tabular foundation model
In plain wordsA neural network pretrained on millions of made-up tables so that, given a new spreadsheet of labeled examples, it can predict the missing answers at once, without being retrained for that table.
A neural network, usually a transformertransformer, n.: The design behind today's chatbots and much of modern AI. It handles text by letting every word look at every other word and decide which of them matter for understanding it., pretrained once on a very large collection of tables — typically millions of synthetic datasets sampled from random structural causal models and other generators — so that at prediction time it takes a whole labeled training table plus unlabeled rows as input and outputs predictions for those rows in a single forward passforward pass, n.: The step where a neural network takes an input and runs it through its layers, one after another, to produce an output — its answer or prediction., with no gradient-based training or hyperparameter tuninghyperparameter tuning, n.: Trying out different settings for a model's training, such as learning speed or model size, and keeping the combination that performs best on data held back for checking. on the new task. This is in-context learningin-context learning, n.: A language model's ability to pick up a new task from instructions or examples placed in its input, on the spot, without any retraining — like a quick study learning a game's rules from one round. applied to tabular datatabular data, n.: Data arranged as a table, like a spreadsheet: each row is a case — a customer, a loan, a patient — and each column a measured property. Much business and scientific data looks like this.: the network has learned to approximate Bayesian posterior prediction under its pretraining distribution, an approach called a prior-data fitted network (PFN). The best-known example, TabPFN, was reported in its 2025 Nature version to match or beat tuned gradient boostinggradient boosting, n.: A method that builds a strong predictor by adding small decision trees one at a time, each trained to correct the errors that remain after all the trees before it. on datasets of up to about 10,000 rows in seconds; later releases extended it, and as of October 2026 TabPFN-3.5 is recommended for up to about a million rows. Strengths are speed on small data, good calibration and no tuning; limits are inference cost on large tables and restrictive licenses for some weights.
On the 3,000-row clinical table, the tabular foundation model beat our tuned boosting baseline without any training.
Origin
Samuel Müller, Noah Hollmann, Frank Hutter and colleagues introduced prior-data fitted networks ('Transformers Can Do Bayesian Inference', 2021–2022) and TabPFN, first as a July 2022 preprint (ICLR 2023) and then, much extended, in Nature (January 2025) under a title calling it a tabular foundation model. Prior Labs, founded in 2024, released TabPFN-2.5 (November 2025), TabPFN-3 (May 2026) and TabPFN-3.5 (September 2026).
Usage note
Other groups released competing tabular foundation models in 2024–2026. As of October 2026 their advantage is clearest on small and medium tables; on very large tables, boosted trees remain strong competitors.
Related
- tabular datatabular data, n.: Data arranged as a table, like a spreadsheet: each row is a case — a customer, a loan, a patient — and each column a measured property. Much business and scientific data looks like this.
- in-context learningin-context learning, n.: A language model's ability to pick up a new task from instructions or examples placed in its input, on the spot, without any retraining — like a quick study learning a game's rules from one round.
- foundation modelfoundation model, n.: A very large AI model, built once at great cost from vast amounts of text, images or code gathered mostly from the internet, that many products then reuse as their starting point, from chatbots to coding tools.
- gradient boostinggradient boosting, n.: A method that builds a strong predictor by adding small decision trees one at a time, each trained to correct the errors that remain after all the trees before it.
- transformertransformer, n.: The design behind today's chatbots and much of modern AI. It handles text by letting every word look at every other word and decide which of them matter for understanding it.
- AutoMLAutoML, n.: Software that automates the trial-and-error parts of building a machine-learning model, such as choosing the algorithm, its settings and the data-preparation steps, so less expert effort is needed.
- gradient boostinggradient boosting, n.: A method that builds a strong predictor by adding small decision trees one at a time, each trained to correct the errors that remain after all the trees before it.
Sources
- Hollmann, N., Müller, S., Purucker, L. et al. (2025). Accurate Predictions on Small Data with a Tabular Foundation Model. Nature, 637(8045), 319–326. doi:10.1038/s41586-024-08328-6.
- Hollmann, N., Müller, S., Eggensperger, K. & Hutter, F. (2023). TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. ICLR 2023. arXiv:2207.01848. arxiv.org/abs/2207.01848
Mentioned in
tabular datatabular data, n.: Data arranged as a table, like a spreadsheet: each row is a case — a customer, a loan, a patient — and each column a measured property. Much business and scientific data looks like this.