ChartDesign: LLM-Based Data Visualization Designer

A new system called ChartDesign post-trains large language models to generate chart design attributes from tabular data. The approach uses vision language models to extract design features from charts in public surveys (PewResearch) and academic repositories (CharXiV), creating a training corpus of data-design pairs. ChartDesign aims to automate chart type selection, axis orientation, font sizes, and layouts, reducing the need for expensive human expertise. The work explores LLMs as chart designers, moving beyond handcrafted heuristics and rule matching that struggle to generalize across domains.