Researchers have developed Fikiria, a novel two-stage approach using Supervised Fine-Tuning (SFT) and a policy gradient method (GRPO) to enhance mathematical reasoning in Swahili. This method incorporates a reward function that balances answer accuracy with the surface fidelity of the reasoning trace to an in-language reference. Trained on a dataset translated from GSM8K using GPT-4 and filtered with SSA-COMET, Fikiria-AfriqueQwen-14B achieved 79.2% accuracy on the AfriMGSM Swahili benchmark. While GPT-4o reached 68.0% under the same strict prompt conditions, it was preferred in a small human evaluation, though its performance was not directly comparable due to differing prompting strategies. AI
IMPACT This research demonstrates a method to improve AI's mathematical reasoning capabilities in low-resource languages, potentially enabling broader access to AI tools.
RANK_REASON The cluster describes a new research paper detailing a novel AI model and methodology for native-language mathematical reasoning. [lever_c_demoted from research: ic=1 ai=1.0]
- AfriMGSM
- Fikiria
- Fikiria-AfriqueQwen-14B
- GPT-4
- GPT-4o
- GSM8K
- IrokoBench
- Jessica Lundin PhD
- SSA-COMET
- Swahili
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
