In practice, multi-objective optimization problems are constructed via computer simulations, and function evaluation is quite costly. It is a popular method for training surrogate models based on historical data to help multi-objective optimization algorithms solve computationally expensive multi-objective problems. However, when the number of optimized objectives increases, it will become more difficult to choose individuals to enhance model quality in model management. To this end, this paper proposes three sampling criteria containing the linear combination confidence lower bound function, the adaptive performance balance criteria, and the scalar deviation matrix to select several individuals to infill the sample set to improve the efficiency of the model-guided multi-objective optimization algorithm for finding the optimal solution from three aspects: the balance between multiple objective estimations and their reliability, the adaptive balance of convergence and diversity of solutions, and current sample distribution, respectively. Meanwhile, non-dominated samples are adopted to drive the current population search, which speeds up searching for the most promising region. Finally, compared with five advanced algorithms, the proposed algorithm is effective on two classical multi-objective test suites and two optimization instances.