DAWN Dataset

DAWN benchmarks AI agents on two core data science tasks: distribution fitting and time series modeling. Both domains are organized into Easy, Hard, and Astro splits, with synthetically generated data to allow robust measurement of model fit.

VESTA: Visual Exploration with Statistical Tool Agents

We propose and benchmark AI agents on DAWN in our arxiv paper: https://arxiv.org/abs/2606.00384

Fitting quantitative models to data is a central step in scientific workflows, yet it remains one of the least automated. Recent agent-based systems leverage language and vision-language models (VLMs) to iteratively propose and refine statistical models, but these systems struggle on more challenging modeling tasks. To address these limitations, we introduce VESTA (Visual Exploration with Statistical Tool Agents), a framework that equips VLMs with a dynamically growing exploration toolkit to guide model refinement through data transformations, hypothesis-driven visualizations, and robust statistical tests. Unlike prior systems that rely on iterative critique alone, VESTA actively explores data before and during refinement by selecting or creating diagnostic tools, which accumulate in the model’s context and can be reused later. We evaluate VESTA against established baselines in three toolkit configurations: no tools, static expert-written tools, and dynamic model-written tools. To support this evaluation, we introduce DAWN (Dataset for Automated Workflows and Numerical Modeling), a benchmark targeting distribution fitting and time series modeling with varying difficulty tiers, and culminating in real-world astronomy tasks including modeling initial mass functions and gravitational-wave chirp signals. We find that VESTA’s dynamic tool creation outperforms prior agentic pipelines, with the largest gains on complex and domain-specific tasks. We further show that dynamically generated tools are substantially more sophisticated than those produced by existing visual tool-creation systems, covering more diagnostic categories per function and strongly preferring visual outputs that the VLM critic can reason over directly.

Distribution Fitting

• dist_easy: Identify the family and parameters of a unimodal distribution. Families include gaussian, lognormal, student-t, exponential, uniform, weibull, laplace, cauchy, and pareto, with priors and sample sizes randomly sampled.
• dist_hard: Mixtures of exactly two distributions from the families above, reflecting real-world phenomena such as gaussian-laplace mixtures (speech signal processing) and lognormal-exponential mixtures (survival analysis).
• dist_astro: Initial mass functions (IMF), describing the distribution of stellar masses at birth. Includes five functional forms: Salpeter (single power law), Kroupa (piecewise power law), Chabrier (log-normal plus power law), and two freeform variants (tight and wide breaks).

Time Series Modeling

• ts_easy: Sequences with simple linear trends (increasing, decreasing, flat) and standard periodic seasonal components, with minimal noise.
• ts_hard: More complex dynamics including random walks (ARIMA processes), sigmoidal (S-curve) functions, ECG heartbeat signals, and sales-forecasting-style seasonality with increasing averages. Requires multi-periodic and Gaussian process modeling (e.g. Matérn, RBF kernels).
• ts_astro: Time series inspired by gravitational wave chirps, where frequency continuously increases as binary star systems converge, including a variant where amplitude additionally decays over time.

Dataset Statistics

Each dataset $D_i$ contains $n \in [600, 1500]$ points sampled from a ground-truth distribution or time series, with parameters drawn uniformly from fixed ranges. For each domain and difficulty split, 50 (Easy/Astro) or 100 (Hard) examples are generated, for a total of 200 distributions and 200 time series.

Citation

If you use DAWN, please cite:

@misc{rudman2026vestavisualexplorationstatistical,
      title={VESTA: Visual Exploration with Statistical Tool Agents}, 
      author={William Rudman and Abhishek Divekar and Kanishk Jain and Sebastian Joseph and Stella S. R. Offner and Matthew Lease and Kyle Mahowald and Greg Durrett and Junyi Jessy Li},
      year={2026},
      eprint={2606.00384},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2606.00384},
}