Researchs | 刘宇哲

Every Cache Entry Earns Its Place: Global Allocation of Resolution and Coverage for KV Cache Compression

As a co-first author, I helped develop GraceKV, which formulates KV cache compression under a fixed cache budget as a global resource allocation problem, adaptively balancing information coverage and local resolution across layers, KV heads, and context positions. It requires no additional training, runs entirely on GPU, and retains useful task behavior even in a 128x stress test, though the loss varies by task.

Project page Paper

ACCT: Asymmetric Cross-LLM Context Transfer

As a principal researcher, I helped develop ACCT, which enables KV cache transfer between models of different scales within the same model family. It selects query-relevant context tokens, allocates a shared KV budget across target layers and heads, and transforms source KV into sparse target KV, enabling the target model to reuse context computation without reprocessing the original context.

Paper