arXiv:2607.18201 2026 Architecture 1 ideas extracted · analyzed Aug 30, 2026
The paper gives a constructive robust-root procedure for a latent tree observed after random graph noise: retain vertices with unusually large degree, take the largest connected component of this high-degree subgraph, and rank candidates by Jordan centrality. The transferable asset is not the random recursive-tree model itself, but the combination of statistically motivated thresholding, connectivity-based denoising, and a graph-geometric center that is stable under sparse perturbations. In neural networks, this can become a noise-robust anchor or positional encoding for graph transformers and message-passing networks, with the anchor computed once before training and used to orient or gate information flow.
Preprocess a noisy input graph into a high-degree core and compute a Jordan-center anchor in that core. Feed each node its distance to the anchor, and optionally use the anchor to bias graph-transformer attention; the hypothesis is that this suppresses spurious low-degree noise and gives the network a stable global coordinate system.