Reduced Order Models (ROMs) are computationally efficient representations of dynamical systems that leverage first principles formulations for response prediction under environmental and operational variability. As such, they serve as critical enablers of digital twinning and higher-level frameworks across a wide range of applications, including decision-making, reliability analysis, and infrastructure management. However, ROMs inevitably encounter scenarios that lie beyond their training envelope, such as extreme loading events, progressive damage, or abrupt changes in boundary conditions. While offline retraining can address such model discrepancies, many decision-making, structural monitoring, and control applications require real-time engineering support. In this work, we addresses a critical gap in response prediction using reduced order models: the ability to maintain predictive fidelity when confronting changing conditions and unexpected phenomena that characterize real-world operation of infrastructure systems. In doing so, our work explores test-time correction strategies of pre-trained ROMs leveraging the online integration of sensing measurements. Specifically, the mismatch between observed system response and ROM predictions will be projected into the model’s latent space and captured through Gaussian Process Regression (GPR). The GPR surrogate serves to account for this discrepancy in a compact, informative representation space. Via further casting the correction problem as a recursive hyperparameter learning task, online technique like Kalman filtering can be utilized, thereby enabling fast and responsive model correction. This approach preserves the computational efficiency and interpretability of the underlying ROM while accommodating unforeseen dynamics through online responsive corrections.