Support-Compiled Feature Folding (SCFF) is a training-free inference scheme for frozen tabular foundation models on wide tables. It rank-orders features by support, puts strong features in a Core and the rest in a Tail, folds both into narrow groups, encodes each group with the same feature encoder, and only adds Tail evidence that passes a support check on residual information. The result is linear interaction cost instead of quadratic, with no new parameters and no ensembling.
On 18 wide-table datasets from fixed AMLB-29, TabZilla, and TabArena snapshots, SCFF improved dataset-macro accuracy and negative log-likelihood on all six evaluated backbones. Median paired GPU-memory savings are 2.09 to 2.36 times, and a maximum peak ratio of 34.3 times was observed. In memory-capped comparisons, savings were reinvested to keep more support-selected features, yielding 4.06 and 3.72 accuracy-point gains over the widest feasible single-leaf baseline on two wide-Core strata.
The interesting part is the conditional Tail: every column is encoded, but Tail evidence must earn its place by contributing residual signal beyond the Core, which addresses the noise and redundancy that usually comes with adding features.