Uniform Scaling Limits in AdamW-Trained Transformers

Collection of 11 new stat‑ML preprints (arXiv 13 May 2026) covering theoretical and methodological advances: provable uniform scaling limits for transformers trained with AdamW (forward–backward ODE/MVODE limits and uniform-initial-condition convergence rates); an interpretable SO(3)-parameterised rotationally anisotropic Gaussian process kernel; Thompson-sampling + Gibbs algorithm for adaptive experimentation under unknown network interference with regret bounds and graph-discovery analysis; the Spatial Adapter — a parameter-efficient post-hoc layer that furnishes closed-form low-rank-plus-no