Handling extremely long documents remains a persistent challenge for large language models (LLMs). Even with techniques such as length extrapolation and sparse attention, models often suffer from ...