Multimodal graph neural network with large language models for node and link prediction.

Introduction: Graph data representation is widely applicable in numerous real-world scenarios, and recent advances in graph neural networks (GNNs) have enabled effective modeling of complex associations in graph-structured data. However, GNNs are often constrained by the over-smoothing problem, which reduces their ability to distinguish node representations. In contrast, large language models (LLMs) are effective at capturing semantic contexts from textual data but are inherently limited in encoding structural information from graphs. Methods: We propose multimodal and semantic-enhanced GNN (M2GNN), an LLM-enhanced graph learning framework for node classification and link prediction on graphs with structural and textual feature modalities. M2GNN is not designed as a general-purpose graph multimodal large language model. Instead, it focuses on a controlled graph-text prediction setting, in which LLM-derived semantic embeddings guide candidate edge refinement, stable positional encodings are applied to the refined graph, and graph-based and language-informed predictors are adaptively fused. Results: Empirical results and ablation analyses show that M2GNN achieves competitive or improved performance on the evaluated benchmarks. The analyses further clarify the contribution of semantic edge refinement, stable structural encoding, and graph-language prediction fusion to node and link prediction performance. Discussion: Rather than claiming novelty for each individual component, this work demonstrates how semantic edge refinement, stable structural encoding, and adaptive graph-language fusion can be integrated into a practical pipeline for graph-text prediction tasks. The findings suggest that combining LLM-derived semantic information with structure-aware graph learning can improve prediction robustness while maintaining a clearly defined task-specific scope.