Accurate modeling of the constitutive behavior of metallic materials is a fundamental prerequisite for precise plastic forming and reliable prediction under complex loading conditions. To address the limitations of existing constitutive modeling approaches in terms of representation accuracy, experimental cost, theoretical assumptions, and extrapolation stability, this study proposes, for the first time, a hybrid modeling framework that integrates an explicit physics-based constitutive model with an implicit neural network model. The virtual fields method (VFM) is incorporated into the training process of the neural network. Full-field strain data covering multiple strain paths are obtained through a small number of uniaxial tensile tests on non-standard specimens. These data are used to identify the parameters of the explicit model via the VFM, and the virtual field principle is further embedded into the neural network training to achieve a complementary fusion between the explicit and data-driven models. The proposed framework enables stress prediction by leveraging the strengths of both model types. The results demonstrate that the approach allows the acquisition of diverse strain-path data from a single experiment, overcoming the limitation of conventional testing which typically captures only one strain path per test. This significantly reduces the number of experiments required to train the neural network model. The incorporation of the virtual field principle effectively compresses the neural network"s parameter space, improving training efficiency and reducing loss values. Moreover, the data-driven compensation component precisely captures the deficiencies of the baseline model in representing anisotropic yielding and nonlinear hardening behaviors, thereby enhancing the prediction accuracy and extrapolation stability. This research provides a novel methodology and technical foundation for high-fidelity modeling and prediction of material plasticity under complex loading paths.