Mathematics | Applied category theory | AI evaluation

Paul Wang

I am a postdoctoral researcher at LIP6, Sorbonne University, in the Multi-Agent Systems team, under the supervision of Vincent Corruble. My postdoc is funded by a Coefficient Giving grant. My current focus is zero-knowledge verification for frontier AI training.

I completed my PhD in mathematics at the Department of Mathematics and Applications of the École Normale Supérieure - PSL in 2025, under the supervision of Silvain Rideau-Kikuchi. My doctoral work studied variations on the group configuration theorem, with applications to definable groups and fields in NTP2 theories.

I am working closely with the GPAI Policy Lab verification programme and with the contributor group listed there, including my co-authors Pierre Peigné and Ky Nguyen. Other research strands listed below are currently on hold or lower priority while this programme is active.

Research ZK paper GPAI programme Teaching CV PDF

Current Focus

I am currently prioritising technical verification for frontier AI training: protocols that can make claims about large training runs checkable while preserving confidentiality.

The main public reference is Zero knowledge verification for frontier AI training is possible, developed with Pierre Peigné and Ky Nguyen, alongside the GPAI Policy Lab verification programme.

Research Threads

These are longer-running threads in my work. For now, they are mostly on hold while I focus on frontier AI training verification.

AI evaluation

Formal and conceptual frameworks for evaluating AI systems beyond static benchmarks, including behaviour, safety, and performance.

Categorical systems theory

Compositional accounts of systems, explanations, nondeterminism, and trajectories.

Model theory

Definable groups and fields, NTP2 theories, valued fields, and variants of the group configuration theorem.

Interpretability and safety

Mathematical tools for local explanations, gluing questions, and robustness of refusal training.

Recent Preprints