1D-Bench: A Benchmark for Iterative UI Code Generation with Visual Feedback in Real-World

arXiv cs.AI · 18 hours ago

Researchers introduced 1D-Bench, a benchmark for evaluating AI models on design-to-code tasks using real e-commerce workflows where models generate React code from UI designs and intermediate representations. The benchmark contains datasets with reference renderings and extracted intermediate representations that may have errors, and evaluates models through iterative component-level editing with execution feedback. Experiments showed that iterative editing improved performance by increasing rendering success rates, though post-training with synthetic repair trajectories and reinforcement learning produced limited gains.