I-Apple Paper Ithi Imibuzo Eyinkimbinkimbi Yosesho Lwe-boolean Iphelele Nge-P

Kwenzekeni

I-Apple Machine Learning Research yashicilela iphepha lika-Amir Aavani ngobunkimbinkimbi bekhompyutha bokuhlola imibuzo ye-Boolean efakwe ngokujulile, engeyona ye-monotonic phezu kwezinkomba ezihlanekezelwe. Iphepha lenza ngokusemthethweni ulimi lokubuyisa olususelwe kumagrafu e-acyclic aqondisiwe, lithi inkinga yalo yokuhlola i-P-Complete ngokuqinile, futhi lethula i-ComputePN, i-algorithm enqumayo esebenzisa izethulo ezingezinhle kanye nokukhumbula nge-DAG.

I-Apple Machine Learning Research ibala iphepha njengoba lashicilelwa ngo-Agasti 2026, u-Amir Aavani njengombhali walo. Iphepha likhuluma nge-inverted-inverted traversal, indlela yokusesha lapho amatemu ekhomba kumadokhumenti aqukethe. I-Apple ibeka inkinga ezungeze ama-ejenti esimanje e-AI asebenzisa ingqalasizinda yokusesha yokugeleza komsebenzi okuyinkimbinkimbi, okungokomfanekiso we-neuro-symbolic. Ngokomthombo, lokho kugeleza komsebenzi kungahlanganisa kube imibuzo ye-Boolean esesidlekeni ehlanganisa ingqondo engeyona ye-monotonic, okufaka ukunganaki. Ngakho-ke leli phepha ligxile ekusetshenzisweni komqondo wokubuyisa okuhlelekile, hhayi imodeli yolimi olusha, umkhiqizo womenzeli, noma isici sokusesha esibheke umthengi.

Iphepha lichaza imikhawulo emibili kumasu okuhlola ajwayelekile. Ithi amamodeli e-stateful Document-at-a-Time iterator aboshwe ngokwesakhiwo ukulinganisa kwefomula ye-NC^1 futhi angabhekana nesimo esibi kakhulu sokuqhuma kwe-O(2^|Q|) lapho umqondo wokuguqulelwa kabusha uqaqelwa esihlahleni. Ithi amamodeli okwenza izinto e-Term-at-a-Time aphindelelayo abhekana nesijeziso sesikhala esingu-Ω(|U|), esichazwa njengokuskena kwendawo yonke, lapho ehlola ukuphika okunengqondo phezu kwendawo yonke yedokhumenti. Lezi izimangalo ezenziwe iphepha abstract. Umthombo onikeziwe awunikezi izibonelo, ukulandelelwa komthwalo womsebenzi, izilinganiso ezisetshenziswayo, noma ukuqhathanisa okubonisa ukuthi iphethini yesimo esibi kakhulu yenzeka kangaki kumasistimu asetshenzisiwe.

I-Apple ithi yenza ngokusemthethweni ulimi lokubuyisa, i-L_R, ngokusekelwe kumagrafu aqondisiwe e-acyclic futhi ifakazela ukuthi ukuhlola imibuzo ngalolu limi ku-P-Complete ngokuphelele. Ibe sethula i-ComputePN, echazwa njenge-algorithm yokuhlola enqumayo, eqaphela u-sparsity. Indlela yehlukanisa ukuphika okunengqondo kusukela ekwenziweni kwezinto kwesikali sendawo yonke ngokusebenzisa ukumelwa okukabili kokuthi Positive-Negative futhi isebenzisa ukugcinwa ngekhanda kwe-DAG komdabu ngakho amagama angaphansi aphindaphindwayo awadingi ukunwetshwa ngokuphindaphindiwe.

Umthombo unikeza isilinganiso sesikhathi sokuhlola esifunwayo esingu-O(|Q| · |U_active|), lapho incazelo ibhekisela kusayizi wombuzo kanye nesethi yedokhumenti esebenzayo. Akubandakanyi ukusetshenziswa kwe-algorithm, ikhodi yomthombo, izinto ezingaguquki ezilinganisiwe, noma izidingo zokusebenza.

Sekuhlangene, incazelo ihlanganisa isitatimende senkinga esisemthethweni sephepha, umphumela walo oyinkimbinkimbi, kanye nendlela yokuhlola ehlongozwayo. Ulimi lokubuyisa lumelwe ngamagrafu e-acyclic aqondisiwe, ubunzima obushiwo buphathelene ne-Boolean logic esidlekeni, futhi i-ComputePN yethulwa njengendlela yokusingatha leso sakhiwo. Isethulo esihle-negethivu sikhuluma ngokungananazi, kuyilapho ukukhumbula ngekhanda kwe-DAG kukhuluma namagama angaphansi aphindaphindiwe. Umthombo uphinde usho isibopho ngokuya ngosayizi wombuzo kanye nesethi yedokhumenti esebenzayo. Ngaphandle kwalawo maphuzu ashiwo wedizayini nobunkimbinkimbi, okokusebenza okunikeziwe akusunguli imininingwane yokusetshenziswa, ukusebenza okulinganiselwe, ukusetshenziswa kokukhiqiza, noma ukuqinisekiswa kwangaphandle. Leyo mibuzo ihlala ihlukile ezimangalweni ezisemthethweni zephepha mayelana nokuhlola.

Imininingwane yomthombo: machinelearning.apple.com ↗

Kungani kubalulekile

Uma izimangalo zephepha zibambe iqhaza ezinhlelweni ezisebenzayo, zingacacisa ukuthi ingqalasizinda yosesho kufanele isebenzise kanjani ingqondo eyinkimbinkimbi yokubuyiswa esetshenziswa ku-AI-ejenti yokuhamba komsebenzi. Umthombo uthi i-ComputePN igwema izindleko ezimbili ezaziwayo: ukunwetshwa komchazi we-logic yombuzo oguqukayo kanye nokwenza yonke indawo yedokhumenti ingananazi. Ayinikezi izilinganiso zokukhiqiza, ngakho-ke umphumela osebenzayo uhlala ungaqinisekisiwe.

Umphumela ubalulekile ngoba ubeka umngcele osemthethweni enkingeni i-Apple exhuma ekutholeni i-AI-ejenti. Ama-ejenti ahlanganisa imiphumela yosesho nezimo ezingokomfanekiso angase adinge okungaphezu kokumatanisa kwamagama angukhiye okulula: angase aveze impambano-ndlela ebekwe esidlekeni, izinyunyana, nokungafakwanga okunengqondo ngokwemvelo yakhe igrafu enamagama angaphansi okwabelwana ngawo. Ukumelwa kwegrafu kungagcina lokho kwabelana, kuyilapho ukunwetshwa kwesihlahla kungakuphinda. Uma i-ComputePN isebenza njengoba kushiwo, iphepha linikeza indlela enesimiso yokuhlola ukucabanga okunjalo ngaphandle kokukhokha ngokuzenzakalelayo izindleko zokunwetshwa kwe-exponential noma izindleko zokwakheka kwendawo yonke. Inzuzo engaba khona esebenzayo ihambisana kakhulu nezinhlelo lapho ubunkimbinkimbi bemibuzo, usayizi wekhophasi, kanye nobungako buhlanganyela.

Ukugwema ukuskena kwawo wonke amadokhumenti kungase kunciphise ingcindezi yenkumbulo yemibuzo enzima, kuyilapho ukubamba ngekhanda amanodi abelwe e-DAG kungavimbela umsebenzi ophindaphindiwe. Isibopho esifunwayo siphinde sigxile kudokhumenti esebenzayo esethi kunomhlaba wonke, okungaba kubalulekile emibuzweni ekhethiwe. Le miphumela ilandela ekwakhiweni kwe-algorithmic echazwe yi-Apple, hhayi kusukela ekuthuthukisweni komkhiqizo okubonisiwe. Umthombo ubika ukuthi akukho ukunciphisa ukubambezeleka, ukukhushulwa kokuphumayo, ukonga izindleko, umphumela wamandla, noma umphumela womsebenzisi.

Leli phepha liwumphumela njengomnikelo wocwaningo, kodwa umthelela walo emphakathini awukaqiniseki. Ayikumemezeli ukwethulwa komkhiqizo noma ithi i-Apple ihlanganise i-ComputePN kusevisi yokusesha ebheke amakhasimende. Futhi ayiqinisekisi ukuthi izinjini zokusesha ezikhona noma amapulatifomu e-ejenti asebenzisa ulimi lombuzo olufana nephepha. Asikho isiqinisekiso esizimele esifakiwe kumpahla enikeziwe. Inani elisebenzayo lizoncika ekutheni imodeli esemthethweni ifana yini nomthwalo wokukhiqiza, noma ngabe ukuguquguquka kwe-algorithm kuyamukeleka, nokuthi iziphatha kanjani uma imibuzo, izinkomba, namasethi asebenzayo kukhulu noma kuxhumeke kakhulu.

Ongakubuka ngokulandelayo

Ubufakazi obulandelayo obubalulekile wukusebenzisa kanye nokulinganisa: amagrafu emibuzo yangempela, osayizi bekhophasi, ukubalwa kwedokhumenti esebenzayo, ukubambezeleka, ukusetshenziswa kwenkumbulo, nokuqhathanisa nezinhlelo ezikhona zeDocument-at-a-Time kanye neTerm-at-a-Time. Umthombo awukhombi isofthiwe, i-API, ukusetshenziswa, indawo yokubuyekezwa kontanga, noma ukwamukelwa usesho noma umkhiqizo we-AI.

Umbuzo wokuqala ukuthi izinzuzo eziyinkimbinkimbi ezifunwayo zihumushela ekusebenzeni kwesistimu elinganisiwe. Ubufakazi obuwusizo bokulandelela buzobandakanya i-benchmark corpora, izinqubo zokukhiqiza imibuzo, ukusatshalaliswa kokujula kwe-DAG nokuhlangana kabusha, osayizi bamadokhumenti wendawo yonke, osayizi abasethiwe abasebenzayo, inkumbulo ephezulu, nokubambezeleka kokuphela ukuya ekupheleni. Ukuqhathanisa kufanele kufake kokubili izisekelo Zombhalo-ngesikhathi kanye Nesikhathi Sesikhathi, kanye namacala aphikisanayo afaka amazwi amancane aphindaphindiwe kanye nokuphika okubanzi. Umthombo wamanje unikeza izimangalo ze-asymptotic kodwa azikho zalezo zilinganiso.

Umbuzo wesibili ukuthi iComputePN ikhona yini njengesoftware esebenzisekayo. Ikhasi le-Apple alixhumi endaweni yokugcina, iphakheji, i-API, ukuqaliswa kobuchwepheshe, noma imiyalelo yokukhiqiza kabusha imiphumela. Futhi ayisho ukuthi indlela ingangezwa yini ezinjinini zenkomba ehlanekezelwe, kungakhathaliseki ukuthi idinga isakhiwo senkomba esisha, noma ingabe isekela izibuyekezo, izinga, ukuhlunga, ukwenza okusabalalisiwe, noma imibuzo ehambisanayo. Lokho okweqiwe kwenza kungenzeki kulo mthombo ukuhlola ukulungela ukuthunyelwa noma ukuhambisana nezitaki ezikhona zokubuyisa.

Umbuzo wesithathu uwukuqinisekisa kanye nobubanzi. Ikhasi libiza umsebenzi njengephepha elishicilelwe kodwa aliyibonisi ingqungquthela, iphephabhuku, inqubo yokubuyekeza, noma impinda yangaphandle. Ukudalulwa kwesikhathi esizayo kufanele kucacise ulimi lokubuyisa olunembile, ukuqagela okusemuva kokumelwa kwedokhumenti esebenzayo, nokuziphatha kwesakhiwo esinegethivu ngaphansi kokufana okuminyene noma cishe kwendawo yonke. Kuzobaluleka futhi ukubona ukuthi ingabe amasistimu we-AI-ejenti akhiqiza ngempela uhlobo lwemibuzo ye-Boolean DAGs echazwe, nokuthi ingabe indlela ehlongozwayo ithuthukisa ukwethembeka noma isungula kuphela indlela yokuhlola yethiyori.

Imihlahlandlela ehlobene nemibuzo