The benchmark contains 320 Change tasks involving existing repositories and 80 Build tasks for longer website-construction projects.
Coding agents can finish frontend changes while breaking existing website behavior
ProductWebBench separates requested functionality from website continuity, finding continuity failures in 28.2% of Change runs that passed every requested-work check.
Chinese Tech
Jiajun Wu · Jian Yang · Yaxin Du · Shukai Liu · Linzheng Chai · Tuney Zheng · +6 more
Beihang University · IQuest Research · Huawei Technologies Ltd.
Research Digest··2 min read
Wu and colleagues introduce ProductWebBench, a benchmark that tests whether frontend coding agents preserve a working website while implementing new requirements.
§