FairTFM: Training Fair Tabular Foundation Models

A recent paper published on arXiv (2608.14211) presents FairTFM, a novel training methodology aimed at embedding fairness constraints within Tabular Foundation Models (TFMs). These machine learning models leverage in-context learning for predictions on new datasets without the need for task-specific training, gaining traction in critical decision-making scenarios. Despite their growing use, the fairness aspects of TFMs remain largely unexamined. The FairTFM approach tackles two significant issues: the scarcity of sensitive attribute access in training datasets and the misalignment of current fairness strategies with the in-context learning framework. By employing synthetic fairness tasks and a fairness-aware architecture featuring a gradient reversal layer, the model learns to produce representations that are unaffected by sensitive attributes. Experiments across 132 fairness tasks demonstrate marked improvements in fairness while preserving predictive accuracy. This research, announced as a cross-type submission, seeks to facilitate equitable predictions in a single forward pass, enhancing its applicability in real-world settings.