tldr: more residual connections-more efficient networks;
but residuals and residual selection matrix have to be normalized otherwise they can overwhelm the useful signal mid-training and get stuck there
https://arxiv.org/abs/2512.24880