Deep generative models have been widely used in molecular generation tasks because they can save time and cost in drug development compared with traditional methods. Previous studies based on generative adversarial network (GAN) models typically employ reinforcement learning (RL) to constrain chemical properties, resulting in efficient and novel molecules. However, such models have poor performance in generating molecules due to instability in training. Therefore, quantitative evaluation of existing molecular generation models, especially GAN models, is necessary. This study aims to evaluate the performance of discrete GAN models using RL in molecular generation tasks and explore the impact of different factors on model performance. Through evaluation experiments on QM9 and ZINC datasets, the results show that noise sampling distributions, training epochs, and training data volumes can affect the performance of molecular generation. Finally, we provide strategies for stable training and improved performance for GAN models.