How AutoGFM Redefines Graph Foundation Models for Multi‑Task, Multi‑Domain Performance

Background

Graph Foundation Models (GFMs) aim to learn a universal graph representation that can be applied to node‑level, edge‑level and graph‑level tasks across domains. Recent GNN‑based GFMs have demonstrated strong cross‑domain knowledge sharing.

Limitations of Existing GFMs

Current GFMs rely on manually designed, fixed GNN architectures. Experiments on typical GNNs (GCN, GAT) across several benchmark datasets show that the optimal architecture varies significantly with the task and data distribution. Fixed architectures therefore limit performance and generalization in heterogeneous graph scenarios. Moreover, differentiable architecture search methods such as DARTS encounter conflicts when multiple tasks require divergent architectures.

AutoGFM: Adaptive Architecture for Graph Foundation Models

AutoGFM introduces three mechanisms to automatically customize GNN architectures for each dataset while preserving a shared foundation.

Decoupled Contrastive Graph Encoder : a sub‑graph‑level discriminative contrastive learning objective that extracts invariant patterns (shared across datasets) and variant patterns (dataset‑specific) from diverse graphs.

Invariant‑Pattern‑Guided Architecture Customization : the invariant patterns are used as signals to condition a neural architecture search, producing a GNN topology (e.g., layer types, aggregation functions, hidden dimensions) tailored to the target dataset.

Curriculum‑Learning‑Based Architecture Constraint : a curriculum schedule gradually increases the weight of dataset‑specific loss, preventing any single dataset from dominating the search and encouraging balanced optimization.

The overall workflow is illustrated in the diagram (Image 1).

Experimental Setup

AutoGFM was evaluated on multiple benchmark tasks covering node classification, link prediction and graph classification across domains such as citation networks, social graphs and molecular graphs. For each task, the architecture discovered by AutoGFM was compared with standard GCN, GAT and manually engineered baselines. All models were trained with the same optimizer (Adam), learning rate 0.001, batch size 32, and early‑stopping based on validation loss.

Results

Across all datasets, AutoGFM‑derived architectures achieved higher accuracy (or AUC) than the baselines, with average improvements of 2–5 % and better generalization to unseen graphs. The contrastive encoder contributed to more stable invariant representations, while the curriculum constraint reduced over‑fitting to dominant datasets. Detailed numbers are shown in Image 2.

Conclusion and Future Work

AutoGFM demonstrates that adaptive architecture search guided by invariant graph patterns and curriculum learning can substantially improve the performance of graph foundation models in heterogeneous multi‑task settings. The method is applicable to downstream scenarios such as academic collaboration network analysis, knowledge‑graph modeling and drug‑molecule property prediction. Future directions include extending the framework to cross‑modal fusion, open‑task adaptation and scalable training on billions of edges.

OpenReview paper: https://openreview.net/forum?id=fCPB0qRJT2

Code example

来源：专知
本文
约1000字
，建议阅读
5
分钟
显著提升图基础模型在多任务、多领域场景下的性能与泛化能力。