A (Travel) Report from a Winter School, a Workshop, and a Refreshed Excitement for Research
2026-04-15
Kai Moltzen, researcher at our institute, recently returned from what he describes as possibly the two most inspiring weeks of his professional career so far. In late March, he attended the ELLIS Winter School on Foundation Models in Amsterdam, followed by the GeoExT workshop at ECIR 2025 - a packed fortnight of new ideas, stimulating discussions, and valuable new connections.
The Winter School, hosted by the Amsterdam Unit of European Excellence Network for AI Research ELLIS, brought together an outstanding lineup of speakers from across the research landscape. Saining Xie, co-founder of AMI Labs (currently one of Europe's most highly valued AI startups, co-founded with Yann LeCun), opened the school with a talk on world models capable of spatial cognition and predicting world states. Hinrich Schütze, a founding figure of neural NLP from LMU Munich, drew a thought-provoking distinction between creative generalization, hallucination, and non-answers in LLMs. When a model is asked about an unknown fact, it may give a plausible but unverifiable answer. This, Schütze argues, is meaningfully different from a straightforward hallucination of a known fact. Pascal Mettes from the University of Amsterdam presented on Hyperbolic Foundation Models, arguing that hyperbolic geometry is better suited than Euclidean space for modeling classification hierarchies. One intuition: if you briefly glimpse something moving on a rooftop, you might only be confident enough to say "animal"; with more certainty, "bird"; with even more, "blackbird." This hierarchical uncertainty framing got Kai thinking about potential integration with geographic cell hierarchies. Frank Hutter, from the University of Freiburg and PriorLabs, presented arguably one of the most exciting talks of the school. His work on TabPFN introduces a tabular foundation model that ingests an entire dataset at inference time for in-context learning, returning predictions in under a second; the corresponding paper was the most-cited Nature publication of last year. Given how much valuable data in industry and finance exists in tabular form, this feels like a breakthrough with enormous practical impact made in Europe. Nuria Oliver from ELLIS Alicante shared striking findings on AI safety: her team's latest paper in Nature Communications demonstrated that large reasoning models can be used as autonomous agents to jailbreak other LLMs - achieving close to a 100% success rate, raising serious questions about the effectiveness of current alignment approaches. Elisa Ricci from the University of Trento rounded out the trustworthy AI thread with a presentation on explainability in multimodal models, focusing on data-free approaches that analyze model weights rather than data-driven activations. This is particularly relevant for making targeted interventions to improve model trustworthiness.
Other highlights included Robin Rombach, co-founder of Black Forest Labs and creator of latent diffusion models, who was interviewed on stage; Ekaterina Shutova's work on multilingual models and the surprising finding that even a single epoch of fine-tuning can rewire a model's cultural value representations; and Robert Geirhos from Google DeepMind, who delivered an insightful half-talk on research philosophy - including the memorable line: "Good ideas fail all the time, but good questions don't."
Beyond the talks themselves, the school offered something equally valuable: the time and space to simply think, and the opportunity to exchange ideas with a diverse group of talented young researchers from across Europe. Kai came home with a long list of research threads he is eager to pursue. And, perhaps most importantly, with a renewed excitement for doing excellent research and a strong conviction in Europe's potential in frontier foundation model research.
Following the winter school, Kai attended the GeoExT workshop at ECIR 2025 in Delft dedicated to advancing the geographic capabilities of machine learning systems. Hearing cutting-edge work in this specific domain and connecting with like-minded researchers was a highlight in its own right, and he is confident that fruitful collaborations will emerge from these new connections.
Two weeks, two events, countless ideas. We look forward to seeing how these experiences feed into the research ahead!