LLM-Driven Theorem Proving Achieves Industrial-Scale Verification on seL4

This paper introduces AutoReal, an LLM-driven theorem proving method designed for real-world industrial-scale systems verification, specifically targeting the seL4 microkernel. The authors address the high cost and expert effort typically required for formal methods (FM) by leveraging recent advances in large language models (LLMs) to automate theorem proving. Unlike previous work that primarily focuses on mathematics-oriented benchmarks or relies on large, closed-source models, AutoReal aims for lightweight local deployment. The methodology incorporates two key improvements: chain-of-thought (CoT)-based proof training, which teaches the LLM the reasoning behind proof steps, and context augmentation, which leverages proof context from the project to enhance LLM-driven proving. Based on this methodology, the authors fine-tune a base model to create AutoReal-Prover, a compact 7B-scale prover. AutoReal-Prover achieves a 51.67% proof success rate on 660 theorems from seL4-designated Important Theories, significantly outperforming prior attempts. To evaluate generalization, AutoReal-Prover is applied to three security-related projects from the Archive of Formal Proofs (AFP), achieving a proof success rate of 53.88% across 451 theorems. This work demonstrates the potential of LLM-driven theorem proving to reduce the cost and complexity of formal verification in industrial-scale projects, paving the way for more widespread adoption of FM in critical systems development.

Transparency Disclosure: This analysis was composed by an AI Large Language Model. Human oversight ensured factual accuracy and editorial integrity, aligning with EU AI Act Article 50 requirements.