Subira ku makuru
Guhanga udushyaAI Understanding ibisobanuro

EDGE impapuro zitanga ubushakashatsi burambye kubakozi ba AI

Urupapuro rwemewe na EMNLP 2026 rutangiza EDGE, uburyo bwo guhugura bufasha abakozi-berekana imvugo gukoresha uburambe bwingirakamaro aho kwishingikiriza kubushake bwo hanze mugihe cyo gufata umwanzuro. Abanditsi bavuga ko batsinze byinshi kuri ALFWorld na WebShop bakavuga ko uburyo bwagumanye inyungu nyinshi nyuma yo gukuraho…

5 min readRead the primary source
Primary-source image accompanying EDGE paper reports more durable exploration for AI agents
Inyandiko y'ibanzeInkomoko yanditse
Umwanditsi
arxiv.org
Ihuza ry'inkomoko
arxiv.orghttps://arxiv.org/abs/2608.21946
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

Kwiga gushimangira
Amahugurwa kubihembo byerekana aho umukozi yiga ibikorwa byongera kugaruka kwigihe kirekire.
Kwibuka (Memory Memory)
Imiterere yabitswe umukozi wa AI akoresha intambwe cyangwa amasomo kugirango atezimbere.
Kuba rusange
Ukuntu icyitegererezo gikora neza kumakuru mashya, atagaragara hanze yamahugurwa.
IsuzumeIkibazo cya AI

Byagenze bite

Uru rupapuro rutangiza EDGE, cyangwa Ubunararibonye-Gukwirakwiza Ubushakashatsi Bwayobowe, urwego rwo guhugura abakozi-b'indimi-bigezweho hamwe no kwiga gushimangira. Ikoresha uburambe bwakuweho nkamahugurwa yigihe gito scafolds, hanyuma igerageza kwinjiza imyitwarire yingirakamaro mubyitegererezo ubwabyo.

Inkomoko ya arXiv yerekana EDGE nkigisubizo cyikibazo cyihariye mugushimangira imyigishirize yimikorere yintangarugero yindimi: uburyo bwubushakashatsi bwingirakamaro muburyo bwimikoranire bushobora gutabwa nyuma yo kuvugurura politiki. Abanditsi bibanze kubakozi bakora ibintu bigoye, birebire-bitambitse, aho inzira igenda ishobora kuba ikubiyemo amakuru yongeye gukoreshwa kubyo kugerageza, ibyo kwirinda nuburyo bwo gukira kunanirwa. Icyifuzo cyabo nyamukuru ni ugukoresha uburambe bwamateka mugihe cy'amahugurwa mugihe bigabanya gukenera kubaza nyuma.

EDGE itandukanya buri tsinda rizunguruka muburyo bubiri bwa trayektori: bimwe bisabwa kuburambe bwakuweho nibindi byakozwe nta. Uru rupapuro ruvuga ko uku kugereranya kugereranya umusanzu mwiza wuburambe kandi bigatuma sisitemu yemera ubuyobozi bwingirakamaro nta yandi mananiza. Ihita itandukanya imyitwarire yavuyemo muri politiki shingiro binyuze mu ntego ya KL ikoreshwa ku nkunga ya politiki yonyine. Mu magambo asobanutse, uburyo bugerageza kumenya imyitwarire yatanzwe hanze ifasha kandi igatoza iyo myitwarire murugero.

Urwego rurimo kandi icyo abanditsi bita banki yuburambe-bwihindurize. Nk’uko inkomoko ibivuga, iyi banki ikomatanya ubuyobozi mu buryo bugaragara bwo gutsindwa kandi ikuraho ibyanditswe bishaje uko politiki ihinduka. Ku bipimo bya ALFWorld na WebShop, abanditsi bavuga ko hari iterambere ryagaragaye kuri GRPO ya 8.3 na 12.5 amanota-yo gutsinda, bakoresheje urugero rwa miliyari 7. Batangaza kandi ko abakozi bahuguwe bagumanye 96.0% byimikorere yabo mugihe uburambe bwo hanze bwakuweho mugihe cyo gufata umwanzuro. Inkomoko ivuga ko code iboneka kandi ko impapuro zemewe mu nama nkuru ya EMNLP 2026.

Ibisobanuro birambuye: arxiv.org ↗

Impamvu ari ngombwa

Niba ibisubizo byatangajwe bifite ibirenze ibipimo bibiri byageragejwe, uburyo bushobora gutuma abakozi barebare batagengwa na sisitemu yo kugarura ibintu kandi bagashobora gukoresha amasomo kuva bagerageje mbere. Ibyo birashobora koroshya gahunda, nubwo isoko idashiraho kwizerwa kwisi cyangwa kwitegura umusaruro.

Ubusobanuro bufatika bwakazi ni ukugerageza gukemura impagarara zimenyerewe mugushushanya abakozi. Gushakisha birashobora gutanga umukozi ufite uburambe bwambere bwambere, ariko sisitemu igomba gushakisha inshuro nyinshi ububiko bwo hanze bushobora gutera ubukererwe, ibiciro remezo nibindi byananiranye. Intego EDGE yavuze ni ugukoresha ibisubizo mugihe cyo kwiga hanyuma ugakora imyitwarire yavuyemo igice cya politiki. Niba bigenze neza, ibyo bishobora kubyara abakozi bitwaje ingamba zabo zo kwiga imbere.

Ibisubizo byo kugumana byatangajwe bifitanye isano cyane niyi ntego. Uru rupapuro ruvuga ko imikorere yagumye kuri 96.0% yurwego rwakuweho nyuma yuburambe bwo hanze bwakuweho mugihe cyagenwe. Ibi nibimenyetso, nkuko abanditsi babivuga, ko uburyo bwakoze ibirenze kuyobora umukozi wigihe gito: inyungu nyinshi zimuriwe murugero. Nyamara, igishushanyo nicyo gisabwa mubushakashatsi bwimpapuro, ntabwo ari igipimo cyigenga cyigenga, kandi inkomoko ntabwo itanga imbonerahamwe yubushakashatsi, igereranya ridashidikanywaho cyangwa ibisobanuro bikenewe kugirango isuzume imbaraga zayo.

Ibisubizo nabyo bifite akamaro kuko bareba imyitwarire yabakozi aho kuba amanota yimvugo-yerekana gusa. ALFWorld na WebShop zirimo imikoranire yintambwe nyinshi, bityo iterambere kuri bo rishobora kuba ingirakamaro kubashakashatsi biga igenamigambi, gukoresha ibikoresho no gukira amakosa. Nubwo bimeze bityo, ibipimo ngenderwaho ntabwo byonyine byerekana ko umukozi yiringirwa mugukingura-kurangiza. Inkomoko ntabwo itanga ibisubizo byoherejwe, ibisabwa byo kugenzura abantu, gusuzuma umutekano, ibiciro, cyangwa imikorere mubihe bigoye cyangwa bihinduka vuba.

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

Ibibazo by'ingenzi ni ukumenya niba EDGE ihuriza ku yindi mirimo, icyitegererezo n'ibidukikije, kandi niba inyungu zayo ziguma mu mibare kandi mu buryo bugaragara mu bizamini byigenga. Uru rupapuro narwo rusize rufunguye umubare wa compte na injeniyeri bigoye uburambe bwa banki yongeyeho.

Ikibazo cya mbere cyo kureba ni kwigana. Inkomoko igaragaza ALFWorld na WebShop ikanatanga ibisubizo ku gipimo cya 7B, ariko ntigaragaza mu magambo ahinnye umubare w'imikorere yakozwe, niba iterambere ryaragaragaye mu mibare, cyangwa uko ibipimo n'ibipimo byagenwe. Abashakashatsi bigenga bazakenera kugenzura ibyagezweho no kumenya niba biterwa nibisobanuro byihariye, igenamigambi ryo kugarura, imibare cyangwa gahunda y'amahugurwa.

Ikibazo cya kabiri ni rusange. EDGE yateguwe hafi yinyigisho zishimangira ubumenyi hamwe nubushakashatsi buyobowe nubushakashatsi, ariko isoko ntisobanura ko uburyo bumwe bukora mumiryango itandukanye yicyitegererezo, umunzani wibipimo, ibidukikije cyangwa ubwoko bwimirimo. Ibisubizo ku bipimo bibiri ntibishobora guhanura imikorere mubikoresho bya software, ubushakashatsi bwakazi, sisitemu ya serivise yabakiriya cyangwa ibidukikije bifatika. Ntibizwi kandi niba uburambe bwimbere bushobora gutuma umukozi adahinduka mugihe ibidukikije bihindutse cyangwa mugihe ingamba zishaje ziba mbi.

Ikibazo cyanyuma kireba ikiguzi nimiyoborere ya banki yuburambe. Uru rupapuro ruvuga ko banki ikomatanya ubuyobozi buva muburyo bwo gutsindwa no gukuramo ibyataye igihe, ariko ibivugwa ntibigereranya ububiko, kubara amahugurwa, kuvugurura inshuro cyangwa ibisabwa kugira ngo bikosorwe. Abashakashatsi n'ababitanga bagomba gusuzuma uburyo gutsindwa byatoranijwe, niba imyitwarire itifuzwa ishobora gutandukanywa hamwe ningirakamaro, nuburyo politiki yavuyemo ishobora kugenzurwa byoroshye. Kugeza ubwo ibyo bibazo bisubijwe, EDGE irasobanuka neza nkigisubizo cyizewe cyubushakashatsi cyatanzwe nabanditsi bacyo, ntabwo ari ibimenyetso byerekana ko abakozi barebare biteguye gukoreshwa bidakurikiranwa. Izi nizo mbibi zibimenyetso biriho hamwe n’ahantu hakenewe ubundi bushakashatsi mbere y’imyanzuro yagutse yerekeye kuramba, kwimurwa, kwiringirwa, cyangwa kwitegura gushobora gukururwa.

Ibijyanye nuyobora & ibibazo

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