LLM mathematical reasoning grounded with formal verification
Allen He · Cambridge journal for junior scientists. · 2025
Allen Heformal statement and verified formal proof to generate high quality solutions.Our experiments and ablation studies on a collection of mathematical competition problems show that, compared to previous work, RGFV improves the accuracy of the answers significantly from 46.7% to 65.8% overall, and even more significantly on the subcategory of algebra, from 66.7% to 96.7%.Moreover, it significantly improves the quality of the generated solutions.The method also helps AI be more reliable and useful for students studying mathematics.