Motivation

Graph machine learning (GML) applies ML to structured data represented as nodes and edges, and underpins applications such as fraud detection, recommendation, drug discovery, weather prediction, and traffic forecasting. The prevailing practice trains a task-specific GNN per dataset: a data-hungry, compute-intensive workflow brittle to shifts in structure, features, or labels. In contrast, foundation models in language, vision, and audio transfer broadly with minimal tuning. 

Graph foundation models (GFMs) extend this paradigm to graphs: pre-train once, then adapt across node/edge/graph tasks, feature/label spaces, and dynamic or multimodal settings. Methodological explorations span GNN-based backbones, Transformers, and LLM pipelines. Industry momentum signals practical readiness and a shift toward graph-centric modeling. Despite progress, key gaps remain: 

1. Methodology: no consensus on model families, objectives, or pretraining for heterogeneous, dynamic, and multimodal graphs; 

2. Scaling: limited understanding of data regimes, generalization and transferability to (large) graphs; 

3. Evaluation: fragmented benchmarks, metrics, and protocols hinder fair comparison; 

4. Transfer: unclear pathways to adapt language/vision/table techniques and generalize across schemas, label spaces, and tasks. 

The goals of our workshop are to build a connected academia–industry community; Identify challenges and opportunities for GFMs, i.e. handling heterogeneity, scaling to massive graphs, and effective pretraining.; Standardize evaluation via shared datasets, metrics, protocols, and baselines.; and catalyze cross-domain transfer from language, vision, and tables, adapted to graph structure.

Call for Papers

We invite long papers (up to 8 pages) for mature, full-length contributions with extensive results and analysis and position papers (2–4 pages) offering concise, opinionated perspectives on defining/deploying GFMs, GFM–LLM/GFM-TFM integration, benchmarking, and lessons from other foundation model domains.

We will manage the paper submission through OpenReview, and the review process is double-blind. Please use the ICML 2026 LaTeX style files. We will select outstanding papers for oral talks, and the award for the best paper will be announced at the workshop. Any LLM involvement must be explicitly disclosed; human authors and reviewers remain fully responsible for all content. AI-generated papers will not be accepted.

Note: This workshop will be non-archival. Dual submissions with papers under review at other venues are permitted.

Important Dates

May 8th, 2026                       Paper Submission Deadline

    May 21st, 2026                          Author Notification

        July 10th, 2026                       Workshop@ICML

Schedule

Organizers