Trustworthy Graph Intelligence at Scale with Language-Augmented Learning
Alessandro Rossi, Chiara Bianchi, Lorenzo De Santis, Martina Esposito, Federico Rinaldi, Elisa Marchetti, Giancarlo Manuele · 2025
Graph Neural Networks (GNNs) have emerged as a dominant paradigm for learning over graph-structured data, achieving state-of-the-art performance across a wide spectrum of domains such as recommendation systems, bioinformatics, social networks, and cybersecurity. However, as GNNs transition from academic prototypes to real-world deployment, two critical challenges have become increasingly prominent: trustworthiness and efficiency. Trustworthiness encompasses the robustness of GNNs to adversarial perturbations, their interpretability and transparency to end-users, and their ability to produce fair and reliable predictions. Efficiency, on the other hand, is concerned with the scalability of GNN models to large graphs, their real-time inference capabilities, and their suitability for deployment in resource-constrained environments. Simultaneously, the emergence of Large Language Models (LLMs) has revolutionized the landscape of artificial intelligence, offering unprecedented capabilities in natural language understanding, contextual reasoning, and zero-shot generalization. Their integration with GNNs represents a new frontier, where the topological reasoning of graphs is fused with the semantic depth of language models to create more powerful and versatile systems. This survey provides a comprehensive and structured overview of the efforts to build trustworthy and efficient GNNs with the aid of LLMs. We begin by formalizing the foundational elements of GNN architectures and outlining the core challenges related to robustness, interpretability, fairness, scalability, and memory efficiency. We then explore how LLMs can be used to augment GNNs—by enriching node and edge features with semantic information, generating explanations for model predictions in natural language, and assisting in graph construction and data augmentation. Emerging architectures that couple GNNs and LLMs—such as dual-encoder, early-fusion, and joint reasoning models—are analyzed in detail, highlighting their capabilities, trade-offs, and underlying design principles. Throughout the survey, we emphasize how these hybrid systems can be designed to balance expressivity with efficiency and reliability with interpretability. We also discuss the key open challenges that hinder the widespread adoption of GNN–LLM architectures, including modality alignment, data scarcity, computational cost, and ethical considerations such as fairness, privacy, and bias mitigation. Finally, we outline promising research directions for the future, including the co-pretraining of graph-text models, the development of scalable and human-aligned multimodal frameworks, and the integration of causal and symbolic reasoning into hybrid GNN–LLM systems. By synthesizing insights from both graph learning and language modeling communities, this survey aims to chart a path toward the next generation of graph-intelligent systems that are not only accurate and scalable but also robust, transparent, and aligned with human values.