Qwen Open-Sources BabelArena: A Multilingual Agent Benchmark

Key Info

Qwen Developers announced BabelArena, a new open-source research benchmark that examines how language barriers break down agent understanding across tool arguments, database entries, and search queries.

Highlights

• The name is inspired by the Tower of Babel story and the problem of language barriers blocking shared goals.
• In one example, Qwen-3.8-Max completed the same shopping task in English and Telugu and retrieved the same four products.
• The English run took 25 tool calls, while the Telugu run took 201, with repeated empty searches against the localized catalog.
• Both runs passed, but the efficiency gap highlights multilingual robustness issues in agents.
• Paper, GitHub repository, and datasets are all released publicly.