The recent advancements in the field of small molecule drug discovery have seen a significant shift towards more robust and scalable frameworks for AI model development and benchmarking. There is a growing emphasis on establishing standardized evaluation practices to ensure the reliability and transferability of AI models in real-world drug discovery scenarios. This trend is exemplified by the introduction of comprehensive datasets and evaluation frameworks that incorporate domain-specific preprocessing and hierarchical curation pipelines. Additionally, there is a notable push towards multi-modal representation learning, where molecular structures are enriched with biomedical text information, enhancing the predictive capabilities of AI models. The integration of visual recognition techniques for molecular structures in real-world documents, such as patents and scientific literature, is also gaining traction, addressing the challenges posed by complex and varied molecular image representations. Furthermore, the field is exploring novel methods for text-molecule retrieval, leveraging optimal transport-based alignments to capture detailed sub-structure information, and advancing fragment-based molecule generation through retrieval augmentation techniques. These developments collectively aim to bridge the gap between AI innovation and practical drug discovery applications, fostering a more efficient and effective drug development pipeline.