Relational Foundation Models Struggle with High-Cardinality Data

Denis Oliveira Correa, Francisco Galuppo Azevedo· September 2, 2026 View original

Key takeaways

  • Current relational foundation models struggle with high-cardinality data due to context window limitations.
  • Simple data pre-aggregation steps can significantly improve model performance on such data.
  • The findings question the immediate readiness of these models for complex real-world applications.
  • Future relational AI development needs to address inherent limitations in handling extensive related records.

Who benefits

FinanceRetailHealthcareLogisticsEnterprise Software

Summary

A study using a synthetic financial dataset, Animus, reveals that current relational deep learning foundation models (RT, Griffin, RelGT) perform poorly when entities have many related records due to constrained neighborhood budgets. Simple temporal pre-aggregation significantly improves performance, questioning the models' readiness for complex real-world data.

Recent relational deep learning architectures, often touted as foundation models for multi-table data, face significant limitations when dealing with real-world datasets characterized by high cardinality. These models typically impose strict neighborhood budgets, forcing the truncation of records when an entity is linked to numerous related entries. Researchers demonstrated this issue using Animus, a synthetic financial dataset where predicting customer income requires aggregating thousands of transactions. On this raw, high-cardinality data, prominent models like Relational Transformer (RT), Griffin, and RelGT achieved very low R-squared scores (<= 0.18). However, a simple, routine temporal pre-aggregation step dramatically improved performance, boosting R-squared to 0.65. This stark difference suggests that while these models are foundational in concept, their current design struggles with the extensive context windows required for real-world relational data. It raises questions about their immediate applicability without significant data preprocessing.

Why it matters

This research highlights a critical practical limitation of current relational foundation models, indicating that they may not be ready for direct application to complex, high-cardinality enterprise data without substantial preprocessing.

How to implement this in your domain

  1. 1Assess existing data pipelines for relational data to identify instances of high-cardinality relationships.
  2. 2Prioritize data preprocessing steps, such as temporal aggregation, before feeding relational data into foundation models.
  3. 3Evaluate the performance of relational foundation models on datasets that accurately reflect real-world complexity and cardinality.
  4. 4Investigate model architectures that can inherently handle larger context windows or more effectively aggregate information from numerous related records.
  5. 5Develop internal guidelines for data preparation when deploying relational AI models to ensure robust performance.

Original post by Denis Oliveira Correa, Francisco Galuppo Azevedo

"arXiv:2609.00460v1 Announce Type: new Abstract: Recent Relational Deep Learning architectures have been proposed as foundation models for multi-table relational data, yet they impose constrained neighborhood budgets that force row truncation when an entity has many related record…"

View on X

Originally posted by Denis Oliveira Correa, Francisco Galuppo Azevedo on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses