Advanced anisotropic yield criteria such as the Cazacu-Plunkett-Barlat (CPB’06) [1] model accurately describe sheet metal plasticity but pose significant challenges for parameter identification due to their nonlinear structure and high dimensionality. While machine learning approaches, including tree-based models (e.g. XGBoost) [2] and neural networks, show excellent performance for simpler criteria such as Hill’48, their generalization capability deteriorates markedly when applied to CPB’06, even with large synthetic datasets. This work compares the performance of XGBoost and neural network models for the machine learning–based calibration of the CPB’06 yield criterion, focusing on strategies that go beyond increasing data volume. Numerically generated data from biaxial tensile tests on cruciform specimens are used to analyze the impact of constitutive complexity on learning performance, with emphasis on parameter sensitivity and training strategy design. Particular attention is given to the role of weakly influential parameters and to aligning learning formulations with the underlying constitutive mechanics through physics-informed constraints. Preliminary results indicate that improved consistency between constitutive modeling and learning strategy can mitigate overfitting and enhance generalization for both learning approaches. The findings suggest that CPB’06 parameter identification is primarily a modeling challenge rather than a data-availability issue.