The Monte Carlo code (CEARPGA I) was developed to generate the elemental library spectra required for implementing the Monte Carlo Library Least-Squares algorithm for prompt gamma-ray neutron activation analysis (PGNAA). The existing big weight problem in which a few histories yield very large weights with very large variance has been investigated thoroughly. It has been found that the expected value splitting technique, a powerful variance reduction technique used in the code is the primary cause of this problem. Two Monte Carlo simulation approaches have been investigated to eliminate the big weight problem while still maintaining high efficiency. They are 1) score importance map with batch tracking and 2) analog linear interpolation. Both approaches were implemented separately in the corresponding Monte Carlo codes on the basis of CEARPGA I code and demonstrated to be feasible for solving the big weight problem. The analog linear interpolation approach was selected to incorporate into the new CEARPGA Monte Carlo code (CEARPGA II). A comparison of the results obtained by both CEAEPGA I and CEARPGA II codes as well as experimentally measured data shows that the big weight problem has been successfully eliminated, the accuracy of the simulation has improved greatly, and the simulated results agree very well with the measured data. In addition, some other important improvements have been made in the CEARPGA II to make it more accurate, efficient and powerful, including: 1) tracking pair production gamma rays, 2) incorporating neutron activation background spectra, 3) generating natural background spectra, 4) taking into account the non-linearity of NaI detectors, and 5) adopting a general geometry package.