arXiv:2611.16104v1 Announce Type: cross Abstract: We present Fikiria, a two-stage SFT+GRPO recipe for Swahili mathematical reasoning. Its reward multiplies a format gate by a weighted sum of answer accuracy and the ChrF\texttt{++} similarity of the Swahili reasoning to an in-language reference chain of thought, so that training targets the surface fidelity of the reasoning trace and not only the final answer.
