GPT-OSS-20B acknowledges your French prompt, then does its thinking in English anyway
GPT-OSS-20B is an open-weight reasoning model that thinks in English. Ask it to reason in French, German, Spanish or Italian, and as AWS's machine-learning blog describes it, the model acknowledges the instruction and then produces its chain-of-thought in English anyway.
On a 50-example multilingual reasoning test, the base model reasoned in the correct target language only 22 per cent of the time. The per-language split is stark: English 100 per cent, German 10 per cent, and French, Spanish and Italian zero.
AWS's own framing of the fix is a clean split between two techniques. Supervised fine-tuning teaches a behaviour the model doesn't have; GRPO optimizes a behaviour the model can already sometimes produce.
AI-generated discussion · Different perspectives
That split is a useful diagnostic, not just a training detail. Zero per cent French is a teaching problem — there is essentially nothing for reinforcement to reward — whereas a behaviour the model manages occasionally is what GRPO can select for and amplify. Match the technique to the gap and you save a lot of compute.
Worth keeping the headline number honest, though. The evaluation is 10 math/logic problems across 5 languages, and what's scored is whether the model reasoned in the right language — not whether its answers were right. A team shipping to non-English users may care a great deal about which language the model thinks in, and a team that only reads final answers may not care at all.