Prior Labs Open-Sources RelArena-α, TabPFN-Rel and RPI — and Flattening Still Beats Specialized Relational Architectures

arXiv 2608.16319 (verified against pypi.org/project/relarena and github.com/PriorLabs/relarena)·medium signalXBlueskyLinkedInCopy link

Prior Labs released three Apache-2.0 pieces: RelArena-α, a unified framework standardizing data loading, evaluation, and tuning for RelBench v1's 21 entity-level forecasting tasks (inspired by TabArena); TabPFN-Rel, a relational harness for TabPFN-3 that improves on RDBLearn via deep feature synthesis with per-task join depth d ∈ {2,3,4}; and RPI, a model-agnostic YAML-driven interface that turns CSV or Parquet files into a task in two lines of code, shipped with an agent skill and example specs for all 21 tasks. TabPFN-Rel currently ranks first on RelArena-α, adding to the evidence that flattening a relational database into a single table remains competitive with specialized relational architectures — `relarena` v0.0.1a1 is already on PyPI.