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

In three linesCRUMB is an inference wrapper for prior-fitted networks (PFNs) that reduces quadratic attention complexity by selecting a distributionally matched training subset via MMD minimisation. Tested on 51 TabArena datasets across TabPFNv2, TabICLv1, TabICLv2, CRUMB outperforms existing context selection strategies and is resilient to covariate drift.

Summary generated by Claude — human-verified