Executable React components, visual layout cloning, and pixel-level UI grounding
Nine benchmarks shape this vertical, led by ScreenSpot Pro (17%), VoxelBench Text (16%), LiveBench 2.0 Design (16%) alongside GameDevBench, ReactBench, Code2Video, Web Clone Bench, React Native Evals, and Next.js Evals. Rather than grading isolated Python algorithms, these suites execute real component trees, check interactive state updates in React and Next.js, ground cursor coordinates on high-density desktop viewports, and render 2D/3D browser scenes.
Thirty-eight of the 76 models hold Supported status because multiple independent evaluators—including ScreenSpot Pro, LiveBench, and framework-specific boards—test the same public checkpoints.