Advancing Open and Reproducible Relational Learning: RelArena-α, TabPFN-Rel and RPI

The Problem: Relational Learning Lacks a Standard Playbook

Relational learning—applying machine learning to data stored in databases with multiple interconnected tables—has no agreed-upon way to evaluate and compare methods. Over recent years, datasets and benchmarks have proliferated, but researchers lack standardized protocols for data loading, evaluation, and hyperparameter tuning. This fragmentation makes it hard to know which approaches actually work best and slows reproducible progress.

Three Tools for the Community

The authors release three open-source components designed to fix this.

RelArena-α is a unified benchmarking framework that standardizes how researchers run and compare relational learning methods. It wraps RelBench v1 (an existing relational dataset collection) with consistent data loading, evaluation protocols, and tuning regimes—mirroring what TabArena did for tabular machine learning. Rather than each team implementing evaluation differently, the framework ensures apples-to-apples comparison.

TabPFN-Rel is a relational wrapper around TabPFN-3, a foundation model for tabular prediction. It currently ranks first on RelArena-α and improves over RDBLearn, a prior relational baseline. The key finding: simply flattening a relational database into a single table and applying a strong tabular model remains surprisingly competitive—or superior—to specialized architectures designed specifically for relational structure. This is both a concrete strong baseline and an empirical claim that challenges assumptions in the field.

RPI (Relational Predictive Interface) is a model-agnostic tool that lets practitioners define prediction problems on any new database and apply any model in RelArena-α to it. This lowers the barrier for practitioners and researchers to adopt relational learning methods without reimplementing infrastructure.

Why This Matters

The release addresses a real bottleneck. Benchmark fragmentation wastes researcher effort and makes published results hard to trust or build upon. By providing standardized infrastructure (RelArena-α) and a strong, simple baseline (TabPFN-Rel), the authors aim to:

1. Enable reproducible comparisons across methods on diverse real relational databases
2. Demonstrate that simple approaches work — the success of table-flattening suggests that specialized relational architectures may be overcomplicating the problem
3. Lower adoption friction — RPI makes it easy for industry and researchers to apply these methods to new databases without boilerplate engineering

The authors emphasize this is an α-release, explicitly inviting community feedback and collaboration. The ambition is not just to release tools but to catalyze the field toward reliable, reproducible progress on a problem (relational learning at scale) that remains largely underexplored compared to image or language domains. By aligning evaluation and publishing a strong baseline, they create a clearer target for improvement.