EvoGUI-Bench: leaderboard
EvoGUI-Bench (Tsinghua University, 2026) tests whether vision-language models understand how GUI actions change interface states, with three probes mined from Mind2Web and WebLINX web trajectories. The EvoGain index averages each probe gain over random choice as a share of the headroom above random, times 100: 0 is random and 100 perfect; higher is better.
Metric: EvoGain index (mean over the three probes of the gain over the random baseline normalized by the headroom above it, times 100: temporal ordering exact-match accuracy, inverse action and value joint accuracy and one-step successor accuracy; 100 is perfect, 0 is random and below-random performance is negative, down to about -42; 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 August 2028. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 60.4 | #144 |
| 2 | Doubao-Seed-2.0-Pro-260215 | 49.8 | #152 |
| 3 | Kimi K2.6 | 43.9 | #178 |
| 4 | Qwen 3.5 122B A10B | 43.8 | #331 |
| 5 | GPT-5.4 | 43.4 | #123 |
| 6 | MiMo-V2.5 | 42.4 | #354 |
| 7 | Qwen 3.6 35B A3B | 42 | #462 |
| 8 | GLM-5V Turbo | 41.9 | #344 |
| 9 | Kimi K2.5 | 41.6 | #256 |
| 10 | Qwen 3.5 35B A3B | 40.5 | #482 |
| 11 | Qwen 3.5 27B | 39.6 | #359 |
| 12 | Qwen 3.5 397B A17B | 39.5 | #246 |
| 13 | Qwen 3.6 27B | 38.5 | #341 |
| 14 | Qwen 3 VL 235B A22B Instruct | 37.5 | #522 |
| 15 | GLM-4.5V | 37.1 | #744 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evogui-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.