KV, a low-rank KV cache compression method achieving up to 20x reduction, with the paper selected as a Spotlight at ICML 2026 ...