Reproduce “Benchmarking Tabular Foundation Models for Conditional Density Estimation in Regression” on OpenResearch

About this paper

Rafael Izbicki, Pedro L. C. Rodrigues

Researchers from Federal University of São Carlos and Inria systematically benchmarked tabular foundation models (TFMs) for conditional density estimation (CDE) against diverse baselines. TabPFN and TabICL variants consistently delivered superior density accuracy across most datasets and sample sizes, exhibiting a notable sample efficiency advantage, particularly highlighted in a photometric redshift estimation case study where TFMs used an order of magnitude less data for comparable or better performance.

Key insights

• Tabular foundation models consistently provide superior density accuracy (measured by CDE loss, log-likelihood, and CRPS) across a broad range of datasets and sample sizes compared to a diverse set of CDE baselines.
• TFMs demonstrate a substantial sample efficiency advantage, often achieving better CDE performance with significantly less data than traditional methods, as evidenced by the photometric redshift estimation case study.
• While TFMs excel in density accuracy, their calibration performance can be less competitive than some end-to-end trained neural baselines at larger sample sizes, suggesting a potential area for further development or post-hoc calibration.

Summary by alphaXiv.