KV² improves KV cache efficiency via selective reconstruction
October 5, 2026
KV² is a query-agnostic compression method that uses a lightweight proxy scorer to identify informative tokens, improving RULER 16K scores by over 40 percentage points at a 2% budget.
HOW THIS AFFECTS YOU
●
builderImplementing this could drastically reduce the cost of serving long-context models.
●
researcherThis method provides a high-accuracy alternative to standard query-agnostic compression.