Large language models are used worldwide, but most safety and reasoning evaluations still happen in English. A new preprint on arXiv argues that this creates a blind spot: models may behave differently in other languages, and current benchmarks do not capture that. The authors propose translating existing English benchmarks as a scalable way to build multilingual test sets.

The abstract introduces a method called "reasoning-guided translation," which appears designed to keep the reasoning demands of the original questions intact during translation. The paper is a preprint, so details on how the method works or how well it performs are not yet available from the abstract alone.

If the approach proves effective, it could help researchers assess whether models reason safely and correctly in languages beyond English, without having to hand-write new benchmarks for every language. The preprint is early-stage, and further results will determine whether the method delivers on that promise.