PolyKV: Heterogeneous Retention and Allocation for KV Cache Compression

In three linesPolyKV optimizes KV cache compression by applying heterogeneous strategies per transformer layer instead of uniform policies. On LLaMA-3.1-8B and Qwen3-8B with 512-token KV budget, PolyKV recovers 54.5% and 25.7% of the LongBench performance gap versus FullKV.

Summary generated by Claude — human-verified