PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Foundation Models: A Comprehensive Survey

Zehong Wang, Zheyuan Liu, Tianyi Ma, Jiazheng Li, Zheyuan Zhang, Xingbo Fu, Yiyang Li, Zhengqing Yuan, Wei Song, Yijun Ma, Qingkai Zeng, Xiusi Chen, Jianan Zhao, Jundong Li, Meng Jiang, Pietro Lio, Nitesh Chawla, Chuxu Zhang, Yanfang Ye

arXiv:2505.15116v1cs.LGcs.AIcs.SI

TL;DR

Graph Foundation Models address the challenge of extending foundation-model capabilities to graph data, whose relational, heterogeneous, and non-Euclidean structure complicates cross-task and cross-domain transfer. This survey synthesizes the field through a modular framework, taxonomy, representative methods, theoretical perspectives, and open challenges, while concluding that GFMs remain constrained by data quality and generalization issues.

  • Problem

    Graph data has relational, heterogeneous, and non-Euclidean properties that make broad transfer across graph tasks and domains difficult to organize and study systematically.

  • Method

    The survey organizes GFMs around backbone architectures, pretraining strategies, adaptation mechanisms, generalization scope, theoretical foundations, benchmarks, and future directions.

  • Results

    The survey consolidates representative GFM methods, transferability and emergent-capability perspectives, and challenges spanning structural alignment, heterogeneity, scalability, evaluation, and data quality.

  • Takeaways & Limitations

    GFMs are positioned as a developing framework for general-purpose intelligence and open-ended reasoning over structured data across graph-centric domains.

  • Takeaways & Limitations

    GFM development remains constrained by incomplete, difficult-to-evaluate graph data and alignment mechanisms that may fail on unseen domains or feature spaces.

Abstract

from arXiv · show

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.

1 Introduction

Graph foundation models extend foundation-model ideas to graph-structured data through large-scale, heterogeneous pretraining and adaptable downstream use. The survey organizes this emerging field around challenges, modular design components, generalization scope, and future research needs.

  • Motivation: GFMs use diverse graph corpora and adaptation methods to support node classification, link prediction, graph classification, and graph-to-text generation across domains.Adaptation may involve fine-tuning, distillation, prompting, or zero-shot inference.
  • Motivation: Foundation models learn transferable capabilities from large-scale, heterogeneous pretraining, while GFMs extend this paradigm to relational, permutation-invariant, non-Euclidean graph data.GFMs target applications including molecular property prediction, knowledge-graph reasoning, social-network analysis, and recommendation.
  • Survey gap: Existing GFM surveys provide fragmented coverage, often emphasizing backbone architectures without comprehensively treating foundational techniques, design challenges, and research directions.The survey positions itself as a unified account of the field’s breadth and depth.
  • Survey framework: The survey decomposes GFMs into backbone architectures, pretraining strategies, and adaptation mechanisms, then classifies them as universal, domain-specific, or task-specific.This framework is used to organize representative methods and their generalization scope.
  • Challenges: The survey identifies feature, structural, and task heterogeneity as core design challenges for learning from graph-structured data at scale.These dimensions capture distinct sources of complexity in building general-purpose graph learning systems.
  • Future directions: Future GFM development requires progress in scalable architectures, graph-data quality, realistic evaluation, adaptation, multimodal representations, and theoretical understanding.The survey specifically highlights transferability, robustness, trustworthiness, zero-shot and prompt-based learning, and cross-domain reasoning as open issues.

2 Background

Graph learning progressed from handcrafted statistical methods and shallow embeddings to message-passing GNNs and, more recently, graph foundation models. GFMs combine large-scale graph pretraining with general-purpose representations intended for transfer across tasks and domains.

  • Historical evolution: Early graph learning relied on heuristic statistical methods, including spectral analysis and graph kernels, for narrowly scoped tasks.These approaches preceded representation-learning methods in the field’s historical development.
  • Historical evolution: Graph embeddings introduced shallow, task-agnostic representations through random walks or matrix factorization, improving structural understanding.Representative methods include DeepWalk, node2vec, and LINE, but these embeddings were largely transductive.
  • Historical evolution: GNNs adopted message passing to aggregate neighboring information and build end-to-end, task-specific models for non-Euclidean graphs.GCNs, GATs, and GraphSAGE exemplify this transition toward deep graph representation learning.
  • Graph foundation models: GFMs are pretrained on large-scale graphs with self-supervised objectives to learn representations that transfer across tasks and domains.Their intended capabilities include structural and semantic modeling together with zero-shot and few-shot generalization.
  • Foundation-model background: Foundation models differ from task-specific systems through diverse pretraining and general-purpose representations that adapt to new tasks with limited supervision.Their capabilities can be transferred through mechanisms such as fine-tuning, prompting, or instruction tuning.
  • Definitions and notations: Graph representations include nodes, edges, attributes, and textual descriptions, with formal definitions for graphs, attributed graphs, and text-attributed graphs.The notation section introduces graph structure through G = (V, E), attributed graphs through G = (X, A), and text-attributed graphs through G = (X, A, D).

3 Challenges in Designing Graph Foundation Models

GFMs must generalize across heterogeneous graph features, structures, and task formulations. Existing remedies address these dimensions, but unified adaptation across real-world graphs and varied tasks remains unresolved.

  • Core challenges: Graph foundation models face three central challenges: feature heterogeneity, structure heterogeneity, and task heterogeneity.These dimensions arise from variation in graph inputs, topologies, and learning objectives.
  • Feature Heterogeneity: Feature heterogeneity reflects domain-specific semantics and inconsistent preprocessing, producing incompatible feature spaces across graph datasets.Cora uses 1433-dimensional keyword vectors, whereas Pubmed uses 500-dimensional TF-IDF vectors.
  • Feature Heterogeneity: Shared latent projections, language-model encoders, and graph-specific projection layers can align features, but often depend on domain knowledge or create scalability bottlenecks.The survey identifies general adaptive feature alignment as an open research direction.
  • Structure Heterogeneity: Structural heterogeneity spans localized social and citation motifs versus long-range molecular patterns, while deeper GNNs risk over-smoothing and over-squashing.Traditional message passing also has locality bias that limits global structure modeling.
  • Structure Heterogeneity: Structure-aware augmentation, graph prompt tuning, and discrete structural codebooks improve adaptability, but current methods do not cover the full spectrum of real-world graph structures.Structural generalization remains limited despite these techniques.
  • Task Heterogeneity: Task heterogeneity spans node, link, graph, knowledge-graph, and molecule-generation objectives, each requiring distinct modeling strategies and inductive biases.Explicit task alignment and implicit task-agnostic representations have both been explored, yet seamless universal adaptation remains challenging.

4.1 Unified Framework

The unified GFM framework replaces task-specific end-to-end training with pretraining on diverse graph data followed by downstream adaptation. It organizes methodology around backbone architecture, pretraining strategy, and adaptation mechanism.

  • Unified Framework: GFMs follow a “pretrain-then-adapt” paradigm rather than training a model end-to-end for one graph task.A backbone is pretrained on a large graph database and then applied or adapted to downstream scenarios.
  • Pretraining: Pretraining learns generalized graph representations from diverse graph datasets using a backbone such as a GNN, Transformer, or language model.The objective minimizes a pretraining loss over source data Dpt.
  • Adaptation: After pretraining, downstream data can support fine-tuning through an adaptation loss to improve task-specific generalization.Adaptation may use full or partial fine-tuning, lightweight tuning, or task-specific prediction heads.
  • Unified Framework: The framework decomposes GFM methodology into backbone architecture, pretraining strategy, and adaptation mechanism.This modular view supports systematic comparison and composability across design choices.
  • Framework Components: Backbones encode nodes, edges, and global context; pretraining learns broad graph patterns; adaptation aligns that knowledge with target tasks or domains.These roles jointly connect general representation learning with specialized application requirements.

4.2 Backbone Architectures

GFMs use three backbone paradigms to combine graph structure and language semantics: graph models as predictors, language models as predictors, and graph-language co-training. These designs trade explicit relational inductive biases against language-based scalability and multimodal integration.

  • GFMs organize backbone designs into graph models as predictors, language models as predictors, and graph-language co-training frameworks.The taxonomy covers architectural foundations, assumptions, and applications.
  • Graph Model as Predictor: Graph models capture relational dependencies through message passing, while Graph Transformers extend this process globally to model long-range dependencies.GCNs use normalized aggregation, whereas GATs assign learned attention weights to neighboring nodes.
  • Graph Model as Predictor: Pure graph backbones exploit structural inductive biases and perform strongly when relational signals are rich, but may struggle with multimodal reasoning or external contextual knowledge.This limitation motivates adding language models as auxiliary modules.
  • Language Model as Predictor: Language models can predict graphs by serializing node identities, features, and neighborhoods into textual sequences processed without explicit graph operations.LangGFM and BeyondText exemplify textual graph representations for reasoning over structure and relational dependencies.
  • Language Model as Predictor: Hybrid language architectures inject graph structure through cross-modal alignment, graph-aware tokenization, or auxiliary encoders that condition language models.These approaches use graph representations to make language-based reasoning more structure-aware.

4.3 Pretraining Strategies

GFM pretraining uses supervised, generative, and contrastive objectives to learn transferable representations from labeled, reconstructed, or paired graph views. These strategies respectively align learning with downstream labels, model graph data distributions, or enforce invariant representations across related views.

  • GFM pretraining comprises supervised, generative, and contrastive paradigms for acquiring transferable knowledge from graph data.The strategies are summarized in Figure 5 and differ in their supervision and learning objectives.
  • Supervised Pretraining: Supervised pretraining optimizes predictions of node-, edge-, or graph-level labels, directly aligning representations with downstream objectives.It can yield faster convergence and improved performance on similar tasks, but depends on costly high-quality labels.
  • Generative Pretraining: Generative pretraining learns broad representations by modeling observed data, using autoregressive or auto-encoding objectives without task-specific labels.Graph auto-encoders reconstruct masked node features, edge features, or structural components from visible graph context.
  • Generative Pretraining: Autoregressive graph generation maps graphs into ordered sequences and generates each node and its connections conditioned on the preceding graph prefix.The hidden state summarizes generation history before predicting the next adjacency slice.
  • Generative Pretraining: Masked graph modeling reconstructs corrupted attributes or adjacency entries using the unmasked graph context.The target may be masked node or edge features, or masked structural components.
  • Contrastive Pretraining: Contrastive pretraining brings positive representations closer and pushes negative representations apart, learning discriminative features without manual annotations.Graph variants use different views or augmentations and can combine local and global information.

4.4 Adaptation

GFMs can be adapted through six paradigms spanning transfer learning, distillation, test-time adaptation, graph prompting, in-context learning, and prototype learning. These strategies differ in supervision, generalization, and computational efficiency.

  • Six adaptation paradigms are identified: transfer learning, distillation, test-time adaptation, graph prompting, in-context learning, and prototype learning.They address different data conditions, supervision levels, and deployment constraints.
  • Transfer Learning: Transfer learning reuses pretrained structural and semantic knowledge for downstream graph tasks with limited labeled data.It initializes the target model from pretrained parameters and supports direct, full, adaptive, and parameter-efficient fine-tuning.
  • Transfer Learning: Parameter-efficient fine-tuning freezes the backbone while adding lightweight adapters, low-rank transformations, or task-specific vectors.These methods reduce trainable parameters and computational cost while aiming to retain generalization.
  • Distillation: Distillation transfers a large teacher’s knowledge to a smaller student through supervised targets, teacher-output matching, and optional feature-level alignment.Graph distillation can preserve topology, attention patterns, and multi-view or multi-task knowledge.
  • Test-Time Adaptation: Test-time adaptation updates a pretrained GFM during inference using unlabeled test data and self-supervised objectives.It is designed for distribution shifts or limited access to labeled target-domain samples.
  • Test-Time Adaptation: A language-model pseudo-label can guide test-time updates, combining language-based supervision with graph reasoning without manual annotations.The procedure generates pseudo-labels with an LLM and optimizes a pseudo-label loss on test data.

5 Universal Graph Foundation Models

Universal GFMs seek generalization across heterogeneous graph domains and tasks through pretraining, unified architectures, and adaptation. The survey organizes unification around task reformulation, architectural generalization, and feature alignment.

  • Universal GFMs aim to generalize across graph domains and tasks despite differences in size, connectivity, attributes, density, and objectives.They follow a pretrain-then-adapt paradigm using diverse graph databases and transferable structural and semantic patterns.
  • Model Unification: Model unification has two forms: explicit task and input reformulation, and implicit architectural generalization with invariance enforcement.These approaches target GNN encoders that transfer across tasks, domains, and graph topologies.
  • Model Unification: Explicit unification reformulates graph tasks into shared prediction formats at link, subgraph, or tree granularity.Link-level methods introduce class nodes, while subgraph-level methods encode ego-graphs for node, edge, and graph tasks.
  • Model Unification: Subgraph-level unification uses local structural context and supports unified node-, edge-, and graph-level prediction frameworks.Contrastive pretraining and prompt learning are used to improve structural capture and alignment across domains.
  • Domain Alignment: Feature alignment addresses heterogeneous graph signals through textual or multimodal alignment and model-based alignment.Shared semantic spaces can be built from textual descriptions, visual content, or combined modalities.
  • Domain Alignment: Model-based feature alignment may fail on unseen domains because projection layers, transformations, and tokens are coupled to pretraining distributions.This can impair semantic consistency or meaningful representations for new feature spaces and structural patterns.

5.3 Language Model-Based Universal GFM

LLM-based universal GFMs convert graphs into textual or structured representations, post-train language models with graph objectives, and adapt them through prompting or fine-tuning. Their central challenge is preserving graph inductive biases during sequential language processing.

  • LLM-based GFMs combine graph tokenization, graph-specific post-training, and downstream adaptation through prompting, in-context learning, or instruction tuning.The pipeline maps graphs to token sequences, optimizes graph-aware objectives, and decodes task-formatted predictions.
  • Graph Conversion: Converting graphs into language-compatible representations must align relational, non-Euclidean structure with LLMs’ sequential and semantic processing.Natural-language and structured conversions offer different trade-offs in interpretability, generalization, efficiency, and application compatibility.
  • Graph Conversion: Natural-language conversion describes nodes, edges, and attributes with templates, hierarchical neighborhoods, and task-specific reasoning prompts.These mechanisms support graph reasoning over relational patterns and multi-hop structural context.
  • Graph Conversion: Structured conversion serializes graphs as JSON, code blocks, or nested lists while preserving topology for parsing and manipulation.Templates, APIs, and recursive context trees support retrieval, traversal, and nested relational reasoning.
  • Downstream Adaptation: Zero-shot and few-shot prompting enables inference without parameter updates, while supervised fine-tuning and preference alignment improve graph-task alignment.Instruction-tuned models can handle varying graph input-output formats and structural complexities.
  • Domain Alignment: Self-supervised topology and feature masked autoencoding aligns LLM representations with graph structure without manually labeled data.LangGFM is reported to achieve strong downstream performance with label efficiency and robustness to domain shifts.

5.4 Graph-Language Co-Training Universal GFM

Graph-language co-training universal GFMs combine GNN structural representations with LLM semantic reasoning. They address information loss from graph linearization through graph encoding, cross-modality projection, and language-based inference.

  • LLMs can lose high-order dependencies, connectivity patterns, and graph invariants when graphs are linearized into sequences.This motivates hybrid architectures that preserve graph inductive biases while enabling semantic generalization.
  • Hybrid Architecture: Graph-language co-training trains a GNN for structural information, projects graph embeddings into LLM token space, and performs language-based inference.The LLM receives projected graph representations alongside task instructions and optional demonstrations.
  • Representation Alignment: Hybrid systems bridge graph embeddings and token-based language models using structured node representations, neighborhood templates, and learned projection modules.Examples include local Neighborhood Detail, broader Hop-Field Overview, and structured graph-token grounding.
  • Domain Alignment: Domain alignment can use supervised end-to-end training, self-supervised text matching, or CLIP-like contrastive learning between graphs and textual summaries.These strategies couple structural features with textual reasoning and can support zero-shot or few-shot transfer.
  • Downstream Adaptation: Zero-shot reasoning and in-context learning enable graph-task inference without additional fine-tuning or by inserting demonstrations into prompts.Example purification selects informative and class-balanced support examples for graph classification.
  • Interpretability: Natural-language outputs combined with structured representations can support human-readable explanations in graph reasoning systems.GraphTranslator demonstrates multi-turn dialog reasoning across behavioral analytics, social networks, and recommendation scenarios.

5.5 Discussion

Universal GFMs are organized into GNN-based, LLM-based, and hybrid models, each trading off structural awareness, semantic flexibility, scalability, and interpretability. Hybrid approaches seek to combine graph-structural reasoning with language-driven generalization.

  • Discussion: The three model families reflect trade-offs among expressiveness, scalability, generalization, and interpretability.GNNs are efficient and structure-aware but semantically limited, whereas LLMs are flexible and language-driven but struggle with graph topology.
  • Graph Model-Based GFMs: GNN-based GFMs capture topology through message passing and localized neighborhood aggregation, supporting efficient computation on large-scale graphs.Their graph-specific inductive biases make them well suited to structure-critical tasks.
  • Language Model-Based GFMs: LLM-based GFMs translate graphs into textual or structured sequences, enabling prompt-based reasoning across heterogeneous tasks and domains.They also support multimodal inputs and offer high interpretability through language-based interactions.
  • Graph-Language Co-Training GFMs: Hybrid GFMs project graph embeddings into LLM token spaces to combine structural awareness with semantic and general-purpose reasoning.This design supports multimodal and cross-domain inference while improving task transferability and expressiveness.

6 Task-Specific Graph Foundation Models

Task-specific GFMs target one graph task across multiple domains, requiring task-aware invariances, cross-domain alignment, and a balance between domain generalization and task adaptation. The survey covers their task formulations, model families, adaptation techniques, and future directions across node, link, graph, question-answering, anomaly-detection, and recommendation tasks.

  • 6 Task-Specific Graph Foundation Models: Task-specific GFMs operate across domains while focusing on a single task, such as node classification, link prediction, or graph generation.They must capture task-aware invariances while aligning disparate graph distributions.
  • Task-Specific Design Principles: Cross-domain generalization requires representations robust to differences in node types, feature distributions, connectivity patterns, and graph structure.Suggested mechanisms include domain-adaptive architectures and domain-specific feature modulation.
  • Task-Specific Design Principles: Task-specific GFMs must balance domain invariance with task adaptation because either extreme can neglect critical task or domain patterns.Overemphasizing domain understanding can favor particular domains, while full generalization can reduce performance by overlooking domain-specific information.
  • Node-level Task: Node classification predicts individual node labels, while node-level GFMs address heterogeneous features and structures through imputation, pretraining, prompting, and self-supervised learning.Representative approaches include generative pretraining, token-pair prompting, and efficient self-supervised frameworks.
  • Model Architectures: GNN, Transformer, and hybrid methods extend task-specific GFMs across diverse datasets by combining structural modeling, common latent spaces, and language-model capabilities.Examples include single-model multi-dataset training, Perceiver-based feature compression, and LLM-assisted graph prediction.
  • Adaptation Mechanisms: Parameter-efficient adaptation can provide an alternative to full fine-tuning for downstream molecular graph classification and regression.The cited approach uses a shared prompt vector and provides derivations supporting its universality and effectiveness.
  • Future Directions: Future graph-level GFMs are directed toward unified architectures, scalable training, rapid adaptation, multimodal evaluation, explainability, and fairness.Proposed directions include structural Graph Transformers, hierarchical pooling, subgraph sampling, prompt-based fine-tuning, unified benchmarks, and adversarial debiasing.

7 Domain-Specific Graph Foundation Models

Domain-specific GFMs learn shared representations within specialized graph domains, where structural and semantic properties require customized architectures, pretraining, and adaptation. The survey reviews approaches spanning molecular, knowledge, temporal, academic, and causal graphs, alongside domain-specific challenges and future directions.

  • Domain-Specific Graph Foundation Models: Domain-specific GFMs generalize across related tasks within one domain, but must capture that domain’s underlying principles and key properties.The survey identifies shared representations as useful across correlated tasks while emphasizing the difficulty of preserving domain-specific structure.
  • Molecular Graphs: Molecular GFMs must model chemical symmetries, diverse atomic and bond features, and interactions ranging from local functional groups to global topology.Representative approaches use either 3D equivariant modeling or 2D connectivity with masked-atom, contrastive, and multi-task pretraining.
  • Knowledge Graphs: Knowledge-graph GFMs should extend beyond binary relations, integrate textual, visual, and numerical data, and improve interpretability, controllability, and scalability.These priorities are especially relevant to sensitive applications such as healthcare and finance.
  • Temporal Graphs: MiNT uses structural and temporal diversity for cross-network pretraining, producing temporally aware node representations that generalize to unseen graphs.The passage reports that MiNT outperforms traditional temporal GNNs trained on individual datasets.
  • Temporal Graphs: Temporal graph GFMs face disentangled spatial-temporal modeling, cross-graph generalization, and flow-aware sequence modeling challenges.Proposed directions include hierarchical temporal abstraction, adaptive serialization, and hybrid GNN–LLM architectures.
  • Academic Networks: LitFM combines citation-aware graph retrieval with instruction tuning to support citation prediction, related-work generation, and literature-review summarization.Its graph retriever uses graph proximity and citation-aware embeddings to address citation hallucination and knowledge incompleteness.
  • Causal Graphs: Causal GFMs must jointly reason over sparse, directional causal structure and noisy, text-heavy evidence while maintaining semantic grounding and structural consistency.Causal graphs encode variables as nodes and cause-effect relationships as directed edges.

8 Theoretical Understandings

Theoretical analyses examine emergence, scaling behavior, and transferability in graph learning, finding stronger evidence in well-structured domains than in general-purpose graph foundation models. They also identify transferable patterns and limitations arising from graph structure, motif semantics, and GNN expressiveness.

  • Emergence and Scaling: Scaling-law evidence is strongest in molecular and atomic graphs, where natural structure supports performance improvements as model capacity and dataset size increase.JMP pretrained on approximately 120 million atomic systems, while DPA-2 used more than 10 million atomic structures for transfer to unseen tasks.
  • Emergence and Scaling: Molecular scaling studies report exponents of 0.17 for the largest dataset and 0.26 for equivariant GNN-based interatomic potentials.Scaling behavior also depends on molecular representation, with graph-based and fingerprint-based encodings showing higher data efficiency than SMILES-based representations.
  • Emergence and Scaling: MiNT improves zero-shot transfer across temporal graphs, with performance increasing consistently as the number of pretraining networks grows.The study pretrained on 64 temporal networks and evaluated transferability on 20 unseen networks.
  • Emergence and Scaling: General-graph scaling remains unresolved: graph size is better measured by nodes or edges, while self-supervised downstream performance remains sensitive to architecture and pretext-task design.SSL loss improves with data and model size, but existing graph SSL methods do not show clear downstream scaling trends.
  • Transferability: Node-, link-, and graph-level transferability depends respectively on neighborhood relations, pairwise proximity, and shared motifs across graphs.Graph-level transfer is additionally constrained by varying motif distributions and the limited expressiveness of standard GNN architectures.
  • Transferability: Transferability is difficult because shared structural patterns can have different meanings across domains, and message-passing GNNs are limited by the 1-WL test.Graphon theory, recurring substructures, and subtree representations are explored as possible foundations for defining transferable patterns.

9 Dataset Resources

Graph-learning benchmarks span diverse domains, structures, and task types, enabling evaluation of both generalizability and domain-specific performance. The resources cover node, link, and graph prediction across e-commerce, academia, knowledge bases, molecular science, temporal and social networks, brain graphs, and image-derived graphs.

  • Domains and Resources: Benchmark resources cover e-commerce, academic citation, knowledge-base, molecular, temporal, social, brain, and image-based graph domains.These datasets differ in scale, features, structural properties, and application scenarios.
  • Task Types: Node classification predicts node labels using node features and neighborhood structure, while link prediction estimates existing or future edges from proximity and feature compatibility.These tasks support applications including research-area classification, product categorization, recommendation, and knowledge-base completion.
  • Task Types: Graph classification predicts labels for entire graphs and requires representations that preserve local substructures and global topology.Molecular property prediction and social-network analysis are representative applications, often using hierarchical pooling.
  • Dataset Organization: Table 14 organizes datasets by domain, task type, structural properties, numbers of nodes, edges, and classes, and the presence of text attributes.This layout supports comparisons across graph scale, supervision targets, and structural characteristics.
  • Recent Benchmark Collections: Recent benchmark collections include 13 text-attributed benchmarks spanning large e-commerce graphs such as Products and smaller academic graphs such as Cora.Other resources provide text-enhanced knowledge graphs and temporal graph benchmarks for semantic link prediction and evolving-network tasks.

10 Open Questions

Open questions concern how to establish scaling laws, obtain sufficient and high-quality graph data, and evaluate GFMs beyond accuracy. They also include improving adaptation, resolving cross-domain transfer conflicts, and understanding robustness and trustworthiness in realistic graph settings.

  • 10.1 How to Enhance Scalability?: Existing GFMs lack the clear scaling trend observed in LLMs, motivating better graph backbones, pretraining objectives, and learning-instance choices.Open issues include GNN limitations, graph objectives focused on low-level semantics, and uncertainty over whether nodes, edges, or whole graphs should be scaled.
  • 10.2 How to Mitigating Data Scarcity?: Graph data scarcity may be addressed through automated collection, synthetic generation, and high-quality data curation.Graph diffusion models can augment structural diversity, while LLMs can generate synthetic text attributes for text-attributed graphs.
  • 10.2 How to Mitigating Data Scarcity?: Graph-data quality remains difficult to evaluate because datasets are often incomplete and quality depends on the selected backbone architecture and pretraining strategy.The survey identifies graph data valuation and quality metrics as open research problems.
  • 10.3 How to Better Evaluate GFMs?: Effective GFM evaluation requires large-scale, real-world-aligned benchmarks spanning diverse graph structures, tasks, and supervision levels.Existing benchmarks may lack transformative applications, practical realism, consistency, and reproducibility.
  • 10.3 How to Better Evaluate GFMs?: Evaluation should extend beyond accuracy to generalization across domains, robustness to adversarial and noisy data, and trustworthiness in high-stakes applications.This requires metrics and benchmarks that explicitly test these dimensions.
  • 10.4 Adaptation and Applications: Advanced adaptation aims to reduce task-specific prompt tuning and enable zero-shot or in-context learning for graphs.The survey points to autoregressive graph-task handling as a possible route toward adaptation without explicit task-specific modification.
  • 10.5 Advanced Theoretical Understandings: Transferability lacks fully developed theoretical explanations, and pattern conflicts arise when identical structures carry different meanings across domains.Triangles may indicate social stability but molecular instability, potentially degrading learned representations.
  • 10.5 Advanced Theoretical Understandings: Robust and trustworthy GFMs must account for long-tail distributions, incompleteness, class imbalance, limited labels, and structural alterations.Stability under structural distribution shifts and adversarial perturbations is identified as a promising direction.

11 Conclusion

The survey unifies Graph Foundation Models through a modular framework and classifies them by generalization scope. It also identifies future priorities for scalable, interpretable, adaptable graph intelligence.

  • The survey decomposes GFMs into backbone architectures, pretraining strategies, and adaptation mechanisms to unify diverse approaches.
  • GFMs are categorized as universal, task-specific, or domain-specific according to their intended generalization scope.Universal models target broad transfer across tasks and domains, whereas task-specific and domain-specific models focus on narrower objectives or application areas.
  • The survey examines design principles, representative methods, comparative strengths and limitations, and theoretical issues including expressiveness, transferability, and generalization.
  • Open challenges include scaling GFMs to massive graphs, integrating multimodal signals, establishing principled evaluation protocols, and explaining transferability and generalization theoretically.
  • Future research should develop more scalable, interpretable, and adaptable architectures, broaden pretraining corpora, and advance theoretical frameworks explaining GFM behavior.The survey envisions GFMs as foundational infrastructure for general-purpose graph intelligence across scientific, industrial, and decision-making settings.
Loading 2505.15116v1…