Collaborative Theorem Proving with Large Language Models: Enhancing Formal Proofs with ProofRefiner
Keywords: Agent-base System, Reasoning
TL;DR: agent-based system for math reasoning
Abstract: Abstract： Theorem proving presents a significant challenge for large language models (LLMs) because formal proofs can be rigorously verified by proof assistants like Lean, leaving no room for errors. Existing LLM-based provers typically operate autonomously, but they often struggle with complex and novel theorems where human insights are crucial. We propose a new framework that positions LLMs as collaborative assistants in theorem proving to address this. This framework enables the seamless integration of LLM inference into the Lean environment, allowing developers to build various proof automation tools. These tools offer features such as suggesting proof steps, completing intermediate goals, and selecting relevant premises, thereby enhancing the theorem-proving process. Users can leverage our pretrained models or integrate their own, supporting local and cloud-based execution. Experimental results demonstrate that our approach is more effective in aiding humans and automating the theorem-proving process than existing rule-based systems. Additionally, we introduce a system called ProofRefiner, which refines questions and answers through dynamic dialogue adjustments to ensure relevance and precision.
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10255
Loading