Microsoft's ChartDesign uses LLMs to automate data visualization
LLMs now generate publication-ready charts 31% more accurately than rule-based systems
ChartDesign is a system that post-trains large language models to act as chart designers, generating design attributes from tabular data. It uses vision-language models to extract data and design from existing charts, then fine-tunes LoRA adapters on Phi3, Qwen3, and InternVL2.5. On a held-out test set, ChartDesign achieves up to 84% accuracy—far surpassing the 53% best baseline—and generalizes to unseen domains. Charts rendered from its specifications are visually appealing and preferred by humans, narrowing the human-AI gap in data visualization.
- ChartDesign uses LLMs (Phi-3, Qwen-3, InternVL2.5) fine-tuned with LoRA to generate chart specs from tabular data
- Achieves 84% accuracy in design generation vs. 53% for best baseline, and generalizes to unseen domains
- System uses vision-language models to extract 10+ design attributes (e.g., axis labels, bar spacing) from public chart repositories
Why It Matters
Automates a tedious, expertise-heavy step in data analysis—freeing analysts to focus on insights, not pixel-level tweaking.