Programmes
Past workshop

Maths for AI Safety Workshop

30th Oct – 1st Nov 2026 · Bonn, Germany

Our first weekend workshop where mathematicians used maths to explore problems in AI safety and form collaborations, for PhDs, postdocs and faculty members.

Photos

Group photo of the workshop participants in the garden of the Universitätsclub BonnMartin Rusch presenting “Why AI safety needs mathematicians”A participant speaking during a discussionLuiza Cristina Corpaci presenting on verification and validationParticipants working together at a laptopJessica Wang talking with participants during a break

Speakers

  • Lionel LevineCornell University

    Lionel is a maths professor at Cornell. His research is aimed at increasing the probability that future AI systems benefit all of humanity and the planet. Some things he has worked on are how to quantify value alignment and character traits in language models, a definition of AGI, a mathematical reasoning benchmark, a metric for belief coherence, and hidden incentives for planning.

    Website
  • Martin RuschSaferAI

    Martin is the Chief Operating Officer at SaferAI. His research interests include AI-driven systemic risks to the financial system and frontier model risk management. Previously managed large-scale product, engineering and data teams in tech companies. Holds a DPhil in theoretical physics from Oxford.

    Website
  • Kyler SiegelImperial College London

    Kyler is an assistant professor of mathematics at Imperial College London and a tenured associate professor (on leave) at the University of Southern California. His background is in symplectic geometry but also has projects in AI safety and interpretability.

    Website
  • Sergei SmirnovUniversity of Helsinki

    Sergei is a PhD researcher in Computer Science at the University of Helsinki, focusing on empirical AI safety and LLM evaluations. He has 5+ years of ML engineering experience, built CivAI’s public AI Values demo, and completed research fellowships at MARS and SPAR.

    LinkedIn
  • Luiza Cristina CorpaciTU Wien

    Luiza is an MSc student at TU Wien with a background covering chip design (Marvell), combinatorial testing (SBA Research, Vienna), and interpretability/explainability (University of Vienna). At AMD, she worked on AI-assisted hardware verification, helping develop pipelines that translate natural-language specifications into code artifacts with formal verification in the loop. Her current research explores the reliability of AI-generated formal specifications, particularly how to detect and characterize semantic drift between intended and formalized meaning.

    Google Scholar
  • Leonardo Costa LesageIliad

    Leonardo is currently a fellow at Iliad, working on synthetic graph models of debate. He participated in the Iliad Intensive in the June cohort. Previously, he spent a year at Oxford studying the chaotic dynamics of the 3-body problem and connecting it with symplectic geometry. This was part of his last year of a double bachelor’s degree in Math and Engineering Physics at CFIS (UPC) in Barcelona. He has always been a math enthusiast and participated in math olympiads in high school.

    LinkedIn
  • Léo DanaInria Paris

    Léo is a French maths PhD student advised by Loucas Pillaud-Vivien and Francis Bach at Sierra, Inria Paris. His research interests are in understanding deep-learning systems through mathematical/theoretical tools, in particular during training and deployment. He was a fellow at MATS 4.0, ML4Good mentor, and teaches the Turing seminar at Master MVA.

    Website
  • David PrinzIliad, London Initiative for Safe AI

    David is a mathematical physicist and Iliad Fellow at the London Initiative for Safe AI. Trained in quantum field theory, with a focus on Hopf algebraic renormalization and the BV-BRST quantization of gauge theories, he now works on the mathematical foundations of machine learning, especially singular learning theory, with an eye towards theoretical aspects of alignment.

    Website
  • Renata PicciottoUniversity College London

    Renata is a postdoc researcher at University College London working in Enumerative Geometry. She is also an Iliad Fellow working on the geometry of ReLU networks.

    Website

Agenda

Fri 30th Oct · Evening
  • Informal dinner for attendees and organisers
Sat 31st Oct · Full day
  • Why AI safety, and where maths comes in
  • Linking areas of maths to open problems
  • Small group sessions in the afternoon
Sun 1st Nov · Until lunchtime
  • Groups present ideas and possible projects
  • Next steps
  • Close, then lunch in town together, at your own cost

Organising team

  • Jessica P. WangFounder & Executive DirectorSafe AI Germany
  • André VidelaDirectorGlasgow Lab for AI Verification
  • Claudius ZibrowiusProfessor in Low-Dimensional TopologyBochum University

Organised by

Safe AI Germany

Financially supported by

BlueDot Impact

Future workshops

Sign up for our newsletter to hear about future workshops.

Hear about future workshops
Maths for AI Safety Workshop photos
Group photo of the workshop participants in the garden of the Universitätsclub BonnMartin Rusch presenting “Why AI safety needs mathematicians”A participant speaking during a discussionLuiza Cristina Corpaci presenting on verification and validationParticipants working together at a laptopJessica Wang talking with participants during a breakA participant speaking in a small-group sessionKyler Siegel presenting on a small language model exampleLeonardo Costa Lesage presenting “Modelling debate with factor graphs”Léo Dana presenting a toy model of activation vectorsThree participants presenting their group’s workA participant presenting working groups to joinA talk on the State of AI ReportSergei Smirnov presenting on fieldbuilding in AI safetyA participant leading a session on category theory at the whiteboardSergei Smirnov presenting “Does thinking longer make LLMs worse?”Three participants in conversation at a tableA participant asking a question during a talkJessica Wang, founder of Safe AI Germany, speaking at the podiumA participant presenting the working groups and their rooms