VESTA: Visual Exploration with Statistical Tool Agents

VESTA is a new AI framework that enhances vision-language models with dynamically generated statistical tools to automate scientific model fitting tasks. The system outperforms prior approaches by actively exploring data through adaptive tool creation rather than relying solely on iterative critique, with particular strength on complex, domain-specific modeling problems.

VESTA represents a meaningful advancement in automating scientific workflows, specifically addressing a persistent automation gap in quantitative modeling. Traditional agent-based systems have relied on language models critiquing and refining models iteratively, but this approach hits diminishing returns on complex tasks requiring domain expertise. VESTA's innovation lies in equipping models with an expanding toolkit that grows dynamically—the system doesn't just critique existing models but actively generates diagnostic visualizations and statistical tests tailored to emerging data patterns.

This approach reflects a broader shift in AI toward more sophisticated task decomposition. Rather than monolithic end-to-end reasoning, modern systems increasingly succeed by breaking problems into specialized subtasks with appropriate tools. VESTA's contribution is making this toolkit creation itself adaptive, allowing the model to synthesize new diagnostic instruments as needed rather than relying on static expert-written options.

The introduction of DAWN, a benchmark spanning distribution fitting, time series modeling, and real-world astronomy tasks, provides measurable validation. Results demonstrate substantial performance gaps between dynamic tool creation and static toolkits, with generated tools exhibiting greater sophistication than prior visual tool-creation systems. This has immediate implications for scientific computing workflows where automation remains limited—physics, astronomy, and materials science could benefit from similar frameworks.

Looking forward, the key question is generalization. VESTA excels within quantitative modeling domains, but whether this dynamic tool-generation approach transfers to other complex reasoning tasks remains unexplored. Success here could influence how foundation models are deployed across specialized scientific domains.

• →VESTA dynamically generates statistical tools rather than relying on static expert-written toolkits, improving performance on complex modeling tasks.
• →The DAWN benchmark introduces a new evaluation framework for automated scientific workflows spanning distribution fitting to real-world astronomy applications.
• →Dynamically created tools outperformed both no-tool and static-tool baselines, with particular gains on domain-specific and challenging tasks.
• →VESTA's approach represents a shift toward adaptive tool-generation as a core component of agentic AI systems for specialized domains.
• →The framework demonstrates that vision-language models can effectively select and synthesize diagnostic tools beyond simple iterative critique patterns.