UI2App: leaderboard
UI2App (Hong Kong University of Science and Technology, 2026) gives a vision-language model uncaptioned screenshots of a real multi-route web application, 45 sets and 327 screenshots in all, and asks it to generate a runnable React application. The Interaction Inference Score (0-100, higher is better) is the recall of the interactions the screenshots imply, each rated working, partial or failed by annotators and weighted by how much state it needs; applications that never build score zero.
Metric: Interaction Inference Score (0-100), the recall of the interactions the screenshots imply, each rated working, partial or failed by three annotators and weighted by state scope 1-3; 45 screenshot sets (327 screenshots) of real multi-route web applications; the model must emit a runnable React application from the screenshots alone; apps that fail to build after three self-debug rounds score 0. Source: arxiv.org. Saturation forecast: Around January 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 39.3 | #161 |
| 2 | Kimi K2.5 (Thinking) | 20.7 | #256 (Kimi K2.5) |
| 3 | Qwen 3.5 397B A17B | 13.2 | #246 |
| 4 | Gemini 3.1 Pro (Preview) | 7.5 | #94 |
| 5 | GPT-5.4 | 6.7 | #123 |
| 6 | GLM-4.6V | 4.5 | #701 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ui2app · How It Works · Data refreshed daily, snapshot 2026-11-13.