CSICL: Gradual Code-Switching Improves Cross-Lingual LLM Performance
A new study released on arXiv (2510.05678) introduces an innovative technique called code-switching in-context learning (CSICL). This method helps large language models (LLMs) process non-English inputs by transitioning them into English, which aids in reasoning. When tested on 4 LLMs, 6 datasets, and 10 languages, CSICL outperformed standard cross-lingual in-context learning methods, achieving an average boost of 6.0 percentage points (pp) for target languages and 4.8pp for languages not seen before. The improvements were especially notable in low-resource situations, with increases of 14.7pp for target languages and 5.3pp for unseen ones. This study, noted as a replace-cross update on arXiv, shows how gradual code-switching can enhance multilingual capabilities without requiring additional training.
Key facts
- CSICL is an inference-time mechanism for cross-lingual representational alignment.
- It gradually transitions from target language to English.
- Tested on 4 LLMs, 6 datasets, and 10 languages.
- Average gains of 6.0pp in target languages and 4.8pp in unseen languages.
- Gains are more pronounced in low-resource settings: 14.7pp target, 5.3pp unseen.
- Improvements generalize across language families.
- The paper is available on arXiv with identifier 2510.05678.
- The method outperforms cross-lingual in-context learning baselines.
Entities
Institutions
- arXiv