Current long-context language models struggle to generate coherent and instructionally-compliant long-form text, despite their ability to process extended input sequences.
