AI may not have a "language wall" at all: by training on one language, it can also improve in other languages

According to DeepProbe's Beating Monitor, the Apple research team conducted a round of reinforcement learning experiments with 9 models in 11 languages to investigate: if the training data is only in one language, can the model's problem-solving ability be transferred to questions in other languages?

The answer is yes, and the effect is quite significant. Training with only one language not only enhances the model's performance in that language but also improves its performance on many other languages.

For example, in a French test, directly training with French led to an average score increase of 25.6 percentage points. Conversely, training solely with Spanish questions without any exposure to French questions still resulted in a 24.6 percentage point improvement when tested in French, with just a 1 percentage point difference.

The model not only learns how to solve questions in a specific language but also grasps problem-solving methods that can be applied across languages. Therefore, when enhancing the model's Chinese reasoning ability, not all reinforcement learning data necessarily needs to be recreated in Chinese.

However, switching the training language should not be taken lightly, as training in a specific language may indeed trigger the degradation of specific abilities. In the same set of reinforcement learning experiments, switching the training language caused some models to significantly deteriorate in performance on tasks in other languages. In a particularly extreme experiment, after training Qwen3-4B in Swahili, a test in English that the model had not encountered during training saw a performance drop of 19.2 percentage points; but when trained with multiple languages together, this test showed an improvement of 4.5 percentage points.

This study primarily validates the automatic judgment of answers to reasoning questions involving mathematics, logic, graphs, and geometry. For these types of questions, the underlying solution often remains unchanged when the questions are presented in a different language. The paper did not verify tasks that rely on language itself, such as reading comprehension, metaphors, semantic judgments, and cultural knowledge.