August 12, 2026 CapKV: Rethinking KV Cache Eviction via a Unified Information-Theoretic Objective 基于信息瓶颈原理统一解释KV缓存驱逐方法，提出CapKV——通过统计杠杆分数最大化对数行列式容量目标，显著提升长上下文推理的性能-内存权衡 kv cache eviction information bottleneck mutual information log-determinant statistical leverage score long-context inference transformer capacity-aware CapKV