TL;DR: The official ScreenSpot-Pro board on Hugging Face (likaixin/ScreenSpot-Pro) lists 33 rows for 32 distinct models. On Nov 20, 2026 none of the 33 is marked verified, 29 come from pull requests that are not on the main branch, and 19 are copies of numbers from a project-run leaderboard site. The top rows are not measured the same way: one model appears twice (rank 3 with a Python tool, rank 4 without), and those two rows are 8.2 points apart. Read the board as a list of reported settings, not as one ranking.
What is ScreenSpot-Pro and what does the Hugging Face board show?
ScreenSpot-Pro is a GUI grounding benchmark. A model gets a high-resolution screenshot of a professional application (IDEs, creative suites, CAD tools, scientific software, office suites) plus an instruction, and has to point at the right element. Scoring is a hit or a miss on the predicted click point. The note on the rank 1 row states 1,581 examples, and I use that as the test size below.
The Hugging Face board is generated from .eval_results/*.yaml files in model repositories. Anyone can add such a file in a pull request, and the board reads the value, an optional source and an optional notes field. I pulled the board JSON from the public endpoint:
import json, urllib.request url = "https://huggingface.co/api/datasets/likaixin/ScreenSpot-Pro/leaderboard" rows = json.load(urllib.request.urlopen(url)) print(len(rows), len({r["modelId"] for r in rows})) # 33 32 print(sum(r["verified"] for r in rows)) # 0 print(sum(1 for r in rows if r.get("pullRequest"))) # 29 That is the whole audit surface: no private data, just what the board itself exposes. Everything below was computed from that response on Nov 20, 2026, and the numbers will move as rows are added.
How many of the 33 rows are verified?
None. The verified flag is false on all 33 rows. That pattern is common on the newer boards (the IFBench and DeepSWE boards I looked at earlier this week looked the same), but it matters more here because provenance is so uneven.
| Provenance | Rows |
|---|---|
| Open pull request (not on main) | 29 |
| On the main branch | 4 |
| Source is the project's own leaderboard site | 19 |
| Source is a paper (Table 6, Table 3, Table 9, abstract) | 6 |
| Source is a model card | 8 |
Two things stand out. First, 19 of the 33 rows cite gui-agent.github.io/grounding-leaderboard as their source. These are not new measurements. They are numbers copied from a project page onto the board, mostly for earlier models (Qwen2.5-VL, UI-TARS, UI-Venus, GTA1, Aguvis, CogAgent, MiniCPM-V-2, Qwen-VL). That is a useful historical reference, but it means a large share of the board has no run attached that a reader can replay from the board alone. Second, only 4 rows are on main. Open pull requests can be edited or closed by their authors, so a row that exists today can change tomorrow.
Why are rank 3 and rank 4 the same model?
Rank 3 is thinkingmachines/Inkling-2-Small at 92.0, and its note says "With Python tool." Rank 4 is the same model at 83.8, and its note carries only the image preprocessing detail: images with a long side over 4,096 px are downsampled to that limit, with aspect ratio kept and ground truth rescaled.
So the board holds two settings of one model, and the gap between them is 8.2 points, or about 130 of the 1,581 examples. The board ranks them as if they were two competitors. A reader comparing rank 3 with rank 5 (75.4) is comparing a tool-assisted number with a plain one, and there is no column to tell them apart.
The top rows show the same issue from a different angle. Ranks 1 and 2 (beipo-ai/NorthCUA-Pro at 93.3 and NorthCUA-Mini at 92.4) carry a note that reads: direct inference on the full screenshot, all 1,581 examples, no iterative zoom or tool-assisted refinement. Rank 6 (NorthCUA-Nano, 72.9) is from the same paper table. Those three rows are the most explicit about their protocol on the board, which makes them easy to interpret. Rank 3 states tool use, a different protocol again. Ranks 1 to 4 therefore carry three different protocol descriptions, and only two of the four say "full screenshot, no zoom".
Does a 0.9-point lead at the top mean anything?
Rank 1 leads rank 2 by 0.9 points. On 1,581 examples that is about 14 examples. A rough binomial 95% half-width at n=1,581 is 1.23 points at a score of 93.3 and 1.31 at 92.4, so the two intervals overlap heavily:
import math n = 1581 def half_width(p): # p in [0,1], normal approximation return 1.96 * math.sqrt(p * (1 - p) / n) * 100 for v in (93.3, 92.4, 92.0, 83.8, 75.4, 70.4, 65.6): print(v, round(half_width(v / 100), 2)) # 93.3 1.23 | 92.4 1.31 | 92.0 1.34 | 83.8 1.82 | 75.4 2.12 | 70.4 2.25 | 65.6 2.34 This is an illustrative sampling band, not a claim about any lab's uncertainty. It treats the 1,581 examples as independent draws, which understates real uncertainty because the examples come from a limited set of applications and screenshots. Even with that generous assumption, ranks 1 to 3 (93.3, 92.4, 92.0) sit within about 1.3 points of each other, and each band is wider than the gap to its neighbor.
The middle of the table is tighter still. Ranks 7 to 10 run from 70.4 down to 65.6, a 4.8-point spread, about 76 examples, while each row has a half-width above 2.2 points. Ranks 7 and 8 (70.4 and 70.3) differ by one tenth of a point, which is under two examples.
Do the stated numbers match a count out of 1,581?
A score is a count of hits divided by 1,581 (if the whole test set was used). I checked which reported values can be reproduced by some whole count at the precision shown. Values at one decimal can almost always be matched, because a 0.1-point step is about 1.6 examples. Four rows are reported as whole numbers (92, 62, 48 and 45), which is coarse rounding and not a problem in itself.
One row does not fit. Rank 25 (ProKope-AI/ProKope-VL-7B) reports 28.33, and its note says "0-shot Accuracy: 28.33%". The nearest counts out of 1,581 give 28.27 (447 hits) and 28.34 (448 hits), so 28.33 is not reachable. That suggests the row was scored on a different number of examples, on a subset, or by a different rule. The row sits near the bottom and does not change any ranking above it, but it shows the board accepts a number without the test-set size attached.
for hits in (447, 448): print(hits, round(hits / 1581 * 100, 2)) # 447 -> 28.27, 448 -> 28.34 What should you do with this board?
Here is the checklist I use before I rely on a grounding board:
- Group by protocol, not by rank. Split rows into "full screenshot, no tools", "with tools or zoom" and "unknown". Compare only inside a group.
- Check the source column. Paper tables, model cards and a project site are three different levels of evidence. A copy of a project-site number is a citation, not a run.
- Check pull-request status. A row from an open PR is a claim waiting for review.
- Convert gaps to examples. Multiply the gap in points by 15.81 to get examples out of 1,581. A gap under 20 examples is inside the noise for most reasonable models of uncertainty.
- Look for duplicate models. The same repository can appear under several settings, so the row count is not the competitor count (33 rows, 32 models here).
- Read the notes field. The facts that decide comparability (tool use, zoom, downsampling, subset) live there, and 26 of the 33 rows have no note at all.
If you are choosing a grounding model for a product, use this board to build a shortlist. The decision should come from a replay on your own applications and screen resolutions, since professional-software screenshots at 4K are a particular regime and your UI may not look like the benchmark's.
FAQ
Is the ScreenSpot-Pro board on Hugging Face official?
It is the official benchmark board attached to the likaixin/ScreenSpot-Pro dataset. "Official" describes the board feature, not a review of each row. None of the 33 rows is marked verified.
How many examples does ScreenSpot-Pro have?
1,581, according to the note on the rank 1 row. I use that as n in the sampling calculation.
Why are there 33 rows but 32 models?
thinkingmachines/Inkling-2-Small appears twice, once with a Python tool (92.0) and once without (83.8).
Are the top scores from the same protocol?
No. Ranks 1 and 2 state full-screenshot inference with no zoom or tool refinement, rank 3 states tool use, and most other rows state nothing.
Can I trust a 0.1-point difference between two rows?
No. At n=1,581, 0.1 points is under two examples, and the rough 95% band is more than 2 points wide for mid-table rows.
Will these numbers change?
Yes. 29 of 33 rows come from open pull requests, so rows can be added, edited or removed. Treat the figures here as the snapshot of Nov 20, 2026.
Links
- Board endpoint: https://huggingface.co/api/datasets/likaixin/ScreenSpot-Pro/leaderboard
- Dataset: https://huggingface.co/datasets/likaixin/ScreenSpot-Pro
- Project leaderboard site cited by 19 rows: https://gui-agent.github.io/grounding-leaderboard/
- NorthCUA paper cited by ranks 1, 2 and 6: https://arxiv.org/abs/2611.02102

