Fast KV Compaction Makes Long Context LLMs Practical

Fast KV Compaction via Attention Matching shows how to compress LLM KV cache in seconds, not hours, while preserving long-context performance.