Subira ku makuru
Guhanga udushyaAI Understanding ibisobanuro

Ubushakashatsi busanga LLMs ikunze kwigirira icyizere mugihe itahuye neza

Isuzumabumenyi ryerekana imiterere 12 yerekana imvugo yerekana ko bakunze kwibeshya umubano utaziguye cyangwa uhinduye umubano uhuza impamvu itaziguye, mugihe bagaragaza ko bizeye cyane imanza nyinshi zitari zo.

5 min readRead the primary source
Primary-source image accompanying Study finds LLMs often overconfident when identifying direct causal links
Inyandiko y'ibanzeInkomoko yanditse
Umwanditsi
arxiv.org
Ihuza ry'inkomoko
arxiv.orghttps://arxiv.org/abs/2608.23660
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

Calibration
Nukuntu amanota yicyitegererezo yerekana amanota ahuye nukuri gukosorwa.
Ibipimo
Ikizamini gisanzwe cyangwa dataset ikoreshwa mugupima no kugereranya imikorere yicyitegererezo.
Icyitonderwa
Umubare wibyiza byahanuwe mubyukuri nibyo.
IsuzumeModeri ya AI Yasobanuwe Ikibazo

Byagenze bite

Ubushakashatsi bushya bwa arXiv busuzuma niba imiterere nini yindimi ishobora kumenya neza isano iterwa no guhuza ikizere. Hafi y'ibice bitandatu byerekana ibishushanyo mbonera, ingamba eshanu zihutirwa hamwe nuburyo bune bwo kwigirira icyizere, abashakashatsi basanze izo moderi zarakunze gukora ibishushanyo byuzuye cyane bifite impande nyinshi zitari nziza.

Urupapuro rusuzuma amabwiriza 12-yafunguye-uburemere bwururimi rwururimi munsi yururimi-rwonyine protocole. Ingero zageragejwe ku bishushanyo bitandatu byerekana impamvu, hamwe ningamba eshanu zihutirwa hamwe nisoko enye zicyizere: ikizere kivugwa mugisubizo, ikizere gikomoka kumyandikire yicyitegererezo, amasezerano mumasezerano n'amasezerano muburyo bw'icyitegererezo. Intego ntiyari iyo gupima gusa niba icyitegererezo cyamenye ko impinduka ebyiri zifitanye isano, ariko niba zarasuzumye neza ko impinduka imwe ari impamvu itaziguye yandi kandi ikerekana icyerekezo cyurwo rubibi.

Isuzuma ryagaragaje ko icyitegererezo cyibutswe-cyiganje. Bagaragaje umubano wabakandida benshi, ariko ibishushanyo byabo byahanuwe byari byinshi cyane kandi birimo umubare munini wibinyoma-byiza. Guhindura ibisubizo byimuye cyane kuringaniza hagati yibisobanuro no kwibuka; ntabwo byakemuye imyumvire yagutse yo guteganya guhuza impamvu. Kongera urugero rw'icyitegererezo byafashaga bike ku bishushanyo binini kandi ntibyakuyeho nabi, nk'uko impapuro zibivuga.

Itandukaniro riri hagati yubusabane nimpamvu itaziguye byari ingingo yo gutsindwa hagati. Ugereranije n'ibishushanyo mbonera byashyizwe ahagaragara, icyitegererezo cyanditse nabi 40.0% by'imibanire itaziguye na 36.0% byahinduwe bitari impande zombi, ugereranije na 28.2% by'ibindi bitari impande. Muri ibyo byiza by'ibinyoma, 80.8% by'imanza zitaziguye na 84,6% by'imanza zahinduwe zidasubirwaho zabonye ikizere mu magambo byibuze 80%. Abanditsi bavuga kandi ko ibyiringiro bishingiye kuri logit byakunze guhurizwa hafi ya 1.0 utitaye ko niba ubuhanuzi ari bwo. Igipimo ngenderwaho-kimenyerewe cyagaragaje kumenyera muri bitanu byerekana-dataset ebyiri, byose birimo dataset ya AsiaM.

Ibisobanuro birambuye: arxiv.org ↗

Impamvu ari ngombwa

Ururimi rwicyitegererezo rugenda rukoreshwa mugutanga igitekerezo cya sisitemu ivumbura imiterere. Ubushakashatsi bwerekana ko ibisubizo byabo bishobora kuba ingirakamaro nkibitekerezo by'agateganyo, byagenzuwe hanze, ariko ntibigomba gufatwa nkibimenyetso bitaziguye byerekana imiterere yabiteye cyangwa byizewe gusa kuko icyitegererezo cyumvikana neza.

Sisitemu yo kuvumbura impamvu ikoresha ibitekerezo byerekeranye nimpinduka zishobora kugira ingaruka kubandi. Niba icyitegererezo cyururimi gitanga impande zigaragaza gusa ishyirahamwe ryagutse, inzira itaziguye cyangwa icyerekezo kitari cyo, icyo gitekerezo gishobora kugoreka isesengura nyuma. Ubushakashatsi rero bukemura ikibazo gifatika cyokwizerwa kubateza imbere nabashakashatsi bashaka gukoresha LLM nkisoko yubumenyi bwambere.

Ibyavuye mu bushakashatsi byerekana kandi impamvu ibisobanuro neza cyangwa ikizere cyumubare bitagomba kwibeshya kubitekerezo byizewe. Icyitegererezo gishobora kumenya ko ibitekerezo bibiri bihujwe mugihe binaniwe kumenya niba ihuriro ritaziguye cyangwa inzira isano itera. Icyizere cyo mumagambo menshi mubyahanuwe muburyo butari bwo butuma iryo tandukaniro rifite akamaro cyane mubikorwa byakazi aho abantu bashobora gukoresha amanota yicyizere kugirango bashyire imbere isubiramo.

Urupapuro ntirwerekana ko LLM ntacyo zimaze kubikorwa byimpamvu. Umwanzuro wacyo ni muto kandi urashobora gukora: moderi irashobora kuba ingirakamaro mugutanga impamvu zoroheje zitera, mugihe izo priors zagenzuwe kubimenyetso byo hanze. Uku gushiraho kurinda uruhare rwindimi mugutanga ibitekerezo mugihe ubitse kwemeza impamvu kuburyo namakuru yashizweho kugirango ashinge imiterere. Irerekana kandi ko igipimo cyikitegererezo cyonyine atari uburinzi buhagije kuri iki cyiciro cyamakosa. Ingaruka rusange kandi ifatika nuburyo bukoreshwa. Abashakashatsi bubaka ibikoresho-byo kuvumbura, sisitemu-ifata ibyemezo cyangwa isuzuma barashobora gukoresha ibisubizo nkumuburo wo gutandukanya isano itera nimpamvu itaziguye no kugerageza kalibrasi neza. Inkomoko ntabwo ishyiraho imikorere mubikorwa runaka byubuvuzi, ubukungu, politiki cyangwa ibikorwa byoherejwe, bityo ibyagaragaye ntibigomba guhuzwa muri ibyo bice nta bimenyetso bifatika.

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

Uru rupapuro rwerekana amasezerano ku bisobanuro cyangwa icyitegererezo nk'ikimenyetso gishobora kuba cyiza kuruta icyizere cyangwa amagambo ashingiye ku bitekerezo, nubwo inyungu zavuzwe zitari zifite imibare nyuma yo gukosorwa. Ibindi bizamini bizakenerwa mumibare myinshi, icyitegererezo hamwe nibibazo nyabyo byisi.

Ikimenyetso cyizewe cyane cyasuzumwe nubushakashatsi ni amasezerano: niba ibisobanuro byinshi byahawe icyitegererezo kimwe, cyangwa moderi nyinshi zahawe imirimo igereranijwe, byatanze urubanza rumwe. Uru rupapuro rutanga ibisobanuro byiza bisobanura kalibrasi no kuvangura amasezerano yihuse kandi yambukiranya icyitegererezo kuruta kugereranya ikizere gisanzwe. Nyamara, izo nyungu ntabwo zari zifite imibare nyuma yo gukosorwa kwa Holm, bityo amasezerano agomba gufatwa nkicyerekezo cyubushakashatsi aho kuba igisubizo cyemewe.

Isuzuma ry'ejo hazaza rigomba gusuzuma niba icyitegererezo gikomeza hamwe na moderi zifunze kandi zifunguye zirenze sisitemu 12 yize, hamwe nubushakashatsi bwiyongereye hamwe nibibazo biterwa nukuri kubireba cyangwa ubushakashatsi. Ibisubizo biriho biva mubishushanyo bitandatu byerekana ibishushanyo hamwe nururimi-byombi byombi protocole. Inkomoko ntabwo itangaza ko uburyo bwageragejwe muburyo bwo gufata ibyemezo bizima.

Ibipimo bimenyerewe nikindi kibazo cyo gukurikirana. Ubugenzuzi bwasanze bushobora kumenyera muburyo butanu-dataset ebyiri, zose zirimo AsiaM. Igisubizo kirashobora kwerekana ko imikorere imwe igaragara yerekana kwerekana ibipimo ngenderwaho, ariko impapuro zerekana ko zishobora kumenyera aho kwerekana ko abanyamideli bafashe mu mutwe ibisubizo. Abasesengura bazakenera gusuzuma umwanda no kumenyera mugihe ugereranije moderi kubikorwa bitera.

Ubushakashatsi nabwo busiga ibintu bitazwi. Ntabwo isobanura niba kugarura hanze, ibikoresho, imyiyerekano cyangwa kugera kubuhanga bwa domaine byatezimbere muburyo bwo kuyobora no guca imanza. Ntabwo yerekana uburyo icyitegererezo gikora mugihe gitanzwe nibimenyetso byubushakashatsi, ntanagereranya ikiguzi cyo gusuzuma abantu gikenewe kugirango bakosore ibyiza. Hanyuma, kubera ko impapuro ari inyandiko nshya yatanzwe, ibyagaragaye ntabwo byashizweho binyuze murungano rusubirwamo mumasoko yatanzwe.

Ibijyanye nuyobora & ibibazo

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