Cache Me If You Can: How Many KVs Do You Need for Effective Long-Context LMs?

arXiv preprint, 2025