM)
Ururimi rwicyitegererezo rwahuguwe kumyandiko minini corpora kubyara no gusesengura inyandiko.
Kwibuka (Memory Memory)
Imiterere yabitswe umukozi wa AI akoresha intambwe cyangwa amasomo kugirango atezimbere.
Ibipimo
Ikizamini gisanzwe cyangwa dataset ikoreshwa mugupima no kugereranya imikorere yicyitegererezo.
IsuzumeIkibazo cya AI

Byagenze bite

Urupapuro rwanditswe na Quang Dao, Purvi Kathalkar na Kenneth Eaton rutangiza Igiti Cyibukwa Cyibiti, cyangwa WMT, sisitemu yo kwibuka muburyo bukoreshwa nabakozi berekana imvugo nini ndende. Itegura amateka yimikorere ya agent mubikorwa, subtasks nibikorwa, hanyuma igenera buri kwibuka amanota yo kugumana imbaraga.

Abanditsi basobanura WMT nka sisitemu yo kwibuka yagenewe abakozi bagomba gutegura, gukoresha ibikoresho no kubona amakuru mu ntambwe nyinshi. Aho gufata amateka yuzuye yo gukora nkumurongo umwe wumurongo, WMT ihagararira muburyo bukurikirana kurwego rwimirimo, subtasks nibikorwa byihariye. Buri kwibuka yakira amanota yo kugumana ashobora guhinduka nkuko umukozi akomeza gukora. Intego yavuzwe ni ugukomeza imiterere yingirakamaro mugihe ugabanya ingaruka zibintu bitagifasha umurimo wubu. WMT ivugurura ayo manota binyuze mubyabaye bishingiye ku byabaye no guhitamo kubora, ukurikije abstract.

Sisitemu irashobora kubika amakuru afatwa nkingirakamaro, yikubye inzira yuzuye muburyo bworoshye, guhagarika ibintu bito-byingirakamaro kandi bikagumana uburyo bwo kubika ibintu byongeye kuba ngombwa. Inkomoko ntabwo itanga impapuro zuzuye za algorithmic, imbibi cyangwa ingero, bityo abstract ishigikira igishushanyo mbonera cyagutse ariko ntabwo ari konte nini yerekana uburyo sisitemu yo gutanga amanota yitwara muri buri kibazo. Uru rupapuro rutanga isuzuma kuri GAIA-Inyandiko ukoresheje Qwen3-8B, Gemma 4 E4B na Llama-3.1-8B. Ugereranije nibyo abanditsi bita umurongo wibukwa, WMT yazamuye neza impuzandengo ya 9.97 ku ijana kandi igabanya gukoresha-token ikoreshwa 32.8%. Ibi nibisabwa nabanditsi ba preprint; inkomoko ntisobanura umubare wimirimo, ishyirwa mubikorwa ryibanze, itandukaniro kurugero cyangwa niba impuzandengo yatangajwe ari mibare.

Abanditsi bavuga kandi ubushakashatsi bwakozwe n'uburozi. Muri ibyo bizamini, WMT yagabanije gutsimbarara no gukwirakwiza amakuru atizewe ugereranije nubundi buryo bwasobanuwe muri abstract. Igisubizo gihuza igishushanyo mbonera cyo kunanirwa muburyo bwo kunanirwa aho ibintu bitari byo cyangwa byinjijwe nabi bishobora gukomeza guhindura ibyemezo nyuma. Inkomoko ntivuga uko uburozi bwatangijwe, umubare w’uburozi bwakoreshejwe, uko intsinzi yapimwe cyangwa niba ibizamini byerekana ibitero bishobora kuba muri sisitemu zoherejwe.

Ibisobanuro birambuye: arxiv.org ↗

Impamvu ari ngombwa

Uru rupapuro rwerekana imbogamizi zifatika z'abakozi ba AI barebare: kugumana amateka menshi birashobora kongera amafaranga yo gufata umwanzuro mugihe nanone ugaragariza umukozi amakuru ashaje, adafite akamaro cyangwa ayobya. Ibisubizo byatangajwe byerekana ko guhitamo kwibuka, kuruta ingano yibuka yonyine, bishobora kuba ingenzi kubikorwa byombi no kwizerwa.

Abakozi barebare bakusanya amateka yo gukora nkuko bateganya, guhamagara ibikoresho no gushyiramo amakuru yo hanze. Inkomoko igaragaza ibibazo bibiri bifitanye isano: ibisobanuro birebire birashobora kuzamura ibiciro, kandi amateka yakusanyirijwemo ashobora kuba arimo amakuru ashaje, ntaho ahuriye cyangwa ayobya. Uburyo bugenzura ibyo kwibuka bikomeza gukora rero byibanda kukibazo gikuru cyibikorwa kubakozi aho kongeramo ibintu byo kwisiga muburyo rusange bwururimi. Kugabanuka kw'ikimenyetso bishobora kuba ngombwa kuko uburebure bwihuse bugira ingaruka kumiterere umukozi wohereje nyuma yicyitegererezo.

Niba byororotse, kugabanuka kwa 32.8% ugereranije nimpapuro umurongo-wibuke shingiro irashobora kugabanya umubare wamakuru yatunganijwe mugihe cyintambwe nyinshi. Inkomoko ntabwo ishyiraho amafaranga yazigamye yo kuzigama, impinduka zubukererwe cyangwa ikiguzi cya sisitemu yose, nyamara, kubera ko ibyo bisubizo byaterwa nicyitegererezo, ibikorwa remezo ndetse no hejuru bisabwa kugirango igiti kibuke. Igisubizo cyukuri kirashoboka cyane kuruta icyifuzo cyo gukora neza. Uru rupapuro rutangaza ko impuzandengo ya 9.97 ku ijana yazamutseho amanota atatu yiswe icyitegererezo kuri GAIA-Text, byerekana ko kugumana kutavangura bishobora kwangiza imikorere. Ibisobanuro byatanzwe ni uko guhitamo gukora bishobora gutuma amakuru ajyanye no kugerwaho atemereye buri ntambwe ibanza kugira ingaruka zingana. Ibyo bisobanuro bikomeje kuba ubushakashatsi, ntabwo ari umwanzuro washyizweho wigenga.

Ibisubizo byuburozi biha akazi akazi ko kwizerwa no kurwego rwumutekano. Niba umukozi umaze igihe kinini agumana amakuru atizewe byoroshye, ikosa ryambere cyangwa amabwiriza yashyizwemo bishobora kugira ingaruka nyuma. Sisitemu yo kwibuka igabanya gutsimbarara no gukwirakwizwa bishobora kugabanya iyo nzira yo gutsindwa. Inkomoko ntabwo yerekana ko WMT irinda uburozi, yemeza ibyemezo byizewe cyangwa igasimbuza ubugenzuzi bwagutse nko kwemeza kwinjiza, uruhushya rwibikoresho no gusuzuma abantu.

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 Agents Quiz

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

Ibyo kureba

Ibyavuye mu bushakashatsi biva mu icapiro rimwe hamwe na GAIA-Inyandiko isuzuma ukoresheje imvugo eshatu zifunguye. Ubundi igenzura rigomba kwibanda ku gipimo ngenderwaho, kugereranya no kwibuka umurongo, ikiguzi cyo kubungabunga WMT ubwayo, ubukana bw'ibizamini by’uburozi niba ibisubizo byimurirwa mu yindi mirimo, icyitegererezo ndetse no kohereza.

Ikibazo cya mbere kubasomyi n'abashakashatsi ni ireme ry'isuzuma. Inkomoko yita GAIA-Umwandiko hamwe nicyitegererezo cyindimi eshatu ariko ntivuga umubare wibikorwa, ibigize, uburyo bwo gukora ibizamini, amanota arambuye cyangwa intera ikizere. Ibisobanuro birakenewe kugirango tumenye niba impuzandengo yunguka ari nini kandi ikomeye cyangwa itwarwa nigice gito cyimirimo. Impapuro zasibwe zishobora gusobanura ibice bya konte ya WMT kubijyanye nimpinduka zavuzwe, ariko abstract ntabwo ivuga muri make ibyo babonye.

Kugereranya nabyo bigomba gusobanurwa neza. "Umurongo wo kwibuka" urashobora kwerekeza kuburyo butandukanye bwo kugumana no kwerekana amateka, kandi ibisubizo birashobora guterwa no guhitamo gushyira mubikorwa nko guhagarika, kuvuga muri make cyangwa kugarura. Ibikorwa bya WMT ubwabyo kuvugurura, gutanga amanota no guhuza-imiterere-yo kwinjira irashobora gusaba kubara cyangwa kubika bitagaragarira mugukoresha-ibimenyetso byonyine. Ikurikiranwa ryingirakamaro ryagereranya iherezo-iherezo ryigiciro nubukererwe, ntabwo umubare wibimenyetso woherejwe kurugero rwururimi. Ubusanzwe nikindi kibazo gifunguye. Isuzuma rikoresha Qwen3-8B, Gemma 4 E4B na Llama-3.1-8B ku gipimo kimwe. Inkomoko ntabwo ishyiraho imikorere kumurongo munini cyangwa nyirubwite, abakozi ba multimodal, domaine yihariye, guhora uhindura ibidukikije cyangwa imirimo aho ikiguzi cyo kugumana ibintu bidasanzwe ariko byingenzi biri hejuru. Ibisubizo birashobora kandi gutandukana nurwego rwabakozi, ibikoresho byashizweho hamwe nubwiza bwamakuru yo hanze yakira.

Ubushakashatsi bwuburozi butuma hasuzumwa neza kuko "amakuru atizewe" akubiyemo ibintu byinshi bishoboka. Ibyingenzi bitazwi harimo niba ibizamini byakoresheje amakosa yimpanuka, byinjijwe nkana cyangwa byombi; igihe amakuru yagumye mumateka; icyabaruwe nko kwamamaza; kandi niba guhagarika imiterere iteye inkeke nabyo bishobora gukuraho amakuru yemewe. Urupapuro rwashyikirijwe arXiv ku ya 21 Kanama 2026, nka verisiyo ya mbere. Inkomoko ntisobanura kwigana kwigenga, gusubiramo urungano, ibimenyetso byoherejwe cyangwa kurekura kode.

Ibijyanye nuyobora & ibibazo

Abakozi ba AIModeri ya AI YasobanuweAmahugurwa ya AIPrompt EngineeringGerageza ibyo uzi - gerageza ikibazo cya AI kubuntuReba ijambo AI mumagambo yacuKurikiza icyerekezo cya AI cyo kurekura
Basanze ari ingirakamaro?