CURE Enhances Tabular Foundation Models for Stream Learning

A new context management policy, CURE (Context management via Uncertainty-aware admission and Redundancy aware Eviction), has been developed to enhance the performance of tabular foundation models (TFMs) in stream learning environments. Stream learning requires models to make predictions on data arriving sequentially, often under distribution shifts. Unlike traditional methods that update model states, TFMs make predictions by conditioning on a labeled context. This shifts the challenge to effectively managing this context. CURE is designed based on three key requirements: preserving recent examples, retaining uncertain examples, and removing redundant ones. The policy implements entropy-gated admission for new examples and redundancy-aware eviction for older ones. Across seven data streams, CURE demonstrated up to a 27.0% relative improvement over classical stream learners, proved robust across various TFM backbones, and outperformed other policy variants. This makes TFMs a natural and effective alternative for stream learning when coupled with intelligent context management.