Junlin Liu

PReM: Learning What to Preserve and When to Refresh for Context Compression

Bohan Yu, Lei Shen, Chenxi Zhou, Chen Han, Junlin Liu, Wenbo Su, et al.

Efficient long-context inference is not only about reducing memory cost, but also about keeping useful contextual evidence accessible as generation proceeds. However, existing compression-oriented approaches, such as key-value (KV) cache compression and context compression, often…Read more