Multi-task learning in mammography has gained attention in recent years as a strategy to improve multi-label breast cancer and density classifications by leveraging shared representations. While previous studies have suggested that combining cancer and density labels in a multi-task learning framework can be beneficial, empirical evidence supporting these claims over single-task baselines remains limited. As breast density is a crucial factor for breast cancer risk but is less explored compared to cancer classification, this study focuses on density classification within the multi-task learning framework. Our findings indicate that single-task models can be beneficial over multi-task models, challenging the assumption that multi-task learning inherently provides performance benefits. Furthermore, we explore various sampling strategies in the multi-task setting to address class imbalance, a prevalent issue in mammography datasets, but find that none surpass the single-task baseline. Our findings suggest that the additional complexity of multi-task learning does not yield substantial benefits for density classification, highlighting the need for further research into alternative strategies for improving mammography classification performance.