C2KV Framework for Position-Agnostic Compressed KV Cache Reuse
July 21, 2026
C2KV introduces a lightweight sidecar Extractor to enable non-prefix KV cache reuse by learning a position-agnostic, compressed manifold. This approach jointly optimizes KV extraction and inference-time concatenation to mitigate the memory and access bottlenecks of long-context serving.
HOW THIS AFFECTS YOU
●
builderYou can reduce long-context inference costs by reusing compressed KV caches without the typical accuracy degradation.
●
researcherThe position-agnostic manifold approach offers a new method for combining KV compression with non-prefix reuse.