Existing ensemble-learning methods for reliability analysis are usually developed by combining ensemble-learning with a learning function. A commonly used strategy is to construct the initial training set and the test set in advance. The training set is used to train the initial ensemble model, while the test set is adopted to allocate weight factors and check the convergence criterion. Reliability analysis focuses more on the local prediction accuracy near the limit state surface than the global prediction accuracy in the entire space. However, samples in the initial training set and the test set are generally randomly generated, which will result in the learning function failing to find the real “best” update samples and the allocation of weight factors may be suboptimal or even unreasonable. These two points have a detrimental impact on the overall performance of the ensemble model. Thus, we propose a general hierarchical ensemble-learning framework (ELF) for reliability analysis, which consists of two-layer models and three different phases. A novel method called CESM-ELF is proposed by embedding the classical ensemble of surrogate models (CESM) in the proposed ELF. Four examples are investigated to show that CESM-ELF outperforms CESM in prediction accuracy and is more efficient in some cases.