Microstructures and Accuracy of Graph Recall by Large Language Models