Abstract: The generating test study is a research in constrained multi-object optimization. It is one of the key technologies in examination management systems. It relates directly to the efficiency and quality of the generating test paper. The test paper generation method suggested in this paper is based on hierarchical adaptive genetic algorithm. It is able to solve the problem of premature convergence or slow convergence in global optimization. On the one hand, the M subpopulations operate on adaptive genetic algorithm and save the intermediate result; then they operate with the top adaptive genetic algorithm until a satisfied paper is found. On the other hand, the minimum weighted mean square error model is used to establish the objective function and to inspect the error between the expectations and the actual value on types, knowledge topics, difficulty and the degree of differentiation of the test paper. It discusses also the error of the answer time, total score and luminosity. It improved the speed of generating test paper of the system. It avoided the problem of premature convergence which often appears in standard genetic algorithm. The high quality of paper generation and the good robustness generated in this algorithm can meet the practical needs of users.