“Scaling law experiments confirm that the performance improvement from Attention Residuals (AttnRes) is consistent across different model sizes.”