偏見消融： 大語言模型的識別效能與認知適配干預研究

打開文本圖片集

Prejudice Mitigation： A Study on the Recognition Efficacy and Cognitive Adaptation Interventions ofLargeLanguageModels

BAIQi-yu，KUANGQing-hua，XIAXue，YANShu-yan，ZHANGShi-jing-yi

Abstract：This study explores whether large language models （LLMs） can accurately detect prejudice in human language and persuade people tochange their prejudiced atitudes.Using the heuristicand systematic models，three experiments were conducted.Theresults indicate thatLLMs exhibit strong capabilities in recognizingand analyzing prejudiceinhumandescriptive texts，withahighlevelofagreement with human evaluations.Specificalypersonalized interventionsbyLLMs，whicharebasedonfactualandcorrectiveinformation，caneffectivelyreduceage-relatedprejudice againstolderadults.Furthermore，intervention strategiesadjustedaccording to individuals’different cognitive stylesfurtherenhancethereductionofageprejudice.ThisresearchnotonlydemonstratestheabilityofLLMstodetect prejudicebutalsovalidates theapplicationof traditionalheuristicandsystematicmodelsinthecontextofLLMs.Itprovides anew perspective for the development of persuasion theory and supports the feasibilityof trainingLLMs as tools for promoting fairness and sustainable development.

Keywords：large languagemodel（LLMs）;persuasion;prejudice;cognitive style

一、引言

大語言模型（LargeLanguageModel，簡稱LLM）的快速發(fā)展正在深刻改變人類社會的各個層面。（剩余17469字）