Fetching the paper…

Discovering the Gems in Early Layers: Accelerating Long-Context LLMs with 1000x Input Token Reduction · Around