2025大语言模型时代质量保障：生成式AI系统评估与验证方法白皮书（英文版） conv.docxVIP

QualityAssuranceintheEraofLLM’sMethodologiesforEvaluatingandValidatingGenerativeAISystems

TableofContents

01 Introduction 03

02 Di?erencebetweentraditionaltesting 04approachandGenAItesting

03 KeychallengesofLLMtesting 05

04 Testingtechniquesandstrategies 06

05 Humanevaluationor 08Human-in-the-Loop(HITL)

06 Keyqualitymetricsandvalidation 09criteriaformeasuringGenAIoutputs

07 Regressiontestingstrategy 11forGenAIsolutions

08 Testautomationof 12LLMoutputs

09 Futuretrendsandconsiderations 13LLMtesting

10 Conclusion 14

?HappiestMinds2025 02

Introduction

AsgenerativeAI(GenAI)systemsarebeingusedmorewidelyinsoftwaredevelopment,traditionaltestingstrategieswillhavetotransform.GartnersrecentsurveyresultsshowthatAIengineeringisestimatedtointroducenewbestpracticesforsoftwareengineeringcompanieswith80%ofthemhavingAI-basedtestingstrategiesinplaceby2025.ThiswhitepaperunfoldsthegapsintestingmindsetsandpresentsstrategiesandtoolsforGenAIapplicationssuccessfultrials.ProvenbythegenerativeAImarketbeingvaluedat$186.33billionby2031,withacompoundannualgrowthrateof34.3%,theneedtocomeupwithreliabletestingapproachesisathighstakes.

GlobalGenerativeAIMarketSize,2021-2023(USDBillion)

186.33Bn

15.95Bn

2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031

Source:/generative-ai-market-478

?HappiestMinds2025 03

Di?erencebetweentraditionaltestingapproachandGenAItesting

TestingGenerativeAI(GenAI)applicationsrequiresanewoutlook,asitdemandsafundamentallydi?erentapproachcomparedtotraditionalsoftwaretesting.Werequireasigni?cantshiftinmindsetwhentestingGenAI-basedsolutionsorplatforms.Herearesomekeydi?erencesinthetestingmindset:

TraditionalSoftwareTestingMindset

Predictiveoutcomes

Scenario-basedtesting

Defectidenti?cation

Comprehensivetesting

GENAITestingMindset

Dynamicr