Abstract: Soybean Glycine max (L.) Merr. is a strategic crop with grain, oilseed, and feed value, yet its improvement has long been constrained by slow yield gains and a narrow genetic base. For China, the high dependence on soybean imports and the low self-sufficiency rate make expanding soybean cultivation and increasing yield per unit area becoming important tasks for ensuring food security, creating an urgent demand for technologies that can improve breeding efficiency. Over the past decade, rapid advances in large-scale sequencing, high-throughput phenotyping, and artificial intelligence algorithms have provided technical conditions for the transformation of breeding approaches. At present, the main constraint on soybean smart breeding is not the insufficiency of any single technology, but the lack of stable connections among data, algorithms, and applications. Whether data can be effectively learned by models, whether prediction results can be translated into breeding decisions such as parent selection, cross screening and target design, and whether new data generated from field validation can be standardizedly fed back, and used to continuously improve models directly affect the operational effectiveness of smart breeding systems. This review summarized research progress in soybean smart breeding from the perspectives of the data layer, algorithm layer, and application layer. It further analyzed the major barriers in data-to-algorithm transfer, algorithm-to-application translation, and application-to-data feedback, and identified key challenges in five aspects: Data, algorithms, applications, platforms and governance. Finally, it discussed the conditions required for soybean breeding to move toward intelligent design breeding. Future efforts should focus on bridging disconnections between layers, making prediction, validation, and data feedback into routine processes, and shifting evaluation criteria from model prediction accuracy to realized genetic gain.