# PluRel: Synthetic Data unlocks Scaling Laws for Relational Foundation Models

**Authors:** Vignesh Kothapalli, Rishabh Ranjan, Valter Hudovernik, Vijay Prakash Dwivedi, Johannes Hoffart, Carlos Guestrin, Jure Leskovec  
**arXiv:** [2602.04029](https://arxiv.org/abs/2602.04029)  
**Date:** 2026-02-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.04029v1/

## Key points

- PluRel is a groundbreaking framework for generating diverse relational databases in a privacy-preserving manner.
- RFMs are crucial for data-driven decision-making but lack publicly available complex multi-table databases hindering their development.
- PluRel enables the synthesis of unlimited relational databases with customizable schemas, connectivity patterns, and data distributions.
- The framework models schemas with directed graphs and inter-table primary-foreign key connectivity with bipartite graphs, offering a step-by-step approach to creating diverse databases while maintaining computational efficiency.
- Scaling the number of synthetic databases using PluRel improves generalization to real databases and serves as a strong foundation for continued training on real-world datasets.
- This innovative approach positions synthetic data scaling as a promising paradigm for advancing RFMs and developing AI systems on enterprise data without compromising privacy.

## Summary

A groundbreaking framework for generating diverse relational databases to train  in a privacy-preserving manner. RFMs are crucial for data-driven decision-making but lack publicly available complex multi-table databases hindering their development. PluRel enables the synthesis of unlimited relational databases with customizable schemas, connectivity patterns, and data distributions. By modeling schemas with directed graphs and inter-table primary-foreign key connectivity with bipartite graphs, PluRel offers a step-by-step approach to creating a wide range of diverse databases while maintaining computational efficiency. The framework's design space supports the generation of synthetic data at scale, allowing researchers to observe new insights such as power-law scaling of RFM pretraining loss with the number of synthetic databases and total pretraining tokens. Furthermore, experiments conducted using PluRel demonstrate that scaling the number of synthetic databases improves generalization to real databases and that synthetic pretraining can serve as a strong foundation for continued training on real-world datasets. This innovative approach positions synthetic data scaling as a promising paradigm for advancing RFMs and lays the groundwork for developing AI systems on enterprise data without compromising privacy. In conclusion,  unlocks new possibilities for  by providing a solution to the scarcity of diverse relational datasets while offering a privacy-preserving method for training models on real-world data. This work contributes significantly to the field of  and relational deep learning, paving the way for more robust and scalable approaches to data-driven decision-making.