Abstract Rice ( Oryza sativa ) is an important staple food, feeding more than half of the global population. A feasible improvement of rice yield is necessary to meet the ever–growing food demands. Genomic selection (GS), as an advanced breeding technique, enables the prediction of phenotypes solely based on genotypic data using a constructed genomic prediction model. A benchmarking study is required to evaluate the prediction performance of various available GS models for accelerating the breeding process. To address this concern, a diverse population containing 688 rice accessions with 66,456 haplotype‐based single nucleotide polymorphisms (SNPs) (HA‐SNPs) were used to construct 17 GS models for grain length, grain width, and 1000‐grain weight separately. The effect of population size on the GS prediction accuracy was slightly greater than that of marker size. Among these models, ridge regression best linear unbiased prediction (RRBLUP) offered several advantages in terms of robustness, simplicity, and computational efficiency. Regarding the predictive ability, machine learning and deep learning models are also viable alternatives. Additionally, the prediction accuracy of genome‐wide association study–detected SNP based GS was significantly higher than that of HA‐SNP based GS. Furthermore, a subset including 63 selected accessions was used to construct the RRBLUP model, and followed by the evaluation of elite germplasm for practical breeding. Taken together, these findings have the potential to reduce the rice yield breeding cycle and enhance the genetic gain.
Hu et al. (2026) studied this question.