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.
![\"\"](\"https://foster-lead-promote.ch/wp-content/uploads/2026/01/CG_EPFL_2026.png\")
\r\nIn all professions including academic careers, satisfaction and a sense of fulfillment are not always at hand, leading sometimes to questioning professional choices. Have you experienced not being completely sure of what is professionally right for you? Have you wondered how you could find it out? Do you feel you might be not well equipped to decide on the next step in your career, or to lack orientation? In this workshop, you will be provided with tools and orientation models that will guide you in your career decisions, and implement them. These are based on an authenticity-oriented perspective aiming at a fulfilling work and self-realization experience.
","image_description":"","creation_date":"2026-07-14T11:21:21","last_modification_date":"2026-07-14T11:23:17","link_label":"Website","link_url":"https://foster-lead-promote.ch/our_services/coachinggroups/","canceled":"False","cancel_reason":"","place_and_room":"EPFL","url_place_and_room":"","url_online_room":"","spoken_languages":["https://memento.epfl.ch/api/v1/spoken_languages/2/?format=json"],"speaker":"Afi Sika KuzeawuRecent reports claim that hydrogen peroxide (H2O2) is spontaneously generated – without external energy input or catalyst – at the air–water interface of microdroplets1, bubbles2, and soap films3. Several mechanisms have been proposed to explain this phenomenon, including: (i) ultrahigh instantaneous electric fields at the air–water interface driving HO− à HO• + e− radicals from hydroxide ions; (ii) water’s comproportionation into H2O•+ and H2O•−; and (iii) partial dehydration of interfacial H+ and HO− followed by radical formation. In contrast, our laboratory experiments reveal that H2O2 formation is due to the reduction of dissolved O2 at water– solid interfaces, i.e., with no role for the air–water interface beyond O2 mass transport4-6. Based on these findings, I will present four diagnostic criteria against which any existing or future mechanistic proposal should be assessed. Our approach reconciles the claims of spontaneous H2O2 formation within established physical chemistry and extends naturally to other claims of anomalous microdroplet reactivity.
\r\n
\r\nReferences:
\r\n(1) Mehrgardi, M. A.; Mofidfar, M.; Zare, R. N. Sprayed Water Microdroplets Are Able to Generate Hydrogen Peroxide Spontaneously. Journal of the American Chemical Society 2022, 144 (17), 7606-7609. DOI: 10.1021/jacs.2c02890.
\r\n(2) Nami-Ana, S. F.; Mehrgardi, M. A.; Mofidfar, M.; Zare, R. N. Sustained Regeneration of Hydrogen Peroxide at the Water-Gas Interface of Electrogenerated Microbubbles on an Electrode Surface. Journal of the American Chemical Society 2024, 146 (46), 31945-31949. DOI: 10.1021/jacs.4c11422 From NLM PubMed-not-MEDLINE.
\r\n(3) Eldeeb, A. M.; Berbille, A.; Dick, J. E. No Microdroplets? No Problem. Soap Films Amplify the Interface and Produce Hydrogen Peroxide. Journal of the American Chemical Society 2025, 147 (37), 33325-33329. DOI: 10.1021/jacs.5c05887.
\r\n(4) Eatoo, M. A.; Mishra, H. Busting the myth of spontaneous formation of H2O2 at the air-water interface: contributions of the liquid-solid interface and dissolved oxygen exposed. Chemical Science 2024, 15 (9), 3093-3103. DOI: 10.1039/d3sc06534k From NLM PubMed-not-MEDLINE.
\r\n(5) Eatoo, M. A.; Wehbe, N.; Kharbatia, N.; Guo, X.; Mishra, H. Why do some metal ions spontaneously form nanoparticles in water microdroplets? Disentangling the contributions of the air-water interface and bulk redox chemistry. Chemical Science 2025, 16 (3), 1115-1125. DOI: 10.1039/d4sc03217a From NLM PubMed-not-MEDLINE.
\r\n(6) Eatoo, M. A.; Mishra, H. Disentangling the Roles of Dissolved Oxygen, Common Salts, and pH on Spontaneous Hydrogen Peroxide Production in Water: No O2, No H2O2. Journal of the American Chemical Society 2025, 147 (39), 35392-35400. DOI: 10.1021/jacs.5c09028.
\r\n
\r\nPapers:
\r\n- Busting the myth of spontaneous formation of H2O2 at the air–water interface: contributions of the liquid–solid interface and dissolved oxygen exposed
\r\n- Disentangling the Roles of Dissolved Oxygen, Common Salts, and pH on Spontaneous Hydrogen Peroxide Production in Water: No O2, No H2O2
\r\n
\r\nBio: Himanshu Mishra is a scientist-entrepreneur from King Abdullah University of Science and Technology (KAUST), Saudi Arabia. His team has significantly contributed to resolving scientific debates surrounding the spontaneous H2O2 formation in water microdroplets, the electrification at water-hydrophobe interfaces, and liquid–solid adhesion forces. Mishra’s team has also pioneered Carbosoil™ technology to transform desert soils for sustainable greening, food production, and carbon sequestration. His start-up Terraxy is now scaling up Carbosoil™ production to advance sustainability projects in the Middle East and beyond. Mishra’s efforts have been featured by the World Economic Forum, Geneva Inventions, and CNN (LINK).
\r\n
Atomically precise carbon nanostructures offer a new route to quantum systems built from the bottom up. By combining molecular design, on-surface synthesis and scanning probe microscopy, we can control atomic structure, electronic states and magnetic interactions within the same platform.
\r\nIn this talk, I will follow the path from bottom-up synthesized graphene nanoribbons [1,2] to open-shell nanographenes [3] and designer quantum spin systems [4]. I will discuss how atomic structure can determine electronic topology and how molecular spins can be coupled into well-defined quantum chains. Recent examples include Haldane chains [4], alternating-exchange spin chains [5] and antiferromagnetic Heisenberg chains [6], where collective and fractional excitations can be probed directly at the atomic scale.
\r\nMore broadly, this work illustrates a materials-science strategy in which structure, functionality and interactions are programmed at the level of molecular building blocks. Rather than optimizing properties in an existing material, the aim is to design the material and its quantum behavior together. Atomically precise carbon nanostructures therefore provide a model platform for exploring how synthesis, structure and emergent functionality can be linked in a predictive way, with implications for the future design of low-dimensional electronic, magnetic and quantum materials.\r\n