ScreenSpot-Pro — Results
| Rank | Model | Score |
|---|---|---|
| 1 | muse-glimmer | 75.4 |
ScreenSpot-Pro
1 phaseActive
High-resolution professional GUI grounding benchmark spanning 23 apps across 5 industries and 3 OSes. CC BY 4.0. Metric: % click-region accuracy.
Overview
ScreenSpot-Pro
Quick answer: ScreenSpot-Pro tests multimodal models' ability to ground GUI instructions (click the right element) on authentic high-resolution screenshots from 23 professional applications across 5 industries and 3 operating systems. It's substantially harder than earlier GUI-grounding benchmarks — the best model in the original paper reached only 18.9% accuracy.
What ScreenSpot-Pro Measures
ScreenSpot-Pro evaluates the GUI grounding capability of multimodal models: given a natural-language instruction and a screenshot, can the model correctly localize the target UI element? Unlike general-purpose GUI benchmarks built on consumer web/mobile apps, ScreenSpot-Pro uses authentic high-resolution screenshots from specialized professional software — spanning CAD, creative, scientific, and other domains — where small target sizes, dense layouts, and unfamiliar interfaces make grounding meaningfully harder.
The paper found existing GUI grounding models perform poorly on this dataset (best baseline: 18.9%), and introduced ScreenSeekeR, a visual search method that narrows the search area using a planner's GUI knowledge, achieving a new state of the art of 48.1% without additional training — evidence that professional, high-resolution GUI grounding remains an open problem even for strong general-purpose models.
Benchmark Specifications
| Field | Value |
|---|---|
| Task category | Reasoning (multimodal GUI grounding) |
| Metric | % accuracy (correct click-region localization) |
| Coverage | 23 applications across 5 industries and 3 operating systems |
| Saturation | Low |
| Created by | Li et al. |
| Source paper | Li et al. 2025 |
| Leaderboard | gui-agent.github.io/grounding-leaderboard |
How ScreenSpot-Pro Is Scored
Models are scored on the percentage of instructions where the predicted click location falls within the correct UI element's bounding region, evaluated across authentic high-resolution screenshots. Because targets in professional software are often small and visually dense, scores in the 40-75% range represent strong performance for current frontier multimodal models, well below saturation.
ScreenSpot-Pro on Benchgen
No Benchgen results yet — be the first to run ScreenSpot-Pro.
ScreenSpot-Pro vs Other Benchmarks
| Benchmark | What it tests | Saturation |
|---|---|---|
| ScreenSpot-Pro | High-resolution professional GUI element grounding | Low |
| OSWorld-Verified | End-to-end real desktop/web computer-use tasks | Low |
| OmniDocBench | Document parsing (PDFs, scanned text) | Low |
ScreenSpot-Pro isolates the grounding step (where to click) rather than full task completion, complementing end-to-end computer-use benchmarks like OSWorld-Verified.
Frequently Asked Questions
What is ScreenSpot-Pro?
ScreenSpot-Pro is a benchmark testing multimodal models' ability to ground natural-language GUI instructions to the correct on-screen element, using authentic high-resolution screenshots from 23 professional applications.What does a good ScreenSpot-Pro score look like?
The original paper's best baseline reached only 18.9% accuracy; a specialized visual-search method (ScreenSeekeR) reached 48.1% — scores above 40% represent strong performance on this still-unsaturated benchmark.Who created ScreenSpot-Pro?
ScreenSpot-Pro was created by Kaixin Li, Ziyang Meng, and collaborators, published in April 2025.Is ScreenSpot-Pro saturated?
No — even specialized methods reach under 50% accuracy, leaving substantial headroom for improvement in professional GUI grounding.Benchmark definition based on Li et al., "ScreenSpot-Pro: GUI Grounding for Professional High-Resolution Computer Use" (arXiv:2504.07981). Last updated 2026-08-10.