Abstract: Graphs provide a fundamental data format for describing entities and their relations while organizing structured knowledge and supporting complex reasoning. Graph Neural Networks (GNNs) explicitly model topological dependencies through message passing and thus offer stable structural inductive biases for node-, edge-, and graph-level predictions. However, their semantic representations typically depend on input features and task-specific supervision, which limits cross-graph transfer, few-shot generalization, and natural language interpretability. Large Language Models (LLMs) possess rich semantic knowledge, as well as in-context learning and multistep generation capabilities; however, they struggle to fully preserve permutation invariance, high-order neighborhoods, and long-range dependencies within a one-dimensional Token sequence. The complementarity between these two families of models has inspired several studies on their integration. Existing studies often organize the literature according to graph types, input formats, or the role of a single model, making it difficult to simultaneously determine which model dominates the prediction and at what level do the two models interact. To this end, this paper proposes a two-dimensional taxonomy organized by the collaboration direction and coupling granularity axes. The collaboration direction axis, determined by the output component and the direction of information flow, comprises LLM-augmented GNNs (LLM4GNN), GNN-augmented LLMs (GNN4LLM), LLM-GNN co-training (Co-training), and unified graph—language architecture (Unified). The coupling granularity axis, defined by the location of information exchange, comprises feature-, Token-, parameter-, and task-level interaction. The two orthogonal axes form a 4×4 method map. This study retrieves and verifies representative studies published from January 2022 to August 2026 using reproducible search databases, keywords, and inclusion and exclusion criteria, giving priority to formally peer-reviewed versions. Preprints that have substantively shaped a paradigm but lack a formal version are explicitly marked accordingly and are not used as primary evidence for performance conclusions. Under this framework, the training cost, online inference cost, applicable tasks, and reproducibility risk are compared across paradigms. Further, classification boundaries are clarified using easily confused method pairs, and practical model selection is facilitated with a horizontal evidence table presented in unified dimensions. Furthermore, key techniques including graph textualization, structural Tokens, cross-modal alignment, and parameter-efficient training are summarized, proceeding from the logic of evolution rather than a method inventory. In addition, the convergence of Graph Retrieval-Augmented Generation (GraphRAG), graph agents, and graph foundation models from separate trajectories toward integration in retrieval, memory, planning, and unified interfaces is analyzed. The differentiated evaluation priorities of different paradigms in terms of task quality, structural understanding, evidence faithfulness, cross-domain generalization, resource efficiency, trustworthiness, and reproducibility are discussed for various scenarios including text-attributed graphs, Knowledge Graph Question Answering (KGQA), recommendation, and molecular and drug discovery. Based on the discussion, it is suggested that unified evaluation should adopt a joint reporting of ″quality—cost—trustworthiness″. Finally, open problems in the field are identified, including large graphs with long contexts; hallucination and evidence consistency; modality alignment; unified evaluation; scaling laws for graph foundation models; graph-agent collaboration; and issues related to privacy, fairness, and security. This survey aims to provide researchers and practitioners with a clearly bounded and auditable method map, as well as a basis for model selection under different data scales, task objectives, and computational constraints.

Key words: Large Language Model (LLM), Graph Neural Network (GNN), graph—anguage integration, Graph Retrieval-Augmented Generation (GraphRAG), graph foundation model, graph agent, co-training

摘要： 图是描述实体及其关系、组织结构化知识并支撑复杂推理的基础数据形式。图神经网络（GNN）通过消息传递显式建模拓扑依赖，在节点、边和图级预测中具有稳定的结构归纳偏置，但其语义表征通常依赖输入特征与任务监督，跨图迁移、少样本泛化和自然语言解释能力有限。大语言模型（LLM）具备丰富的语义知识、上下文学习和多步生成能力，却难以在一维Token序列中完整保持图的置换不变性、高阶邻域与长程依赖。两类模型的互补性推动了大量融合研究，但已有综述多按图类型、输入形式或单一模型角色划分，难以同时回答“由谁主导预测”和“两者在何种层级交互”两个问题。为此，本文提出“协作方向×耦合粒度”二维分类体系：协作方向依据输出主体及信息流划分为LLM增强GNN（LLM4GNN）、GNN增强LLM（GNN4LLM）、LLM-GNN协同训练（Co-training）和统一图-语言架构（Unified）；耦合粒度依据信息交换位置划分为特征级、Token级、参数级和任务级，两轴正交形成4×4方法地图。本文按照可复核的数据库、关键词、纳入与排除标准，对2022年1月至2026年8月的代表性研究进行检索与核验，并优先采用正式同行评审版本；对尚无正式版本但影响范式形成的少量预印本作显式标注，不将其作为性能结论的主要证据。在该框架下，本文不仅比较各范式的训练开销、在线推理开销、适用任务和复现风险，还选取易混淆方法说明分类边界，并以统一维度的横向证据表支撑实践选型。进一步地，本文从演进逻辑而非方法清单出发，总结图文本化、结构Token、跨模态对齐和参数高效训练等关键技术，分析图检索增强生成(GraphRAG)、图智能体与图基础模型从独立路线走向检索、记忆、规划和统一接口融合的转折。结合文本属性图、知识图谱问答（KGQA）、推荐、分子与药物等场景，本文讨论不同范式在任务质量、结构理解、证据忠实度、跨域泛化、资源效率、可信性和可复现性方面的评测侧重，并提出统一评测应采用“质量—成本—可信性”联合报告。最后，本文归纳大图与长上下文扩展、幻觉与证据一致性、模态对齐、统一评测、图基础模型尺度规律、图智能体协同以及隐私公平安全等开放问题。该综述旨在为研究者和工程人员提供边界明确、证据可核验的方法地图，并为不同数据规模、任务目标和算力约束下的模型选择提供依据。

关键词: 大语言模型, 图神经网络, 图-语言融合, 图检索增强生成, 图基础模型, 图智能体, 协同训练

CLC Number: 

• TP389.1