Fetching the paper…

Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning · Around