How Tabular Foundation Models Solve LLM Blind Spots

KumoRFM makes predictions using relationships across database tables. Its research documentation describes a database as a graph in which rows are nodes and links between primary and foreign keys are edges. This lets a prediction use information from related records instead of requiring everything to be summarized into one flat table.

At prediction time, the model receives historical examples with known outcomes alongside the records to predict, much like examples in a few-shot prompt. In its zero-shot use, those examples define the task without task-specific training. The useful distinction is between learning from an isolated table and learning from the connected history surrounding an entity.