Towards Real-World Industrial-Scale Verification: LLM-Driven Theorem Proving on seL4 (arxiv.org)

In a groundbreaking effort to enhance industrial-scale theorem proving, researchers have introduced AutoReal, an innovative method leveraging large language models (LLMs) to make formal verification more accessible and efficient. Traditionally, formal methods are labor-intensive and costly, particularly in projects like seL4, which have relied heavily on expert knowledge. AutoReal addresses these challenges by enabling lightweight local deployment of a compact 7B-scale prover, AutoReal-Prover, which incorporates advancements such as chain-of-thought (CoT) training and context augmentation. These techniques allow the model to provide step-wise explanations of proofs, significantly enhancing user comprehension and engagement. The significance of this development lies in AutoReal-Prover's impressive performance, achieving a proof success rate of 51.67% on critical theorems from the seL4 verification project—surpassing previous attempts, which only recorded a 27.06% success rate. Moreover, it demonstrated its adaptability by successfully proving 53.88% of the 451 theorems from three security-related projects. This research marks a pivotal shift towards practical application of LLMs in formal verification, potentially revolutionizing how industries implement rigorous verification methods in safety-critical systems.