Generative AI offers transformative potential for small-molecule drug discovery, enabling faster and more targeted identification of novel therapeutics. Lead discovery and optimization remain pivotal yet challenging, particularly in designing compounds with precise pharmacological profiles and target interactions. Existing gradient-based and gradient-free methods struggle to meet these demands. We introduce EPOSMol, an evolutionary policy gradient framework for lead optimization that refines molecular structures in latent space. Our approach iteratively samples structures, using oracle tools to evaluate fitness based on desired properties, binding affinity, and target interactions. A flexible reward framework enables adaptive policy updates and seamless integration of ranking tools, while dynamic scheduling of population size and exploration parameters optimally balances global search and local refinement. EPOSMol achieves up to 10× improvement over state-of-the-art techniques in generating target-specific hits and exploring high-potential chemical spaces. This work advances AI-driven drug discovery by combining evolutionary algorithms with generative modeling to enhance lead optimization.