Kubuyekezwa nsuku zonke1866 izindaba eziqinisekisiwe

Izindaba ze-AI. Ngaphandle komsindo.

Ukufakwa kwe-AI ehlolwe ngomthombo yokwethulwa komkhiqizo, ukushintshwa kwenqubomgomo, ucwaningo lwezokuphepha, nomnyakazo womkhakha, okuchazwe ngesiNgisi esilula ithimba lezemfundo elingenzi nzuzo.

Isithangami samasonto onkeWonke amahabhu esihloko

Ingobo yomlando yezindaba

Ingobo yomlando yezindaba ze-AI — Ikhasi 155

Izindaba ze-AI ezihlolwe ngumthombo, ezintsha kakhulu kuqala, zabantu abadinga ukuqonda i-AI ngaphandle kokujaha i-hype.

Source-page capture accompanying Paper Finds Late Layers of a Mixture-of-Experts Model Tolerate Heavy Expert MaskingIndaba ephezuluUkuqamba7 min funda

Iphepha Lithola Izendlalelo Zasemuva Zemodeli Yengxubevange Yochwepheshe Libekezelela Ukufihla ubuso Kwengcweti Okusindayo

Ukuphrinta ngaphambilini kubika ukuthi ukukhubaza ochwepheshe bezinga eliphansi kuzingqimba ezinhlanu zokugcina zemodeli ye-Mixture-of-Experts yezigidigidi ezingu-35 kulondoloze imiphumela esebenziseka kakhulu yokuhumusha ikhodi kunokusabalalisa ukusikeka okufanayo kuzo zonke izendlalelo. Ifaka imodeli eyodwa kanye nebhentshimakhi eyodwa, futhi umbiko ongabonakali ubika ukuthi akukho sisekelo esingafihliwe.Funda indabaUmthombo oqinisekisiwe: arxiv.org

Ukuthola umthombo oqinisekisiwe

Yonke indaba ixhuma ebufakazini obunamandla obutholakalayo: imithombo yoqobo uma itholakala, ngenye indlela ukubika okuchasiselwe ngokucacile.

IsiNgisi esilula

Kwenzekeni, kungani kubalulekile, nokuthi yini okufanele uyibuke - ngaphandle kwejagoni.

Asikho isigcwalisi

Uma isignali incane, asishicileli lutho kunokupheqa okuphakelayo.

Izindaba ezengeziwe

9 izindaba
  2. UkuqambaUkuqamba7 min funda

    I-Benchmark Entsha Ithola Abenzeli Be-AI Bavimbe Ngokungafanele Umsebenzi Ovunyelwe U-28% Wesikhathi

    Ukuphrinta ngaphambili kwethula i-SteerBench-Work, ukuhlolwa kwe-106-scenario yalapho umenzeli we-AI enquma ukwenza noma ukumisa isikhashana ukuze kubuyekezwe. Kuzo zonke izimo ezingamamodeli angama-30, ababhali babika ukuthi ukubamba umsebenzi osuliwe ngokungalungile cishe kwakuvamile izikhathi ezingama-28 kunokuvumela umsebenzi ongaphephile ngokungalungile.arxiv.org
  4. UkuqambaUkuqamba7 min funda

    Iphepha Liphikisa Amasu Okuzivelela Kwehlula i-RL Ekugcineni Impendulo ye-LLM Isetha Okuhlukahlukene

    I-arXiv preprint entsha ithi ama-LLM angemuva kokuqeqeshwa anamasu okuziphendukela kwemvelo - indlela esekelwe kubantu, engenawo ama-gradient ephazamisa izisindo ngokuqondile - yehlula ukufunda okuqinisiwe ku-pass@k kanye nokufakwa kwesixazululo. I-abstract icaphuna imiphumela engcono ye-math-benchmark kodwa amagama awekho amamodeli, ama-benchmarks, noma izinombolo.arxiv.org
  5. UkuphephaUkuphepha7 min funda

    Iphepha Lithi Abasebenzeli Be-AI abazithuthukisayo Bangaguqula Impumelelo Eyodwa Engaphephile Ibe Ikhono Elisebenziseka Kabusha

    I-arXiv preprint entsha ikhombisa imodi ethile yokwehluleka kwe-ejenti: lapho umenzeli ozithuthukisayo ebhala inqubo engaphephile enkumbulweni, ingabuyiswa futhi isetshenziswe kumaseshini akamuva. Konke ukulungiselelwa okuthuthukisiwe okuhloliwe kukhiqize ama-artifact angaphephile, kanye nemisebenzi emithathu enonya engaphezu kokuphindwe kabili kokuphumelela kokuhlasela kwe-carryover.arxiv.org
  6. UkuphephaUkuphepha6 min funda

    Iphepha le-SEAG Liphakamisa Izinkampani Ezizwelayo Zokwehlukanisa Ngaphambi Kokuba Imibuzo Ye-RAG Ifinyelele Kuma-LLM Angaphandle

    Iphrinta ethunyelwe ku-arXiv ichaza uhlaka olushintshanisa amagama abucayi emibuzweni futhi libuyise amadokhumenti eziteketiso ngaphambi kokuwathumela kumodeli yenkampani yangaphandle. Ababhali babika ukunemba okungaphezu kuka-80% kumethrikhi yabo yabasebenzisi abavela ekupheleni ukuya ekupheleni, namazinga okufihla okugcwele aphakathi kuka-74.91% no-77.83% kuwo wonke amamodeli amancane amathathu.arxiv.org
  7. UkuqambaUkuqamba6 min funda

    I-CABS+ Ibika Iphepha Lishibhile, Imodeli Esheshayo Ihlanganisa Kuwo wonke Amasethi edatha angama-27

    Ukuphrinta kusengaphambili okuthunyelwe ku-arXiv kuchaza i-CABS+, indlela yokuhlanganisa imodeli ethatha indawo yokusesha kwegridi ngosesho lwe-coefficient olungenayo i-gradient. Ababhali babika izinzuzo zokusebenza okunamadijithi amabili ngaphezulu kwezisekelo ezimbili, ngaphansi kwekota yememori ye-GPU yesisekelo esisodwa, kanye nokusheshisa okungu-4x ngaphezu kwenye.arxiv.org
  8. UkuqambaUkuqamba6 min funda

    Iphepha Liphakamisa "Izifundo" Ezibuyisiwe Zokuthuthukisa Ukubonisana Kwendawo Kumamodeli Afriziwe Ombono Wolimi

    I-arXiv preprint ichaza i-Spatial Memory Agent, egcina ulwazi oluqinisekisiwe njengezifundo zombhalo ezitholwa ngesikhathi sokubikezela, ifuna izinzuzo kuwo wonke ama-benchmarks wendawo namamodeli amane olimi lombono ngaphandle kokushintsha izisindo zemodeli. Iyabuyekezwa; amagama ayo angabonakali awekho amabhentshimakhi, amamodeli ayisisekelo, noma amamajini.arxiv.org
  9. UkuqambaUkuqamba7 min funda

    I-PROVE-RT Paper Reports 44.7% Impumelelo Yokukhiqiza Ubufakazi Besikhathi Sangempela Ohlolwe Umshini

    I-arXiv preprint yethula i-PROVE-RT, esebenzisa ukubuyisa kanye nokwazisa okufakwe kusiteji ukwenza amamodeli amakhulu olimi abhale imibhalo yobufakazi ye-PROSA/ROCQ yokuhlaziya ukuhlelwa kwesikhathi sangempela. Ababhali babika izinga lempumelelo elingu-44.7% kusethi yokuhlola ekhethiwe, lapho ukwaziswa okuqondile kwehluleka ukukhiqiza ngokuthembekile imishini evumelekile.arxiv.org

Ukwaziswa okukodwa okuwusizo isonto ngalinye

Qhubeka ne-AI ngaphandle kokuhlala kokuphakelayo.

Thola izindaba ze-AI eziqinisekisiwe zeviki, idatha yoqobo, amathuluzi awusizo, ukukhetha okufundayo, nemisebenzi emisha ye-AI.

Send me
One email a week. Switch any time.

Finyelela abantu abafunda i-AI

Ukuqasha uchwepheshe we-AI noma ukwethula umkhiqizo owusizo we-AI? Yibeke phambi kwabantu abeze lapha ukuzofunda nokwenza.

Thumela umsebenzi we-AIThumela ithuluzi le-AI