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

Open original

NVIDIA's Kumo Relational is a pre-trained model that makes predictions like churn, demand, or fraud directly on relational databases, with no per-task training or hand-built features. It was tested on 41 tasks across 15 databases and reportedly beats supervised models and tuned gradient-boosted pipelines.

  • Across 41 benchmark 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 beats AutoGluon ensembles by 8% in MRR, rising to 16% after fine-tuning.
  • It works on the raw tables, keys, and timestamps instead of flattened features, and it scales to databases with 500B+ rows.
  • Its main weak spots: it's strongest where relationships and temporal behavior matter, and large supervised training sets can still win.

NVIDIA's Kumo Relational is a pre-trained foundation model that makes predictions directly on relational databases, with no task-specific training, no feature engineering, and no flattening of tables into a single row per entity. You connect a database, write a predictive query in NVIDIA's Predictive Query Language (PQL) saying what you want to predict (churn, demand, fraud), and the model pulls historical examples as context and returns a prediction in one forward pass. The model was pre-trained once on synthetic relational data generated with a method called PluRel, which builds schemas, foreign-key links, and column values with temporal trends and cross-table dependencies. The authors say this synthetic data is effectively unlimited, which is why they can keep scaling pre-training. The headline numbers come from 41 benchmarks across 15 databases drawn from RelBench V1 and V2, SAP SALT, and 4DBInfer. None of those datasets were in pre-training. Against expert-engineered gradient-boosted tree pipelines, the article reports nearly 4 AUROC points better on average for RelBench V1 classification. On SAP SALT, it beats XGBoost on engineered features by 7 MRR points out of the box and 13 after fine-tuning, and it surpasses AutoGluon ensembles by 8% in MRR. The model is also described as the top in-context learning model on the RelBench leaderboard, and as sitting on the accuracy-efficiency Pareto front on RelArena-α. Treat these as the authors' own claims. They come from the paper linked at arXiv 2604.12596, which I haven't checked, and vendor benchmarks deserve a skeptical read until independent groups reproduce them. Architecturally, the design splits the work in two. A lightweight network handles each table on its own using column and row attention, deciding which fields matter before anything propagates. A larger network then runs graph attention over primary-key and foreign-key links, followed by attention across the context examples. Because that second stage operates on row embeddings rather than every cell, the authors argue it avoids quadratic cost and handles large contexts and deep subgraphs. For temporal tasks, the system replays historical database states, serving the model only data from before an anchor timestamp, which is meant to prevent leakage. The release covers open-source code (structured-data-models and kumo-engine), weights on Hugging Face, a client library, and an NVIDIA NIM, and the authors claim it scales to databases with more than 500 billion rows. Whether it displaces a production XGBoost pipeline will depend on how well those benchmark gains hold on your own schema, which is the test worth running before anything else.

Ask about this article