This paper considers the training of logical (Boolean) capabilities with concentrate on the generalization on the unseen (GOTU) setting, a powerful case of out-of-distribution generalization. That is motivated by the truth that the wealthy combinatorial nature of information in sure reasoning duties (e.g., arithmetic/logic) makes consultant information sampling difficult, and studying efficiently underneath GOTU offers a primary vignette of an ‘extrapolating’ or ‘reasoning’ learner. We then research how completely different community architectures skilled by (S)GD carry out underneath GOTU and supply each theoretical and experimental proof that for a category of community fashions together with cases of Transformers, random options fashions, and diagonal linear networks, a min-degree-interpolator (MDI) is discovered on the unseen. We additionally present proof that different cases with bigger studying charges or mean-field networks attain leaky MDIs. These findings result in two implications: (1) we offer an evidence to the size generalization drawback (e.g., Anil et al. 2022); (2) we introduce a curriculum studying algorithm referred to as Diploma-Curriculum that learns monomials extra effectively by incrementing helps.