Thesis Director: Prof. V. Panaretos,
\r\nMathematics doctoral program
\r\nThesis Nr. 11953
\r\n
\r\nTo take part in the public defense, please contact directly the speaker
Thesis Directors: Prof. S. Süsstrunk, Dr R. Achanta
\r\nComputer and Communication Sciences doctoral program
\r\nThesis Nr. 11581
\r\n
\r\nTo take part in the public defense, please contact directly the speaker
Thesis Directors: Prof. C. R. Binder Signer, Dr M. Duygan
\r\nCivil and Environmental Engineering doctoral program
\r\nThesis Nr. 11896
\r\n
\r\nTo take part in the public defense, please contact directly the speaker
\r\n
\r\nPlease register via the Google Form here:
\r\nhttps://forms.gle/e6UF2Sr9LdcgNLUn7
Wave functions describing bound states of electronic systems satisfy high-dimensional Schrödinger equations whose potentials exhibit Coulomb singularities at particle-collision configurations. These singularities generate cusps that limit the global regularity of the wave function in spectral Barron spaces, which are particularly relevant to the analysis of Fourier-based approximation methods and neural networks.
\r\nIn this talk, we study the effect of extracting a truncated Jastrow factor designed to capture the leading coalescence singularities while preserving the asymptotic behavior of the wave function at infinity. We show that the resulting quotient gains one full order of spectral Barron regularity: whereas the original wave function typically belongs to spaces of order (s<1), the Jastrow quotient belongs to those of order (s<2). An explicit hydrogenic example shows that the threshold (s=2) is natural.
\r\nThe proof relies on conjugating the Hamiltonian by the Jastrow factor and carrying out a global Fourier-space analysis. It combines resolvent estimates, precise control of singular coefficients, and a fixed-point argument separating low and high frequencies.
\r\n
\r\n---
\r\nThe electronic structure reading group brings together researchers and students interested in mathematical aspects of electronic structure problems and adjacent topics, including:\r\n
This workshop is for staff and researchers working in NGOs and non-profits who aim to bring AI into their workflows with an ethical, careful approach. Join us for a hands-on morning of using and deploying the flagship model of the Swiss AI Initiative, learn first hand from the engineers and scientists working on Apertus.
\r\n
\r\nThe workshop runs from 9:30 to 12:30 at the EPFL AI Center Lounge in Lausanne. Doors open at 9:00.
\r\n
\r\nBackground: Apertus is a fully open large language model built in Switzerland, which is fully reproducible and can run on your own infrastructure. As a Foundation Model, it has already been adapted to specific contexts through post-training and fine-tuning processes. In addition to open access at all Swiss academic institutions, a growing number of Swiss and international commercial providers support Apertus for production deployment.
\r\n
\r\nThis workshop is also designed as a two-way conversation: alongside learning how to use and deploy Apertus, we want to hear how your organization is already working with AI, and where today's tools still fall short for you. We would like to hear about the languages, legal contexts, and use cases you operate in. This input will give us the opportunity to prioritize improvements for non-profit organizations on the Apertus roadmap.
\r\n
\r\nAt this workshop, you will have the chance to:\r\n
The talk is jointly organized by the EPFL AI Center and the MLO Lab as part of the AI fundamentals seminar series.
\r\n
\r\nHosting Professor: Prof. Martin Jaggi
\r\n
\r\nTitle
\r\nSpeedrunning SWE-bench, or the $1,000 coding agent
\r\n
\r\nAbstract
\r\nSpeedrunning has become a popular engineering challenge in machine learning: achieve the best performance within a fixed compute budget. We use it as a scientific tool to study how models acquire specialized skills and how far those skills transfer. We speedrun SWE-bench by training coding agents from scratch on demonstrations from stronger models. With about $1,000 of training compute, we reach a 25% solve rate, outperforming GPT-4o. Compared to models with extensive pre-training, our speedrun models solve largely the same SWE-bench problems, and generalize about as well to new bugs and codebases. What, then, does pre-training buy for SWE-bench? Our experiments suggest a one-student hypothesis: pre-training helps a student imitate its teacher, but at matched imitation loss, pre-training choices barely matter. We stress-test this hypothesis by comparing a model trained on the web against a “vintage\" one trained only pre-1931 documents. Up to the SWE-bench scores we reach, pre-training is interchangeable with specialization data. We release nanoswe as an open project for studying specialization through speedrunning.
\r\n
\r\nBio
\r\nRicardo Olmedo is a final-year PhD student at the Max Planck Institute for Intelligent Systems in Tübingen and a Google PhD Fellow. Advised by Moritz Hardt and Bernhard Schölkopf, he works on language model benchmarking and specialization. His research has earned three oral presentations at ICLR and NeurIPS. Most recently, he spent spring and summer 2026 visiting Sanmi Koyejo at Stanford University.