The recent advancements in molecular generative models are shifting towards more functional and text-based representations, leveraging the power of transformers and diffusion models. These approaches aim to enhance the expressiveness and efficiency of molecule generation, moving beyond traditional sequence or graph-based methods. Notably, the integration of large language models (LLMs) with domain-specific feedback mechanisms is proving effective for iterative molecule optimization, significantly improving hit ratios in both single- and multi-property objectives. Additionally, the use of diffusion language models for text-guided multi-property optimization is demonstrating superior performance in maintaining structural similarity while enhancing chemical properties. These innovations not only streamline model architectures but also accelerate generation times, making them highly suitable for scalable, large-scale optimization tasks in drug discovery and material science.