Tag
The paper introduces PAGE, a partition-aware gated KV-cache eviction method that uses a scalar metric to predict input classes and apply eviction only when safe, reducing accuracy degradation in large language models.