<p>Computational sequence-based predictors of protein localization have the potential to accelerate the discovery of protein functions and interactions, thereby advancing our understanding of human biology and disease. While many methods have been proposed, evaluations remain limited by small test sets, coarse-grained cellular compartment labels and single-label classification, despite the fact that nearly half of human proteins localize to multiple compartments. Here we integrate annotations from major protein databases to construct a highly validated, twofold larger benchmark test set of 3,814 human proteins. Using this dataset, we systematically evaluate existing sequence-based predictors and compare combinations of protein language models and aggregation strategies. We find that current models underperform on fine-grained compartments, multilocalizing proteins and pathogenic variants known to mislocalize. Our results reveal fundamental limitations of existing approaches and underscore the need for improved models, standardized benchmark datasets and more rigorous evaluation in subcellular localization prediction.</p>