Dzokera kuNhau
InnovationAI Understanding muchidimbu

Kunyatsogadzirisa maLLMs anowanzo shandura zvinomiririra zvemukati pasina kunatsiridza kuita kwakanangana nebasa

Vatsvaguri vanoona kuti shanduko yemukati inomiririra inoitika panguva yeLLM-tuning haina hukama pamwe chete nezvinoumba izvo zvinotyaira kuita basa.

4 min readRead the primary source
Source-provided image accompanying Fine-tuning LLMs often alters internal representations without improving task-specific performance
Primary-source documentKwakanyorwa
Muparidzi
arxiv.org
Source link
arxiv.orghttps://arxiv.org/abs/2609.21113
Source type
Gwaro rekutanga - chiziviso chepamutemo, bepa, faira, kana peji rebato rekutanga ratinoverenga zvakananga.
ContextNzwisisa izvi mumasekonzi makumi matanhatu

Tanga pano

Matemu akakosha

Kugadziriswa kwakanaka
Kuenderera mberi nekudzidziswa padomeine-chaiyo data kugadzirisa iyo isati yadzidziswa modhi kune rimwe basa.
Mutauro Mukuru (LLM)
Mutauro wemodhi yakadzidziswa pane yakakura text corpora kugadzira nekuongorora zvinyorwa.
Kupatsanurwa
Basa iro modhi inogovera yekuisa kune imwe kana akawanda akatemerwa chikamu.
Zviedze iwe pachakoAI Models Inotsanangurwa Mibvunzo

Chii chaitika

Chidzidzo chitsva chakaburitswa paarXiv chinoongorora hukama huripo pakati pekumisikidzwa-kwakakonzeresa shanduko muLarge Language Models (LLMs) uye izvo zvinokonzeresa zvinokonzeresa kuita basa, sezvakaonekwa neEAP (Edge Attribution Patching).

Chidzidzo ichi chinoongorora magadzirirwo ekugadzirisa zvakanaka magadzirirwo emukati, kunyanya kutarisisa maitiro ekutarisa uye layer-huchenjeri activation. Vachishandisa EAP, vatsvakurudzi vakaona zvinhu zvakati-zvakadai semisoro yekutarisisa uye logit-level activation-inofambisa zvakananga kuita basa.

Vatsvakurudzi vakaona kuti zvinhu izvi zvine chokuita nebasa zvakanyura mukati mezvikamu zvakasiyana, zvichiratidza dhigirii rekushanda kwenzvimbo. Zvisineyi, mitsara inopinda zvakanyanya kumiririra shanduko panguva yekugadzirisa zvakanaka haienderane nezvikamu zvine izvi zvinokonzeresa zvikamu.

Chidzidzo chacho chakawedzera kuongorora kushanda kwe-cross-task performance. Yakaona kuti kunyangwe kana mabasa achigovana dhigirii yakakwira yekupindirana mune yavo EAP-inozivikanwa causal zvikamu, izvi hazvivimbisi kuita kwakanaka kwekufambisa. Mune zvimwe zviitiko, kunyatsogadzirisa pane rimwe basa kwakanyatso kushatisa kuita pane rimwe, kunyangwe yakagovaniswa causal dhizaini.

Kwakabva mashoko: arxiv.org ↗

Nei zvichikosha

Tsvagiridzo iyi inopikisa fungidziro yekuti yakakura yemukati modhi inoshanduka panguva yekumisikidzwa kwakakosha kana inobatsira pakuita basa. Nekuratidza kuti kumiririra kuchinjika kunowanzo kuparadzaniswa kubva kune causal maitiro, chidzidzo chinoratidza kusakosha kwakakosha mune yazvino dzidziso paradigms. Inoratidza kuti kunyatsogadzirisa kunogona kukanganisa kugadzikana kwemuenzaniso nekusaziva, sezvinoratidzwa nekuwana kuti kupindirana kwezvikonzero zvinokonzera pakati pemabasa kunogona kutungamirira mukuderera kwekuita pane kutamisa kwakanaka.

Kubatanidzwa kwekuchinja kwekumiririra kubva mukukosha kwechikonzero kunoratidza kuti maitiro ekugadzirisa zvakanaka azvino anogona kunge ari 'mheremhere,' achigadzirisa zvikamu zvemuenzaniso izvo zvisingabatsire kune zvinodikanwa zvebasa. Izvi zvinopa hwaro hwefungidziro yekuti nei kunyatso-tuna kuchigona kutungamira kune njodzi kukanganwa kana kusatarisira kudonha kwekuita.

Kuwana kuti kupindirana kwezvinokonzeresa zvinokonzeresa zvinogona kutungamira mukuderera kwekuita kwakakosha zvakanyanya pakudzidza-mazhinji-basa. Zvinoreva kuti kungogovana zvikamu pakati pemabasa hazvina kukwana kuti ubudirire uye kuti maitiro emabasa (semuenzaniso, kupatsanurwa kunopesana nechizvarwa) kunotora basa rakakosha mumabatiro anoita zvikamu izvi.

Iri basa rinopa hurongwa hwevagadziri kuti vatarise zvirinani kushanda kwepaipi yavo yekunyatsogadzirisa, zvingangotungamira kune nzira dzekudzidzisa dzakanyanya hunyanzvi dzinotarisa pakugadzirisa chete zvinonyanya kukosha zvikamu zvinokonzeresa.

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 Models Explained Quiz

What is the best response when AI Models Explained makes a mistake in production?

Zvekutarisa zvinotevera

Tsvagiridzo yeramangwana mune dzimwe nzira dzakanangwa dzekugadzirisa zvakanaka dzinoisa pamberi pezvinokonzeresa pamusoro pezvakakura zvinomiririra zvigadziriso, uye kuti izvi zvakawanikwa zvine magadzirirwo emhando dzakasiyana kupfuura ayo akaedzwa muchidzidzo.

Chidzidzo chacho hachitsananguri maitiro chaiwo akaedzwa kana kuwanikwa kwekodhi inoshandiswa pakuongorora kweEAP, ichisiya kushandiswa kunoshanda kune vashandi iye zvino zvisingazivikanwi.

Vacherechedzi vanofanirwa kutarisa kuti zviwanikwa izvi zvinotungamira mukugadzirwa kwemaitiro e 'causally-aware' ekugadzirisa zvakanaka ayo anovavarira kudzikisira machinjiro asina kufanira ekumiririra.

Izvo zvinoramba zvichionekwa kana mhedzisiro iyi ichienderana pamhando dzakasiyana dzemhando uye mavakirwo, kana kuti dzakanangana nemhando dzakaongororwa mutsvakurudzo iyi.

Related guides & Quizzes

AI Models InotsanangurwaKudzidziswa kweAITransformersEdza zvaunoziva - edza yemahara AI quizTarisa kumusoro izwi reAI mune yedu glossary
Wakawana izvi zvinobatsira?