Skip to main content
Back

UC Santa Barbara Proves English-Only Fine-Tuning Hurts Even English AI Accuracy

A 220-run study across Qwen-3 and Gemma-3 finds English-only fine-tuning consistently leaves accuracy on the table, even for English.

Paper, Full Access
  • A study runs 220 supervised fine-tuning experiments to measure multilinguality effects on Qwen-3 and Gemma-3 post-training.
  • Adding even one non-English language to English fine-tuning improves accuracy by 8-9% on target languages.
  • English performance itself improves by 3.4% on math and 0.88% on API calling when non-English data is added.
  • Above 1B parameters, adding more languages never degrades high-resource performance, contradicting curse-of-multilinguality fears.
  • At 6+ training languages, zero-shot cross-lingual transfer matches direct inclusion for high-resource languages.
  • Authors release mAPICall-Bank, a multilingual function-calling dataset across 11 languages with 3,174 training examples each.

Post-training pipelines for large language models have a stubborn habit: they run almost entirely on English. A new controlled study from researchers at UC Santa Barbara and Amazon puts numbers on how much that habit costs, and the answer is that English-only fine-tuning leaves accuracy on the table, including on English evaluations themselves.

The paper runs 220 supervised fine-tuning experiments across two model families and two task types, systematically varying how many languages appear in the training mix. Increasing language coverage generally improves or maintains performance across tasks and model scales, with low-resource languages benefiting the most and high-resource languages plateauing rather than degrading.

The curse that didn't show up

Multilingual pre-training research has long warned about the curse of multilinguality, the idea that adding languages under a fixed parameter budget dilutes performance in the ones you already had. The trade-off is often framed as negative interference between languages competing for capacity, and the natural assumption has been that post-training inherits the same problem.

This study finds it mostly doesn't, at least at the scales where anyone actually deploys models. Across both model families and tasks, performance improves or holds steady as more training languages are added. This pattern is consistent at 1B parameters and above.

One exception is telling. On the smallest model tested, Qwen-3 0.6B, API calling performance initially improves with multilinguality up to four or five languages, then slightly declines, suggesting capacity-driven interference at that scale. Math reasoning showed no such degradation, and neither did any larger model. So the curse is real, but it lives below 1B parameters and mainly bites structured generation tasks.