Leanabell-Prover: Posttraining Scaling in Formal Reasoning
Jingyuan Zhang, Qi Wang, Xingguang Ji, Yahui Liu, Yang Yue, Fuzheng Zhang, Di Zhang, Guorui Zhou, Kun Gai·April 08, 2025
Summary
Research enhances automated theorem proving with Lean 4, using posttraining scaling and continual training with a hybrid dataset. Reinforcement learning, guided by Lean 4 compiler feedback, improves provers, achieving a 59.8% pass rate on MiniF2F, surpassing baseline models. The study addresses challenges in generating proofs using Lean 4 codes, introducing strategies for ATP models' performance enhancement.
Introduction
Background
Overview of automated theorem proving (ATP)
Importance of Lean 4 in formal verification
Objective
Aim of the research: improving ATP with Lean 4
Specific focus: posttraining scaling and continual training with a hybrid dataset
Method
Data Collection
Sources of data for training and testing
Characteristics of the dataset used
Data Preprocessing
Techniques for preparing the data for model training
Handling of missing or irrelevant data
Model Training
Description of the models used
Implementation of posttraining scaling and continual training
Reinforcement Learning Integration
Role of reinforcement learning in enhancing ATP models
Guidance from Lean 4 compiler feedback
Evaluation
Metrics for assessing model performance
Comparison with baseline models
Results
Performance Metrics
Detailed results on MiniF2F
Improvement in pass rate (59.8%)
Comparative Analysis
Comparison with baseline models
Highlighting the effectiveness of the proposed methods
Challenges and Strategies
Challenges in Generating Proofs
Common issues faced in ATP with Lean 4
Strategies for overcoming these challenges
Strategies for Performance Enhancement
Techniques for improving ATP models' performance
Integration of feedback mechanisms
Conclusion
Summary of Findings
Recap of the research outcomes
Implications
Impact on the field of automated theorem proving
Potential for future research
Recommendations
Suggestions for further improvements and applications
Basic info
papers
artificial intelligence
Advanced features
Insights
What are the key implementation strategies used to improve the pass rate on MiniF2F with Lean 4?
What role does posttraining scaling play in the performance enhancement of ATP models?
How does the integration of Lean 4 compiler feedback enhance the architecture of automated theorem provers?
What challenges are addressed in generating proofs using Lean 4 codes, and how are they mitigated?