How do attention mechanisms solve the bottleneck problem in sequence$-$to$-$sequence models?

Group of answer choices

They increase the vocabulary size of the decoder softmax

They incorporate all of the encoder hidden states into the decoder softmax prediction at every time step of the decoder

They add interpretability to the model by allowing users to inspect probabilistic alignments

They dynamically increase the size of the hidden layers in the encoder depending on$$the$$time$$step$.$