<p>Genomic selection (GS) uses genome-wide molecular markers and phenotypic data from a training population to predict breeding values or phenotypes in candidate populations. Unlike marker-assisted selection, GS does not require significance testing of individual markers and is particularly suitable for polygenic traits controlled by many small-effect loci, offering higher prediction accuracy, shorter breeding cycles, and improved efficiency. However, the practical implementation of GS in crops faces several challenges, including low prediction accuracy for traits influenced by genotype-by-environment interactions, difficulty in modeling non-additive effects without overfitting, and the high cost and limited interpretability of machine learning and multi-omics approaches. This review examines key factors that affect GS efficacy, with emphasis on training population design, the incorporation of non-additive effects, and the integration of multi-trait and multi-environment data. We discuss strategies for constructing training populations, compare linear, Bayesian, and machine learning models in terms of predictive performance and interpretability, and address the logistical and economic barriers to multi-omics integration, particularly in hybrid prediction. We also summarize the current status of breeding chip development for major crops. Finally, we highlight future directions, including the development of crop-specific chips, unified analytical platforms, and enhanced model interpretability, to bridge the gap between methodological advances and practical breeding applications.</p>