Abstract: Machine-learned interatomic potentials (MLIPs) are typically trained on datasets that cover only a subset of potential input structures, posing challenges for their generalization to a wider range of systems beyond the training set. However, MLIPs have shown remarkable accuracy in predicting forces and energies in simulations involving complex structures. In this talk, our goal is to explain the good generalization properties observed in MLIPs. We undertake a thorough theoretical and numerical exploration of MLIP generalization in the realm of crystalline defect simulations. We precisely quantify how simulation accuracy is directly influenced by key factors such as the size of training structures, the selection of training observations (e.g., energies, forces, virials), and the level of accuracy achieved during the fitting process. The numerical experiments not only validate existing best practices in MLIP literature but also offer novel insights into dataset design and loss function optimization.