“Transformer models exhibit significantly more superposition than vision models.”