Preprint
Aug 2026
TaskPress: Query-Agnostic KV Cache Compression via Task-Guided Pruning
TaskPress is introduced, a framework for task-guided, query-agnostic KV cache eviction, instead of optimizing the cache for a single query, which constructs a reusable memory representation conditioned on a high-level task guide.
Wonpyo Park, Seung-won Hwang
· 0 citations