Training graph neural networks (GNNs) on small graphs often fails to transfer to larger ones, in part because the computation required for a correct answer may be deeper than the model's fixed architecture allows. A new preprint on arXiv studies this problem for tasks that share a common local iterative computation, suggesting that the effective dependency structure of the task matters for generalization.
The authors propose looped GNNs, which iterate beyond a fixed depth, allowing the model to perform additional computation at inference time. They also emphasize parameter-efficient adaptation, meaning the approach aims to adjust to new graph sizes without requiring a large number of new parameters. The abstract does not provide full architectural or experimental details, but the direction points to a way of decoupling model depth from the size of the input graph.
Because the announcement is brief, the precise mechanism and benchmark results are not yet available. The core claim, however, is that looped architectures can bridge the gap between training and test graph sizes more effectively than fixed-depth models.