GFMate: Empowering Graph Foundation Models with Test-time Prompt Tuning

The key novelty is performing prompt adaptation at inference time using only unlabeled target data, rather than relying on labeled auxiliary tasks or pre-training artifacts. This sidesteps a common trap: methods that work well on the source domain often encode assumptions that break when graphs have different structures or properties.

This work sits alongside a broader shift toward specification-driven, geometry-preserving adaptation methods. GPart (from May 14) replaced LoRA's low-rank bottleneck with isometric partitioning to preserve optimization geometry; GFMate takes a parallel approach for graphs by removing source-domain coupling entirely. Both papers reject the assumption that efficient adaptation must accept fidelity trade-offs. The inventory control paper (also May 14) shows a related pattern: offline meta-training paired with online decision-making. Here, GFMate uses pre-trained graph representations but adapts them purely through unlabeled target inference, avoiding the label scarcity problem that plagues few-shot transfer.

If GFMate maintains performance gains on heterogeneous graphs (e.g., citation networks, e-commerce graphs, molecular structures) without retraining the base model, that confirms the approach generalizes beyond the benchmark used in the paper. If a follow-up shows the method works across different foundation model architectures (not just the one tested), that would signal genuine architecture-agnostic design.