Show Your Work in Swahili: Reinforcement Learning for Native-Language Mathematical Reasoning

We present Fikiria, a two-stage SFT+GRPO framework that directly optimizes the native-language quality of reasoning traces, rather than treating intermediate steps as a byproduct of answer accuracy. Applied to Swahili mathematical reasoning, Fikiria uses a gated ChrF\texttt{++} reward against a Swahili-Swahili reference to jointly optimize native-language surface fidelity and answer correctness, with reasoning-step validity further examined via human evaluation on a sample. Fikiria-AfriqueQwen-14B achieves 79.2\% on AfriMGSM Swahili, matching GPT-4o despite being open-weight and 5x smaller than LLaMA 3.1 70B. Our results show that native-language surface fidelity in reasoning can be directly optimized via gated structured rewards, and that this framework provides a promising approach for multilingual settings where the reasoning language must match the user's language.