Sharing the same parameters between different parts of a model, such as input and output embeddings, to reduce model size.