← News

ResearchSources: arXiv · 16 Nov 2026

PRODiGI infers executable generative programs from data in a single pass

Researchers present PRODiGI, a pretrained model that turns sample data into an explicit, runnable generative program without dataset-specific fitting, and report accuracy and speed gains over competing pretrained models.

In plain words

Density estimation means learning the underlying probability pattern behind a set of data samples, which usually requires fitting a new model for each dataset. PRODiGI instead reads a dataset and outputs an explicit program that can generate samples and be inspected on its own. The authors say this makes tabular generative modeling faster and more interpretable.

What was announced

The model, described in an arXiv abstract, is pretrained on synthetic datasets paired with their ground-truth programs, so at inference it can produce an executable program in a single forward pass. It handles varied generative families and data dimensionalities through template prediction and non-autoregressive decoding of program parameters. Because the output is a standalone program, it supports direct sampling, density and score evaluation, and inspection without the pretrained model.

The authors also introduce program-space fine-tuning, which adjusts differentiable program parameters to maximize the likelihood of real data while leaving the pretrained model unchanged. They report that the fine-tuned variant, PRODiGI-FT, beats existing pretrained models on density and score metrics with multi-fold speedups over its closest competitors, and that fine-tuning cuts generation MMD by 90 percent. The abstract frames the work as a new direction for fast, interpretable tabular generative modeling.

What it does not show

  • The evidence is the paper's abstract only; it reports the authors' own benchmark claims, which are not independently confirmed, and details such as baselines, datasets and evaluation setup are not visible in the excerpt.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.