According to CoinMeta, an Apple research team conducted a reinforcement learning experiment using 9 models and 11 languages to explore whether the problem-solving capabilities of the models could be transferred to other languages if the training data was for only one language. The results showed significant improvements in multiple languages. For example, in French tests, the performance improved by an average of 25.6 percentage points when trained directly in French, and even when trained only in Spanish, the French performance still increased by 24.6 percentage points. The study indicates that the models not only learned problem-solving methods specific to a particular language but were also able to apply these methods to other languages. However, the choice of training language needs to be cautious, as a specific language may lead to a decline in ability. In the experiment, after training in Swahili, the performance in English tests decreased by 19.2 percentage points, whereas multi-language training improved this test by 4.5 percentage points. The paper mainly verified the consistency of solutions to reasoning questions in mathematics, logic, etc., during language conversion, but did not test tasks that depend on the language itself, such as reading comprehension and cultural knowledge.