Graph Foundation Models: A Comprehensive Survey

TL;DR

Graph Foundation Models (GFMs) extend foundation-model principles to non-Euclidean, relational graph data by unifying backbone architectures, pretraining strategies, and adaptation mechanisms in a pretrain-then-adapt framework. The survey introduces a three-way taxonomy (universal, domain-specific, task-specific GFMs) and a modular framework that supports graph neural networks, language models, and their hybrids, enabling cross-domain transfer and open-ended graph reasoning. It analyzes theoretical foundations (transferability, emergent capabilities), benchmarks, and key challenges like structural alignment, heterogeneity, scalability, and evaluation, and outlines future directions toward scalable, multimodal, and theory-grounded GFMs. Resource coverage includes a public repository with datasets, baselines, and pretrained models to foster reproducibility. Overall, GFMs are positioned as foundational infrastructure for scalable, general-purpose reasoning over structured data, with broad implications for molecules, knowledge graphs, biology, and beyond.

Abstract

Graph-structured data pervades domains such as social networks, biological systems, knowledge graphs, and recommender systems. While foundation models have transformed natural language processing, vision, and multimodal learning through large-scale pretraining and generalization, extending these capabilities to graphs -- characterized by non-Euclidean structures and complex relational semantics -- poses unique challenges and opens new opportunities. To this end, Graph Foundation Models (GFMs) aim to bring scalable, general-purpose intelligence to structured data, enabling broad transfer across graph-centric tasks and domains. This survey provides a comprehensive overview of GFMs, unifying diverse efforts under a modular framework comprising three key components: backbone architectures, pretraining strategies, and adaptation mechanisms. We categorize GFMs by their generalization scope -- universal, task-specific, and domain-specific -- and review representative methods, key innovations, and theoretical insights within each category. Beyond methodology, we examine theoretical foundations including transferability and emergent capabilities, and highlight key challenges such as structural alignment, heterogeneity, scalability, and evaluation. Positioned at the intersection of graph learning and general-purpose AI, GFMs are poised to become foundational infrastructure for open-ended reasoning over structured data. This survey consolidates current progress and outlines future directions to guide research in this rapidly evolving field. Resources are available at https://github.com/Zehong-Wang/Awesome-Foundation-Models-on-Graphs.

Figures (6)

• Figure 1: From Task-Specific Graph Models to General-Purpose Graph Foundation Models. This figure contrasts the paradigm shift from traditional Graph Neural Networks (GNNs) to Graph Foundation Models (GFMs). (a) GFMs are pretrained on large-scale graph corpora spanning multiple domains (e.g., social, web, academic, molecular) to acquire broadly transferable representations. Through various adaptation techniques—such as fine-tuning, distillation, prompting, or zero-shot inference—they can generalize across a wide spectrum of downstream tasks, including node classification, link prediction, graph classification, and graph-to-text generation. (b) In contrast, traditional GNNs are typically trained in an end-to-end manner on a single-domain dataset for a specific task, often lacking the scalability and generalization capabilities required for open-world settings. This shift mirrors the transition observed in language and vision domains, where foundation models have redefined the standard for general-purpose intelligence.
• Figure 2: The Evolution of Graph Learning Paradigms. This figure illustrates the historical trajectory of graph learning, highlighting the increasing task-solving capacity over time. (1)Statistical methods (pre-2010s) relied on heuristic-driven techniques, such as spectral analysis and graph kernels, to solve narrowly scoped graph tasks. (2)Graph embeddings (circa 2010) introduced shallow, task-agnostic representations via random walks or matrix factorization, enabling better structural understanding. (3)Graph neural networks (2016 onward) adopted deep learning principles—particularly message passing—to build end-to-end task-specific models capable of capturing semantic dependencies. (4)Graph foundation models (post-2023) represent the latest paradigm, aiming for universal, general-purpose solvers that are pretrained on large-scale graphs and adapted to diverse downstream tasks across domains. This timeline reflects a broader shift from handcrafted, task-bound solutions to scalable, generalizable graph intelligence.
• Figure 3: Key Challenges in Designing Graph Foundation Models. This figure illustrates three fundamental sources of heterogeneity that challenge the development of general-purpose GFMs: (a) Feature Heterogeneity: Graph datasets vary widely in their node features due to differences in domain semantics and data preprocessing. For instance, molecular graphs use atom-level descriptors, while citation networks rely on textual or structural attributes. Such diversity complicates unified representation learning. (b) Structure Heterogeneity: Graphs from different domains exhibit distinct topological patterns—molecules contain recurring chemical motifs such as benzene rings, whereas social networks emphasize relational structures like triadic closures. Capturing such heterogeneous structures requires models to be structurally adaptive. (c) Task Heterogeneity: Graph learning tasks span multiple granularities—including node-level classification, link prediction, and graph-level classification—each requiring different inductive biases. A universal GFM must effectively generalize across these varied task formulations. Together, these challenges underscore the need for robust, flexible architectures and training paradigms capable of aligning graphs across both structure and semantics.
• Figure 4: Backbone Architectures of Graph Foundation Models. We illustrate three representative paradigms: (a) Graph Model as Predictor, where GNNs serve as the primary reasoning engine with optional auxiliary modules; (b) Language Model as Predictor, where LLMs interpret graph-structured inputs converted into text or structured prompts; (c) Graph-Language Co-Training, which jointly optimizes GNNs and LLMs through alignment for enhanced generalization across tasks and modalities.
• Figure 5: Pretraining Strategies of Graph Foundation Models.(a) Supervised Pretraining: Models are pretrained using labeled subgraphs derived from large-scale graphs, where task-specific labels guide the learning process via supervised objectives. (b) Generative Pretraining: Models learn to reconstruct masked or corrupted parts of the graph, such as node features or adjacency information, often using autoregressive or autoencoding paradigms to capture high-order dependencies. (c) Contrastive Pretraining: Models are trained to distinguish between similar (positive) and dissimilar (negative) node or subgraph pairs across different views or augmentations of the graph, thereby learning invariant and transferable representations.

Theorems & Definitions (6)

• Definition 2.1: Graph
• Definition 2.2: Attributed Graph
• Definition 2.3: Text-Attributed Graph (TAG)
• Definition 2.4: Graph Neural Network (GNN)
• Definition 2.5: Large Language Model (LLM)
• Definition 2.6: Graph Foundation Model (GFM)