WorldGUI & GUI-Thinker

Dynamic GUI benchmark testing agents from any starting point: unlike static-screenshot suites, WorldGUI initializes each desktop task with varied real initial states to probe how agents handle post-action dynamics and unexpected UI conditions, across widely-used desktop and web applications. Ships with GUI-Thinker, a self-reflective computer-use agent framework that critiques and revises its own actions before and after execution.

From Mike Zheng Shou's ShowLab — the evaluation companion to the ShowUI GUI-agent model line.