EvoGUI-Bench - Inverse Action and Value: leaderboard

Metric: Joint action and value accuracy (%; share of 1,000 consecutive screenshot pairs for which the model predicts both the click, type or select action and its typed or selected value; on EvoGUI-Bench, 3,000 zero-shot visual question answering probes mined from Mind2Web and WebLINX web trajectories across 120 domains, screenshots at max_image_side 512, temperature 0; higher is better). Source: arxiv.org. Saturation forecast: Around 2031. 28 models tracked.

Top models

#ModelScore
1Gemini 3 Flash (Preview)65.4
2Qwen 3.5 397B A17B56.6
3Qwen 3.6 35B A3B56.6
4Qwen 3.5 122B A10B56.4
5Doubao-Seed-2.0-Pro-26021555.7
6Qwen 3.6 27B55.1
7Qwen 3 VL 235B A22B55.1
8Qwen 3.5 35B A3B54.8
9Qwen 3.5 27B54.8
10MiMo-V2.554.5
11GLM-5V Turbo54
12Kimi K2.552.8
13Kimi K2.652.6
14GPT-5 Mini52.5
15GLM-4.5V52.3

Interactive version: theaggregate.ai/benchmark?slug=evogui-bench-inverse-action-and-value · How It Works · Data refreshed daily, snapshot 2026-09-29.