Gradual Language Blending Helps AI Models Reason Better Across Non-English Languages

Large language models excel at many tasks, but they struggle unevenly across different languages. The core problem: most LLMs rely heavily on English-centered thinking patterns deep in their representations, even when processing other languages. A new study proposes a simple inference-time fix that yields substantial gains.

A Gradual Bridge Between Languages#

Rather than abruptly switching between a non-English language and English, the researchers developed code-switching in-context learning (CSICL)—a method that gradually scaffolds reasoning by transitioning from a target language toward English. This gentle progression aligns non-English inputs with the model's English-dominant reasoning space, rather than forcing an immediate translation.

Performance Gains Across the Board#

The team evaluated CSICL on 4 different LLMs, 6 datasets, and 10 languages. The results were consistent:

• Target languages (those used during training): +6.0 percentage points improvement
• Unseen languages (new at test time): +4.8 percentage points improvement
• Low-resource languages: +14.7 percentage points in target languages and +5.3 percentage points in unseen languages

Improvements held up across different language families, suggesting the technique is broadly applicable rather than tailored to specific linguistic groups.

Why This Matters#

The findings underscore that multilingual AI systems can be made more equitable without retraining. By leveraging code-switching—a natural phenomenon where bilingual speakers blend languages—during inference alone, this work moves toward LLMs that serve speakers of all languages more fairly and effectively.