International collaboration in AI with a visiting professor from Columbia University…

Micah Goldblum, assistant professor of electrical engineering at Columbia University in New York, has just spent four weeks at the Polytechnic as a visiting professor in the Alliance program.
He had previously given two lectures at l’X as part of the Master of Science & Technology program “Data Science and AI for Business X-HEC” during the first semester of the 2025-2026 academic year.

This stay at the Center for Applied Mathematics (CMAP*), as a guest professor of the international Alliance network to which the X belongs, was devoted to two research projects in AI carried out in collaboration with El Mahdi El Mhamdi, professor at the École Polytechnique. 

The research work of Micah Goldblum, professor at Columbia University, focuses on deep learning, particularly large-scale AI systems and the scientific understanding of their mechanisms.

Before joining Columbia University, Micah Goldblum worked as a postdoctoral researcher alongside Yann LeCun and Andrew Gordon Wilson at New York University (NYU) on a wide variety of topics in the field of Machine Learningand in particular the concept of generalization, the quantification of uncertainty and learning approaches aimed at overcoming the imbalance of classes in a dataset.

Micah Goldblum says he “ AI-oriented in a very short time » after his doctorate in mathematics, while remaining very attached to his original discipline. So, when asked what he appreciates about the École Polytechnique in particular, he quotes, without hesitation, “ the importance that the École Polytechnique places on fundamentals and mathematical rigor ».

Two joint research projects: AI Safety and optimization for LLMs

At the École Polytechnique, he and researcher El Mahdi El Mhamdi began a collaboration around two research projects – one dedicated to the safety of artificial intelligence (AI Safety) and the other to training algorithms for large language models (Large language model – LLM) – which also involve other CMAP research colleagues.

« The first project focuses on methods that aim to optimize the use of LLM training data, as the highest quality data becomes scarce with increasing computing resources. “, explains Micah Goldblum. The second project concerns the optimization of LLMs: “ The instability observed in large language models shows the difficulty of training LLMs reliably. We seek to understand what causes this instability and to develop stable training algorithms. »

Committed to strengthening international research collaboration and involving other scientists, both in the United States and France, Micah Goldblum and El Mahdi El Mhamdi have opened their research projects to several doctoral students. Thus, two doctoral students from Columbia joined them at the Polytechnic School during the last week of Professor Goldblum’s visit in order to interact with Professor El Mhamdi’s team and other CMAP researchers on campus.

Professor Goldblum’s experience as a researcher in the United States allows him to observe not only common scientific practices and lines of research, but also cultural differences in research.

« In France, it is common to spend several years working on the same subject, which has the advantage of allowing you to deepen your expertise and study a problem for a long time. On the other hand, it happens that a field that was fascinating when you started studying it becomes less relevant over time. In the United States, we generally change research directions more quickly “, he observes, before adding that each approach has advantages and disadvantages, depending on the field. 

For Micah Goldblum, in the field of AI today, many researchers “seem to assume that the current form of large language models is already a permanent form.” However, he identifies “many points where these models are very suboptimal, for example their architectures, the modes of their training, or the integration of text and other parameters“, and consider thatthere is a huge field of research in AI outside of language models, which many young researchers leave aside ».

Alliance Visiting Professorships

The Alliance international network offers faculty members from partner institutions – Columbia University in New York, École Polytechnique, Sciences Po and Paris 1 Panthéon-Sorbonne – the opportunity to participate in a transatlantic exchange as a visiting professor, to strengthen links between institutions and initiate possible collaborations.

The period spent at the partner institution as a visiting professor varies and can range from a few weeks to an entire semester. During their exchange, participants can continue their research, participate in seminars and give lectures. The Visiting Professorship Alliance receives funding from the École Polytechnique Foundation as part of its “Serving Science” campaign.

* CMAP: a CNRS mixed research unit, École Polytechnique – Institut Polytechnique de Paris