NVIDIA Kumo Relational claims first few-shot foundation model to outperform supervised ML on relational benchmarks

NVIDIA introduced Kumo Relational, a pre-trained relational foundation model designed to make predictions directly on enterprise databases without task-specific training or feature engineering. Given a predictive query, the model gathers up to 10,000 historical examples as context and predicts outcomes such as churn, demand, or fraud in a single forward pass.

Kumo Relational is positioned as the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks. Across 41 tasks on 15 databases, it outperforms supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines. On the SAP SALT enterprise benchmark, it surpasses AutoGluon ensembles by 8% in mean reciprocal rank (MRR).

The model scales to databases with 500 billion or more rows and is accessible to both humans and agents through NVIDIA’s open-source, GPU-native structured-data-models library or as an NVIDIA NIM. Kumo Relational operates directly on relational databases, respecting tables, foreign keys, and timestamps, and adapts to new tasks through in-context learning by reading up to 10,000 context examples at inference time.

Each context example pairs a local subgraph around an entity at a point in time with its ground-truth target, such as whether a user churned or an item sold a specific number of units. The system uses a Predictive Query Language (PQL) to turn natural language queries into predictive tasks, and automatically generates context by replaying historical database states to prevent data leakage.

Reasoning within tables is handled by a lightweight network that alternates column and row attention to filter noise, while reasoning across tables uses a larger network with graph attention over primary-key/foreign-key relationships and cross-sample attention across context examples. This design avoids quadratic cost and scales to large contexts and deeper subgraphs.

Kumo Relational was pre-trained on entirely synthetic relational data designed to reflect real-world databases with diverse predictive queries, with pre-training proceeding in stages from single tables to increasingly complex relational structures.