Panel: On Linear Representations and Pretraining Data Frequency in Language Models When Attention Sink Emerges in Language Models: An Empirical View Common Functional Decompositions Can Mis-attribute Differences in Outcomes Between Populations U-shape