WebUIProof: Benchmarking WebUI Code Generators with UI-Agent Execution Harness
- Source
- arxiv.org
- Author
- Yun-Yun Tsai, Yuning Mao, Shiqi Wang, Junfeng Yang, Sinong Wang
- Date

Context
AI-written web interfaces can render and still not work, and a called WebUIProof tests that gap by actually using the page. Earlier benchmarks, the authors say, mostly check that code builds or that a screenshot looks plausible, which misses whether buttons, controls and simulations behave as specified.
WebUIProof gives each task a structured specification and executable interaction tests, covering general interfaces such as dashboards and games and 3D interactive simulations. A UI loads the page in a headless browser, finds elements, acts on them, observes what changes and checks the specified assertions, repeating a plan, act, observe loop.
Across eight commercial models, the authors observed frequent failures on interaction requirements even when pages rendered, especially for 3D simulations. They also report that using the test outcomes as reward signals to train compact models, Qwen2.5 14B and MIMO 7B, improved functional completion and reduced build failures. The abstract gives no numbers for either result, so effect sizes need the full paper.
- WebUIProof argues that build success and screenshots miss functional correctness, so it pairs structured specs with dense executable interaction tests covering general web UIs and 3D interactive simulations such as particle systems and physics dynamics.
- Its UI- runs tests in a headless browser through a plan, act, observe loop: it finds DOM elements, performs actions, watches the resulting UI and DOM changes, and checks the specified assertions.
- Across eight commercial LLMs, the authors saw frequent failures on interaction requirements even when pages rendered fine, with 3D simulation interfaces the weakest area.
- The same harness doubles as a reward source: training compact models such as Qwen2.5 14B and MIMO 7B with RL rewards from the executable tests improved functional completion and reduced build failures, per the authors.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
articleFramework and Benchmark for Code-Driven Agentic Testing in Web DevelopmentBin Hong, Zhenchao Zhang, Jiyuan He, Kai Zhang, Zhenya Huang
articleWebUniverse: Scaling Verified Training Data for Agentic Web DevelopmentYaorui Ding, Fengqing Jiang, Yichen Feng, Yuetai Li, Zhangchen Xu, Kaiyuan Zheng, Xuanyu Lu, Bhaskar Ramasubramanian, Basel Alomair, Luyao Niu, Radha Poovendran
articleTestSuiteBench: Coding Agents Struggle to Generate Strong Test Suites from SpecificationsKazuki Egashira, Niels M\"undler-Sasahara, Mark Niklas M\"uller, Martin Vechev
Checking sign-in…
Loading comments…