WikiSkill Uyasebenza

I-WikiSkill iwuhlaka lokuguquka kwamakhono e-Agent olwenziwe yi-Google Research. Lolu hlaka lusebenzisa ukwakheka kwezingqimba ezintathu: i-Raw Layer egcina imikhondo eluhlaza, i-Wiki Layer eqhubeka iqongelela ulwazi oluhlelekile, kanye ne-Skills Layer egcina amakhono asebenzisekayo. Lokhu kwenza ulwazi lwe-Agent luguquke lube ulwazi oluhlala isikhathi eside, bese kukhula kulo amakhono angasetshenziswa kabusha. Emjikelezweni ngamunye wokuphindaphinda, i-Wiki Maintainer ihlaziya imikhondo futhi ibuyekeze isizindalwazi solwazi, kanti i-Skill Proposer iphakamisa izibuyekezo zamakhono ngokusekelwe olwazini. Ngemva kokuqinisekiswa, izibuyekezo ziyamukelwa noma zibuyiselwe emuva. Ucwaningo lubonisa ukuthi i-9B model ene-WikiSkill ingadlula i-27B model engenazo ezinye izithuthukisi, futhi amakhono angadluliselwa phakathi kwamamodeli.

Ukuhlukaniswa kolwazi ngezingqimba ezintathu: imikhondo eluhlaza, ulwazi oluhlelekile namakhono asebenzisekayo kugcinwa ezingqimbeni ezintathu ezihlukene, ngayinye inomsebenzi wayo.
Ukuziguqula kwamakhono: umjikelezo wokuphindaphinda wenza amakhono e-Agent akhule futhi athuthukiswe ngokuqhubekayo ngokusekelwe kokuhlangenwe nakho.
Ukuqoqwa kolwazi oluhlala njalo: i-Wiki Layer iqhubeka iqongelela ulwazi emijikelezweni ehlukahlukene; ukwenqatshwa kwekhono akususi ulwazi oluqoqiwe.
Ukudluliselwa phakathi kwamamodeli: amakhono aguqukile angasetshenziswa kabusha kumamodeli ahlukene, ngisho naphakathi kwemindeni yamamodeli ehlukene.
Ukuhlunga ngesango lokuqinisekisa: amakhono angabazekayo ahlolwa kudathasethi yokuqinisekisa, bese kwamukelwa kuphela izibuyekezo ezithuthukisa ukusebenza.

Ukwakheka kwezingqimba ezintathu: i-Raw Layer igcina unomphela imikhondo yokwenza engaguquki, esebenza njengesisekelo samaqiniso; i-Wiki Layer ihlanganisa imikhondo ibe amaphethini ahlelekile namarekhodi okuguquka, iqhubeke iqongelela ulwazi futhi ayilokothi ibuyiselwe emuva; i-Skills Layer igcina amakhono asebenzisekayo asebenzayo njengamanje, isekele izibuyekezo ezinemibandela nokubuyiselwa emuva, futhi isebenzise i-PURPOSE.md ukulandelela umsuka wamakhono.
Umjikelezo wokuguquka onezinyathelo ezine: i-Inference Agent isebenzisa amakhono amanje kudathasethi yokuqeqesha futhi ikhiqize imikhondo; i-Wiki Maintainer isampula imikhondo, yenze ukuhlaziya imbangela eyisisekelo, bese ibuyekeza ikhathalogi yamaphethini namarekhodi e-Wiki; i-Skill Proposer ifunda i-Wiki nemikhondo ngendlela ye-ReAct, iphakamise ukwakhiwa kwekhono elilodwa noma i-patch; indlela ye-Gating ihlola ikhono eliphakanyisiwe kudathasethi yokuqinisekisa, futhi iyalamukela uma amaphuzu enyuka, kungenjalo ilibuyisele emuva, kuyilapho i-Wiki ingathinteki.
Ukuphikelela kolwazi: i-Wiki Layer isebenzisa i-skill-impact.md ukurekhoda i-diff yesiphakamiso ngasinye, amaphuzu okuqinisekisa kanye nomphumela wokwamukelwa noma ukwenqatshwa, kwakheka umlando wokucwaninga onobuqiniso. Lokhu kwenza ama-Proposer alandelayo agweme izindlela ezahluleka ngaphambilini, futhi kudale inzuzo yolwazi eqoqekayo emijikelezweni ehlukahlukene.
Ukuhlukaniswa kokutholwa nokusetshenziswa kwamakhono: ucwaningo lubonisa ukuthi ukutholwa kwamakhono, okuwukukhipha amasu kokuhlangenwe nakho, nokusetshenziswa kwamakhono, okuwukusebenzisa amasu ngesikhathi sokucabanga, kungamakhono amabili ahlukene. Amakhono aguqulwe yi-WikiSkill angadluliselwa phakathi kwamamodeli, futhi ngezinye izikhathi amakhono aguqulwe amanye amamodeli adlula lawo aguqulwe yimodeli uqobo.

Qalisa indawo yokusebenza: dala izinkomba ezintathu, i-raw/, i-wiki/ kanye ne-skills/, uqale ngeqoqo lamakhono elingenalutho.
Yenza i-training rollout: i-Inference Agent isebenzisa amakhono amanje kudathasethi yokuqeqesha, bese ibhala imikhondo yokwenza ku-raw/.
Buyekeza isizindalwazi solwazi: i-Wiki Maintainer isampula imikhondo ephumelele nehlulekile, yenze ukuhlaziya imbangela eyisisekelo, idale noma ibuyekeze amafayela wamaphethini ku-wiki/patterns/, bese yengeza ku-logs.md naku-skill-impact.md.
Phakamisa isibuyekezo sekhono: i-Skill Proposer ifunda inkomba ye-Wiki, amarekhodi omlando womthelela nemikhondo eluhlaza, bese iphakamisa ukwakhiwa kwekhono elilodwa noma i-patch ngendlela ye-ReAct.
Ukuqinisekisa nokulawula ngesango: ikhono eliphakanyisiwe lihlolwa kudathasethi yokuqinisekisa; uma amaphuzu enyuka, lamukelwa njengekhono elisha ku-skills/.
Phindaphinda umjikelezo: phinda lezi zinyathelo kuze kube ukusebenza kokuqinisekisa kuyazinza noma kufinyelelwa enanini lemijikelezo elibekiwe, bese uthola iqoqo lamakhono aguqukile elizosetshenziswa ekucabangeni.

Inzuzo yolwazi oluhlala njalo: i-Wiki Layer iqongelela ulwazi emijikelezweni ehlukahlukene; lapho ikhono lenqatshwa, ulwazi alulahleki, ngakho izibuyekezo ezilandelayo zingaphinda zisebenzise yonke imininingwane yomlando.
Ukuhlukaniswa kolwazi namakhono: “ukwazi ukuthi yini” kuhlukaniswa “nokwazi ukuthi kwenziwa kanjani”; amakhono angabuyiselwa emuva kodwa ulwazi luhlala njalo, kugwenywe ukulahleka kokuhlangenwe nakho kanye ne-patch yekhono.
Imodeli encane idlula enkulu: i-Qwen-3.5-9B ene-WikiSkill ingadlula i-Qwen-3.6-27B engenazo ezinye izithuthukisi, isebenzise ulwazi ukuvala umehluko wosayizi.
Ukudluliselwa nokusetshenziswa kabusha phakathi kwamamodeli: amakhono aguqukile angadluliselwa phakathi kwemindeni yamamodeli; amasu atholwe amanye amamodeli kwesinye isikhathi adlula amakhono aguqulwe yimodeli uqobo.
Ukugwema ukuphinda amaphutha: i-skill-impact.md irekhoda ngokuphelele umlando weziphakamiso nezizathu zokwenqatshwa, ngakho i-Proposer ayiphindi izindlela esezahluleka.

Iphepha lobuchwepheshe ku-arXiv: https://arxiv.org/pdf/2608.27454

Ukuzenzakalela komsebenzi wasehhovisi: ukuphatha ukubalwa kwamafomula e-Excel ayinkimbinkimbi, ukuhlanzwa kwedatha nokuxhumanisa amaspredishithi ahlukene; i-Agent ifunda izinqubo ezikhethekile zespredishithi njengokuthi “qala ngokwenza i-recalculate, bese ufunda”.
Imibuzo nezimpendulo emibhalweni emide: ukwenza ukusesha nokuzulazula ngezinyathelo eziningi kumadokhumenti e-PDF/Word anamakhulu wamakhasi; amakhono okusesha aguqukile aqondisa i-Agent ekutholeni izigaba ezibalulekile kahle, kugwenywe ukulahleka kumongo omude.
Ukucabanga ngezibalo nokuxazulula izinkinga: ukuguqula ulwazi lokuxazulula izinkinga zezibalo ezisezingeni lemincintiswano lube amasu angasetshenziswa kabusha, kuthuthukiswe ukunemba ngokuqhubekayo.
Ukusesha ulwazi ku-inthanethi: ukuthuthukisa inqubo yokusesha eyimizuliswano eminingi, kukhule inqubo ejwayelekile ethi “ukunweba amagama angukhiye → ukuqinisekisa imithombo eminingi ngokuphambana → ukuhlanganisa ulwazi”, kuthuthukiswe ukubuyiswa nokunemba kwemibuzo eyinkimbinkimbi.
Imisebenzi yomzimba esebenzisanayo: ukwenza imisebenzi eminingi yokunyakaza komzimba endaweni ebonakalayo, kusetshenziswa amaphethini anjengokuthi “gwema ukuphinda umjikelezo wesenzo” kanye “hlola isimo sokuqedwa komsebenzi” ukuze kuthuthukiswe izinga lokuphumelela komsebenzi.