跳到主要导航 跳到搜索 跳到主要内容

QDTSynth: Quality-Driven Formal Theorem Synthesis for Enhancing Proving Performance of LLMs

  • Lei Wang
  • , Ruobing Zuo
  • , Gaolei He
  • , Jianlin Wang
  • , Zhengfeng Yang*
  • *此作品的通讯作者
  • East China Normal University
  • Henan Univeristy

科研成果: 书/报告/会议事项章节会议稿件同行评审

摘要

Automated Theorem Proving is an important and challenging task. Although large language models (LLMs) have demonstrated remarkable potential in mathematical reasoning, their performance in formal theorem proving remains constrained by the scarcity of high-quality supervised fine-tuning (SFT) data. To address this limitation, we propose a Quality-Driven Theorem Synthesis method (QDTSynth) in Lean4. During the statement synthesis, we enhance Monte Carlo Tree Search (MCTS) with an adaptive adjustment mechanism that dynamically optimizes the search strategy based on the synthesis of statements. In addition, we propose diversity screening and the self-assessment method to select theorems that exhibit both diversity and high quality from the initially synthetic statements, enabling the synthesis of a high-quality Lean4 theorem dataset. After fine-tuning three open-source large language models on our synthetic dataset, experiments on the miniF2F benchmark demonstrate that QDTSynth significantly improves the performance of various open-source LLMs in theorem proving tasks. Our work offers a promising new direction for the future synthesis of high-quality formal mathematical theorems.

源语言英语
主期刊名Long Papers
编辑Wanxiang Che, Joyce Nabende, Ekaterina Shutova, Mohammad Taher Pilehvar
出版商Association for Computational Linguistics (ACL)
14683-14698
页数16
ISBN(电子版)9798891762510
DOI
出版状态已出版 - 2025
活动63rd Annual Meeting of the Association for Computational Linguistics, ACL 2025 - Vienna, 奥地利
期限: 27 7月 20251 8月 2025

出版系列

姓名Proceedings of the Annual Meeting of the Association for Computational Linguistics
1
ISSN(印刷版)0736-587X

会议

会议63rd Annual Meeting of the Association for Computational Linguistics, ACL 2025
国家/地区奥地利
Vienna
时期27/07/251/08/25

指纹

探究 'QDTSynth: Quality-Driven Formal Theorem Synthesis for Enhancing Proving Performance of LLMs' 的科研主题。它们共同构成独一无二的指纹。

引用此