Deep learning-based surrogate models are increasingly adopted to approximate and replace high-fidelity numerical solutions in computational mechanics. Among these, generative latent-variable models, such as variational autoencoders, are often adopted to learn low-dimensional representations and probabilistic formulations. However, the challenges in applying such variational generative models to eigenvalue problems in mechanics are still open. In this work, the behavior of variational generative surrogates for modal analysis of micro electromechanical systems (MEMS) structures is examined. Dynamic analyses are considered to match performance indicators linked to sense and drive frequencies, to also avoid or reduce to a minimum cross-talk and internal resonance issues. High-fidelity finite element simulations are used to compute eigenfrequencies and relative eigenmodes, handled exclusively as high-resolution images. No explicit low-dimensional geometric parameterization is therefore considered. A conditional variational autoencoder is allegedly considered to be able to learn a latent representation linking geometry, mode shapes, and frequencies. Despite standard stabilization strategies, the model consistently exhibits posterior collapse, with the latent variables contributing marginally to the reconstruction. This behavior is examined by considering the physical relationship between geometry and modal response. In the considered setting, characterized by consistent mode ordering and the absence of degeneracies or explicitly modeled variability, the geometry-to-modal-response relationship is strongly constrained and effectively deterministic. In this case, the response is nearly deterministic, making stochastic latent representations difficult to identify and of limited practical use, having disregarded any variability in the structural geometry due to microfabrication imperfection and uncertainty. The obtained results motivate a theory-guided surrogate design perspective, in which generative modeling is more naturally placed in the geometry or design space, while the physics response for given geometries is captured by deterministic, structure-aware predictors. This provides guidance on when generative surrogates are likely to be beneficial and when deterministic formulations are more consistent with the governing mechanics.