A11y-Compressor Framework Boosts GUI Agent Efficiency
A11y-Compressor has been developed by researchers as a framework aimed at improving the efficiency of GUI agent observations through the conversion of linearized accessibility trees into more compact, structured forms. The implementation, known as Compressed-a11y, utilizes a streamlined pipeline that features modal detection, redundancy elimination, and semantic organization. When evaluated using the OSWorld benchmark, Compressed-a11y achieved a reduction of input tokens to 22% of their original count, while also enhancing task success rates by an average of 5.1 percentage points. This initiative tackles issues related to redundancy and the absence of spatial relationship data in conventional accessibility tree formats.
Key facts
- A11y-Compressor transforms linearized accessibility trees into compact structured representations.
- Compressed-a11y uses modal detection, redundancy reduction, and semantic structuring.
- Tested on OSWorld benchmark.
- Reduces input tokens to 22% of original.
- Improves task success rates by 5.1 percentage points on average.
- Addresses redundancy and lack of spatial relationships in accessibility trees.
- Published on arXiv under Computer Science > Computation and Language.
- arXiv ID: 2605.00551.
Entities
Institutions
- arXiv