Dellu ci xibaar yi
YeesalAI Understanding

Këyitu PhysElite dafa wax ni LLM bi gëna mag moo facc 33.7% ci jafe-jafe physique ci niveau Olympiad

Benchmark bu bees bu 11,586 jafe-jafe physique yu sukkandiko ci diagram, dafa wane ni li gëna am doole ci 18 xeeti làkk yu bari yuñ natt, amul lenn ludul 33.7% ci tontu yu jaar yoon.

5 min readRead the primary source
Primary-source image accompanying PhysElite paper reports that leading multimodal LLM solved 33.7% of Olympiad-level physics problems
Këyitu xët bu njëkkSource biñ enregistre
Siiwalkat
arxiv.org
Lëkkalekaayu cosaan
arxiv.orghttps://arxiv.org/abs/2608.25097
Xeetu balluwaay
Këyitu njëkk - ab yëgle ofisel, këyit, dosiye, wala xëtu pàrti bu njëkk bi ñuy jàng ci saasi.
KontekstXam lii ci 60 seconde

Tambalil fii

Term yu am solo

Modelu làkk bu mag (LLM)
Benn xeetu làkk buñ tàggat ci corpus mbind yu bari ngir sos ak jàngat mbind.
Référence
Test buñ yamale wala ensemble done yuñ jëfandikoo ngir natt ak méngale liggéeyu model bi.
Dataset
Dajale misaal yuñ yamale wala yuñu yamalewul yuñ jëfandikoo ngir tàggat, saytu wala saytu.
Nattal sa boppModèlu IA leeral quiz

Lu xew

Gëstukat yi dugal nañu PhysElite, muy benn référence multimodal bu am ñaari làkk, ñu defaree ko ngir natt xeeti làkk yu bari ci jafe-jafe physique yu tollu ci Olympiad. Done yi dañu am 11,586 jafe-jafe, ay diagram yuy wane, ay pexe yu bawoo ci làkku chinois-anglais ak tontu yu mujj. Këyit dafay wane njariñu jàngat yiñ def ci 18 xeeti làkk yu bari yu am source ubbeeku ak source tëju, ak model bi gëna am doole buy yegg ba 33.7% ci tontu yu jaar yoon.

Këyit dafay wane PhysElite ni référence ngir xalaat ci physique ci niveau Olympiad ci xeetu làkk yu bari yu bari. Ñu ngi ko jox arXiv ci 25 ut 2026. Bindkat yi dañu wax ni physique yi fi nekk dañu yam ndax amu ñu ay jafe-jafe yu doy te duñu tënk bu baax leeral yi ñuy gis, xam-xam physique, ak anam yu ñuy saafara jéego-ci-jéego.

Sunu sukkandikoo ci abstract bu këyit bi, PhysElite amna 11,586 jafe-jafe Olympiad ci làkku Chinois ak Angale. Jafe-jafe bu nekk dafa ànd ak diagram buñ mëna gis, muy génne ci ñaari làkk ñu xaaj ko ay jéego, ak tontu bu mujj. Woykat bi dafa wax ni dañu génne done yi, waaye bind biñ joxe du am lëkkalekaay bi ñuy yóbbu wala leeral lisence bi, formaa bi, wala li ñuy laaj ngir mëna ci dugg.

Auteur yi jàngat nañu 18 xeeti làkk yu bari, boole ci sistem open-source ak closed-source yépp. Royuwaay bi gëna am doole amna 33.7% ci tontu yu jaar yoon. Gëstukat yi amal nañu itam jàngat ci wàllu jéego ngir xam fu model yi jaarul yoon ci biir seen xalaat. Source bi joxeewul tur model yi, limu jéem yi ci jafe-jafe bu nekk, sàrti poñ yi, wala resultaa yi leer ci xeetu jafe-jafe bi.

Ay leeral ci cosaan: arxiv.org ↗

Lu tax mu am solo

Resultaa bi dafay wane ni performance yu am doole ci physique yi fi nekk wala ci test yu gëna bari mën nañu baña tekki ci pexe yu wóor ci jafe-jafe physique yu jafe, yu aju ci diagram, yu bari jéego. PhysElite dafa bëgga gëna wane ñakk kattan yooyu ci jàngat tontu yu mujj yi ak etape yu benn-benn ci doxalinu xalaat bu benn model.

bi dafay wax ci benn ñakk kattan ci ni ñuy faral di xoole sistemu xalaat yu IA yi. Benn model mën na joxe ay leeral yu neex wala mu def lu baax ci laaj yu gàtt te ba leegi mën na ñàkk sudee dafa wara tekki ab diagram, tànnee njàngale yu am solo ci wàllu physique, def ay deduction yu bari yu lëkkaloo, ba noppi am tontu bu dëggu. PhysElite dañu ko tëral ngir jàppale jafe-jafe yooyu, du ñu jàppee physique ni tontub laaj ci bind kese.

33.7% ci resultaa yi ñu joxe dañu am njariñ ndax dañu tëral ay yamaleg li sistem yiñ natt wane ci liggéey bu jafe lool bii. Wanewul ni xeetu làkk yu bari yi mënu ñu saafara physique Olympiade, te wanewul ni xeetu làkk bu nekk dafay def ci benn niveau. Dafay wane, buñu sukkandikoo ci këyit bi, ni sistem bi gëna am doole biñ jàngat mënul saafara jafe-jafe yu bari ci anam wu jaar yoon.

Evaluation bi ci jéego yi mën na gëna am njariñ ci gëstukat yi ak jëfandikukat yi, moo gën benn poñ bu mujj. Sudee jafe-jafe yi dañu gëna bari ci tekki diagram yi, tànneef equation yi, def xayma yi, wala boole resultaa yi ci digg yi, developpeur yi mën nañu gëna mëna xool bu baax ñakk kattan yooyu. Source biñ joxe waxul ban etape moo jur njuumte yu gëna bari, kon valeur diagnostic bi mënul baña xoolaat ci detay.

Ngir askan wi, li ñu gis defay artu ci jappe sistem yu juntuwaay yu bees yi ni ay weccitu xalaat yu xarañ ci physique. Sistem buy joxe ab pexe mu lalu ci jéego yu bari te jaarul yoon, mën na nekk lu jafe ci ku nekkul eksper ngir mu saytu ko. Design bilingual ak visuel bu bi mën na jàppale itam ñu wane ay jafe-jafe yu nëbbu ci jàngat yi lalu ci bind Àngle wala formaa tontu yu mujj yu gàtt.

Interactive Mechanism

Mekanism buy weccoo xalaat: naka lay doxee

Saytu xarala yu bees yi ci ginaaw yokkute bii ci anam wu weccoo xalaat.

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.
Saytu konsept buy weccoo xalaat+10 Points
AI Models Explained Quiz

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

Li nga wara seetaan ci topp

Laaj yi gëna am solo ñooy ndax setu done yi ak kodu jàngat bi mën nañu leen jëfandikoo bu baax ngir ñu mëna jëfandikoo seen bopp, naka la performance di wuutee ci topic physique yi ak formaa yi ñuy xool, ak ndax model yu ëlëg yi dina ñu gëna mëna jëfandikoo ci bi te duñu yéem ci pexe yuñ jottali. Source bi xamul model yiñ natt wala joxe leeral ci jàngat bi, kon 33.7% biñ wax waru ñu ko jëfandikoo ngir rang sistem yuñ tànn.

Baamtu moom boppam mooy test bi ci saasi. Gëstukat yi dina ñu wara xoolaat jafe-jafe yiñ génne, diagram yi, pexe yi, ak anam yi ñuy joxe poñ yi ngir xam ndax bi amul benn lu jaxasoo, mbir yuñ ñaari yoon, wala yeneen mbir yu mëna indi jafe-jafe ci resultaa yi. Boroom bi dafay firndeel génnug done yi waaye du joxe leeral yu doy ngir jàngat ni ñu ko mëna jëfandikoo wala ñu mëna ko defaraat.

Coverage yi ci kanam war nañu wane ni 33.7% ci lim yi di xaajaloo ci physique yi, niveau jafe-jafe yi, xeetu diagram yi, lakk yi, ak formaa tontu yi. Dafay baax itam nga xam ndax model yi dañuy jot benn laaj, ndax jumtukaay yu melni kalkulatër ñu ngi nangu, ak ni ñuy jëfandikoo xaalis bu néew. Benn ci leeral yooyu feeñul ci balluwaay biñ joxe.

Referanse bi mën na gëna am njariñ sudee jàngat yi ci topp dañuy méngale xeetu model yi ci diir bu gàtt ba noppi ñu joxe tontu bu mujj bu jaar yoon ak wóor ci wàllu jéego. Gëstukat yi dañu wara xool itam ndax model yi mën nañu xàmmee lu wóorul, xàmmee jéego yu jaarul yoon ci digg bi, wala laaj leeral sudee diagram bi leerul. Abstract bu këyit bi waxul ci jëf yooyu.

Li gëna am solo buñu xamul mooy naka la PhysElite di wane liggéeyu physique dëgg. Jafe-jafe Olympiad yi dañuy jafe lool te mën nañu baña méngoo ak tàggat yi ñuy def ci klaas, jàngat ci labo, jëmmal ingenieur, wala jëfandikoo gëstu. Source bi dafay taxawal yaatuwaayu bi ak resultaa biñ xamle, waaye taxawul ni performance ci PhysElite di waxee performance ci yeneen jekkal yi.

Buñu leen boole, tegtal yiñ am dañuy jàppale joxe leeral yu gàtt ci resultaa bi. 33.7% ci lim bi bokk na ci liñu xamle ci tontu-jaar ci 18 xeeti làkk yu bari yuñ natt ci bii. Dañu ko wara jàng ci wetu jafe-jafe ñaari làkk yi ci done yi, diagram yu ñuy gis, derivations, ak tontu yu mujj, ak itam jàngat bu wuute ci niveau de processus. Ndax source biñ joxe du am tur model, sàrti poñ, lim jéem, xaaj topic, wala kodu jàngat, jàngkat yi mënu ñu jëfandikoo abstract bi kese ngir tabaxaat méngale bi wala xam ban wàll ci liggéey bi moo jur resultaa bi. Kon leeral yooyu ñàkk dañu am solo ngir xam rapoor yi ci topp ci PhysElite.

Gid ak quiz yu ci méngoo

Model IA leeral nañu koTransformatërTaggat ci IAJikko yu AINatt li nga xam — natt quiz IA bu amul faydaSeetal benn baat IA ci sunu glossaireToppal toppukaayu génne xeetu IA
Gis nga lii am njariñ?