Proposing the In-context Autoencoder (ICAE) to compress long contexts efficiently in large language models.