To solve the sampling problem in multi-response modeling and support the efficient and accurate establishment of multiple agent models, a Bayesian support vector machine-based modified multi-response modified expected improvement for global fit (MR-MEIGF) sampling criterion is proposed. First, the gradient of candidate samples is calculated by the Bayesian support vector machine model, and the neighborhood is constructed to obtain the local exploitation criterion based on the projection of the gradient of the neighborhood. The predicted variance of the sample points obtained by the model is used as the global exploration criterion, and the two are combined to obtain the hybrid sampling criterion for individual responses. And then, through the local index, the importance of each response is quantified, and the MR-MEIGF sampling criterion that takes into account the accuracy of multiple response models is further obtained so as to realize the comprehensive optimization of multiple responses. Based on the MR-MEIGF criterion, the newly added sample points are selected in the candidate pool. Three 2-dimensional and three 6-dimensional cases are combined to form a multi-response problem, and compared with the sequential space-filling method, the traditional one-time space-filling method as well as other multi-response adaptive sampling methods, to validate the effectiveness of the proposed sampling method, and compare the performance of the Bayesian support vector machine model with the Kriging model on the 6-dimensional case.