Web agents are typically evaluated on English-language interfaces, leaving their performance on multilingual websites largely unexplored. The paper highlights that this gap is especially significant for low-resource languages, where the need for such evaluation is most acute.
To address this, the paper introduces LoWAgent, a benchmark designed for low-resource multilingual web agent evaluation. The benchmark aims to provide a standard way to measure how well agents operate in these languages.
The significance lies in broadening the scope of web agent testing. Without such benchmarks, agents may be optimized for English and fail in real-world multilingual contexts. LoWAgent is a step toward more inclusive evaluation.