Tencent Hy introduces WebCraftBench for AI-built web apps

Tencent Hy introduces WebCraftBench. Agents use the live app, coverage-guided exploration finds parts that were never reached, and the benchmark scores aesthetics, usability, and whether the original request was met. Tencent Hy reports that, on 197 human-validated pairs, it matches human preference 85.3% of the time. The paper is arXiv:2609.15387.

Arena.ai says WebCraftBench was built from 369 requests collected from Tencent Hy's internal replica of Code Arena, covering informal language, incomplete instructions, different project scopes, and diverse preferences. Arena.ai also says that, across 17 models, its rankings correlate 0.89 with the Code Arena leaderboard.