This study proposes an improved adaptive kriging model-based FE model updating framework to increase the updated FE model’s accuracy and reduce computational costs. The proposed adaptive kriging model uses an efficient hybrid expected improvement and minimizing surrogate prediction sampling strategy to generate new training samples. The k-fold cross-validation is introduced to assess the kriging model’s accuracy. Benchmark functions and model updating of a full-scale bridge structure validated the proposed adaptive kriging model. The results demonstrate that the hybrid sampling strategy balances the global and local exploration of the design space and provides reasonable training samples, increasing the accuracy of the adaptive kriging model. The cross-validation significantly saves computational costs. After model updating, the frequency relative errors of the continuous bridge are under 2%, and the modal assurance criterion values are over 0.96. The updated FE model can serve as the baseline FE model of the bridge.