基于Puma算法引导帕累托前沿的高效多目标提示优化方法

打开文本图片集

关键词：提示优化；多目标优化；美洲狮算法；帕累托最优；大型语言模型中图分类号：TP301.6 文献标志码：A 文章编号：1001-3695（2025）10-019-3041-12doi：10.19734/j.issn.1001-3695.2025.03.0074

Puma-guided Pareto optimization for eficient multi-objective prompt generation

Dong Xiangqian1，Xiao Zheng2，3t （1.LargeModelApcon&ResechCte，eduNesUisityed844，na;2.hooftellgentfcu &InformationEnginering，icanTecholo&usinessColge，Chgu3Cha;olofIfoatioEnneia dong Vocational College of Hotel Management， Dongguan Guangdong 52396o， China）

Abstract：Existing promptoptimizationmethodssuferfrom limitations inscalabilityandadaptivity.Toaddress theseissue， this studydevelopedamulti-objectivepromptoptimizationframework，Puma-MOPT，basedonthePumaalgorithm.Theframwork integratedtheadaptive phase switchingand globalsearch capabilitiesofthePumaalgorithm with PromptWizard’sprompt generationand evaluation mechanism toenableautomatic promptsearchandmulti-objective trade-off.To improvesearchefficiencyand enhance generalizationinfew-shotscenarios，Puma-MOPT incorporatedasemanticsimilarityconstraintandemployedanadversarial filtering technique.Experimentalresultsinfivedomainsincluding mathematical reasoning，medical questionanswering，andcode generation demonstratethattheframework significantlyoutperformsbaseline methodssuchas NSGA-II，MOEA/D，EvoPrompt，andPromptWizard onmultipleevaluation metrics.Puma-MOPT providesaneficient， robust，and general solution for large language model （LLM） prompt engineering.

Key Words：promptoptimization；multi-objectiveoptimization；Pumaalgorithm；Paretooptimality；large languagemodel

0 引言

大语言模型（large languagemodel，LLM），如GPT-4、Deep-Seek，在智能问答、代码生成、医疗辅助诊断等领域展现出革命性的潜力。（剩余33259字）