Test-Time calcul scalage
Test-time compute scaling mooy jox model bi diir bu gëna bari ngir xalaat ak xayma suy tontu benn laaj, du def ko lu gëna mag ci diiru tàggat yaram.
Résumé
It is the breakthrough behind 'reasoning models' that can solve hard math and coding problems by deliberating before responding.
Plongeur bu xóot
Ay at ci ginaaw, yokkuteg IA dafa tekki ni dafay yokk tàggat yaram: gëna bari done, gëna bari paramet, gëna tàggat ordinatër. Test-time calcul scaling dafay yokk ñaareelu axe, di dugal xaalis bu bari ci inference. Du joxe tontu ci saasi, ab xeetu xalaat dafay defar ay xalaat yu gudd ci biir, di banneexu jéego yi, di xool liggéey bi, ak di dellu ginaaw. Pexem yi bokkunaci seen xalaat bu yaatu, jël misaalu pexe yu bari ak tànn bi gëna baax (kombis-sa-bopp wala gëna-ci-N), ak seetlu garab gu ñuy teg ci xeetu verifier wala xeetu neexal. OpenAI's o1 ak o3, DeepSeek-R1, ak __AIU_PROTECTED_8_'s xalaat yu yaatu siiwal lii: njub ci math ak prograam ci joŋante dafay tëb bu baax soo bàyyee model bi 'xalaat lu gëna yàgg,' njub ci njaay ci jafe-jafe ak njëg ngir njuumte.
Gis-gis xarala
Royuwaay bi dañu ko tàggat ak jàng buy dooleel ngir génne ay token xalaat yu am njariñ, su ko defee nga xaaj 'budget xalaat.' Token yu bari dañu koy may mu dindi jafe-jafe yi, jàpp ay njuumteem, ba noppi xool boppam. Best-of-N sampling ak seetlu buñ teg ci saytukat bi dañuy yokk calcul parallel: defar jéem yu bari, jox leen poñ, tëye ki jël raw gàddu gi. Li gëna am solo mooy model yu ndaw yi am test-time compute yu bari mën nañu méngoo ak model yu gëna mag yiy tontu ci saasi, di soppali courbe coût bi.
njeextalu pexe
Gaawaay ak yaatuwaay
Liggéeyukaay yi ci làkk yi mën nañu gëna gaaw te duñu yàq deggoo gi.
Dugg ak yegg
Dafay yaatal jëfandikoo gi ci làkk yi ak ci anam yi ñuy jokkoo.
dogal yu gëna leer
Ekip yi mën nañu gëna yàgg ci àtte ci jamono ji otomatisation di liggéey ci baamtu.
Ëlëgu Test-Time
Test-time calcul mooy levier bi njëkk ci scaling ci wetu tàggat yaram. Xaarandi budget adaptive fu model bi di dogal ni ñuy xalaatee ci jafe-jafe, xalaat bu gëna yomb jaaraleko ci distillation chaine yu gudd ci yu gëna gàtt, ak 'agent' loops yuy boole xalaat ak woote jumtukaay ak seetlu web. Lu inference hardware di gëna xarañ, xalaat bu jaar yoon mooy nekk default ci liggéey yu am solo yu melni gëstub science, ingenieur losisel, ak waajal bu jaffe, fekk seetlu yu gaaw yi dañuy gaaw te yomb.
Doxal ci àdduna dëgg
OpenAI xeetu o1 ak o3 yi dañuy xalaat ci jafe-jafe math yu tollu ci Olympiad jéego par jéego, di raw bu baax xeetu tontu yi ci saasi ci AIME ak ci xeetu joŋante yi.
DeepSeek-R1 dafa jëfandikoo njàngum dooleel ngir jàngal xalaat yu gudd, di wane bu baax njariñu njubte bu rëy ci xayma inference bu gëna bari.
Claude's mode xalaat bu yaatu dafay may developpeur yi ñu def budget token suko defee model bi di xalaat lu gëna yàgg ci kodage yu jafee xam wala liggéey jàngat balaa ñuy tontu.
AlphaCode ak yeneen sistem yu mel noonu, dañuy jël ay junni prograam yuñ tànn ci jamonoy test, ba noppi segg leen, daal di leen rang ngir saafara jafe-jafe yi ci prograam bi.
Risk yi ak balustrade yi
Lépp lu jaarul yoon mën na dugg ci rapoor yi, jàppale ci liggéey bi, wala ci njariñu gëstu bi.
Sensibilite bu gaaw mën na jur njariñ yu wuute ci laajte yu noonu mel.
Done yu am solo mën nañu feeñ sudee seytu jëfandikoo gi néew doole.
Roadmap ngir samp gi
Mandargal formaa génne gi, melokaan bi, ak standard kalite yi laata ngay dugal ko.
Tontu yu am solo ak balluwaay yu wóor saa yu dëggu bi di am solo.
Fexeel am barabu xool nit ñi ngir am njariñ yu am solo.
Toppal anami gacce yi ak di faral di tàggataat ay laaj wala def-liggéey.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Test-Time Compute Scaling quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Yokkateg waxtu test
Laaj yi ñuy faral di laaj
What is Test-Time Compute Scaling?
Test-time compute scaling mooy jox model bi diir bu gëna bari ngir xalaat ak xayma suy tontu benn laaj, du def ko lu gëna mag ci diiru tàggat yaram. Mooy jéego bi am ci ginaaw 'modèlu xalaat' yi mëna saafara jafe-jafe math ak kodage yu jafe suñu xalaatee bu baax balaa ñuy tontu.
Luy 'test-time calcul'?
Test-time (inference-time) calcul mooy liggéey bi ñuy def ngir génne ab tontu, wuute na ak calcul bi ñuy jëfandikoo ci diiru pre-training.
naka lay def ba xeetu xalaat yu melni o1 di jëfandikoo calcul time-time bu gëna bari?
Dañuy génne xalaat yu yaatu jéego par jéego, di xool ak di xool saafara yi balaa ñuy jël dogal ci tontu bu mujj.
Lan mooy xéewal gi gëna am solo ci diiru test-time escalier?
Bayyi benn model mu xalaat lu gëna yàgg mooy gëna yombal njub ci jafe-jafe yu metti yi waaye dafay yokk diiru tontu bi ak njëgu ordinatër bi.
Luy 'gën-ci-N' échantillonnage?
Best-of-N dafay defar ay jéem-jéemi saafara yu bari ci paralel ba noppi jëfandikoo ab poñkat wala saytukat ngir tëye ki gëna am doole, benn xeetu test-time scaling.
Lan moo waral ñuy jàppee xayma waxtu test ni 'axe scaling' bu bees?
Ay at ci ginaaw scaling dafa tekki daw yu gëna mag ci tàggat yaram; test-time xayma yokk ñaareelu anam ngir yokk kàttan, ci xayma lu gëna bari ci inference.