Subira ku makuru
Guhanga udushyaAI Understanding ibisobanuro

Ubushakashatsi buvuga ko isuzuma rya LLM ryumuntu rihinduka mugihe amakadiri yibibazo ahindutse

Inyandiko ya arXiv ivuga ko uburyo bwa GPT-4o bwigana bwabantu buterwa ahanini nuburyo ibibazo byimiterere byateganijwe kandi bihujwe, byerekana ko amanota yose hamwe ashobora kubura imyitwarire yingenzi.

5 min readRead the primary source
Source-page capture accompanying Study says LLM persona evaluations change when question frames shift
Inyandiko y'ibanzeInkomoko yanditse
Umwanditsi
arxiv.org
Ihuza ry'inkomoko
arxiv.orghttps://arxiv.org/abs/2607.02368
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.
Kuba rusange
Ukuntu icyitegererezo gikora neza kumakuru mashya, atagaragara hanze yamahugurwa.
Byihuse
Iyinjiza amabwiriza nibisobanuro byatanzwe muburyo bwo kubyara.
IsuzumeModeri ya AI Yasobanuwe Ikibazo

Byagenze bite

Igitabo cyanditswe na Yuan Yuan gisuzuma niba imiterere yimbere ijyanye na LLM umuntu ari ibintu bihamye cyangwa ibihangano byerekana uko isuzuma ryakozwe. Ukoresheje GPT-4o mu kwigana abantu b'Abanyamerika n'Abashinwa-Abanyamerika, ubushakashatsi bwasesenguye ibisubizo ku ngingo 50 y’ibibazo mpuzamahanga by’ibidukikije by’ibidukikije mu buryo butandukanye. Itangaza ko gukusanya amanota atanu hamwe nubusabane bwa geometrike mubisubizo bitwara ukundi mugihe ikizamini cyo gusuzuma gihindutse.

Uru rupapuro rwiga ikibazo cyihariye mugusuzuma imiterere nini yururimi: niba umuntu ugaragara wikitegererezo ashobora kugabanuka kugirango atange amanota y'ibibazo cyangwa niba isano iri mubisubizo byihariye nayo ikubiyemo amakuru afite ireme. Ikoresha ibibazo bya IPIP-50, byashizweho kugirango bipime ibipimo binini bitanu, kandi byubaka murwego-rwerekana isano iva mubisubizo by'icyitegererezo. Iyo mibare noneho isesengurwa nkingingo zifatika nziza-isobanutse, cyangwa SPD, manifolds - imibare igereranya imiterere. Inkomoko yerekana ibi nkuburyo bwo kubika amakuru amanota asanzwe yerekana.

Ubushakashatsi bukoresha GPT-4o mu kwigana abantu b'Abanyamerika n'Abashinwa-Abanyamerika. Abashakashatsi bakoresha uburyo bwo gutondekanya ibibazo no kugereranya uko bigenda iyo ibipimo byo gusuzuma byateganijwe, bidahuye, cyangwa bisangiwe. Ibisobanuro bitanga ibisubizo bibiri bitandukanye kuri izo mpinduka. Ibiranga byose, bigaragazwa n amanota manini atanu, byerekana ko byagabanutseho 21% munsi ya randomisation ariko bisobanurwa nkibikoresho-bikomeye. Imiterere ya geometrike yerekana ko 42% yagabanutse munsi yikintu kidahuza hanyuma igakira cyane - kugeza kuri 84% - mugihe ibibazo bikoresha amakadiri asangiwe. Ibisobanuro bivuga ko uku gukira kurenze 76% ibisubizo bihuye nibintu byegeranijwe.

Urupapuro ruranga ibisubizo nkibimenyetso kubintu bibiri bitandukanijwe bigize imvugo ya LLM yumuntu: ikadiri-ikomeye igiteranyo hamwe na geometrie ishingiye. Konti yimpapuro, ibice bya geometrike nuburyo bwo guhuza ibisubizo aho kuba imico ihamye idahinduka munsi ya buri kiganiro. Inkomoko ntabwo itanga abstract-urwego rurambuye rukenewe kugirango rwigenga gusuzuma igipimo cyibigereranyo cyangwa kubara neza. Ntabwo ivuga umubare wibisubizo byintangarugero, ibisobanuro byuzuye bikoreshwa mukurema abantu, protocole ya randomisation, igenamiterere rya decoding, cyangwa uko amanota yijanisha yasobanuwe.

Ibisobanuro birambuye: arxiv.org ↗

Impamvu ari ngombwa

Urupapuro rwibanzeho ni uko LLM isuzuma ryabantu rishobora kuba ryumva imiterere y'ibibazo, ntabwo ari icyitegererezo cyangwa abantu bahita bapimwa. Niba byigana, ibi birashobora kugira ingaruka kuburyo abashakashatsi basobanura ibirego bijyanye numuntu wintangarugero, guhuzagurika, gutandukana kwimico, nimyitwarire ijyanye numutekano. Ntabwo ishyiraho ko LLMs ifite imiterere yumuntu, kandi inkomoko ntabwo yerekana ko imiterere yatangajwe muri rusange irenze igeragezwa ryageragejwe.

Ikibazo gifatika ni gupima. Icyitegererezo kirashobora gutanga amanota asa muri rusange amanota atanu mugihe ahindura umubano mubisubizo bya buriwese mugihe ikibazo cyabajijwe cyangwa cyateguwe muburyo butandukanye. Niba ubwo buryo bukora mugupima kwagutse, umugenzuzi wanditse amanota atanu gusa yincamake ashobora kwemeza ko umuntu atuje mugihe yabuze impinduka muburyo bwo gusubiza icyitegererezo. Uru rupapuro rero ruvuga ku isuzuma-rimenyekanisha ryerekana ibyerekezo rusange hamwe na geometrie y'ibisubizo. Ibyo nibisobanuro byuburyo buvuye kubitabo, ntabwo byumvikanyweho.

Ibi bifite akamaro kubushakashatsi kumyitwarire yicyitegererezo kuko abantu basanzwe bakoreshwa mukwiga guhuzagurika, guhagararira umuco, kubogama, no guhuza. Isuzuma-ryerekana isuzuma rishobora kwerekana ko itandukaniro ryagaragaye rituruka ku gishushanyo mbonera aho guturuka ku buryo burambye buranga. Irashobora kandi gufasha abashakashatsi gutandukanya ihinduka rihamye mubisubizo ugereranije nimpinduka muburyo ibisubizo bihura. Inkomoko ntigaragaza ingaruka kubicuruzwa byoherejwe, ibyemezo byabakoresha, cyangwa impanuka z'umutekano, bityo izo ngaruka zikomeza kuba nziza.

Ibyavuye mu bushakashatsi ntibigomba gusomwa nkikimenyetso cyerekana ko GPT-4o ifite imiterere-muntu cyangwa ko abantu bigana Abanyamerika n’Abashinwa-Abanyamerika bahuye n’abaturage nyabo. Ubushakashatsi bugerageza icyitegererezo cyibibazo byatanzwe mubisobanuro byasobanuwe. Ntabwo ishyiraho kandi ko isesengura rya SPD-rinyuranye risumba buri muntu wese wasuzumye, cyangwa ko raporo yakize yatuma habaho guhanura imyitwarire nyayo. Umusanzu wingenzi wasobanuwe ninkomoko ni igitekerezo cyo gutandukanya ubwoko bubiri bwibimenyetso byo gupima no kuburira ko guteranya bishobora guhisha imiterere-shingiro.

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

Intambwe ikurikiraho ni kwigana kurugero, amabwiriza yumuntu, indimi, ibibazo, hamwe nuburyo bwo gutoranya. Abasomyi bagomba kandi gushakisha amakuru yuzuye yubushakashatsi inyuma yimihindagurikire y’ijanisha ryavuzwe, harimo ingano yerekana urugero, ibisobanuro, igenamiterere-ibisekuruza, hamwe n'ibipimo nyabyo byakoreshejwe mu kugereranya geometrike. Ubushakashatsi ni integuro ya arXiv, bityo imyanzuro yayo igomba gufatwa nkibisabwa ubushakashatsi butegereje ubugenzuzi bwagutse.

Kwisubiramo bizerekana niba imibare yatangajwe 21%, 42%, 84%, na 76% imibare ikomeye. Kugereranya kwingenzi kwaba kurimo izindi ngero zindimi, verisiyo zitandukanye za GPT-4o, ubundi buryo bwo kubaza abantu, gusubiramo inshuro nyinshi, hamwe nibibazo birenze IPIP-50. Abashakashatsi bagomba kandi gusuzuma niba ingaruka zimwe zigaragara mugihe gahunda yibibazo ihinduwe idahinduye amagambo, mugihe amagambo hamwe nimiterere yumuco bitandukanye bitandukanye, kandi mugihe abasesengura bakoresha imvugo yigenga aho kwigana abantu.

Impapuro zuzuye hamwe nubushakashatsi bwakurikiyeho birashobora gusobanura niba ingaruka za geometrike zigaragaza imitungo ifatika yimyitwarire yicyitegererezo cyangwa sensitivite yatangijwe no kubaka matrica ihuriro no guhitamo ibipimo byinshi. Inkomoko yinkomoko ntisobanura igipimo nyacyo cyintera, ibizamini byibarurishamibare, ibyingenzi, cyangwa ibigereranyo bidashidikanywaho. Ibyo bisobanuro birakenewe mbere yo kugereranya ijanisha ryatangajwe murwego rusange hamwe na geometrike cyangwa guhitamo umubare wabapima uburemere bagomba kugenera ibimenyetso.

Kubakoresha bifatika, isomo ryihuse ni rito ariko rifite akamaro: gukora ikizamini cyumuntu umwe ntigomba gufatwa nkibisobanuro byuzuye byimyitwarire ya LLM. Amakipe asuzuma icyitegererezo arashobora gushaka gutandukanya ibibazo no gutondekanya, kubika ibintu-kurwego rwibisubizo, no gutanga raporo idashidikanywaho aho kwishingikiriza gusa kumurongo wihariye. Niba inzira nkizo zitezimbere guhanura imikoranire nyayo ntabwo bizwi. Urupapuro ni integuro, kandi inkomoko ntamakuru atanga kubijyanye no gusuzuma urungano, kwigana byigenga, ingaruka zoherejwe, cyangwa rusange muburyo butandukanye uretse igeragezwa rya GPT-4o.

Ibijyanye nuyobora & ibibazo

Moderi ya AI YasobanuweImyitwarire ya AIPrompt EngineeringGerageza ibyo uzi - gerageza ikibazo cya AI kubuntuReba ijambo AI mumagambo yacuKurikiza icyerekezo cya AI cyo kurekura
Basanze ari ingirakamaro?