
What happened
A survey count of 208 deep-learning theorem-proving papers found only 46 target Lean, versus 24 for Isabelle, 18 for Mizar and 17 for Coq. Lean research went from zero in 2021 to 6 in 2022 and 17 in 2023, exceeding all other systems combined that year.
Why it matters
The countries that built the foundational tools do not dominate current research.
Summaries like this, every morning.
The count draws on the DL4TP bibliography attached to a COLM 2024 survey, which labels each paper by the proof assistant it targets. Because a single paper can target several systems, it is counted once per system, and the list is not exhaustive.
The body also traces where the early work came from. A small group of researchers around Cambridge and Edinburgh produced LISA in 2021, then Thor and Draft, Sketch, and Prove in 2022, and Magnushammer and Llemma in 2023. A separate lineage starts from OpenAI's GPT-f in 2020 and runs through miniF2F and LISA. These threads suggest the field's direction was shaped by a handful of people before it spread to large corporate labs.
The country sections complicate any simple national scoreboard. Japan has no identified national project naming theorem proving, yet a Kyoto University-led effort produced Prover Agent, an 8B-parameter model that scored 88.1% on miniF2F. China's State Council issued a 2025 AI action document, but it does not name theorem proving; corporate labs such as DeepSeek and universities moved on their own. The U.S. and U.K. also show no named national project, though U.K. work built heavily on Isabelle, a tool co-maintained with Germany.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.