Think Multilingual, Not Harder: A Data-Efficient Framework for Teaching Reasoning Models to Code-Switch
Researchers created a framework to teach AI models beneficial code-switched reasoning behaviors efficiently. They analyzed existing models' code-switching patterns and developed a taxonomy based on function, form, and coherence. By applying specific fine-tuning methods using diverse data, they showed that strategically teaching models to use certain languages as a matrix language significantly improves their reasoning performance.