Can AI agents learn to build better statistical tools on the fly?

Scientists spend enormous time manually fitting statistical models to data. VESTA equips vision-language models with the ability to dynamically create and reuse diagnostic tools—visualizations and statistical tests—as they iteratively refine models. Tested on a new benchmark (DAWN) spanning distribution fitting to real astronomy tasks like modeling gravitational-wave signals, VESTA's adaptive toolkit approach substantially outperforms systems relying only on static tools or critique alone.