A11y-Compressor cuts a desktop accessibility tree to 22% of its tokens and gains 5.1 points on OSWorld; phoneshell cut one iPhone read from 5.82s to 0.89s by dropping two attributes by hand

Both projects found that most of an accessibility tree is dead weight. A11y-Compressor learned a pipeline to remove it and got more accurate as a side effect. phoneshell dropped two named attributes by hand and only ever claimed the time back, never the accuracy, because it never measured it.

A11y-Compressor (arXiv:2605.00551) restructures a linearized accessibility tree into a form it calls Compressed-a11y and reports, on the OSWorld desktop benchmark, input tokens cut to about 22% of the original alongside an average 5.1 percentage point gain in task success. It has been tested on desktop Linux and Windows applications only; the authors state plainly that mobile interfaces have not been investigated. phoneshell's own tree reads, measured on a physical iPhone and recorded in FINDINGS.md section 2, got most of their cost back a smaller and more specific way: excluding the visible and accessible attributes from a WebDriverAgent read cut one heavy third-party screen from 5.82 seconds to 0.89 seconds, at a stated cost of one element out of 24 going undescribed. Separately, deduplicating unchanged screens across conversation turns cut a measured three-turn task from 18,584 to 1,278 cache-write tokens. Both projects pulled on the same lever, that most of what an accessibility tree exposes is redundant for a given step, but only one of them turned pulling it into a measured accuracy gain, because only one of them was set up to measure accuracy at all.

What changed, and when

A11y-Compressor was posted to arXiv in May 2026 under the title "A Framework for Enhancing the Efficiency of GUI Agent Observations through Visual Context Reconstruction and Redundancy Reduction." Its pipeline runs a linearized tree through modal detection, redundancy reduction and semantic structuring, in that order, rebuilding the spatial relationships a flat text dump throws away before handing the result to the agent. We found no public code or dataset release to accompany the preprint as of this writing, so the 22% and 5.1-point figures are the paper's own reported numbers, not independently rerun here.

phoneshell's attribute exclusion and screen-dedup changes predate that paper and are documented in FINDINGS.md sections 2 and 8, both committed before A11y-Compressor's submission. Neither change in phoneshell ran an accuracy comparison; both are wall-clock and token-cost measurements on the same tree content, read once with the expensive attributes and once without, on the same screen.

What each project actually measured

What this means for someone tuning a tree read

The honest reading is not "phoneshell should adopt A11y-Compressor's pipeline" or "A11y- Compressor's gain would vanish on iOS." Neither claim is supported by what either project has actually run. What is supported: an accessibility tree, on both platforms, carries enough redundant structure that cutting most of it costs little and in at least one documented case (A11y-Compressor, desktop, OSWorld) costs nothing and buys accuracy back as well. phoneshell's own number that should worry anyone copying its attribute exclusion wholesale is the one element out of 24 it could no longer describe. That tradeoff was accepted for a 6.5x time saving on a heavy screen without ever checking whether the missing element was ever the one a task needed. A project that wants A11y-Compressor's kind of result, cost down and accuracy up rather than cost down and accuracy unmeasured, would need to run the same paired comparison phoneshell ran for vision in experiment 2 (FINDINGS.md §26): same tasks, same model, tree with the excluded attributes against tree without, scored from device state. Nobody has run that comparison for phoneshell's own attribute exclusion yet.

Check it yourself

git clone https://github.com/Blomega/phoneshell && cd phoneshell
grep -A6 "visible.*and.*accessible" FINDINGS.md     # section 2, the 5.82s to 0.89s read
grep -A8 "Token cost" FINDINGS.md                     # section 8, the 18,584 to 1,278 token dedup
sed -n '1,40p' phoneshell/wda/client.py               # the attribute exclusion itself

A11y-Compressor's own pipeline has no public repository we could find as of this writing, so its 22% and 5.1-point numbers cannot currently be rerun outside the paper; the arXiv preprint (2605.00551) is the only available source for them.

What would prove this wrong

• If A11y-Compressor or a reimplementation is released and run against a WebDriverAgent tree dump from this repository, and task success drops rather than holds or rises, that would mean iOS's tree is already close enough to the compressed form that restructuring adds overhead instead of value. We predict a smaller gain than 5.1 points on iOS specifically because phoneshell's default depth-20 cap already removes much of the redundancy A11y-Compressor's redundancy-reduction stage targets on deeper desktop trees. Untested by us, falsifiable by anyone who runs it, through 31 December 2027.
• If a paired accuracy comparison on phoneshell's own 80 tasks shows the attribute-excluded tree and the full tree disagree on zero tasks, the one-element loss recorded in section 2 would be confirmed as immaterial to outcomes on this task set, and the caution above would be withdrawn.