research_project · DiggingBeagle record
PointerBench
GUI agent grounding benchmark with explicit coordinate-frame and target-category analysis.
- Entity kind
- research_project
Record description
PointerBench is a fine-grained graphical target localization benchmark consisting of 1,500 annotated screenshot-and-instruction items across 100 professional software applications. It distinguishes user-facing GUI targeting from instruction-following and tests coordinate transformations explicitly. It does not observe actual account permissions, side effects or task completion.
The 1,500-task static benchmark includes distinct targeting settings. Reported Fable 5 performance was 83.2% overall and 69.6% on text, while GPT-5.6 Sol scored 77.7% overall and 53.4% on text. Coordinate-frame conventions accounted for major raw-score changes for some other models; none of these percentages is a payment or GUI side-effect authorization rate.