NVIDIA Unveils Untrained Predictions with Kumo Relationnel

NVIDIA details Kumo Relationnel, a pre-trained foundation model for relational databases that promises predictions without specific training. The company highlights measured gains on RelBench and SAP SALT, large-scale inference capability, and stability under constraints.

Measured Gaps on SAP SALT, RelBench, and RelArena-α

NVIDIA reports that Kumo Relationnel surpasses established benchmarks. On SAP SALT, the company announces an 8% advantage in MRR over AutoGluon sets, which extends to 16% after adjustment. On the same benchmark, the claimed gap reaches approximately 25% against recent tabular foundation models. On RelBench V1, Kumo Relationnel would exceed expert-optimized gradient trees by about 4 AUROC points on average. NVIDIA also indicates that the model outperforms XGBoost by 7 MRR points in its initial configuration and by 13 MRR points after adjustment on SAP SALT. Overall, the evaluation covers 41 benchmarks across 15 relational databases, including those from RelBench V1 and V2, SAP SALT, and 4DBInfer, which were not seen during pre-training. On RelArena-α, the model is presented as positioned on the Pareto front of precision-efficiency from initialization.

Declared Robustness with 0.2% of Data and 75% of Links Removed

According to NVIDIA, Kumo Relationnel maintains its functionality with only 0.2% of the training data used as context. The company also claims stability when up to 75% of foreign key links are removed. The model is further announced as scalable to databases exceeding 500 billion rows.

Usage Presented Without Training or Feature Engineering

The highlighted usage flow involves connecting a relational database, describing the target, and obtaining a prediction in one pass, without specific training or feature engineering or flattening. Predictions concern issues such as churn, demand, or fraud, with up to 10,000 historical examples taken into context at inference. NVIDIA offers Kumo Relationnel via a native GPU open-source library and in the form of NVIDIA NIM. The model is designed to be invoked by agents through a declarative predictive query language, PQL.

Positioning Against Existing Pipelines Like XGBoost

NVIDIA asserts that Kumo Relationnel, as a foundation model for structured data, claims better performance than XGBoost-based pipelines on the cited benchmarks while simplifying production deployment. The argument draws parallels to the evolution observed in NLP, where LLMs have rendered several pipeline steps obsolete, suggesting a similar movement in relational data. The company presents Kumo Relationnel as a few-shot model outperforming supervised learning on 41 tasks covering 15 databases, including relational deep models and expert-designed sets.

Direct Processing of Relational Graphs and Leak Control

Kumo Relationnel operates directly on tables, foreign keys, and timestamps. Tasks are formulated in PQL, which describes the target without specifying the method. For temporal scenarios, the system replays past states by centering subgraphs on an entity up to an anchor timestamp, calculates the target on subsequent events, and shows the model only the prior data. NVIDIA presents this mechanism as a prevention of data leaks. Context examples associate a subgraph with a ground truth target, which can range from user churn to a quantity sold, such as 142 units for an item.

Attention-Based Architecture and Synthetic Pre-Training PluRel

The architecture combines a lightweight network processing each table with alternating attention on columns and rows, and a larger network propagating row embeddings through primary key and foreign key relationships, coupled with inter-sample attention. NVIDIA describes relational reasoning linking, for example, a user profile and its transaction history. The use of row embeddings aims to avoid quadratic costs and allow for extended contexts as well as deeper subgraphs. The model follows a multi-phase pre-training on synthetic data designed to replicate the structure of real databases. Due to insufficient public diversity, the approach relies on PluRel, to which team members have contributed, to generate schemas, structuring foreign key links, and column values derived from Structural Causal Models, with temporal trends and dependencies. NVIDIA indicates that the generation is lightweight and that the pre-training loss follows a power law as synthetic databases are added, with gains transferred to real databases. When aiming for maximum accuracy, NVIDIA anticipates the possibility of fine-tuning the model on a given dataset and task.