[arXiv]score: 0.12
RDB-PFN: First Relational Foundation Model Trained on 2M Synthetic Tasks
May 29, 2026
RDB-PFN is a relational database foundation model pre-trained on over 2 million synthetic single-table and relational tasks generated from Structural Causal Models, enabling in-context learning on new databases without fine-tuning. It outperforms baselines on 19 real-world relational prediction tasks in few-shot settings.
cs.LGcs.AIcs.DB
HOW THIS AFFECTS YOU
●
builderWorth watching as a potential drop-in for tabular/relational prediction tasks where labeled data is limited, though production readiness is unconfirmed.
●
researcherThe synthetic pre-training approach via relational prior generators offers a replicable path to foundation models in data-scarce structured domains.