Computer-aided synthesis planning (CASP) has been helping chemists to synthesize novel molecules at an accelerated pace. The recent integration of deep learning with CASP has opened up new avenues for digitizing and exploring the vastly unknown chemical space, and has led to high expectations for fully automated synthesis plannings using machine-discovered novel reactions in the "future". Despite many progresses in the past few years, most deep-learning methods only focus on improving few aspects of CASP (e.g., top-k accuracy). In this work, we target specifically the efficiency of reaction space exploration and its impact on CASP. We propose NeuralTPL, a template-oriented generative approach, that performs impressively across a range of evaluation metrics including chemical validity, diversity, and novelty for various tasks in CASP. In addition, our Transformer-based model bears the potential to learn the core reaction transformation as it can efficiently explore the reaction space. We then perform several experiments and conduct a thorough analysis regarding the three metrics and demonstrate its chemical value for improving the existing deep-learning-driven CASP algorithms.