We know that there exist predicates which are easier to evaluate than generate examples of (e.g. RSA-2048), such that, given that the model is actively looking for such predicates, it will be able to use them to make evaluation easier than generation.
We know that there exist predicates which are easier to evaluate than generate examples of (e.g. RSA-2048), such that, given that the model is actively looking for such predicates, it will be able to use them to make evaluation easier than generation.