A repeated execution of a shared transformer block that progressively refines predictions across iterations.