A11y-Compressor: 78% Token Reduction for GUI Agents with +5.1pp Task Success
arXiv·medium signal
Transforms linearized accessibility trees into compact structured representations via modal detection, redundancy reduction, and semantic structuring. Experiments on the OSWorld benchmark show Compressed-a11y reduces input tokens to 22% of the original while improving task success rates by 5.1 percentage points on average. Addresses the core bottleneck that accessibility trees — the primary text-based observation format for GUI agents — suffer from redundancy and lack spatial relationship information.