Crystalis framework boosts LLM multi-view chart generation to 75% success

Large language models (LLMs) are decent at generating single charts, but building coordinated multi-view visualizations (CMVs)—where multiple views share data, encodings, and interactions—has been a major challenge. Errors in one component can silently break others, making end-to-end generation unreliable. To solve this, researchers from Zhejiang University and elsewhere propose Crystalis, a framework that breaks CMVs into structured queries over a dependency graph spanning data, visualization, and interaction components across three abstraction levels (requirement, specification, executable object).

Crystalis uses two complementary mechanisms: progressive nucleation, which crystallizes each query vertically from requirement to object following dependency order, and semantic annealing, which enforces horizontal consistency across queries via layered logical checks. On a 12-task benchmark using five frontier LLMs, Crystalis achieved up to 75% end-to-end (E2E) success—compared to just 8.3% E2E with a standard agentic coding baseline using the same foundation model. A user study with 12 practitioners validated the decomposition and iterative refinement workflow. The paper is available on arXiv (2607.24766) and marks a significant step toward reliable, LLM-generated multi-view dashboards.