arXiv:2607.23354 2026 Architecture 1 ideas extracted · analyzed Aug 30, 2026
The paper defines a globally coupled weighting for a finite metric space by solving Zw = 1, where Z is an exponentially decaying distance-similarity matrix. The resulting coefficients and total magnitude provide a principled measure of geometric diversity: redundant points interact and can receive less aggregate influence, while isolated or boundary points can receive more. This can transfer to set, point-cloud, and graph neural networks as a geometry-aware pooling layer replacing mean, max, or independently scored attention. The main limitation is quadratic pairwise-distance cost and a dense linear solve, so exact experiments should begin with small sets and later test low-rank approximations.
Replace mean or max pooling over a set of learned element embeddings with pooling based on the metric-magnitude weighting. Pairwise distances create a globally coupled correction for redundancy, so geometrically isolated or boundary elements can contribute differently from dense clusters of nearly duplicate elements.