NVIDIA Kumo Relational Outperforms Supervised ML on Relational Data Without Task-Specific Training

Details

NVIDIA released Kumo Relational, a pre-trained relational foundation model that performs direct predictions on enterprise databases without task-specific training, feature engineering, or table flattening.
The model outperforms supervised relational deep learning models and expert-designed gradient boosting pipelines across 15 databases and 41 tasks, achieving an 8% average MRR advantage over the AutoGluon ensemble on the SAP SALT benchmark.
Inference uses up to 10,000 historical examples as context in a single forward pass, respecting table structures, foreign keys, and timestamps while preventing data leakage via time-based context filtering.
The model supports databases with over 500 billion rows, processing data at up to 5 GB/s via memory-mapped I/O, and is available via the open-source structured-data-models library or NVIDIA NIM.
Performance details include a 4-point average AUROC advantage over gradient boosting trees on RelBench V1 classification tasks and a 25% MRR advantage over the latest table foundation models on SAP SALT.
The system is robust to noise, maintaining stable performance even when 75% of foreign key links are removed or noisy columns are added, though it may be slightly outperformed by task-specific supervised models on tasks with very large training sets.

Key point