AdaTab: Adaptive Evidence Routing for Context-Conditioned Tabular Learning

Context-conditioned tabular models have emerged as a promising paradigm for structured data prediction, enabling flexible inference from examples provided in context. However, existing methods typically construct contexts using full-context conditioning, random sampling, or fixed nearest-neighbor retrieval, overlooking that different test instances may require different amounts and types of evidence. This can introduce redundant or uninformative examples, increasing inference cost and potentially degrading prediction. We propose AdaTab, a utility-aware adaptive evidence routing framework that formulates context construction as budget-constrained evidence acquisition at inference time. For each test instance, AdaTab estimates the marginal utility of candidate examples and selects the most informative evidence. It then dynamically expands or terminates context construction according to predictive uncertainty and the available context budget. AdaTab further incorporates missingness patterns and feature availability to improve robustness in realistic tabular settings. Without modifying backbone parameters or relying on dataset-specific schemas, AdaTab can be applied as a general inference-time enhancement for context-conditioned tabular models. Experiments across multiple structured datasets demonstrate improved predictive performance under limited context budgets and a better trade-off between accuracy and inference cost.