“The Attention Residuals (AttnRes) architecture consistently outperforms the baseline Transformer architecture across all tested compute budgets.”