Synthetic tabular data is useful for sharing datasets without exposing private records, but many generation methods are computationally heavy. A new paper on arXiv proposes a lightweight algorithm that models tables with both continuous and discrete variables, aiming to strike a balance between simplicity and usefulness.

The method works by first capturing each variable's empirical marginal distribution, then applying a chosen quantile function to transform the data into a latent space. This avoids complex joint-distribution modeling while still preserving the overall shape of the original data. The authors describe the approach as lightweight, suggesting it may be especially practical for settings with limited compute or where a quick baseline is needed.

Because the paper is an abstract-level announcement, details on performance and comparisons with other methods are not included here. The contribution appears to be a pragmatic alternative to more resource-intensive generative models, though its trade-offs in fidelity and utility would need to be assessed from the full paper.