Test-time Graph Extrapolation via Progressive Anchor-Guided Expansion

Graph generative models have made significant progress in synthesizing graphs with complex structural properties and strict validity constraints. However, even state-of-the-art models fail to extrapolate to larger graphs. When using models to generate graphs larger than those seen during training, statistical fidelity rapidly deteriorates, and validity approaches zero. Here, we propose Pro-Guid, a test-time extrapolation framework based on structure-aware graph anchoring and conditional synthesis. Starting from a seed graph, the method iteratively selects valid anchor subgraphs and conditionally synthesizes expansion structures that get merged back into the main graph. By constraining the size of anchor graphs and using a repaint-like approach to guide expansion, Pro-Guid avoids overburdening the model with processing overly large structures that degrade generation quality.