In the Transformer architecture, the primary mechanism that connects the encoder and decoder is called ____. This mechanism allows for parallelization and has improved the efficiency of training models compared to traditional methods.

The Transformer’s encoder and decoder are linked by the attention mechanism, specifically multi‑head attention. This lets each token look at every other token at once, so the whole sequence can be processed in parallel. Because the calculations are independent, the model trains faster and uses less memory than older recurrent approaches. For example, when translating a sentence, the decoder can attend to all encoder outputs simultaneously, speeding up both training and inference. Thus, attention is the key innovation that makes Transformers efficient and scalable.