通过递进知识更新和自一致性增强大语言模型推理能力

打开文本图片集

关键词：大语言模型；复杂多步推理；思维链；推理增强；动态知识更新 中图分类号：TP183 文献标志码：A 文章编号：1001-3695（2025）12-023-3707-09 doi：10. 19734/j.issn.1001-3695.2025.05.0126

Enhancing reasoning ability of large language models through progressive knowledge updates and self-consistency

ChangXiaonan，ZhangLong†，MaFuxiao （SchoolofComputer&InformationEngineering，TianjinNormal University，Tianjin3Oo387，China）

Abstract：This paperaddressed thelimitations oflarge language models incomplex multi-stepreasoning tasks.Existing methods，such aschainof thought（CoT），enhancereasoning capabilitiesby guiding the model to generatereasoning steps；however，theyoften encountererrors inthegeneratedintermediatesteps andinformationomissions.Whenanypartofthereasonng processfails，itfrequentlyleadstoincorrctfinalanswers.Toovercomethesechallnges，thispaperproposedanovelreasoning method called progressive consistent reasoning （PCR）. PCR constructed a dynamic knowledge repository -a structured listof informationthatwascontinuallupdatedduringthereasoningprocess—byextracting explicitkeyinformation fromthe original problem toestablishan initialknowledgebaseanddecomposingthe problem into multiple sub-problems.Aftersolving each sub-problem，PCR dynamicallyupdatedtheknowledgerepository by extracting new informationfromthesub-problemanswers.The modelthenreconsideredtheoriginalproblembasedontheupdatedknowledgebasetogeneratecompleteanswerattempts foreach stage.Finally，itemployedanaggregationstrategytointegratethecandidatesolutionsfromeach stage，resulting ina morerobustand accurate final answer.Compared toother methods，PCR improved performanceonvarious complex reasoning benchmarks，such as GSM8K and CSQA，by 11. 9% ，5. 73% ，and 3. 45% ，0. 95% over traditional chain of thoughtandself-consistencyreasoning methods，respectively.TheresultsdemonstratethatPCRefectivelyreduces theimpact of erorsinintermediatestepsandinformationomisionsonthefinaloutcomes，therebyenancingthestabilityandaccuracyof reasoning.

Key words：large language models；complex multi-step reasoning；chainofthought；reasoning enhancement;dynamic update

0 引言

随着深度学习技术的快速发展，大型语言模型在自然语言处理（naturallanguageprocessing，NLP）领域取得了革命性突破。（剩余17928字）