Towards Fair In-Context Learning with Tabular Foundation Models

TL;DR

This work investigates fairness in transformer-based tabular foundation models performing in-context learning, evaluating three models (TabPFNv2, TabICL, TabDPT) across diverse fairness benchmarks. It introduces three preprocessing strategies—correlation removal, group-balanced demonstrations, and uncertainty-based demonstration selection—showing that uncertainty-based sampling consistently improves group fairness with minimal utility loss. The study reveals that correlation removal can backfire due to leakage when applied to both training and test data, and that uncertainty-based methods offer a robust, model-agnostic path to fair ICL without retraining. Overall, the findings suggest practical, scalable fairness improvements for tabular ICL and provide code to facilitate reproducibility and adoption in real-world tasks.

Abstract

Transformer-based tabular foundation models have recently demonstrated promising in-context learning (ICL) performance on structured data, emerging as competitive alternatives to gradient-boosted trees. However, the fairness implications of this new paradigm remain largely unexplored. We present the first investigation of fairness in tabular ICL, evaluating three recently proposed foundation models--TabPFNv2, TabICL, and TabDPT--on multiple benchmark datasets. To mitigate biases, we explore three pre-processing fairness-enhancing methods: correlation removal (decorrelating input features from the sensitive attribute), group-balanced sample selection (ensuring equal representation of protected groups in context examples), and uncertainty-based sample selection (prioritizing context examples with high sensitive-attribute prediction uncertainty). Our experiments show that the uncertainty-based strategy consistently improves group fairness metrics (e.g., demographic parity, equalized odds, and equal opportunity) with minimal impact on predictive accuracy. We release our code to facilitate reproducibility https://github.com/patrikken/Fair-TabICL.

Figures (13)

• Figure 1: Comparing the fairness-utility Pareto-front of different fairness interventions using tabpfn on the ACSIncome, ACSMobility, and ACSTravelTime datasets.
• Figure 2: Comparing the fairness-utility tradeoffs of tabular foundation models (tabicl, tabdpt, and tabpfn) under uncertainty-based in-context sample selection (uncert_tabpfn) for different coverage ($\epsilon$ controlling the tradeoff). Results with other datasets can be found in the Appendix (Figure \ref{['fix:app:trade_off_tabicl_vs_tabpfn']}).
• Figure 3: Ablation on the in-context example set size. Analyzing the impact of the in-context set size on the fairness and accuracy of ICL prediction with tabpfn.
• Figure 4: Evaluating variants of the cr on the ACSIncome dataset with tabpfn and tabicl. Applying cr to the training and testing data exacerbates unfairness, while applying the transformation only to the training set improves fairness.
• Figure 5: tabpfn vs. tabicl vs. tabdpt. Comparing the fairness-accuracy tradeoffs of tabular foundation models under uncertainty-based fairness interventions. tabpfn generally provides better fairness accuracy tradeoffs.