The branching random walk in a uniform magnetic field : magnetization concentration and overlap distributions

What the math gives to ML

The paper develops a hierarchical Gibbs model in which leaf energies are sums of Gaussian increments along tree paths, giving an ultrametric covariance structure rather than independent scores. At low temperature, Gibbs mass becomes sparse and concentrates on a hierarchically organized set of leaves, while explicit Gaussian concentration controls free-energy fluctuations. This suggests replacing flat mixture-of-experts routing with a tree-structured stochastic router whose correlated logits encourage coarse-to-fine expert specialization. The most direct test is an equal-budget MoE comparison measuring loss, load balance, routing sparsity, and seed stability.