This document introduces a novel framework called Graph Pointer Neural Networks (GPNN) for learning node representations in heterophilic graphs. GPNN incorporates a graph pointer generator into the GNN architecture to distinguish crucial information from distant nodes and perform non-local aggregation selectively. Experimental results show that GPNN outperforms previous state-of-the-art methods, achieving an average lift of 6.3% on node classification tasks. The document analyzes GPNN's performance on heterophilic graphs and its ability to mitigate over-smoothing issues.