Seldon

Abstract

Tabular data determine most enterprise decisions: credit approvals, fraud screens, demand forecasts, clinical risk scores. For a decade, gradient-boosted decision trees have held this domain almost uncontested, and large language models have not changed that. A new paradigm is now reshaping the landscape: Tabular Foundation Models (TFMs), neural predictors pre-trained once on synthetic priors and deployed at inference time as in-context predictors.

In this report we introduce Seldon, Neuralk's tabular foundation model reaching state-of-the-art performance. We first formalize what TFMs approximate, and survey the 2022–2026 landscape. We then present results on TabBench, our open evaluation suite of 189 classification problems spanning retail, healthcare, finance, energy, and other industries. On that benchmark the top tier (Seldon, TabPFN v3 and TabICL v2) is statistically indistinguishable and sits clearly ahead of every tuned tree ensemble and every other open TFM.

We then turn to private industrial data spanning five sectors (retail, behavioural, equity, transportation, and energy), 22 problems in total. The picture is different from TabBench: the strict separation between TFMs and tree ensembles disappears, tuned XGBoost and LightGBM stay competitive on every sector, and yet Seldon takes the best mean rank on the pool, narrowly ahead of XGBoost. Within the three top-tier TFMs the gap is sharper: Seldon wins outright on 50% of the industrial pool against 18% for TabPFN v3 and 9% for TabICL v2. This is the result that motivates Seldon as an industrially focused TFM. Finally we describe the public Seldon API, which removes the GPU and scaling friction that currently makes TFMs hard to adopt.

Keywords: tabular foundation models · in-context learning · prior-fitted networks · industrial machine learning