Think Multilingual, Not Harder: A Data-Efficient Framework for Teaching Reasoning Models to Code-Switch

Eleanor M Lin, David Jurgens

Keywords: Multilingual, code-switching, reasoning, large language models

TL;DR: We introduce a framework for teaching reasoning models to code-switch (i.e., mix languages) during reasoning, improving reasoning performance in a data-efficient manner for lower-resource languages.

Abstract: State-of-the-art reasoning models code-switch (i.e., mix languages) in their reasoning, a behavior which has previously been viewed as an undesirable error. However, more recent works attempt to control code-switching to improve reasoning performance. In this work, we introduce the first linguistically and behaviorally motivated fine-tuning framework for teaching reasoning models to code-switch for better reasoning performance. First, we perform a systematic analysis of reasoning traces from diverse models, languages, tasks, and domains to understand the types of code-switching behaviors found in base reasoning models. Then, we develop fine-tuning interventions that teach reasoning models to code-switch based on our observations of helpful behaviors in the base models. We find that training reasoning models on examples of code-switched reasoning can boost reasoning performance in a data-efficient manner. Interestingly, we also find that code-switching behaviors in reasoning models can be modified by fine-tuning for tasks that do not directly demonstrate code-switching in reasoning (e.g., a traditional machine translation task). Our work suggests that code-switching should not always be viewed as an undesirable reasoning behavior, and that data-efficient interventions can instill helpful forms of this behavior in reasoning models.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 119