Subira ku makuru
Guhanga udushyaAI Understanding ibisobanuro

Urupapuro rwa PhysElite ruvuga ko kuyobora LLM ya multimodal byakemuye 33.7% byibibazo bya fiziki yo mu rwego rwa Olympiad

Ibipimo bishya by’indimi ebyiri 11,586, bishingiye ku gishushanyo mbonera cy’ibibazo bya fiziki byerekana ko bikomeye muri 18 bigeragezwa by’ururimi rwimvugo nyinshi byageze kuri 33.7% gusa.

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
Inyandiko y'ibanzeInkomoko yanditse
Umwanditsi
arxiv.org
Ihuza ry'inkomoko
arxiv.orghttps://arxiv.org/abs/2608.25097
Ubwoko bw'inkomoko
Inyandiko y'ibanze - itangazo ryemewe, impapuro, dosiye, cyangwa urupapuro rwambere-dusoma mu buryo butaziguye.
ImirongoSobanukirwa ibi mumasegonda 60

Tangira hano

Amagambo y'ingenzi

Ururimi runini (LLM)
Ururimi rwicyitegererezo rwahuguwe kumyandiko minini corpora kubyara no gusesengura inyandiko.
Ibipimo
Ikizamini gisanzwe cyangwa dataset ikoreshwa mugupima no kugereranya imikorere yicyitegererezo.
Dataset
Icyegeranyo cyingero zubatswe cyangwa zitubatswe zikoreshwa mumahugurwa, kwemeza, cyangwa kugerageza.
IsuzumeModeri ya AI Yasobanuwe Ikibazo

Byagenze bite

Abashakashatsi berekanye PhysElite, igipimo cyindimi ebyiri kigenewe gupima imiterere yindimi nini kubibazo bya fiziki yo mu rwego rwa Olympiad. ikubiyemo ibibazo 11,586, igishushanyo mbonera, ibisubizo byubushinwa-Icyongereza, nibisubizo byanyuma. Uru rupapuro rutanga ibisubizo bivuye mu isuzuma rya 18 rifungura-isoko-rifunze-indimi nyinshi zerekana imvugo, hamwe nicyitegererezo gikomeye kigera kuri 33.7%.

Uru rupapuro rwerekana PhysElite nk'igipimo ngenderwaho cya fiziki ya Olempike yo mu rwego rwa fiziki itekereza ku buryo bunini bw'ururimi runini. Yashyikirijwe arXiv ku ya 25 Kanama 2026. Abanditsi bavuga ko ibipimo ngenderwaho bya fiziki biriho ari bike kubera ko bitarimo ibibazo bihagije bigoye kandi ko bidakubiyemo amakuru yose y’amashusho, ubumenyi bw’ubumenyi bwa fiziki, hamwe n’uburyo bwo gukemura ibibazo.

Dukurikije ibisobanuro by'uru rupapuro, PhysElite ikubiyemo ibibazo 11.586 byo mu rwego rwa Olympiad mu rwego rw'igishinwa n'icyongereza. Buri kibazo cyahujwe nigishushanyo mbonera, inkomoko yindimi ebyiri zacitsemo intambwe, nigisubizo cyanyuma. Inkomoko ivuga ko yasohotse, nubwo inyandiko yatanzwe ntabwo ikubiyemo aho ujya cyangwa ngo isobanure uruhushya rwayo, imiterere, cyangwa ibisabwa.

Abanditsi basuzumye imiterere 18 yindimi nyinshi, harimo sisitemu ifunguye kandi ifunze-sisitemu. Icyitegererezo gikomeye cyageze kuri 33.7% ibisubizo byukuri. Abashakashatsi bakoze kandi intambwe-ntambwe yo gusuzuma inzira kugirango bamenye aho moderi zananiranye murwego rwo gutekereza. Inkomoko ntabwo itanga amazina yicyitegererezo, umubare wibigeragezo kuri buri kibazo, amategeko yo gutanga amanota, cyangwa ibisubizo birambuye kubwoko bwibibazo.

Ibisobanuro birambuye: arxiv.org ↗

Impamvu ari ngombwa

Igisubizo cyerekana ko imikorere ikomeye kuri fiziki iriho cyangwa ibizamini rusange byo gutekereza ntibishobora guhindurwa kubisubizo byizewe kubibazo bitoroshye, biterwa nigishushanyo, ibibazo byinshi bya fiziki. PhysElite igamije gutuma izo ntege nke zigaragara mugusuzuma ibisubizo byanyuma hamwe nicyiciro cyihariye cyibikorwa byo gutekereza.

Ibipimo byerekana intege nke zifatika muburyo sisitemu yo gutekereza ya AI isuzumwa. Icyitegererezo kirashobora gutanga ibisobanuro neza cyangwa gukora neza kubibazo bigufi mugihe bikiri binaniranye mugihe bigomba gusobanura igishushanyo, guhitamo amahame yumubiri afatika, gukora ibicuruzwa byinshi bifitanye isano, no kugera kubisubizo nyabyo. PhysElite yateguwe hafi yizo mbogamizi aho gufata fiziki nkibisubizo byanditse gusa.

Ibisubizo byatangajwe 33.7% ni ingaruka kuko ishyiraho imipaka igaragara kubyo sisitemu yapimwe yerekanye kuri iki gikorwa cyoroshye cyane. Ntabwo yerekana ko imvugo yindimi nyinshi zidashobora gukemura ibibazo bya fiziki ya Olympiade, kandi ntibisobanura ko buri cyitegererezo gikora kurwego rumwe. Irerekana, ukurikije impapuro, ko na sisitemu ikomeye yasuzumwe itakemuye ibibazo byinshi bipimo neza.

Intambwe-urwego rwo gusuzuma rushobora kuba ingirakamaro kuruta amanota imwe yanyuma kubashakashatsi nabakoresha. Niba kunanirwa bibaho cyane cyane mugusobanura ibishushanyo, guhitamo ibingana, gukora ibarwa, cyangwa guhuza ibisubizo hagati, abitezimbere barashobora kwibasira izo ntege nke neza. Inkomoko yatanzwe ntisobanura ibyiciro byatanze amakosa menshi, bityo igipimo cyo gusuzuma agaciro ntigishobora gusuzumwa muburyo burambuye.

Kubaturage, ibyagaragaye biratanga umuburo wo kwirinda sisitemu rusange ya AI nkibisimburwa byizerwa mubitekerezo bya fiziki. Sisitemu itanga igisubizo cyemewe ariko kitari cyo gukemura intambwe nyinshi zirashobora kugorana noxpert kugenzura. Ibipimo byerekana indimi ebyiri n'amashusho bishobora kandi gufasha kwerekana imbogamizi zihishwa nisuzuma ryibanze ku nyandiko y'Icyongereza cyangwa imiterere ngufi-yo gusubiza.

Interactive Mechanism

Uburyo bukoreshwa: Uburyo bukora

Shakisha ikoranabuhanga ryihishe inyuma yiri terambere.

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.
Kugenzura Ibitekerezo Byagenzuwe+10 Points
AI Models Explained Quiz

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

Ibyo kureba

Ibibazo nyamukuru ni ukumenya niba kode yamakuru nisuzumabumenyi ishobora kugerwaho bihagije kugirango yigane yigenga, uburyo imikorere itandukana kubijyanye na fiziki hamwe nuburyo bugaragara, kandi niba imiterere yigihe kizaza itera imbere mubipimo bidashingiye kubisubizo byafashwe mumutwe. Inkomoko ntisobanura imiterere yapimwe cyangwa ngo itange isuzumabumenyi, bityo raporo 33.7% ntabwo igomba gukoreshwa kugirango itondekanye sisitemu runaka.

Kwigana kwigenga nikizamini ako kanya. Abashakashatsi bazakenera kugenzura ibibazo byasohotse, igishushanyo, ibisubizo, hamwe nuburyo bwo gutanga amanota kugirango bamenye niba igipimo kitarangwamo ibidasobanutse, ibintu byigana, cyangwa ibindi bintu bishobora kugira ingaruka kubisubizo. Inkomoko yemeza irekurwa rya ariko ntabwo itanga amakuru ahagije kugirango isuzume niba igerwaho cyangwa yororoka.

Ibizaza ejo hazaza bigomba kwerekana uburyo igishushanyo cya 33.7% kigabanyijemo amasomo ya fiziki, urwego rugoye, ubwoko bwigishushanyo, indimi, nuburyo bwo gusubiza. Ni ngombwa kandi kumenya niba moderi yakira ibisobanuro bimwe, niba ibikoresho nka calculatrice byemewe, nuburyo inguzanyo igice gikemurwa. Nta na kimwe muri ibyo bisobanuro kigaragara mu isoko yatanzwe.

Ibipimo bishobora kurushaho gutanga amakuru niba nyuma isuzuma rigereranya verisiyo yicyitegererezo mugihe kandi ugatanga raporo yanyuma-igisubizo cyukuri hamwe nintambwe yo kwizerwa. Abashakashatsi bagomba kandi gusuzuma niba icyitegererezo gishobora kumenya ukudashidikanya, kumenya intambwe ndende zitari zo, cyangwa gusaba ibisobanuro mugihe igishushanyo kidasobanutse. Urupapuro rwerekana ntabwo rwerekana iyo myitwarire.

Icyingenzi kitazwi nuburyo uhagarariye PhysElite nakazi ka fiziki nyayo. Ibibazo bya Olympiad biragoye nkana kandi ntibishobora kwerekana imyitozo yo mwishuri, isesengura rya laboratoire, igishushanyo mbonera, cyangwa imyitozo yubushakashatsi. Inkomoko ishyiraho ibipimo ngenderwaho hamwe nibisubizo byatangajwe, ariko ntabwo ishyiraho uburyo imikorere kuri PhysElite iteganya imikorere murindi miterere.

Ufatiye hamwe, ibisobanuro bihari bishyigikira ibisobanuro bike kubisubizo. Imibare 33.7% ni iy'ibisubizo byatangajwe-bisuzumwa byerekana imiterere 18 yindimi nyinshi zapimwe kuri iki gipimo. Bikwiye gusomwa hamwe nibibazo byindimi ebyiri za , igishushanyo mbonera, inkomoko, nibisubizo byanyuma, kimwe no gusuzuma intambwe-ntambwe yo gusuzuma. Kuberako isoko yatanzwe ntabwo ikubiyemo amazina yicyitegererezo, amategeko yo gutanga amanota, kugerageza kubara, guhagarika ingingo, cyangwa kode yisuzuma, abasomyi ntibashobora gukoresha abstract yonyine kugirango bongere kugereranya cyangwa kumenya ibice bigize umurimo byatumye ibisubizo. Ibyo kubura ibisobanuro rero nibyingenzi kugirango dusobanukirwe na raporo zikurikira kuri PhysElite.

Ibijyanye nuyobora & ibibazo

Moderi ya AI YasobanuweAbahinduraAmahugurwa ya AIImyitwarire ya AIGerageza ibyo uzi - gerageza ikibazo cya AI kubuntuReba ijambo AI mumagambo yacuKurikiza icyerekezo cya AI cyo kurekura
Basanze ari ingirakamaro?