barestateAtom on ScreenSpot-v2.
A compact model for finding the right control on a screen. Barestate measured 75.94% one-shot grounding accuracy across 1,272 targets.
Two configurations, reported separately.
| Configuration | Correct targets | Accuracy |
|---|---|---|
| One pass | 966 / 1,272 | 75.94% |
| Two passes with ReGround | 984 / 1,272 | 77.36% |
The frozen baseline was reproduced by Barestate on 25 September 2026. Every item is included, with errors counted as misses. There were no inference errors in these runs. Success means the predicted point falls inside the target box.
What was evaluated.
The full ScreenSpot-v2 dataset includes 334 desktop, 437 web, and 501 mobile targets. The dataset revision is 5efbb1f1b5463a575f2eb7bc30fe29e49c15f93c. Each item supplies a screenshot and an instruction identifying one control. The score measures target location; it does not measure complete computer-task success.
A small local package.
The complete package is 169,049,084 bytes, or 161.2 MiB, including weights, tokenizer, processor, and configuration. The native Apple Silicon runtime processes grounding locally. A separate planner or external agent connection has its own data boundary.
Leaderboard context.
Checked on 4 October 2026, the public ScreenSpot-v2 leaderboard contains 28 entries, with 19 scores above 75.94%. Atom’s measured score would fall in the top-20 range if included in that table. This is a comparison with published scores; Atom is not currently listed or independently certified by the leaderboard.