Score: 3

Leanabell-Prover: Posttraining Scaling in Formal Reasoning

Published: April 8, 2025 | arXiv ID: 2504.06122v3

By: Jingyuan Zhang , Qi Wang , Xingguang Ji and more

Potential Business Impact:

Makes computers prove math ideas much faster.

Business Areas:
Natural Language Processing Artificial Intelligence, Data and Analytics, Software

Recent advances in automated theorem proving (ATP) through LLMs have highlighted the potential of formal reasoning with Lean 4 codes. However, ATP has not yet be revolutionized by the recent posttraining scaling as demonstrated by Open AI O1/O3 and Deepseek R1. In this work, we investigate the entire posttraining of ATP, aiming to align it with breakthroughs in reasoning models in natural languages. To begin, we continual train current ATP models with a hybrid dataset, which consists of numerous statement-proof pairs, and additional data aimed at incorporating cognitive behaviors that emulate human reasoning and hypothesis refinement. Next, we explore reinforcement learning with the use of outcome reward returned by Lean 4 compiler. Through our designed continual training and reinforcement learning processes, we have successfully improved existing formal provers, including both DeepSeek-Prover-v1.5 and Goedel-Prover, achieving state-of-the-art performance in the field of whole-proof generation. For example, we achieve a 59.8% pass rate (pass@32) on MiniF2F. This is an on-going project and we will progressively update our findings, release our data and training details.


Page Count
23 pages

Category
Computer Science:
Artificial Intelligence