NVIDIA KumoRFM Beats Supervised ML on Relational Data Zero-Shot

NVIDIA's KumoRFM is a relational foundation model that makes zero-shot predictions over multi-table relational data without task-specific training. According to the research documentation, it achieves 76.71 average AUROC zero-shot across 12 entity-classification tasks, surpassing LightGBM at 62.44, an LLM baseline at 68.06, and supervised relational deep learning at 75.83. Fine-tuning raises performance further to 81.14 AUROC. Zero-shot inference takes approximately one second and one line of PQL, compared to 30 minutes and 56 lines for a relational deep learning GNN pipeline, and 12.3 hours and 878 lines for a manual data-scientist workflow. The model works natively with declared schemas, relationships, entity tables, fact tables, context rows, and prediction rows without requiring callers to flatten connected tables into a single feature table. Benchmarks used approximately 100 tables and 2 billion rows across 15 graphs of real-world data, and the model was not trained or tuned on those benchmark datasets. During zero-shot inference, model parameters stay frozen while the new database, target definition, and historical labeled context are supplied dynamically. RelBench provides standardized evaluation infrastructure for relational ML, addressing earlier reproducibility problems caused by proprietary or ad-hoc data extracts. Databricks and NVIDIA separately announced an expanded collaboration allowing enterprises to fine-tune open models on proprietary data and deploy them to production within Databricks, using NVIDIA NeMo, NGC containers, and Nemotron cookbooks on serverless NVIDIA GPUs, with enterprise data governance maintained throughout.