《大语言模型》最新综述：能力与局限性分析.pdf

ASurveyonLargeLanguageModelswithsomeInsights

ontheirCapabilitiesandLimitations

AndreaMatarazzoRiccardoTorlone

ExpediaGroupRomaTreUniversity

ItalyItaly

a.matarazzo@riccardo.torlone@uniroma3.it

5Abstract

2

0Therapidadvancementofartificialintelligence,particularlywiththedevelopmentof

2LargeLanguageModels(LLMs)builtonthetransformerarchitecture,hasredefinedthe

ncapabilitiesofnaturallanguageprocessing.Thesemodelsnowexhibitremarkableperfor-

amanceacrossvariouslanguage-relatedtasks,suchastextgeneration,questionanswering,

J

translation,andsummarization,oftenrivalinghuman-likecomprehension.Moreintrigu-

3

ingly,LLMshavedemonstratedemergentabilitiesextendingbeyondtheircorefunctions,

]showingproficiencyintaskslikecommonsensereasoning,codegeneration,andarithmetic.

L

CThissurveypaperexploresthefoundationalcomponents,scalingmechanisms,and

.architecturalstrategiesthatdrivethesecapabilities.EmphasizingmodelslikeGPTand

s

cLLaMA,weanalyzetheimpactofexponentialdataandcomputationalgrowthonLLM

[performance,whilealsoaddressingthetrade-offsassociatedwithscaling.Wealsoex-

1amineLLMapplicationsacrosssectors,suchashealthcare,finance,education,andlaw,

vhighlightingtheiradaptabilityandpotentialtosolvedomain-specificchallenges.

0

4CentraltothisworkarethequestionsofhowLLMsgeneralizeacrossdiversetasks,

0exhibit