- April 2, 2026 → Apple Machine Learning Research Apple publications — Apple’s personalized GRPO paper date
- May 16, 2025 → Apple Machine Learning Research Apple publications — Apple’s “English Accent” multilingual naturalness paper date
Konstantin Dobler still remembers the moment the results came back sideways. The lead author of a new Apple-led study on Group Relative Policy Optimization had expected native-language reasoning to fall far behind English. It didn’t.
Dobler explained that the team expected a much wider gap, and the finding reshaped how they read every later benchmark. Reinforcement Learning with Verifiable Rewards, usually tuned with GRPO, has become the default recipe for sharpening reasoning after pretraining Apple Machine Learning Research. Almost every published result, though, comes from English prompts.
What the researchers measured
They ran a large empirical sweep across many base models, training languages, and reward designs arXiv preprint. The team tracked reasoning quality when models learned to “think” in a language other than English, then tested whether the skill transferred.
The headline result surprised them.
The native-language surprise
The gap was smaller than anyone expected. The authors noted, “we further observe strong crosslingual transfer,” and the gains traveled across many model families Apple Machine Learning Research.
But the averages hid something uncomfortable.
Where it backfires
In several configurations, training in a single language inflicted severe regressions on out-of-domain abilities in other languages Apple Machine Learning Research. The effect was fiercely model- and language-dependent — no universal rule emerged. A setup that helped one language could quietly damage math reasoning elsewhere.
That is the trap. A team shipping a multilingual assistant might celebrate one early win and miss the collateral damage.
Apple’s longer multilingual arc
This is not Apple’s first pass at the problem. A May 16, 2025 study Apple Machine Learning Research found that even multilingual models echo English-centric grammar and produce unnatural output in other tongues. A separate April 2, 2026 paper Apple Machine Learning Research showed GRPO’s group-based normalization quietly smooths over conflicting individual preferences.
Why it matters for builders
For teams localizing reasoning models, the lesson is blunt: evaluate broadly, not just on the target language. zBrandco’s own coverage of a multilingual retrieval study found English tops PII leak rates in mixed-language RAG systems English Tops PII Leak Rate in Multilingual RAG Study, another reminder that language shifts surface hidden failures.
The open question is measurement. If a model improves in one language while regressing in another, which score do you ship?
