NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

NVIDIA Kumo Relational is a pre-trained foundation model designed to make predictions directly on enterprise relational databases. When given a predictive query, it retrieves up to 10,000 historical examples as context and produces outcomes such as churn, demand, or fraud in a single forward pass. The model eliminates the need for task-specific training, feature engineering, or flattening database tables, streamlining the prediction process for structured data.

Kumo Relational is the first few-shot foundation model to surpass supervised machine learning on common relational benchmarks. Across 41 tasks spanning 15 databases, it outperforms supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines. On the SAP SALT enterprise benchmark, it exceeds AutoGluon ensembles by 8% in mean reciprocal rank (MRR), demonstrating superior accuracy without requiring fine-tuning.

The model operates directly on relational databases, preserving tables, foreign keys, and timestamps without manual aggregation. It adapts to new tasks through in-context learning, reading up to 10,000 context examples at inference time. Each example pairs a local subgraph around an entity with its ground-truth target, enabling the model to understand values within tables, relate records across tables, and connect context examples to the entity being predicted.

Kumo Relational is pre-trained on synthetic relational data reflecting real-world datasets and can be fine-tuned for maximum accuracy. It evaluates on 41 benchmarks across 15 databases from RelBench V1 and V2, SAP SALT, and 4DBInfer, spanning domains like e-commerce, medical, and enterprise. The model outperforms supervised models in both classification and regression, including relational deep learning and expert-engineered ensembles.