NVIDIA Kumo Relational: Outperforming Supervised ML on Relational Data Without Training

NVIDIA Kumo Relational is a pre-trained relational foundation model designed to make predictive queries directly on enterprise databases. By gathering up to 10,000 historical examples as context, Kumo Relational predicts outcomes such as churn, demand, or fraud in a single forward pass, requiring no task-specific training, feature engineering, or table flattening.

Kumo Relational is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks. Across 41 tasks on 15 databases, it beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines. Furthermore, on the SAP SALT enterprise benchmark, it surpasses AutoGluon ensembles by 8% in mean reciprocal rank (MRR). The model scales to databases containing over 500 billion rows and is accessible through NVIDIA's open-source, GPU-native structured-data-models library or as an NVIDIA NIM.

The Shift to Structured Data Foundation Models

Structured data foundation models are driving a paradigm shift in how predictive modeling is performed on relational databases, mirroring the impact of large language models (LLMs) on natural language processing (NLP).

Traditionally, predictive modeling on relational databases required extensive feature engineering, task-specific training, and tuned pipelines. Structured data foundation models simplify this process by offering three key advantages:

Better Accuracy Today

As drop-in replacements, these foundation models already surpass many existing production pipelines, particularly XGBoost. On RelBench V1 classification tasks, Kumo Relational outperforms gradient-boosted trees using expert-engineered features by nearly 4 AUROC points on average. On the SAP SALT benchmark, it beats XGBoost on engineered features by 7 MRR points out of the box, and by 13 MRR points after fine-tuning. These results support the replacement of production XGBoost pipelines with Kumo Relational.

Simplicity and Agility

A single pre-trained model replaces multiple separate pipelines needed for individual tasks. This allows a new prediction to be framed as a query rather than a complex project, enabling ML teams to tackle use cases that were previously too labor-intensive.

Enabling Predictive Agents

Structured data foundation models allow AI agents to anticipate future events, such as predicting customer churn, and subsequently act on those predictions. Kumo Relational is built to be called by agents, using its Predictive Query Language (PQL) to provide a simple interface for asking predictive questions.

Kumo Relational Architecture and Operation

Unlike most structured data foundation models that require flattening real-world data into a single table, Kumo Relational operates directly on the relational database structure, respecting tables, foreign keys, and timestamps. The model is pre-trained on synthetic relational data and adapts to new tasks using in-context learning, which reads up to 10,000 context examples during inference.

Predictive Query Language (PQL)

Tasks are defined using the Predictive Query Language (PQL), a declarative interface that specifies what to predict rather than how to compute it. For temporal tasks, Kumo Relational replays historical database states. It samples entity-centric subgraphs up to an anchor timestamp and computes the target from events occurring after that timestamp, which prevents data leakage.

Multi-Stage Reasoning Process

To generate a prediction, Kumo Relational performs three distinct steps:

1. Automatic Context Generation: The system samples entity-centric subgraphs based on the PQL query and prepares the necessary historical context.
2. Reasoning Within Tables: A lightweight network processes each table independently. It alternates between column attention (relating values within a column) and row attention (relating features within a row), filtering noise to determine relevant columns and rows for the current prediction.
3. Reasoning Across Tables: A larger network distributes row embeddings across the relational structure using graph attention over primary-key/foreign-key relationships. This is interleaved with cross-sample attention across context examples, enabling relational reasoning, such as linking a user's profile to their transaction history. This design avoids quadratic cost and scales to large contexts and deeper subgraphs.

Training and Implementation

Kumo Relational is pre-trained on entirely synthetic relational data that simulates real-world databases and diverse predictive queries. The generation process builds upon the PluRel approach, which extends single-table data generators to full relational databases.

Synthetic Data Generation

PluRel generates each database in three stages: defining a schema, establishing foreign-key links that cluster like real data, and generating column values using Structural Causal Models that include temporal trends and cross-table dependencies. This lightweight generation allows for effectively unlimited pre-training data, and research shows that pre-training loss follows a power law as synthetic databases are added.

Deployment and Usage

Kumo Relational is available via two primary methods:

• Structured-Data-Models Library: The model can be loaded in memory through NVIDIA's GPU-native structured-data-models library, allowing direct prediction from a GPU-backed notebook.
• NVIDIA NIM: For deployment as a separate service, Kumo Relational can be deployed as an NVIDIA NIM, which is accessed via the open-source structured-data-models-client.

Performance and Benchmarks

Kumo Relational was evaluated on 41 benchmarks across 15 relational databases from RelBench V1 and V2, SAP SALT, and 4DBInfer, covering domains such as e-commerce, medical, and enterprise. None of these datasets were used in the pre-training phase.

Benchmark Results

• RelArena-α: On the RelArena-α leaderboard, Kumo Relational sits on the accuracy-efficiency Pareto front out of the box. It achieves an Elo of 1,842 with a median inference time of about 17 seconds per 1,000 samples. This makes it approximately 430 times faster than the highest-scoring entry (rt-plurel) at similar accuracy.
• SAP SALT: On the SAP SALT enterprise benchmark, Kumo Relational surpasses large AutoGluon ensembles by 8% in mean reciprocal rank (MRR) and recent tabular foundation models by about 25%. Fine-tuning the model extends its lead over AutoGluon to 16%.

Robustness

The model demonstrates strong robustness, performing well when provided as little as 0.2% of the training data as context. It remains stable even when up to 75% of foreign-key links are removed, and it shows minimal degradation when many noisy columns are added.

Limitations

Kumo Relational performs optimally on data that possesses meaningful relationships and temporal dynamics, such as tracking user behavior over time. For tasks involving extremely large training sets, task-specific supervised models may still achieve higher accuracy than in-context predictions, given that Kumo Relational's benchmarks used a maximum of 10,000 context examples. Furthermore, predictive signals must be reachable within a limited number of hops; users should start with a focused query over a few relevant tables and expand the graph as necessary.