Researchers have developed Crystalis, a framework that enables large language models to generate coordinated multi-view data visualizations with 75% end-to-end success across benchmark tests.

The framework addresses a core limitation in current AI systems: while models can create individual charts, they struggle with complex visualizations where multiple views share data and interact with each other. Errors in one component often cascade through the system, breaking the entire visualization.

Crystalis decomposes complex visualizations into structured queries across three component types—data, visualization, and interaction—operating at three abstraction levels from requirements to executable code. The system uses two complementary mechanisms: progressive nucleation builds each query vertically from concept to implementation, while semantic annealing maintains consistency across components through layered logical checks.

Testing across five frontier language models on a 12-task benchmark, Crystalis substantially outperformed an agentic coding baseline that achieved just 8.3% end-to-end success with the same foundation model.

The research team, led by Dazhen Deng from an undisclosed institution, conducted a user study with 12 practitioners who confirmed the usability of the decomposition and iterative refinement workflow.

Breaking the visualization complexity barrier

The framework targets what researchers call "coordinated multi-view visualizations"—dashboards where charts share data flows and users can interact across different views. Traditional approaches fail because tight coupling between data transformations, visual encodings, and interaction logic creates fragile dependencies.

Rather than pursuing end-to-end analytical quality, which depends heavily on domain expertise, the researchers focused on structural correctness—ensuring the generated visualizations function as intended without runtime errors.

The query-centric modeling approach represents each visualization component as a structured query over a dependency graph, allowing the system to track relationships and validate consistency at each abstraction level.

The research appears in a paper submitted to arXiv, marking a significant step toward more reliable AI-generated business intelligence tools and data analysis workflows.