NVIDIA's Kumo Relational model beats supervised machine learning benchmarks on relational data without any task-specific training. The system works directly on raw relational tables, no feature engineering, no labeled examples, no fine-tuning pipeline required.
The architecture treats relational databases as graph structures, propagating signals across foreign key relationships to generate predictions. This matters because most enterprise data lives in relational schemas, and the standard workflow of flattening tables into feature vectors for gradient boosting or neural nets destroys structural information that Kumo preserves natively.
The full post details the benchmark comparisons, the specific datasets used, and how Kumo handles temporal leakage, a problem that silently corrupts most off-the-shelf relational ML pipelines. Read it for the methodology, not just the headline number.
[READ ORIGINAL →]