LLM's tend to "average out" language, making it less nuanced and more predictable. Combined with outright mistranslations, I don't think it'd perform better than what "reasoning mode" already does.