Scalable Tabular AI Inference for Large Datasets with Prototype Selection

Pretrained tabular foundation models like TabPFN offer strong predictive capabilities but struggle with large datasets due to inference context limitations. Researchers have developed Balanced Adaptive Prototype Selection (BAPS), a novel framework designed to overcome this challenge. BAPS constructs a compact, yet highly informative, context for the model by preserving key data characteristics such as representative structure, decision boundaries, local density, class balance, and feature-space diversity. This approach allows TabPFN to process significantly larger datasets without requiring any modifications or retraining of the original model. Experimental results on million-row datasets like HIGGS and SUSY demonstrated that using just 512 prototypes, BAPS achieved an approximate 1,953-fold context compression. Crucially, this compression maintained strong predictive performance and reliable calibration, all while running on standard CPU hardware without GPU acceleration. These findings highlight context construction as a practical and effective method for extending the utility of pretrained tabular foundation models to massive datasets.