# Contribution Statement

## Paper: HubKV: Redundancy-Penalized KV Cache Compression via Submodular Marginal Discounting

### One-Sentence Contribution
We reformulate KV cache compression as a score-weighted local coverage problem and propose HubKV, a hardware-parallel token ranking method that approximates submodular marginal gains through local max-pooling and head-wise selectivity calibration.

### Key Claims
1. Current Top-K eviction strategies optimize a modular objective, ignoring local semantic redundancy confirmed by high spatial autocorrelation.
2. Submodular maximization better models diminishing returns of redundant tokens through facility location objectives.
3. HubKV provides an efficient O(N·k) parallel proxy (SMD core) for submodular marginal gains, reducing local redundancy collapse in the cache.
4. Outperforms modular baselines in redundancy-heavy scenarios (long-code comprehension, summarization) without inference overhead.

### Target Venue
ACL 2025 (8 pages, long paper)

### Current Status
- Paper draft: 60% complete (Methods done, Experiments missing, Related Work missing)
- Experiments: Not yet run
- Code: kvpress library ready at /home/scm/Project/kv-llm