Graphical user interface (GUI) agents follows user instructions by observing screenshots and selecting actions such as clicking, typing, or scrolling. They are commonly based on Transformer architectures that combine visual tokens from screenshots...
Graphical user interface (GUI) agents follows user instructions by observing screenshots and selecting actions such as clicking, typing, or scrolling. They are commonly based on Transformer architectures that combine visual tokens from screenshots with textual tokens from instructions to predict the next action. The main efficiency challenge arises in multi-step tasks, where a goal requires a sequence of GUI operations instead of one isolated action. Such tasks require the model to remember prior screenshots, previous actions, producing long multimodal contexts and increases inference cost during both prefilling and key- value (KV) cache decoding. This thesis addresses the efficiency problem with a multi-stage context pruning framework for GUI agent inference. The method first removes less important historical screenshot blocks before exact prefilling using a Taylor-based importance score. It then applies token-level pruning and hybrid decoding to reduce decode-time KV-cache cost while preserving action- relevant history. Experiments on AGENTNETBENCH and OSWORLD show that the framework maintains performance under compressed KV-cache budgets. Specifically, on AGENTNETBENCH, even at only 20% KV retention, it achieves 78.3% step success with Qwen3VL 8B and 57.2% with EvoCUA 8B, substantially preserving action quality under severe compression. As the budget increases to 40% and 60%, performance rises to 81.2% and 82.0% for Qwen3VL 8B, and to 63.6% and 64.8% for EvoCUA 8B, approaching or slightly exceeding full-cache performance at higher retention. These results indicate that GUI agent contexts should be compressed according to interaction structure rather than treated as flat token sequences.