mathematical reasoning

Retrieval that teaches models to reason by analogy

RA-RFT trains retrievers to find useful reasoning analogies, then fine-tunes models with those demonstrations.