EquiTabPFN: A Target-Permutation Equivariant Prior Fitted Networks

However, these models overlook However, row-order symmetry is not the only symmetry a crucial equivariance property: the arbitrary relevant to tabular data. Another key symmetry pertains ordering of target dimensions should not influence to feature order, where the arrangement of columns should model predictions. In this study, we identify not influence model predictions. Recent work (Müller et al., this oversight as a source of incompressible 2024; Hollmann et al., 2025) has addressed this challenge by error, termed the equivariance gap, which introduces employing bi-attention mechanisms similar to those studied instability in predictions. To mitigate these in earlier work (Kossen et al., 2022). This approach alternates issues, we propose a novel model designed to preserve attention over rows and columns, making the models equivariance across output dimensions. Our equivariant to feature permutations and better suited for experimental results indicate that our proposed handling another inherent symmetry of tabular data.