New AI Method Dramatically Improves Mathematical Reasoning in 17 African Languages

Large language models (LLMs) have demonstrated impressive capabilities in mathematical reasoning, but this proficiency often falters when applied to low-resource languages. This disparity creates a significant barrier to equitable AI access and development globally. Researchers have introduced a novel technique called Crosslingual On-Policy Self-Distillation (COPSD) to bridge this gap by enabling models to transfer their high-resource reasoning skills to languages with less available data.

COPSD operates on a self-distillation principle, where the same model acts as both student and teacher. The student model processes a problem in a low-resource language, while the teacher model benefits from additional crosslingual context, including an English translation and reference solution. This setup allows for dense, token-level supervision during training, which is more stable and effective than traditional outcome-only reinforcement learning methods.

The effectiveness of COPSD was rigorously tested, notably on 17 distinct low-resource African languages. The experiments revealed consistent and substantial improvements in mathematical reasoning across various model sizes. COPSD significantly outperformed existing methods like Group Relative Policy Optimization (GRPO), demonstrating its robustness and efficiency in enhancing AI performance for these languages.

Further analysis indicated that COPSD not only boosts reasoning accuracy but also improves adherence to answer formats and strengthens test-time scaling. The method showed particular promise for generalizing to more challenging multilingual reasoning benchmarks, with the most significant gains observed in the lowest-resource languages. This research is critical for fostering inclusive AI development, ensuring that the benefits of advanced language models are accessible to diverse linguistic communities across Africa.

The availability of the code and data associated with COPSD at the provided GitHub link underscores the commitment to open science and facilitates further research and application of this promising technique within African contexts and beyond. This advancement paves the way for more equitable AI tools that can support education, research, and innovation in numerous African linguistic communities.