Iterative labeling have heavy assumptions about the primitive is using.

[dmartinez05]

This primitive relies on the that the primitive has a stored fitted version inside that contains the predict_probamethod.
So far, the only primitives that are aligned with these expectations are the Sklearn primitives, other ones such as XGboost are not:

https://gitlab.com/datadrivendiscovery/common-primitives/-/issues/154
https://gitlab.com/datadrivendiscovery/common-primitives/-/issues/155

I think this is the main reason:
https://github.com/autonlab/autonbox/blob/master/autonbox/iterative_labeling.py#L118