基于大语言模型的多模态互动内容人机协作创作

打开文本图片集

关键词：自动游戏创作；程序化内容生成；大语言模型；视频游戏描述语言；协作创作中图分类号：TP18 文献标志码：A 文章编号：1001-3695（2025）12-008-3582-12doi：10.19734/j.issn.1001-3695.2025.05.0160

Abstract：This paper proposedahuman-AIcooperative workflow framework toaddress theisues of weak usercontroligh expertisebariers，andlackofuser interfaceininteractivecontent generationusinglargelanguage models.Theframework enhancedusercontrolthrough user-definedtaskdecompositionandoptimized thegenerationprocess withavisual debugginginterface.Thisresearch also implemented aworkflow instanceforgenerating rulesand levelsusing the videogame description languageanddevelopedaWeb-basedcooperativecreationplatform.Experimentalresultsdemonstratethat theframewor ffectivelygenerates varioustypesof interactivecontent across diferentlargelanguage models，inreasingtheaveragesuccessrate by 20% compared to baseline methods.The userstudyyieldsasystemusabilityscale score of78.75，which indicates the platformhas goodusability，caneectivelysupportnon-expertusers，improvethequalityandcontrolabilityofmultimodalinteractive content generationandlower the creation barier.Aditionally，this paperalsorealized3Dsene generationbyintegrating video game description language withunreal engine and Godot engine，and generated JavaScript-based real-time strategy games byadjusting the workflow.It shows the extensibility of the framework.

Key words：automatic game creation；procedural content generation；large language model（LLMs）；video gamedescription language（VGDL）；cooperative creation

0 引言

近年来，大语言模型（LLMs）的发展推动了其在各个领域的应用，如智能合约代码生成[1]、数学问题求解[2]等，也被应用于互动内容的生成，例如MarioGPT[3]生成2D《马里奥》的游戏关卡，Huang等人[4]生成了3D《我的世界》的场景。（剩余30297字）