航空运输的大型语言模型综述（英）Y-Liu-2024

Y.LitJoumal of the Air Trarsport Research Society 2 (2024)100024Table 1Overview of the twelve large language models assessed in our study.ILM NameLicenseSizeDateCountryClaude-2Partial130BNot specifiedUSACohereCommercial52B2020CanadaERNIE Bot 3.5Commercial130B2019ChinaFalcon-180BOpen source180Bup-to-dateAbu DhabiGPT 3.5CommercialUp to 175BJan.2022USAHunYuan-V15.8CommercialOver 100B2019ChinaLlama-2-70bOpen source70Bup-to-dateUSAMistral-7B-Instruct-v0.2Open source7BsepL.2,2021FrancePaLM 2Commercial340B2022USAQwen-7BOpen source7B2019ChinaVicuna-33BDpen④rce33B2021USAi-34BOpen souree34B2023Chinacantly accelerate research and discovery.Moreover,in the legal domainmodels,airport codes,and engine specifications,demonstrating theirLLM AI tools like ROSS are being utilized for contract analysis and legalpotential utility in accessing technical aviation data.research.By quickly sifting through and interpreting legal documents,Understanding of Industry Dynamics.Questions like ECHALL andLLMs can provide attomeys and legal professionals with insights and rel-CRFUEL reflect the models'comprehension of industry trends,eco-evant case laws,thereby reducing research time and improving the ac-nomic strategies,and advancements in aviation technology,high-curacy oflegal advice.In the financial industry,LLMs are applied in risklighting their potential as analytical tools for industry professionals.assessment,fraud detection,and automated financial advising,offeringOperational Applications.Operational applicability was testedmore efficient and sophisticated financial services.These examples un-through experiments generating results on CRSPEE,CRCABI,derscore the transformative impact of LLMs across various professionalELOADF,and FSEATS,which emphasize the models'capabilities infields,highlighting their role in driving innovation,enhancing precision,optimizing and planning various aspects of flight operations.and improving decision-making processes.Strategic Insights.Answers for the questions CRALLI,EROUTE,andTable 1 provides an overview on the twelve models which we com-FALLIA could illustrate the LLMs'effectiveness in providing strategicpare in our study.These models have been chosen mainly based on threeinsights into airline alliances and network planning.criteria:model size,license,and the recency of the training data usedMarket Analysis.Questions such as FTOPAL and FFASTF target theto construct the model.Concerning the model size,these LIMs rangemodels'proficiency in market analysis,including airline rankingsfrom rather small models (of about 7 billion parameters)to consider-and record-setting flight assessments.ably large models (with 340 billion parameters).Size is a primary differ-Environmental and Economic Considerations.The experimentsentiator among these models:Larger models generally show enhancedalso evaluated the LLMs'understanding of sustainability and cost-understanding and generative capabilities,though they require signifi-efficiency in aviation,as repres