This paper deals with the problem of on-line model identification of multivariable processes with nonlinear and time-varying dynamic characteristics. In this respect, two adaptive learning approaches for multi-input, multi-output (MIMO) radial basis function (RBF) neural networks, i.e. growing and pruning algorithm for radial basis function (GAP-RBF) and minimal recourse allocation network (MRAN) are employed to identify MIMO time-varying nonlinear systems. The unscented Kalman filter (UKF) is proposed as a new learning algorithm for both GAP-RBF and MRAN approaches. Some desired modifications on the growing and pruning criteria in the original GAP-RBF have been proposed to make it more adequate in online identification. The performances of the algorithms are evaluated on a highly nonlinear and time-varying CSTR benchmark problem. Simulation results demonstrate the better performance of the modified GAP-RBF (MGAP-RBF) neural network with respect to the original GAP-RBF and MRAN algorithms.