Recent research sheds light on the strengths and weaknesses of encoder-decoder and decoder-only models architectures in machine translation tasks.