Dzokera kuNhau
InnovationAI Understanding muchidimbu

Preprint inokurudzira kureba-kuziva nzira yekudzidzira yeGUI vamiririri

A new arXiv preprint inokurudzira kudzidzisa GUI vamiririri vane trajectory-level mhinduro, inopa mubairo kupfupika kwakabudirira kuuraya uye kusiyanisa kutadza nekusiyana kwavanoita kubva pakubudirira maitiro.

5 min readRead the primary source
Primary-source image accompanying Preprint proposes length-aware training method for GUI agents
Primary-source documentKwakanyorwa
Muparidzi
arxiv.org
Source link
arxiv.orghttps://arxiv.org/abs/2608.21830
Source type
Gwaro rekutanga - chiziviso chepamutemo, bepa, faira, kana peji rebato rekutanga ratinoverenga zvakananga.
ContextNzwisisa izvi mumasekonzi makumi matanhatu

Tanga pano

Matemu akakosha

Mutauro Mukuru (LLM)
Mutauro wemodhi yakadzidziswa pane yakakura text corpora kugadzira nekuongorora zvinyorwa.
Kusimbisa Kudzidza
Kudzidziswa nemasaini masaini apo mumiririri anodzidza zviito zvinowedzera kudzoka kwenguva refu.
Kupatsanurwa
Basa iro modhi inogovera yekuisa kune imwe kana akawanda akatemerwa chikamu.
Zviedze iwe pachakoAI Agents Quiz

Chii chaitika

Vatsvagiri vanokurudzira Kureba-Kuziva Kusiyanisa Kudzidza kune GUI Agents, kana LACL-GUI, yekusimbisa-yekudzidza dhizaini kune multimodal GUI vamiririri. Iyo nzira inowedzera trajectory-level yemhando masiginecha kune mhedzisiro-yakavakirwa manejimendi, ichivavarira kuita kuti hunhu hunobudirira huve hupfupi uye kupa misiyano yakanaka pakati pekuedza kwakakundikana. Iro bepa rinoshuma kuvandudzwa kwekuita kwakafanana pamusoro pemaitiro ekutanga pane GUI-agent mabhenji.

Yekutanga sosi ndeye arXiv preprint yakatumirwa Nyamavhuvhu 22, 2026, yakanzi "Kupfuura Kubudirira uye Kukundikana: Kureba-Kuziva Kusiyanisa Kudzidza kweGUI Agents." Inobata AI zvakananga: multimodal hombe yemutauro modhi maajenti anoshanda kuburikidza negraphical mushandisi nzvimbo. Vanyori vanoronga kusimbisa kudzidza senzira huru yekudzidzisa kune aya masisitimu uye vanotarisa kuti chiratidzo chekudzidziswa chinogadzirwa sei kana mumiriri anoedza basa. Muchigadziro ichocho, chinotariswa nebepa hachisi chimiro chitsva kana chinotarisana nemushandisi. Inzira yakarongwa yekugadzira kudzidza kubva muzvinyorwa zvemabasa akaedza.

Pepa racho rinoti nzira dzinoshandiswa zvakanyanya seBoka Relative Policy Optimization inogona kutambura kubva kune yainodaidza kuti mubairo-gradient kusarongeka, kuburitsa kusashanda kana kusagadzikana optimization. Inoisa basa rayo mukati mekuedza nguva pfupi yadarika kugadzirisa dzidziso yekusimbisa ine mibairo inogoneka seyakasiyana kana kupatsanura-kwakavakirwa zvinangwa. Zvinoenderana neabstract, iwo maitiro anogona kuvandudza kugadzikana nekudzivirira dambudziko gradient maitiro, asi ivo vanowanzo tarisa chete yekupedzisira mhedzisiro yetrajectory. Musiyano une basa nekuti iwo iwo ekupedzisira label anogona kuvanza mutsauko wekuti mumiririri akasvika sei pazviri. LACL-GUI saka inobata trajectory sechikamu chechiratidzo chekudzidzisa.

LACL-GUI inoratidzwa seyakasiyana-siyana yekusimbisa-kudzidza-ine-inogoneka-mibayiro hurongwa iyo inowedzera ruzivo nezve kunaka kweiyo yakazara chiito kutevedzana. Mukati menzira dzakabudirira, inogadza zvido zvakagadzirirwa kufarirwa kuurayiwa kwenguva pfupi. Mukati menzira dzakakundikana, inosiyanisa kuedza maererano nekusiyana kwavo kubva kune nzira dzakabudirira. Iyo abstract mishumo miyedzo paGUI-agent mabhenji uye inoti nzira yacho yakagadzira mamwe masaini ekudzidza anoshanda uye anogara achivandudza mashandiro pamusoro pemaitiro ekare. Izvo hazviratidze mabhenji, magadzirirwo emodhi, kuverenga kwebasa, mhedzisiro yenhamba, kana bvunzo dzenhamba. Izvi zvinoita kuti chimiro chegwara chive pakati pechinangwa cheiyo nzira. Mhedzisiro yakashumwa ndeye maitiro ekudzidza pamabhenji, nehumbowo hwekuyedza hwakasiiwa kune ruzivo rwebepa.

Kwakabva mashoko: arxiv.org ↗

Nei zvichikosha

GUI vamiririri vakagadzirirwa kupedzisa mabasa munzvimbo dzedhijitari, saka kudzidzisa kushanda nesimba uye kuvimbika kunokanganisa zvakananga kana vachigona kubatsira kupfuura kuratidzwa. Mupiro wepakati webepa inzira yekuburitsa rumwe ruzivo kubva kune zvese zvakabudirira uye zvisina kubudirira kuedza pane kubata chimwe nechimwe mhedzisiro sekubudirira kana kukundikana. Nekuti iyo sosi haipe mazita ebhenji, zvibodzwa, zvibodzwa, kana humbowo hwekutumira, huyero hunoshanda hwekuvandudzwa kwakataurwa hunoramba husina kujeka.

Tsvagiridzo inogadzirisa kushaya simba kwekongiri muvamiriri vekudzidzisa vanofanirwa kuita akawanda interface zviito. Mhedzisiro yebhinari inogona kuratidza kuti kuedza kwakabudirira kana kukundikana, asi haingoratidzi pachayo kana kuedza kwakabudirira kwakashandisa matanho asina kufanira kana kuti kukundikana kumwe kwaive pedyo nekupedzwa kupfuura imwe. Iyo nzira yakatsanangurwa inobata iwo misiyano sekutarisa kunobatsira, zvichigona kupa iyo optimizer ruzivo rwakawanda kubva kune yega yega yakarekodhwa trajectory. Chikumbiro ichi zvinoenderana nekuti izvo zvaunofarira zvinotsanangurwa uye kushandiswa sei panguva yekugonesa. Idzo sarudzo dzemagadzirirwo akakosha mamiriro ekududzira akashumwa kuvandudzwa kwekuita.

Kune vanogadzira GUI vamiririri, iyo pfungwa inogona kukosha nekuti mabasa edhijitari anowanzo sanganisira kutevedzana kwete kufanotaura kumwe chete. Nzira yekudzidzisa inokurudzira nzira dzakapfupika dzakabudirira inogona kuderedza kudyidzana kusingakoshi, nepo kurapwa kwegiredhi kwekutadza kunogona kubatsira mumiririri kudzidza kubva kune-vapotsa pane kuvaisa mumapoka nekuedza kwakashata. Izvi zvinokonzeresa magadzirirwo enzira yacho, kwete zvakawanikwa zvakazvimiririra zvichiratidzwa nekwakabva kunze kwezvido zvevanyori. Iko kuumbwa zvakare kunosiya kuvhurika kuti yakawanda sei bhenefiti inowanikwa kana mabasa achisiyana mukuomerwa kana kana interface maitiro achinja. Kunobva hakupedzi mibvunzo iyoyo.

Mhedzisiro yacho inonzwisiswa zvakanyanya sekufambira mberi kwetsvakiridzo pane humbowo hwechigadzirwa chakagadzirira kuendesa. Iyo abstract inoti LACL-GUI inogara ichinatsiridza mashandiro pamusoro penzira dzakapfuura, asi haipe masaizi ekuita, basa-rakananga mhedzisiro, compute zvinodiwa, kana kuenzanisa ruzivo. Izvo hazviratidziwo kuti nzira yacho inovandudza kuchengetedzwa, generalization, kuwanikwa, kana kuvimbika mune chaiyo-yepasirese interfaces. Izvo zvipingamupinyi zvinoita kuti basa rive rakakosha pakunzwisisa gwara rekudzidziswa uku richisiya ruzhinji rusina chokwadi. Mumashoko anoshanda, pfungwa yacho inobatanidza kushanda zvakanaka nemhando yekutarisa. Izvo hazvisi zvega zvinosarudza kuti mumiriri anofanira kuenzanisa sei kumhanya, kungwarira, uye kudzoreredza. Kunyatsoongorora kwaizoda kupatsanura mupiro wenzira kubva kune kugona kweiyo yepasi modhi uye nemabasa akasarudzwa. Kuparadzaniswa ikoko hakuna kutsanangurwa muchidimbu chiripo.

Interactive Mechanism

Interactive Mechanism: Iyo Inonyatsoshanda

Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.

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.
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?

Zvekutarisa zvinotevera

Iyo yakakosha yekutevera ndeyekuti LACL-GUI yakataurwa yakawana inobata munzvimbo dzakasiyana dzeGUI, mamodhi akawanda, kureba kwebasa, uye kuongororwa kwakazvimiririra. Vaverengi vanofanirwawo kutarisa humbowo nezve mutengo wekudzidzisa, maitiro ekufungidzira, kuberekana, uye kana mapfupi akabudirira trajectories anoramba ari echokwadi pasi pekuchinjika kwakachinja kana mamwe mabasa akaomarara. Kwakabva hakuratidze kuti nzira yacho inoburitswa pachena, yakagadzirira kugadzira, kana yepamusoro kunze kwezviyedzo zvinotsanangurwa mubepa.

Kudzokorora kwakazvimirira kunofanirwa kuyedza kana iyo yakashumwa mukana unobva pakureba-kuziva kutarisiro pachayo kana kubva kune dzimwe sarudzo mukuseta yekudzidzisa. Kuenzanisa kunobatsira kwaizoparadzanisa yakabudirira-trajectory yekuda, kutadza-mhando chiratidzo, uye chiripo chinosiyanisa chinangwa. Kwakabva hakutauri kana kubviswa kwakadaro kwakasanganisirwa, saka mupiro wechikamu chimwe nechimwe unoramba uri mubvunzo wakavhurika. Kuyedzwa kwakadaro kwaizojekesa kana nzira yacho ichichinja chete optimization dynamics kana kuburitsawo maitiro anoramba akavimbika kunze kwekumisikidzwa kwekuongorora. Kunobva parizvino kunosiya musiyano iwoyo usina kugadziriswa.

Generalization ndeimwe yakakosha isingazivikanwe. GUI vamiririri vanogona kusangana neakasiyana marongero, misangano yekudyidzana, horizons yebasa, uye interface shanduko. Sosi inotaura chete kuti miedzo yakaitwa paGUI-agent mabhenji; haigadziri kushanda pane zvisingaonekwi, mafambiro ekushanda kwenguva refu, mapeji ekuona ane ruzha, kana mabasa apo nzira yakapfupika isiri iyo yakachengeteka kana yakavimbika nzira. Ongororo dzenguva yemberi dzinofanira saka kuyera kurongeka pamwe nekuverenga kwekuita, kudzoreredza kubva mukukanganisa, uye kusimba kwekuchinja.

Kuwanikwa kwebepa uye mamiriro ekuita zvinodawo kuongororwa. Sosi yacho inozivisa kutumira kwearXiv uye zvinongedzo kune bepa, asi haitaure kuti kodhi, data rekudzidzisa, nzvimbo dzekutarisa, kana mumiriri inowanikwa pachena. Izvo zvakare hazvipi ruzivo nezve rezinesi, Hardware, nguva yekudzidziswa, nyaya dzekukundikana, kana kutariswa kwevanhu. Kusvikira izvo zvinyorwa zvashumwa, mhedziso yakasimba yakatsigirwa ndeyekuti vanyori vanokurudzira uye kuenzanisa nzira inogona kubatsira yekudzidzisa, kwete kuti vamiririri veGUI vanoishandisa vakagadzirira kutumirwa kwakafara. Izvo zvisipo zvekugadzira zvingaitawo kuti zvive nyore kuongorora nzira uye kubereka kuenzanisa kwakataurwa. Kusavapo kwavo kubva kumanyuko kunoganhurira zvingagumiswa pamusoro pokurera kunoshanda.

Related guides & Quizzes

AI AgentsAI Models InotsanangurwaKudzidziswa kweAITransformersEdza zvaunoziva - edza yemahara AI quizTarisa kumusoro izwi reAI mune yedu glossaryTevedza iyo AI modhi yekuburitsa tracker
Wakawana izvi zvinobatsira?