---
title: "Kumo's KumoRFM-2 Beats Supervised Models on 41 Database Benchmarks Without Training"
subtitle: "A pre-trained foundation model reads directly from multi-table SQL databases and beats supervised graph models on 41 predictive tasks without task-specific training."
url: "https://alphasignal.ai/news/kumo-s-kumorfm-2-beats-supervised-models-on-41-database-benchmarks-without"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-14T00:00:00.000Z"
updated: "2026-08-23T23:34:09.137Z"
type: "Paper"
topics:
  - "Llms"
  - "Data"
subtopics:
  - "Small Models"
  - "Synthetic Data"
upvotes: 2384
source_url: "https://arxiv.org/abs/2604.12596"
comments: 0
truncated: true
---

# Kumo's KumoRFM-2 Beats Supervised Models on 41 Database Benchmarks Without Training

> A pre-trained foundation model reads directly from multi-table SQL databases and beats supervised graph models on 41 predictive tasks without task-specific training.

Foundation models have swallowed text, images, and audio, but the messiest and arguably most valuable data in the enterprise, the relational database, has stubbornly resisted the same treatment. A team from Kumo, Stanford, and collaborators just released [KumoRFM-2](https://arxiv.org/abs/2604.12596), a pre-trained model that operates directly on connected SQL tables and, for the first time, beats carefully tuned task-specific supervised models on standard benchmarks without any task-specific training.

Across 41 challenging benchmarks the model outperforms supervised and foundational approaches by up to 8%, the first time a few-shot foundation model has been shown to surpass supervised approaches on common benchmark tasks, with performance further improving under fine-tuning. The practical upshot: churn, fraud, LTV, and demand-forecasting models that used to take a data scientist days of feature engineering can now be answered in seconds against a live warehouse.

## Why flat tables were holding foundation models back

Tabular foundation models like TabPFN have shown that a pre-trained transformer can solve small classification tasks in a single forward pass. The problem is that real business data almost never lives in a single table. Users link to orders, orders link to items, items link to categories, and the predictive signal is spread across those joins along with timestamps that must not leak future information into training examples.

The dominant workaround has been propositionalization: flatten the graph into one wide table via hand-crafted features or automated schemes like Deep Feature Synthesis. This introduces an expressivity gap, since task-agnostic aggregations fail to capture complex, task-specific relational structures or interactions across rows. The paper illustrates the gap with a toy example where the label depends on two binary features co-occurring in the same child row. Column-wise encoders collapse to AUROC 0.5 (random guessing) while KumoRFM-2 hits 1.0.

## A database-native architecture

KumoRFM-2 treats a relational database as a temporal heterogeneous graph where rows are nodes and primary-foreign key relationships are edges, each stamped with a time so that historical snapshots can be replayed without leakage. Compared to v1, the new model performs in-context learning over two separate attention stages, one for intra-table processing and one for aggregating information across tables. This design enables task-conditioned feature extraction at both the table and database level, improving predictive performance.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [KumoRFM-2: Scaling Foundation Models for Relational Learning](https://arxiv.org/abs/2604.12596)
- [KumoRFM-2 HTML paper](https://arxiv.org/html/2604.12596v1)
- [Introducing KumoRFM: A Foundation Model for In-Context Learning on Relational Data](https://kumo.ai/company/news/kumo-relational-foundation-model/)
- [kumo-rfm GitHub repository](https://github.com/kumo-ai/kumo-rfm)
- [Predictive Query Language paper](https://arxiv.org/abs/2602.09572)