“The Attention Residue architecture computes a layer's output by applying an attention mechanism over the hidden states of all previous layers, not just the immediately preceding one.”