CRUMB: Efficient Prior Fitted Network Inference via Distributionally Matched Context Batching

The article introduces CRUMB (Clustered Retrieval Using Minimised-MMD Batching), an inference wrapper designed to enhance the efficiency of prior-fitted networks (PFNs) by clustering test queries and selecting a distributionally matched subset of training data, thus mitigating the computational burden of self-attention mechanisms. Evaluated on the TabArena benchmark across three PFN architectures (TabPFNv2, TabICLv1, TabICLv2), CRUMB demonstrates superior performance compared to existing context selection methods while remaining architecture-agnostic and requiring no retraining. This approach is significant for practitioners as it enables scalable and efficient inference on large datasets, addressing the challenges of covariate drift in real-world applications.