Physics-Informed Neural Networks (PINNs) offer a promising mesh-free alternative for solving boundary value problems in continuum mechanics. However, their application to the Navier-Cauchy equations of linear elasticity is often hindered by convergence issues, particularly in scenarios involving stress concentrations, singularities, or mixed Dirichlet-Neumann boundary conditions. While variational approaches, such as Robust Variational PINNs (RVPINNs), offer theoretical stability via dual norms, their significant computational overhead—primarily due to the inversion of the Gram matrix—limits their scalability in industrial simulations. This work proposes a hybrid training methodology designed to solve 2D linear elasticity problems with improved stability and computational efficiency. Our hybrid approach integrates two mechanisms: (1) Physics-Informed Sampling: A generator that leverages preliminary field analysis to automatically densify collocation points in critical zones—such as corners or load application points—where stress gradients are most pronounced. This extends recent advances in residual-based adaptive sampling to mechanics-specific constraints. (2) Adaptive Loss Balancing: An algorithm that dynamically re-weights the residuals of the equilibrium equations against boundary constraints, effectively mitigating the optimization stiffness inherent in mechanical simulations. We validate the proposed method on benchmark problems in solid mechanics, simulating stress distributions in 2D elastic bodies. Comparative analysis demonstrates that our approach achieves L2 error metrics competitive with RVPINNs while maintaining the low VRAM usage and faster training times of standard collocation PINNs. These results indicate that the framework can be extended to computationally demanding problems in material processing, such as elasto-plasticity.