The application of machine learning in material property prediction has accelerated the discovery of new materials. Since a single material possesses multiple key properties, multi-task learning models that predict these properties simultaneously offer substantial benefits. Due to experimental costs and technological challenges, molecular datasets often exhibit a sparse data distribution. This sparsity and feature complexity pose challenges for multi-task learning models in effectively sharing features across tasks. However, these properties are often interrelated, indicating potential latent interactions. For instance, the dipole moment can aid in predicting the polarizability of a molecule, enhancing the predictive performance of multi-task learning methods. To harness these correlations, we present CTL-MoL, a novel multi-task learning framework that consists of: (1) a Cross-task Attention Module, which utilizes attention mechanisms to uncover latent inter-task interactions and strengthen cross-task features; and (2) a Task Relational Module, which identifies and builds task relevance to improve targeted feature integration. Experimental results demonstrate the effectiveness of our method in predicting molecular properties using the QM9 and OGB-MolPCBA datasets.