LOCKS Method Uses Per-Page SVD to Cut KV Cache Reads by 10x at 1M Context

A new technique called LOCKS addresses a core bottleneck in large-language-model inference, where loading the full key-value cache from GPU memory becomes prohibitively expensive at long contexts exceeding one million tokens. Unlike prior methods such as ShadowKV and RocketKV that rely on a single shared low-rank projection basis, LOCKS computes a separate SVD summary for every 16-token page of keys, capturing each page's unique subspace orientation. At each decode step, LOCKS scores all pages using only these compact summaries — storing roughly 10% of the original KV data — without reading the full K or V tensors, then routes attention only to the top-ranked pages. The paper includes formal proofs showing that any fixed shared projection has inherent blind spots, explaining why existing methods degrade at small token budgets rather than from poor tuning. Benchmark results indicate this per-page approach achieves approximately a 10x reduction in KV cache reads while maintaining strong coverage of the most relevant attention mass.