EvoGenUI-Bench: leaderboard
EvoGenUI-Bench (Peng et al., 2026) tests language models as multi-turn generative UI assistants that must keep an executable web interface working as requests evolve over five turns, across information presentation, interaction and tool-grounded tasks. Turn pass rate in percent averaged over the three suites, where a turn passes when its executed artifact meets every hidden requirement checked by a MiMo-V2.5 evaluator; higher is better.
Metric: Turn pass rate (%; mean of the three suites' shares of requested turns whose artifact passes every hidden requirement; 150 five-turn generative-UI tasks (750 turns) in three suites of 50: information presentation, executable interaction and tool-grounded external state; each turn's artifact is executed in a browser and a MiMo-V2.5 evaluator checks hidden requirements from screenshots, source and DOM evidence, MiMo-V2.5 actor traces and runtime logs; turns left unexecuted after an earlier build failure or invalid output count as failures; temperature 0). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 74.9 | #79 |
| 2 | GPT-5.5 | 61.7 | #49 |
| 3 | Claude Haiku 4.5 (20251001) | 51.1 | #536 |
| 4 | Qwen 3.6 Plus | 44 | #222 |
| 5 | Gemini 3 Flash (Preview) | 41.7 | #144 |
| 6 | Gemini 3.1 Pro (Preview) | 23.6 | #94 |
| 7 | Qwen 3 Coder 480B A35B Instruct | 23.2 | #643 |
| 8 | GLM 4.5 Air | 21.1 | #868 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evogenui-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.