Crystalis: LLM Framework for Multi-View Visualization Generation

A novel framework named Crystalis facilitates large language models in producing coordinated multi-view visualizations (CMVs), which integrate shared data flows and interactions across views. This system tackles the issue of tight coupling at the field level among data transformations, visual encodings, and interaction coordination, where a mistake in one area can compromise others. Crystalis employs query-centric CMV modeling to break down a CMV into structured queries that utilize a dependency graph comprising three types of components (Data, Visualization, Interaction) and three levels of abstraction (requirement, specification, executable object). To maintain structural integrity, two complementary mechanisms function within this framework. The research focuses on whether LLMs can consistently generate structurally accurate CMVs and the necessary abstractions for this, rather than aiming for complete analytical quality. The study is available on arXiv with the identifier 2607.24766.