Towards Advanced Mathematical Reasoning for LLMs via First-Order Logic Theorem Proving

The article introduces DREAM, a self-adaptive framework aimed at enhancing the mathematical reasoning capabilities of large language models (LLMs) through first-order logic (FOL) theorem proving. It addresses the limitations of existing models, such as Deepseek-Prover-V2-7B, which achieved only 4.2% accuracy on a new theorem proving dataset, by implementing an Axiom-Driven Strategy Diversification and Sub-Proposition Error Feedback. This approach demonstrates an improvement in performance by 0.6% to 6.4% and provides a curated dataset of 447 mathematical theorems in Lean 4 format, which is significant for practitioners focused on advancing LLMs' reasoning abilities in complex mathematical contexts.