Support-Compiled Feature Folding: More Evidence at Lower Memory Across Tabular Foundation Models
SCFF claims lower memory use while improving accuracy and NLL across six tabular-model backbones.
The paper describes a training-free inference method that folds support-ranked features through bounded encoder leaves instead of full-width pairwise mixing. In the reported wide-table benchmark slice, it improves dataset-macro accuracy and NLL on all evaluated backbones. The authors report median paired GPU-memory savings of 2.09x to 2.36x, with separately observed peak ratios up to 34.3x. Under a peak-memory cap, SCFF keeps more selected evidence and beats the widest feasible single leaf by 4.06 and 3.72 accuracy points on the named TabICLv2 and TabPFN-3 strata. ArXiv · AI/CL/LG's note
The paper describes a training-free inference method that folds support-ranked features through bounded encoder leaves instead of full-width pairwise mixing. In the reported wide-table benchmark slice, it improves dataset-macro accuracy and NLL on all evaluated backbones. The authors report median paired GPU-memory savings of 2.09x to 2.36x, with separately observed peak ratios up to 34.3x. Under a peak-memory cap, SCFF keeps more selected evidence and beats the widest feasible single leaf by 4.06 and 3.72 accuracy points on the named TabICLv2 and TabPFN-3 strata. ArXiv · AI/CL/LG's note
score 4