Papers target KV cache compression for long-context LLM inference

Four new arXiv papers attack the KV-cache memory bottleneck for long-context LLMs from different angles: attention-preserving vector quantization, INT2 rotation-based quantization, anchor-residual compression, and online compaction for agents. Community KLD benchmarks on Qwen 3.6 27B and Gemma 4 31B report KVarN 6-bit beating q8_0, with the precision tail dominating quality.