Unified retention and compensation for KV cache eviction via coverage calibration

CORE combines offline set-level log-determinant coverage distillation with a lightweight indexer that drives both retention and evicted-mass redistribution, outperforming baselines by up to 3.78 points at 90% compression.

Research Lab
Shuxin Liu · Qing Liu · Yi Du · Ou Wu

Hangzhou Institute for Advanced Study · University of Chinese Academy of Sciences

Research Digest··2 min read
Liu et al.

The authors start from an exact identity: eviction error equals evicted attention mass times the directional gap between the evicted centroid and the retained output.

Why this paper

From University of Chinese Academy of Sciences and Hangzhou Institute for Advanced Study

In one line

CORE unifies KV-cache retention and compensation by calibrating one allocation distribution that jointly selects Top-B states and writes latent memory from excluded mass.

What we could check

  • ·No code link found
  • ·No weights link found
  • ·No dataset link found
  • ·No compute details found
  • ·No stated limitations found
  • ✓Reports numbers on named benchmarks

Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.

§

Research Digest

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.