Since both encoder and decoder models are learned (meaning they can be re-trained), the same encoder or decoder architecture can be specialised for different tasks.