Ku laabo Warka
Hal-abuurnimoAI Understanding warbixin kooban

Waraaqu waxay soo jeedinaysaa isbarbardhigga xawaaraha si loo cabbiro abaal-marin-hagaajinta qaababka fidinta

Cilmi-baarayaashu waxay soo jeedinayaan is-waafajinta xawaaraha ku saleysan abaalmarinta, hab-raac-bilaash ah oo si toos ah u cusbooneysiiya moodooyinka fidinta meelaha xawaaraha, waxayna soo sheegaan, gaaraan natiijooyin la mid ah ama ka wanaagsan hababka suurtagalka ah ee ku saleysan qiimaha tababarka.

5 min readRead the primary source
Primary-source image accompanying Paper proposes velocity matching to scale reward fine-tuning for diffusion models
Dukumeentiga isha aasaasiga ahIsha la duubay
Daabacaha
arxiv.org
Xidhiidhka isha
arxiv.orghttps://arxiv.org/abs/2608.23664
Nooca isha
Dukumeentiga aasaasiga ah - ogeysiis rasmi ah, warqad, xereyn, ama bogga xisbiga koowaad waxaan si toos ah u akhrinay.
Dulucda sheekadaKu fahan tan 60 ilbiriqsi gudahood

Halkan ka bilow

Qodobbada muhiimka ah

Hagaajinta
Sii wadida tababarka xogta gaarka ah ee domain-ka si loo waafajiyo qaabka horay loo tababaray ee hawl gaar ah.
Xoojinta Waxbarashada
Tababbarka abaal-marintu waxay muujisaa halka wakiilku ku barto ficillada kordhinaya soo laabashada muddada dheer.
Qaabka Faafida
Nashqada wax soo saarka ee barata in dib loo rogo buuqa si loo habeeyo sawirada, maqalka, ama waxyaabaha kale.
Is tijaabiMoodooyinka AI Kedis La Sharaxay

Maxaa dhacay

Koox ka kooban todoba cilmi-baarayaal ayaa soo gudbiyay warqad arXiv ah oo soo jeedinaysa abaal-marin ku salaysan is-waafajinta xawaaraha (RVM), hab loogu hagaajinayo qaababka fidinta fidinta xagga dookhyada aadanaha ama yoolalka gaarka ah ee hawsha. RVM waxay si toos ah u cusboonaysiisaa goobta xawaaraha moodeelka halkii ay dib-u-dhis ugu samayn lahayd suurtagalnimada in ay beeniso raad-raacyada ama qiyaasidda fursadaha dhammaadka.

Isha ayaa ah soo gudbinta arXiv oo ku taariikhaysan Agoosto 24, 2026, cinwaankeeduna yahay "Barashada Xoojinta Xoojinta ee Qaababka Faafinta iyadoo loo marayo Iswaafajinta Xawaaraha." Qorayaashu waxay soo bandhigaan hagaajinta abaal-marineed si ay ula qabsadaan moodooyinka faafinta ee dookhyada aadanaha iyo ujeedooyinka gaarka ah ee hawsha. Sheegashadooda dhexe ayaa ah in dariiqooyinka jira ay amaahdaan mashiinnada siyaasadda-jilicsan ee loogu talagalay moodooyinka is-difaacida, abuurista kakanaanta dheeraadka ah sababtoo ah moodooyinka faafintu ma bixiyaan fursadaha la heli karo ee muunado la soo saaray.

Habka la soo jeediyay, isbarbardhigga xawaaraha abaal-marin-ku-salaysan, ayaa lagu tilmaamay mid-raac-raac-la'aan ah. Halkii laga dhisi lahaa suurtagalnimada beddelka diidmada stochastic ama qiyaasida suurtagalnimada dhammaadka dhammaadka leh ee caddaynta hoose, RVM waxay si toos ah uga shaqeysaa goobta xawaaraha. Marka loo eego qoraalka aan la taaban karin ee warqadda, cusbooneysiintu waxay xoojinaysaa jihooyinka la xidhiidha jiilalka abaal-marinta sare leh waxayna xakamaysaa jihooyinka la xidhiidha jiilalka abaal-marinta hooseeya.

RVM waxaa ku jira erey barroosin oo ikhtiyaari ah oo loogu talagalay in lagu xakameeyo ka leexashada xawaaraha tixraaca. Qorayaashu waxay sidoo kale yiraahdaan qaab-dhismeedku wuxuu ka soo kaban karaa hababka hagaajinta wanaagsan ee dhowaan, oo ay ku jiraan RAM iyo DiffusionNFT, sida kiisas gaar ah. Taas ayaa ka dhigaysa soo jeedinta qaabaynta midaynta iyo sidoo kale qaanuun cusub oo cusub, in kasta oo isha la keenay aanu sharaxin ka-soo-jeedinta saxda ah ama xaaladaha ay ku jiraan kuwa u dhigma.

Cilmi-baadhayaashu waxay ka warbixiyaan imtixaannada kala duwan ee noocyada kala duwan ee faafinta-qaabaynta abaal-marin-hagaajinta hawlaha. Waxay yiraahdeen RVM waxay la tartameysay ama ka sarreysay habab siyaasad-si-celis ah oo ku saleysan raad-raac iyadoo ay u baahan tahay kharash tababar oo aad u yar. Jiilka fiidiyowga, waxay soo sheegaan in abaalmarinta doorbidida caadiga ah ay soo saari karto muuqaal nadiif ah laakiin wax soo saar ku dhow; waxay soo bandhigaan abaal-marin firfircooni-raac ah waxayna yiraahdaan waxay wanaajisay dhaqdhaqaaqa iyadoo sidoo kale hagaajinaysa waxqabadka guud ee VBench. Dhibcaha tirada, magacyada moodeelka, miisaaniyada tababarka ama miisaska isbarbardhigga laguma darin isha la bixiyay.

Faahfaahinta isha: arxiv.org ↗

Maxay muhiim u tahay

Waraaqdu waxay ku doodaysaa in cusboonaysiinta xawaaraha tooska ah ay fududayn karto oo ay yarayn karto kharashka abaal-marin-hagaajinta muuqaalka baaxadda weyn iyo nidaamyada fidinta muuqaalka. Tijaabooyinkeeda muuqaalku waxay sidoo kale tilmaamayaan ka-ganacsiga abaal-marinnada doorbidida caadiga ah: waxa laga yaabaa inay door bidaan wax-soo-saar muuqaal ah oo nadiif ah oo leh dhaqdhaqaaq yar.

Haddii sheegashada warqaddu ay noqoto mid guud, si toos ah u wanaajinta qaabka fidinta ee matalaadda xawaaraha hooyo waxay ka dhigi kartaa abaal-marin-habayn fudud si loo cabbiro. Muhiimadda la taaban karo maaha oo kaliya shaqo lumis cusub: habku waxa uu bartilmaameedsanayaa xisaabinta iyo korka algorithm ee la xidhiidha wanaajinta siyaasadda ku salaysan suurtagalnimada. Qiimaha tababbarka hoose wuxuu ka dhigi karaa doorbid ama la qabsiga hawsha gaarka ah mid aad u fudud oo ay heli karaan kooxaha ka shaqeeya sawir-qaadayaasha waaweyn iyo fiidiyoowga.

Warqaddu waxay sidoo kale diiradda saaraysaa naqshadeynta abaalmarinta lafteeda. Tijaabooyin muuqaal ah oo qorayaasha ah, abaal-marin xoogga la saarayo nadaafadda muuqaalka ayaa lagu soo waramayaa inay door bidday wax soo saar yar oo dhaqdhaqaaq yar leh. Abaalmarintooda raadraaca firfircooni leh waxaa loo soo bandhigay hab lagu cabbiro dhaqdhaqaaqa si wax ku ool ah, iyadoo weli la hagaajinayo natiijada guud ee VBench ee warqadda. Tani waxay tusinaysaa in hagaajinta nidaamka wax-soo-saarka ay ku xidhan tahay oo keliya qaanuunka cusboonaysiinta laakiin sidoo kale waxa habka wanaajinta la waydiiyo inuu qiimeeyo.

Ereyga barroosinka ikhtiyaariga ah waa arrimo sababtoo ah hagaajinta abaal-marintu waxay ka dhaqaajin kartaa moodal-dhaqanka tixraaca. Isha ayaa sheegaysa in barroosinku uu xakameeyo ka leexashada xawaaraha tixraaca, laakiin ma sheegayo sida xakamayntaasi u saamayso tayada, kala duwanaanta, xasiloonida ama khatarta ah in abaalgudku ku dhaafo. Faahfaahintaasi waxay go'aamin doontaa in RVM ay faa'iido u leedahay la qabsiga la kantaroolo halkii ay ka ahaan lahayd hagaajinta bartilmaameedka oo keliya.

Natiijadu waxay ahaanaysaa sheegasho cilmi baaris ah oo laga helay hal warqad arXiv, ma aha caddaynta in tababbarka qaab-faafintu uu si ballaaran isu beddelay. Nuqulka la keenay ma tilmaamayo moodooyinka la tijaabiyay, xog-ururinta, baaxadda abaal-marinaha, xisaabinta kaydinta, kala duwanaanshaha tirakoobka ama kiisaska guul-darrooyinka. Sidoo kale ma cadeeynin in habka uu si siman ugu shaqeeyo sawirada, fiidyaha iyo codsiyada kale ee faafinta, ama haddii faa'iidooyinkeedu ay ku xiran yihiin naqshado abaal-marineed oo gaar ah.

Interactive Mechanism

Farsamaynta Is-dhexgalka: Sida Dhabta Ay U Shaqeyso

U baadh tignoolajiyada hoose ee ka dambeeya horumarkan si isdhexgal leh.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Hubinta Fikradda Is-dhexgalka+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Maxaa la daawan doona xiga

Su'aalaha ugu muhiimsan ayaa ah in qiimaha la sheegay ee RVM iyo faa'iidooyinka tayada ay ku hayaan dhammaan tijaabooyinka si madax-bannaan loo soo saaray, shaqooyinka abaal-marinaha, cabbirrada moodeelka iyo hawlaha jiilka, iyo haddii habka uu iska ilaalin karo hagaajinta abaalmarinta cidhiidhiga ah ama marin-habaabinta. Isha la keenay ma bixiso natiijooyin tirooyin, tafaasiisha fulinta, xisaabinta miisaaniyadaha ama caddaynta ansixinta dibadda.

Soo saarista waa in ay marka hore diirada saarto isbarbardhigga warqada iyo hababka siyaasad-ku-salaysan ee ku salaysan jihada. Jeegaga waxtarka leh waxaa ku jiri doona qaab isku mid ah iyo hawl isku mid ah oo hoos yimaada miisaaniyada xisaabinta ee iswaafaqsan, sababtoo ah "kharashka tababarka oo aad hoos u dhacay" ayaa macno leh kaliya marka xisaabinta ay ku jirto dhammaan tababarka iyo shaqada qiimeynta. Isha la keenay ma bixinayso saamiga qiimaha tirooyinka, markaa xajmiga faa'iidada la sheegay lama garanayo.

Nakhshadeynta abaalgudku waxay u qalantaa qiimeyn gooni ah oo laga soo bilaabo cusbooneysiinta xawaaraha. Qorayaashu waxay yiraahdeen, marka cusbooneysiinta xawaaraha la fududeeyo, kala duwanaanshaha khasaaraha gaarka ah wuxuu ka yar yahay abaalmarinta iyo naqshadeynta barroosinka. Sheegashadaasi waxay soo jeedinaysaa in horumarku aad ugu xidhan yahay sida jiilalka abaalgudka sare iyo kuwa hooseeya loo calaamadeeyay ama loo buun buuniyay. Tijaabooyin madax banaan sidaas darteed waa inay beddelaan shaqooyinka abaal-marinta oo ay cabbiraan haddii faa'iidooyinku sii socdaan wixii ka baxsan ujeedooyinka la doortay.

Jiilka fiidyaha, goobjoogayaashu waa inay hubiyaan in abaal-marinnada raadraaca firfircooni ay horumariyaan dhaqdhaqaaq macno leh ama ay kordhiyaan isbeddel muuqda oo keliya. Isha ayaa ka warbixisa dhaq-dhaqaaq wanaagsan iyo natiijada guud ee VBench oo la hagaajiyay, laakiin ma bixiso cabbirada hoose ama ma qeexdo hababka guul darada. Qiimayntu waa in ay eegtaa tayada muuqaalka, joogtaynta ku meel gaadhka ah iyo kala duwanaanshiyaha wada jir ah, oo ay ku jiraan kiisaska mooshinku aanu la rabin ama khilaafaya waxa loogu talagalay.

Muhiimadda ballaaran ee warqaddu waxay ku xirnaan doontaa hirgelinta iyo faahfaahinta xaqiijinta ee aan ka jirin isha la bixiyay. Waxyaabaha muhiimka ah ee aan la garanayn waxaa ka mid ah cabbirka xawaaraha saxda ah, goobaha barroosinka, moodeelka iyo kaydinta xogta, muddada tababarka, soo saarista isbarbardhigga la soo sheegay, iyo haddii RVM ay weli xasiloon tahay maaddaama ujeedooyinka abaalgudku isbeddelayaan. Waraaqaha daba-galka ah, koodka la sii daayay iyo ku celcelinta madax-bannaan ayaa gacan ka geysan doona in la ogaado in soo jeedintu tahay hab miisaan guud ama natiijo ku xidhan tijaabooyin gaar ah.

Tilmaamaha la xidhiidha & su'aalaha

Moodooyinka AI ayaa la sharaxayTababarka AITransformersMustaqbalka AITijaabi waxaad taqaan - isku day kedis AI oo bilaash ahKa raadi erey AI qaamuuskeenaRaac qaabka AI raadraaca sii deynta
Tan faa'iido ma u heshay?