The answer is yes, and the effect is quite noticeable. Training with only one language still makes the model stronger across many other languages as well.

For example, on a French test, training directly in French improved the average score by 25.6 percentage points. If you don’t train in French at all and only use Spanish questions for practice, you can still improve the score on the French test by 24.6 percentage points—just a 1-point difference.

What the model learns isn’t only problem-solving for a single language; it also learns some problem-solving methods that can be transferred and continued to be used in other languages. So in the future, if you want to strengthen the model’s Chinese reasoning, you may not need to remake all reinforcement learning data into Chinese.

But you also can’t just change the training language at will. A specific language really can trigger a degradation of certain abilities. Even with the same reinforcement learning, if you switch the training language, some models actually perform noticeably worse on other languages and tasks. In the most extreme set of experiments, after Qwen3-4B was trained in Swahili, a set of English tests the model hadn’t seen during training dropped by 19.2 percentage points compared with the original model. However, when training multilingual together, that test improved by 4.5 percentage points instead.

The main thing this paper validates is reasoning problems where answers in math, logic, diagrams, geometry, and the like can be automatically judged as correct or incorrect. When these questions are changed to another language, the underlying solution approach often doesn’t change. For tasks that truly depend on language itself—reading comprehension, metaphor, semantic judgment, and cultural knowledge—the paper does not validate them.

Cao Yuan, a former senior research scientist at Google DeepMind and co-founder and CEO of Unreasonable Labs, believes that AI for Science is entering a breakout period. But for AI to truly independently make scientific discoveries on the level of a Nobel Prize, it will still take at least another twenty or thirty years. Cao Yuan previously worked on projects like Gemini, and his current startup direction is also to help AI discover new knowledge.

He thinks the biggest bottleneck right now is verification. With code, you can run tests immediately, and math can be checked step by step with Lean proofs. But biology, materials, and physics ultimately have to be done with real experiments. One experiment can be very expensive, and it might also take a long time. AI is hard to do continuous, rapid trial and error the way a coding agent can.

What’s even harder is “creating new concepts.” Cao Yuan believes today’s models are good at searching for answers within existing knowledge, definitions, and theorems, but they still can’t, like top scientists, abstract new mathematical objects, definitions, and theories. He calls this kind of “concept abstraction” the “last mile” of AGI.