Dzokera kuNhau
ChengetedzoAI Understanding muchidimbu

Kudzidza kunowana mhinduro-inoziva AI monitors inogona kupotsa kufunga kwakakanganisika

Preprint inoshuma kuti kupa ma AI monitors mhinduro yakavimbika kwakavandudza kugona kwavo kuona zvisizvo mhedziso asi zvakaderedza kugona kwavo kuona zvikanganiso zvakavigwa mukati memhinduro chaidzo.

5 min readRead the primary source
Source-provided image accompanying Study finds answer-aware AI monitors can miss flawed reasoning
Primary-source documentKwakanyorwa
Muparidzi
arxiv.org
Source link
arxiv.orghttps://arxiv.org/abs/2609.00264
Source type
Gwaro rekutanga - chiziviso chepamutemo, bepa, faira, kana peji rebato rekutanga ratinoverenga zvakananga.
ContextNzwisisa izvi mumasekonzi makumi matanhatu

Tanga pano

Matemu akakosha

Mutauro Mukuru (LLM)
Mutauro wemodhi yakadzidziswa pane yakakura text corpora kugadzira nekuongorora zvinyorwa.
Nguva Yekutenda
Huwandu hwehuwandu hunogona kunge huine hukoshi hwechokwadi hwemetric yemodhi yakayerwa.
Chain-of-Pfungwa
Chimiro chekufunga uko iyo AI modhi inobvisa dambudziko kuita nhanho dzepakati.
Zviedze iwe pachakoAI Ethics Quiz

Chii chaitika

Chikwata chakaongorora kana kuwana mhinduro yakavimbika kunobatsira maAI masisitimu kuratidza kufunga kunogadzirwa nemamwe maAI modhi. Vachishandisa 237 nhanho-nhamba mhinduro kumibvunzo makumi manomwe nepfumbamwe yefizikisi, vaongorori vakawana kuti mhinduro yekuwana yakagadziridzwa mhedziso vachitarisa kupfuura kuongororwa kwakazvimirira kwegakava rinotsigira.

Iyo preprint, yakaunzwa kuarXiv muna Nyamavhuvhu 31, inoongorora cheni-ye-ye-pfungwa yekutarisa, nzira iyo imwe AI system inoongorora nzira yekufunga inogadzirwa neimwe. Vanyori vakaunganidza 237 nhanho-nhamba mhinduro kumibvunzo makumi manomwe nepfumbamwe yeHumanity Yekupedzisira Exam fizikisi kubva kumatatu emiganhu. Maronda aya aive asina zvikanganiso zvakaiswa; vatsvakurudzi vakanyora vakazvimiririra zvese kana mhinduro yega yega yekupedzisira yaive yechokwadi uye pakaitika nhanho yekutanga yenhema.

Chiyero chereferensi chakasanganiswa zvirevo zvefizikisi, gakava rakazvimiririra remutauro mukuru-muenzanisi uye mutongo wakavharwa nekwakabva. Iyi nzira yakaratidza makumi maviri nemana akakosha nzira: mhinduro dzine mhinduro dzekupedzisira dzaive dzechokwadi kunyangwe kufunga kwaisanganisira kukanganisa chaiko. Bepa rinotsanangura zvikanganiso izvi sezvakajairwa uye kazhinji zvisingatakuri-kutakura pane kuita nemaune mhandu, izvo zvinoganhurira kuti zvakananga mhedzisiro yacho inogona kuitwa sei kune huvengo maitiro.

Vatariri vasere veAI vakaongorora maitiro pasi pemamiriro matatu: nemhinduro isina kusimbiswa, nemhinduro yakasimbiswa, kana mushure mekuita kuzvipira kweupofu vasati vaona mhinduro. Nechitupa, zvinoreva kurongeka kwakaringana kwakawedzera kubva ku0.637 kuenda ku0.796, nepo chaiyo yekutanga kukanganisa kwakawedzera kubva ku0.261 kuenda ku0.379. Izvo zvakawana zvaisareva kuti mamonitor akave nani zvakafanana pakuzvimiririra.

Musiyano waive wakajeka mukurangarira. Certification yakawedzera mugove wezvisizvo-mhinduro traces akacherechedzwa semhosho kubva 0.653 kusvika 0.951, asi yakaderedzwa kurangarira pane dzakakosha tw kubva 0.521 kusvika 0.438. Mushure mekuzvipira kweupofu, vatariri vakaratidza mhinduro ichangobva kucherechedzwa 93.8% yezvakambopfuura zvisirizvo-mhinduro traces, zvichienzaniswa ne18.0% chete yezvakakomba traces. Vanyori vanorondedzera kuti vatariri vasere vose vakaratidza nzira imwechete yekusiyanisa, nemubvunzo-bootstrap 95% nguva yekuvimba ye [+ 0.256, + 0.506].

Kwakabva mashoko: arxiv.org ↗

Nei zvichikosha

Zvakawanikwa zvinopokana fungidziro yakajairika mukutarisisa-kwe-kufunga: kuti mhinduro chaiyo yekupedzisira humbowo hunobatsira hwekuti nzira yekufunga yaive yakanaka. Nekutarisa kweAI, chidzidzo ichi chinoratidza kuti ongororo dzinogona kuwedzeredza kugona kwekutarisa kana mamonitor achibvumidzwa kuvimba nemhedzisiro inovimbwa.

Chirevo chepakati ndechekuti mucherechedzo weAI unogona kunge uchitarisa kuti chikonzero chekufunga chinobvumirana nemhinduro pane kuona kuti kufunga kwacho kuriko here. Mhinduro isiriyo inogadzira kusawirirana kuri pachena kuti kuwana mhinduro kunogona kufumura. Mhinduro chaiyo, nemusiyano, inogona kuita kuti tsvakiridzo isina kunaka iite seyakavimbika kunyangwe nhanho yekutanga iri yenhema. Musoro wepepa unobata mutsauko uyu: mhinduro haisi iyo nharo.

Izvo zvine basa kune kuchengetedzwa kwekuongorora kwakavakirwa pane zvinogamuchirwa zvinobuda. Kana mamonitor akapihwa mhinduro yakavimbika yereferensi, maitiro avo akayerwa anogona kuratidza mhedziso-kusawirirana kutarisa panzvimbo yekuzvimiririra kuongororwa kwemaitiro. Vanyori vanotsanangura iyo pateni seyakanaka analogue yemubairo yekubira: chinobuda chinogona kugutsa muongorori nepo maitiro akaigadzira anoramba asina kunaka. Uku kuda kubva kupreprint, kwete humbowo hwekuti masisitimu akaiswa parizvino ari kushandisa kufunga kwakakanganisika mune imwe nzvimbo yepamusoro-soro.

Kunetseka kunoshanda hakugumiri kufizikisi kana kugadzirisa matambudziko. Chero nzira yekutarisa inobata chigumisiro chakabudirira sehumbowo hwenzira yakachengeteka inogona kutarisana nenzvimbo yakafanana yebofu, kunyanya kana kufunga kwepakati kwakaoma kusimbisa zvakananga. Sosi yacho hairatidze kuti mhedzisiro yacho inoendeswa kune mamwe madomasi, asi inopa kongiri yekuyedza kesi yekuparadzanisa mhedzisiro yekusimbisa kubva kune process verification.

Kune masangano anoongorora masisitimu eAI, zvakawanikwa zvinonongedza kukuzivisa mhinduro sechinhu chakakosha chekuyedza. Monikiritsi inoita zvakanaka kana ichikwanisa kuenzaniswa nemhedzisiro yakasimbiswa inogona kunge ichiri isina kusimba pakutsvaga zvikanganiso mumateki anopera nemazvo. Musiyano iwoyo unogona kukanganisa kuti vanogadzira vanoturikira sei zvibodzwa zvekuchengetedza, sarudza nzira dzekutarisa uye sarudza kuti ndedzipi mamwe macheki anodiwa usati wavimba nebasa rinogadzirwa nemuenzaniso.

Interactive Mechanism

Interactive Mechanism: Iyo Inonyatsoshanda

Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.

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.
Interactive Concept Check+10 Points
AI Ethics Quiz

Why can ethical evaluation not be reduced to one model score?

Zvekutarisa zvinotevera

Iro basa ndere preprint yakavakirwa pamhinduro dzefizikisi kubva kumatatu emuganho modhi, mamonitor masere uye makumi maviri nemana akatsoropodza anoteedzera apo mhinduro yaive yechokwadi asi kufunga kwaive nechikanganiso. Kumwe kuyedza kunodiwa pazvidzidzo zvakasiyana, modhi, mhando dzezvikanganiso uye zvigadziriso zveanopikisa.

Chinonyanya kukosha chisingazivikanwi ndeye generalization. Chidzidzo ichi chinoshandisa mibvunzo yefizikisi, mhinduro kubva kumhando nhatu dzemuganho uye mamonitor masere eLLM. Kwakabva hakutauri kana iwo maitiro akafanana akaratidzwa mumasvomhu, kukodha, kuongororwa kwemutemo, kufambiswa kwesainzi kana mabasa emumiriri wepasirese. Izvo zvakare hazvitaridzi kuti mashandiro anochinja sei nemhuri dzakasiyana dzemhando, tarisa zvinokurudzira kana kuteedzera mafomati.

Iwo makumi maviri nemana anoteedzera akakosha pakupokana kwebepa, asi idiki diki remhinduro dzakaunganidzwa. Kwakabva hakupi ruzivo rwakakwana kuti uone kuti zvinomiririra sei mhinduro dzemhinduro dzese kana kuti mhinduro yaizochinja sei nemuenzaniso mukuru uye wakasiyana. Vaverengi vanofanira naizvozvo kubata zvikamu zvakashumwa sehumbowo kubva muongororo iyi, kwete semitengo yepasi rose yevatariri veAI.

Vatsvagiri vanonyatsotaura zvikanganiso zvakadzidzwa zvaive zvakajairika, kazhinji zvisingatakure uye kwete zvemhandu. Izvo zvinosiya zvakavhurika kana kuedza nemaune kuvanza chikanganiso kungaita kuti mhinduro-yekuziva yekutarisa zvakanyanya kana kushomeka. Iyo sosi zvakare haitauri mhedzisiro yekutumirwa, kuenzanisa kwevanhu-kutarisa kana mhedzisiro muhurongwa hwekushanda.

Basa remangwana rinofanirwa kuyedza mamiriro ekutarisa umo muongorori asingagamuchire mhinduro yakavimbika, anofanira kuzvipira asati aona mhedzisiro, kana kushandisa yakazvimirira cheki yezvichemo zvepakati. Inofanirawo kuongorora kana vatariri vachikwanisa kusiyanisa nekuvimbika mhedziso yakasvikika kuburikidza nekufunga kwakanaka kubva kune yakasvika nenzira isina kunaka. Kusvikira iyo mibvunzo yapindurwa, mucherechedzo wepamusoro pamhinduro-inoziva yekutarisa haifanirwe kuverengerwa sehumbowo hwekuti AI system yekufunga maitiro inonzwika.

Related guides & Quizzes

Tsika dzeAIAI Models InotsanangurwaAI AgentsEdza zvaunoziva - edza yemahara AI quizTarisa kumusoro izwi reAI mune yedu glossaryTevedza iyo AI regulation tracker
Wakawana izvi zvinobatsira?