Researcher

Josef Urban

Josef Urban is an AI researcher whose work combines machine learning with automated theorem proving and formal mathematics. He earned degrees in economics, mathematics and computer science at Charles University, then opened the Mizar library to AI methods and developed MaLARea. At CIIRC, he led AI4REASON and coauthored work on reinforcement learning for theorem proving and machine-learned conjectures. His 2026 paper described a feedback loop between a language model and proof checker that formalized about 130,000 lines of topology. Since January 2026, he has been a researcher in Data Science and AI at Chalmers and leads the ERC-funded NextReason project.

Born
1973
Works
196
Citations
4,117
h-index
34

Years link to their sources. Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

  1. 1995–2004

    Three degrees at Charles University

    He earns a B.S. in economics at Charles University's Faculty of Social Sciences in 1995 and an M.S. in mathematics at its Faculty of Mathematics and Physics in 1998. He completes a computer science Ph.D. at the Faculty of Mathematics and Physics in 2004.

  2. 2003–2013

    Opening Mizar to AI

    In 2003, he makes the Mizar library of formal mathematics available to automated theorem proving and AI methods, and co-founds the Prague Automated Reasoning Group. By 2013, his MaLARea system combines proof deduction with machine-learning guidance and solves most problems in CASC's Large Theory division.

  3. 2006–2013

    Posts in Miami, Prague and Nijmegen

    From 2006 to 2008, he is a Marie-Curie fellow in the University of Miami's Department of Computer Science. By 2013, he is a postdoctoral researcher in the Foundations Group of ICIS at Radboud University Nijmegen, after an assistant professorship in Charles University's Department of Theoretical Computer Science and Mathematical Logic.

  4. October 2016

    Leading AI4REASON at CIIRC

    By October 2016, he is a principal researcher at the Czech Institute of Informatics, Robotics and Cybernetics and leads the ERC Consolidator project AI4REASON.

  5. 2018

    Reinforcement learning for theorem proving

    In 2018, he coauthors Reinforcement Learning of Theorem Proving with Cezary Kaliszyk, Henryk Michalewski and Mirek Olšák. Its prover uses reinforcement learning from earlier attempts to guide connection-style proof search and solves over 40% more problems than a baseline under the same number of inferences.

  6. 2025

    Learning conjectures from scratch

    In 2025, he coauthors Learning Conjecturing from Scratch with Thibault Gauthier. Their self-learning method generates induction predicates for 16,197 problems derived from the OEIS and solves 5,565, compared with 2,265 solved by CVC5, Vampire or Z3 in 60 seconds.

  7. 2026

    130k lines of formal topology

    In 2026, he submits a paper describing a feedback loop between a large language model and a proof checker that autoformalizes about 130,000 lines of topology, with most lines produced over two weeks across December 2025 and January 2026.

  8. Since 2026

    Researcher at Chalmers

    In January 2026, he joins Chalmers as a researcher in Data Science and AI. In October 2026, he leads NextReason, an ERC-funded project on theorem proving, conjecture generation, neuro-symbolic reasoning and learning-based methods for mathematical reasoning.

Career

  1. Since 2026Chalmers University of TechnologyResearcher
  2. 2014–2025Czech Technical University in Prague
  3. 2012–2014Universität Innsbruck
  4. 2010–2018Radboud University NijmegenPostdoctoral Researcher
  5. 2006–2008University of MiamiMarie-Curie Fellow
  6. From 2003Prague Automated Reasoning GroupCo-founder
  7. 2003–2008Charles UniversityAssistant Professor
  8. Czech Institute of Informatics, Robotics and CyberneticsPrincipal Researcher

Education

  1. 2004Charles UniversityPh.D. in computer science
  2. 1998Charles UniversityM.S. in mathematics

Most-cited works

  1. A FORMAL PROOF OF THE KEPLER CONJECTURE2017 · Forum of Mathematics Pi · 356 citations
  2. Mizar: State-of-the-art and Beyond2015 · Lecture notes in computer science · 229 citations
  3. Learning-Assisted Automated Reasoning with Flyspeck2014 · Journal of Automated Reasoning · 145 citations
  4. Premise Selection for Mathematics by Corpus Analysis and Kernel Methods2013 · Journal of Automated Reasoning · 128 citations
  5. MPTP 0.2: Design, Implementation, and Initial Experiments2006 · Journal of Automated Reasoning · 120 citations

Sources and credits

Photo: We have no published photo of Josef Urban yet. Send us a free-licensed photo.

  1. Josef Urban's home page ciirc.cvut.cz. Checked 21 October 2026.
  2. Josef Urban on Machine Learning and Automated Reasoning - Machine Intelligence Research Institute intelligence.org. Checked 21 October 2026.
  3. [1805.07563] Reinforcement Learning of Theorem Proving arxiv.org. Checked 21 October 2026.
  4. [2503.01389] Learning Conjecturing from Scratch arxiv.org. Checked 21 October 2026.
  5. [2601.03298] 130k Lines of Formal Topology in Two Weeks: Simple and Cheap Autoformalization for Everyone? arxiv.org. Checked 21 October 2026.
  6. 2025 - a year of extraordinary progress | University of Gothenburg gu.se. Checked 21 October 2026.
  7. Josef Urban | Chalmers chalmers.se. Checked 21 October 2026.
  8. NextReason: Next Generation Computer Assisted Reasoning supr.naiss.se. Checked 21 October 2026.