Tabular foundation models that learn from their own mistakes

A researcher at Leaflet has published a paper showing that foundation models trained on tables can repair themselves. The trick: after the model makes predictions, you compare its output against the original table, flag the rows it got wrong, and feed those examples back into training. The model gets better at its own mistakes — no human labelers needed.

This is a small but meaningful step. Tabular data — spreadsheets, databases, CSV files — is the bread and butter of how businesses actually work. It's not the flashy stuff. But it's the stuff that keeps the lights on. And for years, the big models have been mediocre at it. They hallucinate rows, misread headers, can't do basic joins. This paper suggests a practical path toward fixing that.

The broader picture: if models can self-improve on tabular data at scale, the bottleneck shifts from data labeling to data quality. Bad tables make bad models. Clean tables make better ones. And the models start helping you clean them.

Why this matters for us: a lot of the small businesses, bodegas, and family operations we know run on Excel and Google Sheets — if models get good at that stuff, the tools built on top of it stop being expensive and start being useful.