Search papers, labs, and topics across Lattice.
This paper addresses the challenge of context management in tabular foundation models (TFMs) for stream learning, where predictions must adapt to sequentially arriving data under distribution shifts. The authors introduce CURE, a context management policy that incorporates uncertainty-aware admission and redundancy-aware eviction to effectively manage the context for TFMs. Experimental results demonstrate that CURE achieves up to a 27.0% relative improvement over traditional stream learning methods across seven datasets, showcasing its robustness across various TFM architectures.
CURE redefines context management for tabular foundation models, achieving significant performance gains in stream learning by intelligently managing uncertainty and redundancy.
Tabular stream learning requires predictions on sequentially arriving examples under distribution shift. While standard methods adapt by updating model states, tabular foundation models (TFMs) make predictions conditioned on a labeled context in an in-context manner, making them a natural alternative for stream learning. This shifts the challenge from how to update the model to how to manage the context. We propose a future information view that yields three practical requirements for context management: preserve recent examples, retain uncertain examples, and remove redundant examples. We instantiate these requirements as CURE (Context management via Uncertainty-aware admission and Redundancy aware Eviction), a context-managing policy with entropy-gated admission and redundancy-aware eviction. Across seven streams, CURE shows up to 27.0% relative improvement over classical stream learners, remains robust across multiple TFM backbones, and ranks first among other policy variants. Code and datasets are available at https://github.com/morcellinus/CURE-ICML-FMSD.
