AI probably doesn't have a language barrier at all: practice only one language, and others will get better

Comparative news. According to monitoring, the Apple research team conducted a round of reinforcement learning experiments with 9 models and 11 languages, and wanted to find out: if the training data were in only one language, could the problem solving ability learned by the model be used for questions in other languages? The answer is, yes, and the effect is pretty obvious. By training in only one language, the model will also improve in many other languages. For example, in the French test, when training directly in French, the score increased by an average of 25.6 percentage points; no French training at all, only Spanish questions were used to practice, and the final French exam could improve by 24.6 percentage points, which is only 1 percentage point difference. The model learned not only questions in a certain language, but also some problem solving methods that can be continued to be used in different languages. Therefore, if you want to improve the model's Chinese reasoning in the future, not all of the reinforcement learning data will have to be converted back to Chinese. However, training languages cannot be changed at will; specific languages may indeed trigger the deterioration of specific abilities. With the same set of reinforcement learning and another training language, some models will clearly regress in other languages and tasks. In the most exaggerated set of experiments, after Qwen3-4B was trained in Kiswahili, a set of English tests that had not been seen during training dropped 19.2 percentage points from the original model; when training multiple languages together, this test increased by 4.5 percentage points. This paper mainly verifies reasoning questions where answers such as mathematics, logic, graphs, and geometry can automatically determine right or wrong. After these questions are changed from one language to another, the underlying solutions often remain the same. The paper did not verify the tasks of reading comprehension, metaphor, semantic judgment, and cultural knowledge that really depend on the language itself.