Relational foundation models are making feature engineering obsolete before most data teams noticed
A class of models is now making predictions directly from raw relational databases, with no training loop, no hyperparameter tuning, and no feature engineering. Published research confirms the capability is real, and its implications for applied machine learning are significant.
Jure Leskovec’s description of what a Relational Foundation Model requires from its user is short enough to be surprising. No training, no hyperparameter optimization, no feature engineering. A raw relational database and a task specification are sufficient to generate predictions. That is not a description of a faster or more convenient version of the existing machine learning workflow. It is a description of a different workflow entirely.
The claim is not speculative. Research published on arXiv, including work associated with Kumo.ai’s KumoRFM project and corroborated by material from Google Research, confirms that relational foundation models can perform in-context learning directly on raw relational data. The approach requires no fine-tuning step and no hand-crafted feature pipelines. The model receives the database structure and a specification of the prediction task, and it generalizes from there.
This matters because the feature engineering step Leskovec describes as unnecessary has historically been one of the most expensive parts of applied machine learning. Data scientists working on relational prediction problems, the kind that arise routinely in commerce, healthcare, and logistics, spend substantial time deciding which columns to use, how to aggregate across tables, and which transformations to apply before a model ever sees the data. That work requires domain knowledge, iteration, and headcount. A model that sidesteps it entirely changes the calculus of who can deploy predictive systems and how quickly.
No training is necessary no hyperparameter optimization is necessary no feature engineering is necessary um all you need is a raw data database and a way to specify Jure Leskovec
The technical mechanism behind the claim is in-context learning applied to relational structure. Rather than being trained on a specific dataset for a specific task, the foundation model has been pre-trained on a broad distribution of relational data and can handle a new database at inference time, reading the schema and the task specification as context. This mirrors what large language models do with text, but the input is structured relational data rather than natural language. The arXiv preprint describing the approach makes clear that the zero-shot generalization is genuine: the model is not simply retrieving a memorized solution for a known schema.
The practical boundary conditions on this capability are still being established. Relational databases vary enormously in size, complexity, and cleanliness, and no published benchmark covers the full range of real-world conditions a production deployment would encounter. Leskovec’s framing, that all that is needed is raw data and a way to specify the task, captures the user-facing simplicity of the interface without resolving every question about where the model’s performance degrades. Those limits will be documented over time, as external research groups and practitioners test the approach against their own data.
What the external record does confirm, clearly enough to report, is that the core claim holds. Zero-shot prediction on relational databases, without a training run and without feature engineering, is now a documented capability rather than a roadmap item. The research community has reached the same conclusion through independent paths, which is the kind of corroboration that moves something from interesting assertion to credible finding.
The broader significance is that the data preparation and modeling work that currently employs entire teams of specialists is being compressed in the same direction that AI has compressed other forms of technical labor. Leskovec’s description of the interface, raw data in, predictions out, is simple enough to sound like an overstatement. The published research suggests it is not. Whether enterprise data teams treat that compression as a threat to existing workflows or as a tool that expands what a smaller team can attempt, the capability is already in the literature. The adaptation decisions will follow.