Deep learning methods built for relational databases finally get a fair fight.
Researchers have published one of the first systematic benchmarking studies comparing deep learning models designed specifically for relational databases (RDBs). The study evaluates several methods across five RDBs, running both classification and regression tasks on each. To make comparisons honest, the team refactored all models to run under consistent experimental protocols — something the field had notably lacked before this work.
The headline finding: the relational transformer (RT) approach outperformed both graph-based RDB models and TabPFN 2.5, a leading tabular foundation model, on the tested tasks. That second result is the more striking one. Even on single-table tasks — where tabular foundation models are supposed to shine — dedicated RDB methods held their own. Expanding learning across multiple linked tables (what the paper calls higher "hops") improved performance, though with diminishing returns as computational costs climbed.
Most enterprise data lives in relational databases, not tidy flat files, so methods that can natively traverse table relationships have practical appeal at scale. The benchmark is also open-source, which means other researchers can plug in new methods rather than cherry-pick favorable comparisons — a genuine service to a field where apples-to-apples evaluation has been the exception.
Whether RT-style approaches can hold up outside controlled benchmarks, on messier production schemas, is the question this paper leaves open.