Mathematical Structures in Generative AI and Machine-Assisted Reasoning
Large language models (LLMs) built on the transformer architecture now produce formally verified proofs of difficult mathematical results, yet we possess only fragments of a mathematical theory of how they represent and manipulate mathematical content. The course treats both objects, the transformer and Lean, as mathematical entities and studies their interaction using the tools of: information theory, Markov processes and their generalizations, high-dimensional geometry, mathematical logic and type theory, and computational complexity. The guiding questions are:
(Q1) What does a transformer compute, in the sense of complexity theory and formal language theory, and how does it encode mathematical concepts?
(Q2) What is the mathematical structure of Lean (type theory, categorical and operadic semantics, decision procedures), and how is that structure reflected in the representations of a language model?
(Q3) Where exactly are the bottlenecks when transformers are coupled to Lean to produce mathematical texts?
(Q4) Are the current transformers the best possible architectures for mathematical texts?
The pace is deliberately slow: every result used is stated precisely and either proved or reduced to a clearly identified result in the literature.
(Q1) What does a transformer compute, in the sense of complexity theory and formal language theory, and how does it encode mathematical concepts?
(Q2) What is the mathematical structure of Lean (type theory, categorical and operadic semantics, decision procedures), and how is that structure reflected in the representations of a language model?
(Q3) Where exactly are the bottlenecks when transformers are coupled to Lean to produce mathematical texts?
(Q4) Are the current transformers the best possible architectures for mathematical texts?
The pace is deliberately slow: every result used is stated precisely and either proved or reduced to a clearly identified result in the literature.
Lecturer
Serguei Barannikov
Date
18th September ~ 25th December, 2026
Location
| Weekday | Time | Venue | Online | ID | Password |
|---|---|---|---|---|---|
| Friday | 13:30 - 16:55 | Shuimo-LG24 | - | - | - |
Prerequisite
Measure-theoretic probability, linear algebra, and basic logic. Helpful but not required: computational complexity and Python.
Video Public
No
Notes Public
No
Language
English
Lecturer Intro
Prof. Serguei Barannikov earned his Ph.D. from UC Berkeley and has made contributions to algebraic topology, algebraic geometry, mathematical physics, and machine learning. His work, prior to his Ph.D., introduced canonical forms of filtered complexes, now known as persistence barcodes, which have become fundamental in topological data analysis. More recently, he has applied topological methods to machine learning, particularly in the study of large language models, with results published in leading ML conferences such as NeurIPS, ICML, and ICLR, effectively bridging pure mathematics and advanced AI research.