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

NVIDIA Kumo Relational is a pre-trained foundation model that makes predictions directly on relational databases without requiring task-specific training or feature engineering. It functions as a few-shot model, using up to 10,000 historical examples as in-context learning to predict outcomes like churn, demand, or fraud in a single forward pass. The model has been shown to outperform supervised machine learning methods, including gradient-boosted trees and AutoGluon ensembles, on common relational data benchmarks. Kumo Relational operates directly on database schemas, avoiding the need to flatten tables, and is designed to be used by both data scientists and autonomous agents.