Multiple singing voice separation (MSVS) is closely related to speech separation, yet presents greater challenges due to the highly correlated nature of singing voices and the scarcity of multiple singing datasets. Existing studies on MSVS can be broadly classified into two categories: choral music separation and popular music separation. The latter remains underexplored and continues to exhibit limited performance. In this work, we address these limitations by introducing: (1) a data mining strategy for constructing highly correlated training mixtures, (2) a reverse attention mechanism to suppress highly correlated regions between outputs, and (3) a magnitude penalty loss that penalizes spectrogram regions containing energy that should exclusively belong to the other output. Experimental results demonstrate that our approach achieves substantial performance gains over prior methods.