Discovering the Gems in Early Layers: Accelerating Long-Context LLMs with 1000x Input Token Reduction Paper • 2409.17422 • Published Sep 25 • 24 • 5