**[AI May Not Have a "Language Barrier": Training in One Language Improves Others Too]** According to monitoring by Dongcha Beating, Apple's research team conducted a round of reinforcement learning experiments using 9 models and 11 languages to investigate: if the training data consists of only one language, can the problem-solving ability learned by the model be applied to other languages? The answer is yes, and the effect is quite significant. Training in just one language also improves the model's performance in many other languages. For example, in French tests, directly training in French improved scores by an average of 25.6 percentage points; without any French training and only using Spanish training data, the French test scores still improved by 24.6 percentage points—just 1 percentage point less. The model doesn't just learn how to solve problems in a specific language; it also acquires problem-solving methods that can be transferred to other languages. This means that in the future, if we want to improve a model's reasoning ability in Chinese, not all reinforcement learning data needs to be converted into Chinese. However, the choice of training language cannot be random, as certain languages may indeed trigger a decline in specific abilities. In the same reinforcement learning setup, changing the training language can cause some models to perform significantly worse in other languages and tasks. In the most extreme experiment, after training Qwen3-4B in Swahili, its performance on an unseen English test dropped by 19.2 percentage points compared to the original model; however, when trained in multiple languages simultaneously, the same test saw a 4.5 percentage point improvement. This paper primarily validates reasoning tasks such as math, logic, graphs, and geometry, where answers can be automatically judged as right or wrong. For these tasks, the underlying methods often remain unchanged when switching languages. However, for tasks that truly rely on language itself—such as reading comprehension, metaphors, semantic judgment, and cultural knowledge—the paper does not provide validation. [Original Link]