In-context Autoencoder for Context Compression in Large Language Models

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