Complex material models with internal state variables are essential for capturing inelastic deformation in solid mechanics, yet their numerical integration within finite element simulations remains computationally demanding, especially in parametric studies and large-scale analyses. This work introduces machine-learning-based surrogate formulations for elastoplastic constitutive modeling implemented in the JAX computational framework. We investigate two complementary and conceptually extreme learning strategies. In the first, full physical knowledge of the constitutive model and its internal variables is assumed to be available, and deep learning is employed to accelerate classical integration schemes through physics-informed training [1]. In the second, the internal state variables are treated as latent and inaccessible, and neural operators are trained in a purely data-driven manner to learn the constitutive response directly from strain–stress histories [2]. The proposed approaches are validated on benchmark constitutive problems and integrated into prototype finite element implementations. Comparative studies assess accuracy, robustness, and computational efficiency relative to classical integration algorithms. The results demonstrate that the learning-based surrogates accurately reproduce reference material responses while providing substantial computational acceleration, highlighting their potential for large-scale simulations, parametric studies, and stochastic material modeling.