No Need to Train Your RDB Foundation Model

2026-02-14 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, short

Summary

The paper "No Need to Train Your RDB Foundation Model" (arXiv:2602.13697, submitted 14 Feb 2026, last revised 4 Jun 2026) introduces a principled family of RDB encoders designed to work with existing single-table in-context learning (ICL) foundation models without requiring training or fine-tuning. This approach addresses the challenge of applying ICL to multi-table relational databases (RDBs) by compressing variably-sized RDB neighborhoods into fixed-length ICL samples. A key finding is that ICL-specific compression should be constrained within high-dimensional RDB columns where entities share units and roles, rather than across heterogeneous columns. The authors demonstrate that encoder expressiveness is maintained even without trainable parameters. They also developed scalable SQL primitives to implement the encoder stage, resulting in the open-source RDBLearn foundation model, which shows robust performance on unseen datasets out of the box. This work was accepted to ICML 2026.

Key takeaway

For Machine Learning Engineers and Data Scientists aiming to deploy foundation models on complex relational databases, this work means you can extend existing single-table in-context learning models to multi-table RDBs without costly retraining. You should consider integrating the open-source RDBLearn foundation model, which provides scalable SQL primitives for encoder stages and delivers robust, out-of-the-box predictive performance on unseen datasets, significantly reducing development time and computational resources.

Key insights

A new RDB encoder enables existing single-table ICL foundation models to operate on multi-table RDBs without retraining.

Principles

• ICL compression for RDBs must be within high-dimensional columns with shared units.
• Encoder expressiveness is maintained without requiring trainable parameters.

Method

Compress variably-sized RDB neighborhoods into fixed-length ICL samples, constraining compression to high-dimensional columns with shared units and roles.

In practice

• Use RDBLearn for out-of-the-box RDB predictive modeling on unseen datasets.
• Integrate scalable SQL primitives for efficient RDB encoder implementation.

Topics

• Relational Databases
• Foundation Models
• In-Context Learning
• RDBLearn
• Predictive Modeling
• Multi-table Data

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Relational Foundation Models for Enterprise Data with Jure Leskovec - #768 — Relational Foundation Models enable accurate, no-training predictions on multi-table enterprise data by learning directly from graph-structured relationships.
• Context Payload Optimization for ICL-Based Tabular Foundation Models — Optimizing context payloads is crucial for balancing accuracy, latency, and cost in ICL tabular foundation models.
• Scaling seismic foundation models on AWS: Distributed training with Amazon SageMaker HyperPod and expanding context windows — Optimizing distributed training infrastructure with SageMaker HyperPod and direct S3 streaming dramatically accelerates large 3D Vision Transformer model training.
• Decoupled and Divergence-Conditioned Prompt for Multi-domain Dynamic Graph Foundation Models — DyGFM enables multi-domain dynamic graph foundation models by decoupling semantics and dynamics, and managing divergence.
• Why Foundation Models Haven’t Replaced Classical Machine Learning — Classical ML remains vital for proprietary, structured data, requiring context-aware agentic systems to navigate fragmented enterprise ecosystems.
• TRivia: Teach a 3B Model to Parse Tables — Labeled data scarcity and privacy concerns hinder open-source table recognition model performance against proprietary systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.