基于不确定性估计的微调代码生成模型与大语言模型的协同方法

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）10-007-2947-09

doi：10.19734/j.issn.1001-3695.2025.04.0082

Coordination of fine-tuned code generation models and large language model via uncertainty estimation

Hong Shaodong 1a，1b ，Shen Guoweila，1bt，Luo Sufen²，Liu Tao² （1.a.NatioalKyLbatorofublcigat，nistrofducatiEgeinRsechCentefTetCoputing&oete ligence，Guizhou University，Guiyang55025，China；2.GuizhouProvincialJudicialPoliceAcademyGuyang5505ia）

Abstract：Thecomplementary mechanismbetween fine-tunedcode generation modelsandlargeLLMremains underexplored， leading toambiguousdecisionboundaries intheircollaboration.Thispaperproposedamethod named Coral tocoodinate finetunedmodelsandLLMsbasedonuncertaintyestimation.Thismethodanalyzedthecomplementaritybetweenthetwo models andquantified theirdecision boundaries.Coraladopted theconceptof expectedcalibrationeror tocompareuncertainty estimationmethodsandselectedastable methodfor thefine-tuned model.This enabledthefine-tuned modeltooutputuncertainty scoresreflectingpredictionconfidence.CoralcalculatedanuncertaintythresholdbymaximizingBLEUscoresonavalidation dataset，whichquantified thedecision boundarybetweenthemodels.Basedonthethresholdand theuncertaintyscores，the method identifiedIDand OODdata.TheLLMhandledOODdata to improve thegeneralizationof the fine-tuned model.Evaluationon two benchmark datasets showthat Coraloutperforms theuseof either modelalone inboth BLEUandExact Match metrics.The results indicate that Coral effectively coordinates the fine-tuned model and LLM.

tey Words：large language model；fine-tuned model；code generation；uncertainty estimatio！

0 引言

代码生成是软件工程和人工智能社区中长期存在的任务。（剩余23468字）