In encoder-decoder architectures, the outputs on the encoder blocks act as being the queries towards the intermediate representation of your decoder, which supplies the keys and values to determine a representation in the decoder conditioned over the encoder. This attention is called cross-aware