FineState-Bench: leaderboard

FineState-Bench (2026) shows a GUI agent a screenshot and asks it to locate a control and then point where an action sets the control to an exact target state, such as a slider value or a selected option. Exact-state success rate averaged over mobile, web and desktop screens, in percent; higher is better.

Metric: Exact-state success rate at interact (ES-SR@Int, %): the predicted operation point falls inside the state-changing region of the target control and the resulting state equals the instructed exact goal state (a slider value, toggle, selected option), mean of the mobile, web and desktop rates, on FineState-Bench (2,209 instances, 23 UI component types); single-point protocol on static screenshots, no localization hints; higher is better. Source: arxiv.org. Saturation forecast: Around April 2028. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1UGround-V1-7B22.8
2Gemini 2.5 Flash10#463
3OS-Atlas-Base-7B10
4Claude 3.5 Sonnet5.6#752
5ShowUI-2B4.6
6GPT-4o4.3#751
7Jedi-7B-1080p3.5
8CogAgent-9B-202412203.1

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=finestate-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.