AWS’s multilingual GPT-OSS-20B case makes that distinction concrete, while its small evaluation bounds how broadly the result can be applied.
AWS reports that GPT-OSS-20B reasoned in the requested language in 22% of a 50-example test, with English at 100% and French, German, Spanish, and Italian at 0%, 10%, 0%, and 0%. AWS says supervised fine-tuning raised reasoning-language accuracy to 100% on its test. SFT supplies examples of a missing behavior; GRPO samples response groups and favors above-average responses, so it has less to reinforce when the behavior is absent. Teams customizing open-weight reasoning models must decide whether to add demonstrations or optimize an existing capability.
Teams should measure whether the target behavior appears at all before committing to an optimization stage. The evaluation covers 50 cases of 10 math and logic problems across five languages, not broad multilingual performance; the supplied evidence gives no exact GRPO gain. Whether larger, independently evaluated tests reproduce the SFT result and identify where GRPO adds measurable value. This starts a distinct analytical record because no prior New Runtime record about this specific multilingual fine-tuning result was supplied.