Machine learning for formal verification
Title: Machine learning for formal verification
DNr: Berzelius-2026-322
Project Type: LiU Berzelius
Principal Investigator: Hamed Nemati <hnnemati@kth.se>
Affiliation: Kungliga Tekniska högskolan
Duration: 2026-11-15 – 2027-06-01
Classification: 10201
Keywords:

Abstract

With the recent advances of AI in mathematics, it has become increasingly clear that many open problems in mathematics and theoretical computer science may eventually be solved not by humans, but by computer programs. Moreover, understanding those proofs is no easy task for humans. In fact, it will take teams of top mathematicians weeks and months if not years to fully grasp the entirety of the AI generated proofs of the hundreds of the open questions. Yet, we are able to guarantee the correctness of these proofs, because we have specialized tools (e.g. the theorem prover Lean) that mechanically check each proof step for us. Formal verification, or theorem proving in particular, has more applications than in just mathematics. It has been long used for software and hardware design verification, where we formally model their behavior and prove the desirable properties stated as theorems about these models. Unlike mathematics, where we have access to relatively large corpus of either informal or formal data, the scarcity of program verification data has impeded the application of machine learning to verify programs autonomously. However, program verification has never been more important, as the coding tools have made producing plausibly correct programs easier than ever; yet without an expert checking the code line by line, it is far from certain that the AI-generated code is correct and unlike in the human case, the accountability is often missing. This project aims to investigate several promising approaches to synthetic data generation for applying machine learning to formal verification, including a recent paper on self-play training on zero data. We aim to gain a better understanding of the problem of generating formal data, where the exact syntax is defined, versus natural language data, where ambiguity is a part of the beauty of languages. We will use the requested resource to test and develop efficient and scalable prototypes in this specialization area that will hopefully allow us to rely less on centralized computing operated by large data center providers, but on local resources that we have more control of.