# ScreenSpot-v2 ![Category](https://img.shields.io/badge/Category-Reasoning-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Accuracy-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Medium-d29922) ![Created](https://img.shields.io/badge/Created-2024-555) [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://arxiv.org/abs/2401.10935) [![GitHub](https://img.shields.io/badge/GitHub-niuzaisheng%2FScreenSpot-181717?logo=github&logoColor=white)](https://github.com/niuzaisheng/ScreenSpot) [![Dataset](https://img.shields.io/badge/Dataset-HuggingFace-ff9d00?logo=huggingface&logoColor=white)](https://huggingface.co/datasets/likaixin/ScreenSpot) > **Quick answer:** ScreenSpot-v2 tests a multimodal model's ability to ground natural-language instructions to the correct on-screen element across mobile, web, and desktop GUI screenshots. It's a refined, re-annotated version of the original ScreenSpot benchmark introduced alongside SeeClick. LFM2.5-VL-3B averages 80.7% across platforms as of August 2026. ## At a Glance **What it tests:** Whether a vision-language model can locate and click the correct GUI element (button, icon, text field) given a natural-language instruction, across mobile, web, and desktop screenshots. **Why it matters:** GUI grounding accuracy is a prerequisite for autonomous computer-use and screen-navigation agents — a model that can't reliably locate the right element can't reliably complete multi-step UI tasks. **Known limitations:** As models have improved on the original ScreenSpot benchmark, scores have crept toward saturation; the harder, higher-resolution ScreenSpot-Pro benchmark was introduced separately to keep headroom for frontier models. ## What ScreenSpot-v2 Measures ScreenSpot-v2 is a refined and re-annotated version of the original ScreenSpot benchmark, which was introduced alongside the SeeClick GUI-grounding agent. It presents a model with a screenshot and a natural-language instruction (e.g., "click the settings icon") and scores whether the model's predicted click location falls within the correct on-screen element's bounding box. The benchmark is split across three platforms — Mobile, Web, and Desktop — each reported separately and then averaged. Because it draws directly from GUI grounding — the core skill needed for computer-use and screen-navigation agents — ScreenSpot-v2 is widely cited by vision-language model labs (including Liquid AI, Alibaba's Qwen team, and Google's Gemma team) as a proxy for real-world agentic screen-understanding capability. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Reasoning (GUI grounding) | | Metric | % accuracy (click-region grounding), per-platform and averaged | | Saturation | Medium | | Created by | Cheng et al. (re-annotated from the original ScreenSpot / SeeClick benchmark) | | Source paper | [Cheng et al. 2024 (SeeClick)](https://arxiv.org/abs/2401.10935) | | GitHub | [niuzaisheng/ScreenSpot](https://github.com/niuzaisheng/ScreenSpot) | | Dataset | [likaixin/ScreenSpot on HuggingFace](https://huggingface.co/datasets/likaixin/ScreenSpot) | ## How ScreenSpot-v2 Is Scored For each instruction, the model outputs a predicted click coordinate or bounding box; a prediction is scored correct if it falls within the ground-truth element's region. Scores are reported per platform (Mobile, Web, Desktop) and averaged into a single headline number. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [LFM2.5-VL-3B](/models/liquid-ai/lfm2-5-vl-3b) | 80.7% (avg) | [LFM2.5-VL-3B launch blog](https://www.liquid.ai/blog/lfm2-5-vl-3b) | 2026-08 | *Score sourced from Liquid AI's LFM2.5-VL-3B announcement, August 2026.* ## ScreenSpot-v2 on Benchgen > No Benchgen results yet — [be the first to run ScreenSpot-v2](/benchmarks/cheng-et-al/screenspot-v2). ## ScreenSpot-v2 vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **ScreenSpot-v2** | GUI element grounding across mobile/web/desktop | Medium | | [ScreenSpot-Pro](/benchmarks/li-et-al/screenspot-pro) | High-resolution GUI grounding in 23 professional apps | Low | ScreenSpot-Pro was introduced as a deliberately harder, higher-resolution successor once ScreenSpot-v2 scores began climbing toward saturation for frontier models. ## Run ScreenSpot-v2 on Your Model Benchgen lets you evaluate your own vision-language model on GUI grounding tasks and track screen-understanding accuracy across model versions. ## Frequently Asked Questions

What is ScreenSpot-v2?

ScreenSpot-v2 is a GUI element grounding benchmark that tests whether a vision-language model can locate the correct on-screen element given a natural-language instruction, across mobile, web, and desktop screenshots.

What does a good score look like on ScreenSpot-v2?

LFM2.5-VL-3B reports an 80.7% average as of August 2026; scores above 80% are considered strong for current small vision-language models.

Who created ScreenSpot-v2?

ScreenSpot-v2 is a re-annotated version of the original ScreenSpot benchmark introduced alongside the SeeClick GUI agent (Cheng et al., 2024).

Is ScreenSpot-v2 saturated?

It's approaching saturation for frontier models, which is why the harder ScreenSpot-Pro benchmark was introduced as a successor with more headroom.

How does ScreenSpot-v2 differ from ScreenSpot-Pro?

ScreenSpot-v2 covers general mobile/web/desktop screenshots at standard resolution; ScreenSpot-Pro is a harder, higher-resolution benchmark spanning 23 professional applications.

Rank	Model	Score
1	lfm3-vl-2b	94.6
2	lfm2-5-vl-3b	80.7

ScreenSpot-v2 — Results

ScreenSpot-v2 — Phases

ScreenSpot-v2