\begin{abstract}
Key-Value (KV) caching enables efficient autoregressive inference, but its memory cost grows linearly with context length.
Most score-based KV compression methods retain tokens via greedy Top-K selection, implicitly treating token utility as modular.
We show that high-scoring tokens are often spatially clustered and that attention heads differ in their score selectivity, which can cause Top-K retention to waste cache budget on locally redundant context.
Motivated by this observation, we formulate an ideal score-weighted local coverage objective with monotone submodular structure to model diminishing returns in locally redundant cache entries.
We then propose \textbf{HubKV}, a single-pass marginal-gain proxy that detects local score hubs, softly discounts neighboring tokens, and applies compression-ratio-gated head-wise calibration while preserving the parallel Top-K pruning interface of existing KV compression systems.
We validate HubKV on four backbones across 16 diverse tasks.
On Qwen3-8B, HubKV improves FastKVZip and KVZip by 3.23 and 6.53 average score points at 95\% prefill compression, and by +1.71/+1.87 points on average across budgets.
Decoding-stage experiments further show that the same score correction remains effective on reasoning datasets AIME25 and MATH.
\end{abstract}