基于大语言模型的多模态互动内容人机协作创作

针对大语言模型在互动内容生成中可能存在的用户可控性弱、专家知识门槛高及缺乏用户界面的问题，提出了一种基于大语言模型的人机协作工作流框架。该框架通过用户自定义的任务分解提升可控性，并辅以可视化调试界面优化生成过程。基于该框架，实现了一个面向视频游戏描述语言的规则与关卡生成工作流实例，并开发了基于Web的协作创作平台。实验结果表明，该工作流使用不同的大语言模型均能有效地生成不同类型的互动内容，成功率相较于基线方法平均提升了20%。用户调研的系统可用性量表分数为78.75，证明该平台具备良好的可用性，能有效支持非专家用户创作，并有效提升多模态互动内容的生成质量与可控性并降低创作门槛。此外，通过将视频游戏描述语言分别集成至虚幻引擎与Godot引擎实现了3D场景生成，并通过修改工作流生成了基于JavaScript的即时战略游戏，展示了该框架的拓展性。

This paper proposes a human-AI cooperative workflow framework to address the issues of weak user control, high expertise barriers, and lack of user interface in interactive content generation using large language models. The framework enhances user control through user-defined task decomposition and optimizes the generation process with a visual debugging interface. This research also implements a workflow instance for generating rules and levels using the Video Game Description Language and develops a web-based cooperative creation platform. Experimental results demonstrate that the framework effectively generates various types of interactive content across different large language models, increasing the average success rate by 20% compared to baseline methods. The user study yields a System Usability Scale score of 78.75, which indicates the platform has good usability, can effectively support non-expert users, improve the quality and controllability of multimodal interactive content generation and lower the creation barrier. Additionally, this paper also realizes 3D scene generation by integrating Video Game Description Language with Unreal Engine and Godot Engine and generates JavaScript-based real-time strategy games by adjusting the workflow, which shows the extensibility of the framework.