Gradual Code-Switching In-Context Learning for Cross-Lingual Transfer of Large Language Models

ACL ARR 2026 January Submission9258 Authors

Keywords: code-switching, in-context learning, cross-lingual transfer, large language models

Abstract:

While large language models (LLMs) have achieved notable progress in multilingual settings, their performance remains uneven across languages as LLMs often rely on English-centric latent representations.
In this work, we introduce code-switching in-context learning (CSICL), a simple yet effective prompting strategy that gradually transitions from a target language to English within demonstrations and instruction to facilitate their latent reasoning in English.
By explicitly scaffolding the reasoning process through controlled code-switching, CSICL acts as an implicit linguistic bridge that enhances cross-lingual alignment.
We conduct extensive experiments across 4 LLMs, 6 datasets, and 10 languages, spanning both knowledge-intensive and reasoning-oriented domains.
Our results demonstrate that CSICL consistently outperforms X-ICL baselines, achieving 6.0%p and 4.8%p higher performance in both target and unseen languages, respectively.
The improvement is generalized across diverse language families and even more pronounced in low-resource settings, with gains of 14.7%p in target and 5.3%p in unseen languages.
These findings establish code-switching as a robust and effective approach for overcoming the cross-lingual misalignment during inference, moving LLMs toward more equitable and effective multilingual systems.

Paper Type: Long

Research Area: Multilinguality and Language Diversity

Research Area Keywords: code-switching, multilingualism, cross-lingual transfer, multilingual representations, chain-of-thought, prompting, mixed language, less-resourced languages, transfer

Contribution Types: NLP engineering experiment, Approaches to low-resource settings

Languages Studied: English, French, Korean, Yoruba, Chinese, Spanish, Indonesian, Turkish, Swahili, Telugu, Japanese, Dutch

Submission Number: 9258