A principled framework for uncertainty decomposition in TabPFN

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Fortini, Sandra, Ng, Kenyon, Petrone, Sonia, Rousseau, Judith, Wei, Susan
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908813017219072
author Fortini, Sandra
Ng, Kenyon
Petrone, Sonia
Rousseau, Judith
Wei, Susan
author_facet Fortini, Sandra
Ng, Kenyon
Petrone, Sonia
Rousseau, Judith
Wei, Susan
contents TabPFN is a transformer that achieves state-of-the-art performance on supervised tabular tasks by amortizing Bayesian prediction into a single forward pass. However, there is currently no method for uncertainty decomposition in TabPFN. Because it behaves, in an idealised limit, as a Bayesian in-context learner, we cast the decomposition challenge as a Bayesian predictive inference (BPI) problem. The main computational tool in BPI, predictive Monte Carlo, is challenging to apply here as it requires simulating unmodeled covariates. We therefore pursue the asymptotic alternative, filling a gap in the theory for supervised settings by proving a predictive CLT under quasi-martingale conditions. We derive variance estimators determined by the volatility of predictive updates along the context. The resulting credible bands are fast to compute, target epistemic uncertainty, and achieve near-nominal frequentist coverage. For classification, we further obtain an entropy-based uncertainty decomposition.
format Preprint
id arxiv_https___arxiv_org_abs_2602_04596
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle A principled framework for uncertainty decomposition in TabPFN
Fortini, Sandra
Ng, Kenyon
Petrone, Sonia
Rousseau, Judith
Wei, Susan
Machine Learning
Methodology
62F15
TabPFN is a transformer that achieves state-of-the-art performance on supervised tabular tasks by amortizing Bayesian prediction into a single forward pass. However, there is currently no method for uncertainty decomposition in TabPFN. Because it behaves, in an idealised limit, as a Bayesian in-context learner, we cast the decomposition challenge as a Bayesian predictive inference (BPI) problem. The main computational tool in BPI, predictive Monte Carlo, is challenging to apply here as it requires simulating unmodeled covariates. We therefore pursue the asymptotic alternative, filling a gap in the theory for supervised settings by proving a predictive CLT under quasi-martingale conditions. We derive variance estimators determined by the volatility of predictive updates along the context. The resulting credible bands are fast to compute, target epistemic uncertainty, and achieve near-nominal frequentist coverage. For classification, we further obtain an entropy-based uncertainty decomposition.
title A principled framework for uncertainty decomposition in TabPFN
topic Machine Learning
Methodology
62F15
url https://arxiv.org/abs/2602.04596