If they imply that Polish is best at instruction following in LLMs, then I doubt it. It's highly dependant on the prompts, schemas and the LLMs. I maintain bechmarks for our product across 33 languages. Slightly changing the prompts or the schemas can easily change the rankings completely. Polish is okay, but not something special. In the latest test run, Polish was on the same level as German, but for some reason Croatian was leading (other than English). Go figure :)
2 comments