🧵 6. One interesting thing Synthefy shows here is that Nori does not have to remain a black box: they can probe its own predictions, isolate how much each feature contributes, and rebuild that behavior as a much more transparent additive model.

In their credit-default example, “just 6 of 23 features recover about 95% of Nori’s full-model accuracy”, while also showing exactly which inputs push the prediction up or down.

That is a pretty important result for tabular foundation models, because you can use a powerful pretrained model to discover the signal, then turn what it learned into a smaller model that a person can actually inspect feature by feature.

For areas like credit, insurance or healthcare, that could make the difference between getting a strong prediction and getting a strong prediction you can actually explain and defend.

Check out this official blog -

https://t.co/OWZJZuLvPu