Automated citation systems play a crucial role in efficiently managing the growing body of scientific literature. The SCIDOCA 2025 Shared Task aims to advance research in scientific document analysis and citation management through three key subtasks: Citation Discovery, Masked Citation Prediction, and Citation Sentence Prediction. In this paper, we present our approach to tackling these challenges using attention-based network architectures to model citation relationships between texts and candidate references. We propose two methods: the first constructs a graph representation of the text and candidate references, leveraging a Graph Attention Network (GAT) to capture structural dependencies. The second method retrieves co-citing texts and applies an attention mechanism to enhance citation relevance by attending to both the co-citing texts and the candidate reference. Experimental results demonstrate the potential of our approaches across all three subtasks.