LiveEvalBench Judges Generated Frontends With a Build Engineer, Code Engineer, and UI Tester Instead of Static Rubrics

arXiv 2608.03689·medium signalXBlueskyLinkedInCopy link

The argument is that frontend artifacts are interactive rather than static and admit many equally valid implementations, so static benchmarks misjudge them. LiveEvalBench runs evaluation as a collaborative review workflow where three specialized agent roles gather evidence across deployment, code inspection, and browser interaction, with an adaptive protocol pairing shared cross-model rubrics against implementation-grounded per-artifact criteria. New evaluator roles and dimensions can be added without redesigning the pipeline, and the authors report close alignment with human expert judgment; code is at github.com/wyysteelhead/LiveEvalBench.