GTFormer proposes replacing Transformer self-attention with a hierarchical graph neural network, achieving zero-shot length ...