Nkhani

CLEVI, 79.07% pa GAIA pogwiritsa ntchito from scratch mtundu wake womwe...ili m’gulu la 2.5% lapamwamba mwa mitundu 3,090

Gwero: Nyuzipepala ya Electronic News, mtolankhani Lee Won-ji

Gwero: Nyuzipepala ya Electronic News, mtolankhani Lee Won-ji

Kutchula nkhani yonse ya “CLEVI, 79.07% pa GAIA pogwiritsa ntchito from scratch mtundu wake womwe...ili m’gulu la 2.5% lapamwamba mwa mitundu 3,090”

Tsiku lofalitsidwa: 2026-04-07

Ulalo: https://www.etnews.com/20260407000203

cip-5.5-agent·cip-5.5-mm stack ya mitundu 5 ya ma agent onse odziyimira pawokha, onse apeza mfundo zoposa 70

Poganizira kutayika kwa chidziwitso, chiwongola dzanja cha mayankho olondola ndi choposa 98%, kuposa anthu (92%)

Chithunzi cha m’nkhani

CLEVI, kampani yoyambitsa ya AI, idagwiritsa ntchito from scratch mtundu wake womwe ndipo idapeza chiwongola dzanja cha mayankho olondola cha 79.07% pa benchmark yapadziko lonse ya ma agent a AI yotchedwa GAIA, ndikukhazikika m’gulu la 2.5% lapamwamba mwa mitundu 3,090 yonse yolembetsedwa.

Malinga ndi kufotokozedwa kwa akatswiri a gawoli, GAIA imayamikiridwa pamsika wa ma benchmark a AI chifukwa imayimira mokhulupirika luso la “ma agent a AI othandiza”. Benchmark iyi, yomwe idapangidwa limodzi ndi Meta AI, HuggingFace ndi University of Paris-Saclay, idatulutsidwa koyamba mu Novembala 2023.

Pali zinthu zitatu zazikulu zomwe zimasiyanitsa GAIA ndi ma benchmark ena. Choyamba, popeza mayankho olondola amasungidwa mwachinsinsi, overfitting n’zosatheka. Chachiwiri, chifukwa imafuna kulingalira kovuta kwa multistep ndi multimodal, kupeza mfundo zambiri sikungatheke pogwiritsa ntchito kungofananitsa ma pattern osavuta. Chachitatu, mitundu yoposa 3,090 padziko lonse imapikisana pamikhalidwe yofanana kudzera pa leaderboard yovomerezeka ya HuggingFace.

Test set ili ndi mafunso 301 onse. Level 1 imafuna kugwiritsa ntchito chida chimodzi komanso kulingalira kosavuta, Level 2 imafuna kuphatikiza zida zingapo komanso kulingalira kwapakati, ndipo Level 3 ndi mafunso ovuta kwambiri omwe amafuna kukonza ndi kuchita kwa agent m’magawo osachepera 5. Panthawi yomwe benchmark idatulutsidwa, ma GPT model (okhala ndi ma plugin) anali pafupifupi 15%, koma pano magulu otsogola akweza chiwerengerocho kufika pa 70–80%.

Pakati pa izi, CLEVI idatsindika kuti chinthu chofunika kwambiri paukadaulo pa zomwe yakwanitsazi ndi chakuti sinagwiritse ntchito konse ma API a ma LLM akunja monga GPT, Claude kapena Gemini. Chodziwika pa CLEVI ndi kugwiritsa ntchito mitundu iwiri yomwe idapanga yokha pogwiritsa ntchito njira ya from scratch.

cip-5.5-agent ndi mtundu wa AI wa agentic womwe umakhala ngati ubongo waukulu wa njira ya ma agent yomwe imakonzekera (planning), kuchita (execution), ndi kutsimikizira (verification) ntchito zovuta mwaokha. cip-5.5-mm ndi mtundu wamphamvu wa multimodal wogwira ntchito zosiyanasiyana womwe umamvetsetsa ndi kupanga ziganizo pa mafayilo osiyanasiyana monga mawu, zithunzi, ndi makanema. Popeza mafunso ambiri a GAIA ali ndi zolowetsa zosakhala zolemba monga mafayilo a PDF, zithunzi, ndi mawu, luso la multimodal limeneli linali chinthu chachikulu chomwe chinathandiza kupeza zigoli zapamwamba.

CLEVI inagwiritsa ntchito mitundu iwiri imeneyi yomwe inapangidwa yokha kukhazikitsa masanjidwe 5 osiyanasiyana a ma agent pa GAIA, ndipo onse anapeza zigoli zoposa 70.

Malinga ndi kufotokoza kwa kampaniyo, zotsatira zosangalatsa zinapezeka pakuwunikanso pambuyo pake mkati mwa CLEVI. Mwa mafunso onse 301, umboni wa mayankho olondola a ena sunapezekenso pa intaneti yapagulu yomwe ilipo panopa. Izi zikutanthauza kuti zambiri zomwe kale zinkapezeka pa intaneti kapena mu injini zosakira zinachotsedwa kapena kusinthidwa ndipo sizingapezekenso. Pamene zotsatirazo zinawunikidwanso pogwiritsa ntchito zambiri zomwe anthu angatsimikizire poyera panopa, chiŵerengero cha mayankho olondola a CLEVI chinapezeka kukhala choposa 98%. Izi zaposa kale avareji ya anthu, yomwe ndi 92%.

Woimira CLEVI anafotokoza kuti, “CLEVI inalowa m’gulu la 2.5% lapamwamba pa benchmark yapagulu, pomwe ma agent onse 5 anapeza zigoli za 70+, pogwiritsa ntchito mitundu yake yokha yopangidwa from scratch ndi mayankho ake a AI agent, popanda kusakaniza mitundu yakunja. Zikuwoneka kuti zigoli zake zovomerezeka zitha kuwonjezeka kwambiri m’tsogolo kudzera pakukonza mitundu ndi mayankho a ma agent ake. Kupambana kumeneku kuli ndi tanthauzo lalikulu chifukwa kunayezedwa pa benchmark yapagulu yapadziko lonse ndipo kukuwonetsa ‘chidaliro chotsimikiziridwa’; ndi kupambana kochokera ku mtundu wa kampani ya ku Korea wopangidwa from scratch; ndi zotsatira za stack ya mitundu yodziyimira payokha, osati kusakaniza mitundu yakunja; komanso chifukwa, poganizira za kutayika kwa zidziwitso pa mafunso ena, kuli ndi tanthauzo loposa zigoli zokha.”

Bwererani ku chipinda cha nkhani
CLEVI

Chinenero ndi dera

Zinenero zomasuliridwa ndi makina zalembedwa. Kupezeka kumatsatira phukusi lofalitsidwa la tsambali.

Zinenero 136

Amalimbikitsidwa

1

Asia Kum’mawa

7

Asia Kum’mwera chakum’mawa

11

Asia Kum’mwera

18

Asia Wapakati

5

Kum’mawa Kwapakati ndi Kaukasas

10

Ulaya Kumadzulo ndi Kum’mwera

16

Britain ndi Ireland

4

Ulaya Kumpoto ndi Baltic

10

Ulaya Wapakati ndi Balkan

14

Ulaya Kum’mawa

5

Afirika Kum’mawa ndi Nyanga

8

Afirika Kumadzulo ndi Wapakati

9

Afirika Kum’mwera

8

Amereka

5

Pasifiki

5