The field of large language models is moving towards improving inference efficiency and cache management. Recent developments focus on reducing memory overhead and computational costs associated with long-sequence inference. Techniques such as sparse attention mechanisms, dynamic KV cache placement, and query-aware unstructured sparsity are being explored to achieve this goal. These innovations have the potential to significantly boost throughput and reduce latency in various applications, including retrieval-augmented generation and streaming video understanding. Noteworthy papers in this area include: SamKV, which achieves significant reductions in sequence length without accuracy degradation, ZigzagAttention, which reduces latency and improves performance by enforcing exclusive retrieval or streaming heads, Cold-RL, which introduces a learned eviction policy for NGINX that outperforms classical baselines, Accelerating LLM Inference via Dynamic KV Cache Placement, which investigates dynamic KV cache placement to maximize aggregated bandwidth utilization, LEAD, which incorporates learned models within DHT structures to optimize range query performance, Rethinking the Potential of Layer Freezing, which provides a systematic solution to the challenges of layer freezing, SparK, which applies unstructured sparsity to prune KV cache channels, and StreamMem, which proposes a query-agnostic KV cache memory mechanism for streaming video understanding.