Graph neural networks (GNNs) are usually judged by their final-layer output, which can obscure how task-relevant information changes as messages propagate through the network. A new arXiv preprint argues that this black-box evaluation misses critical failure modes that occur earlier in the architecture.
The authors introduce a post-hoc diagnostic that freezes a trained GNN and fits probes at each layer. By measuring the entropy of these probes, the method reveals where and why the network loses or distorts the information needed for a given task. This layer-wise view could help practitioners identify whether a GNN is failing due to oversmoothing, poor feature propagation, or other intermediate bottlenecks.
Because the diagnostic is applied after training, it does not require retraining or architectural changes. The abstract does not yet report specific experimental results, but the approach offers a new way to audit GNN behavior beyond final accuracy metrics.