Efficient Blending of Large Language Models

Abstract

Due tothelimitedcapabilitiesofsingleLargeLanguageModels(LLMs),multipleLLMscanbe employedintandemforbetterreliabilityofanswers.Blendingreferstocombiningthestrengths of variousLLMstomakeuseoftheircomplementarycapabilitiesforgeneratinghigh-quality responses.Itisanon-trivialproblem,andthetaskbecomesevenmoredifficultwhenaiming for minimallatencyandsupervisingtheblendingcomponents.Thestandardframework,LLM- Blender, approachesthisinthreestages:responsegeneration,candidateselectionviaranking, and responsefusionthroughsummarization.However,thispipelinefacestwocriticallimita- tions—high latencyduetorepeatedrankingsteps,andheavyrelianceonexternal,supervised componentsincludingalearnedencoderforrankingandaseparatesequence-to-sequencesum- marizer forfusion. In thisthesis,weproposenovel,efficientalternativestoovercomethesechallenges.Thisthesis comprises twoworks.First,weshowthatreducingthefrequencyofrankingwithinmulti- turn conversationssignificantlyimproveslatencywithminimaldegradationinoutputquality. Second, weintroduceapeer-review-basedresponsefusionmechanism,whereLLMscollectively evaluateandreviseeachother’sresponses,removingtheneedforanyexternallytrainedrankers or summarizers.Thiscollaborativemethodenablesfullyself-containedLLMblendingwithout additional trainingorsupervision. WeassessourproposedmethodsonthetaskofConversationalQuestionAnsweringacrossfive multi-turnconversationalbenchmarks—ConvQuestions,Atlas-Converse,CoQA,QuAC,and DoQA—using tendiverse,publiclyavailableopen-weightLLMs.Experimentalresultsdemon- strate thatourpeer-review-drivenframeworkwithreducedrankingachievesqualityonparwith existing approacheswhilebeingsubstantiallymoreefficient.Ourworkpresentsasteptoward scalable, modularLLMensemblingforreal-worldopen-domaindialoguesystems.