Researchers have built a benchmark to expose how badly AI models fail at recreating interactive dashboards from visual inputs.
The paper introduces Dashboard2Code, a task that asks a model to explore a live interactive dashboard — clicking, filtering, interacting — and then generate the code needed to reproduce it. To measure performance, the team built DashboardMimic: 180 manually verified dashboard-code pairs built with Plotly and Dash, spanning three difficulty levels and eight real-world interaction patterns. The evaluation framework combines code analysis with dynamic interaction testing, and the authors report it aligns closely with human judgments.
Most AI coding benchmarks still treat "generate a chart" as the ceiling. Real-world data tools are interactive — filters, dropdowns, drill-downs — and this benchmark is a rare attempt to measure whether models can handle that complexity. The persistent gap between open-source and closed-source models on high-complexity dashboards also signals that this is not a problem scale alone will solve.
For context: models that ace static chart generation still stumble when the target moves. If your team is betting on AI to automate dashboard development, this research suggests you should budget for a human in the loop for anything non-trivial.