Amazon Fine-Tunes AI to Reason in Multiple Languages

When you ask an AI to solve a problem in French, you expect it to think in French, not switch to English halfway through. But many AI models, like GPT-OSS-20B, are trained mostly on English and struggle to reason in other languages. In tests, the base model reasoned in the correct language only 22% of the time—and that was mostly English examples. For non-English users, this is frustrating and limits the model's usefulness.

Amazon researchers tackled this by fine-tuning the model on a multilingual dataset. They used a method called supervised fine-tuning (SFT), which teaches the model a new behavior by showing it examples. They also experimented with reinforcement learning (GRPO) to further refine the model's responses. The key finding: SFT did most of the work, and GRPO helped when there was room for improvement. The training was done on Amazon SageMaker HyperPod, a cloud service for running AI workloads.

The results showed that after fine-tuning, the model reasoned in the correct language far more often. This means that businesses and developers can now deploy AI that communicates naturally with users in their own language, without the awkward mid-thought language switches. It's a step toward making AI more accessible and useful for people around the world.

While the technique is promising, it requires technical expertise and cloud resources. But for companies looking to serve global audiences, it's a valuable approach. The takeaway: with the right fine-tuning, open AI models can become truly multilingual, breaking down language barriers in AI interactions.