Beyond temporal encoding: Efficient and hierarchical model-level online adaptation for tabular distribution shift

Tabular models suffer severe performance degradation under temporal distribution shift. Existing approaches address this mainly at the feature level, for example through temporal encoding (TARS, TE) or feature modulation, but require costly end-to-end tuning and retraining for each dataset and cannot adapt to novel shift patterns at test time. A complementary model-level approach is proposed: a three-tier architecture inspired by prefrontal cortex plasticity and informed by a neural-manifold view of multi-timescale adaptation. The slow tier provides frozen pretrained priors (foundation models or tree ensembles), the intermediate tier performs fixed-weight ensemble blending, and the fast tier applies a novel dual-buffer k-nearest-neighbor correction in learned embedding space. This fast tier combines a short buffer for recent drift with a long buffer for global stability, using drift-aware blending to respond automatically to distribution change. Unlike feature-level methods, the proposed approach adapts continuously to evolving distributions without retraining. On eight TabReD benchmarks, with additional validation on thirteen temporal and domain-shift datasets, the architecture achieves the best reported RMSE on four of five TabReD regression datasets, on three of them by margins beyond run-to-run variation (reductions of up to 2.7% over the strongest baseline on each dataset) and competitive or superior AUC (0.9652, 0.8697, and 0.6373) on the three binary classification datasets. As a downstream benefit of the improved point predictions, adaptive conformal inference yields prediction intervals 3-33% narrower at equivalent 90% coverage on the regression benchmarks. All of these gains derive from the same mechanism: model-level online residual correction over frozen priors, without test-time retraining.