<p>This study evaluates the impact of experimental design strategies on surrogate-assisted Bayesian FEM updating using stochastic modal parameters. A unified Bayesian FEM updating framework is developed, focusing on structural parameters inferred using modal frequencies and mode shapes treated as stochastic quantities. Modal data is obtained from repeated ambient and earthquake recordings and consolidated using a three-layer stochastic framework. The use of consensus modal parameters derived from multiple measurement runs enables a robust and uncertainty-aware representation of modal characteristics, which enhances the stability and reliability of the Bayesian updating process. Two experimental design strategies are implemented: a space-filling Latin Hypercube Sampling (LHS)-based design and an adaptive Expected Improvement (EI)-driven sampling strategy that iteratively refines the surrogate model. The results show that EI-based adaptive sampling concentrates training points in regions associated with low model-data mismatch and high posterior probability, leading to improved surrogate accuracy within the posterior region, more accurate posterior frequency predictions, and tighter posterior parameter distributions with fewer finite element model evaluations. The study demonstrates that posterior-aware adaptive experimental design significantly improves the efficiency and reliability of surrogate-assisted Bayesian FEM updating, outperforming conventional space-filling designs when stochastic modal data are employed. The findings emphasize the importance of local surrogate accuracy within the posterior region and provide guidance for uncertainty-aware structural model calibration and future reliability-oriented applications.</p>