GUI-KV: Efficient GUI Agents via KV Cache with Spatio-Temporal Awareness
GUI-KV is a plug-and-play method to compress Key-Value (KV) caches for GUI agents by exploiting spatial and temporal redundancy in screenshots. It uses spatial saliency guidance based on visual token importance and temporal redundancy scoring to select only the most informative tokens. This allows agents to maintain near full cache accuracy with modest budgets, significantly reducing decoding computational load.