WebGen-Bench

Website-generation-from-scratch benchmark for coding agents: 101 natural-language instructions for building interactive, functional websites, graded by 647 UI-agent-executed test cases — the agent actually operates the generated site rather than string-matching its code. Frontier agents scored well under half of tests passed at release, and the benchmark spawned the WebGen-Agent follow-up line.

From Zimu Lu and Hongsheng Li (MMLab-CUHK), sibling to the group's MathCoder2 math line.