Attention Retrieves, MLP Memorizes: Disentangling Trainable Components in the Transformer

Re-assign community