01.04.2026

ELLIS Winter School on Foundation Models

In late March, I attended the Foundation Models Winter School, hosted by the European Laboratory for Learning and Intelligent Systems (ELLIS) in Amsterdam. The first day started with Saining Xie (NYU Courant, AMI Labs) explaining why visual representation still matters, Hinrich Schütze (ELLIS Munich) shedding light on internal representations, and Pascal Mettes (ELLIS Amsterdam) introducing hyperbolic foundation models—which almost felt like a political campaign against Euclidean space.

Over the following days, Yossi Gandelsman (Reve, former UC Berkeley) shared his research on model hijacking and steering, Konstantinos Derpanis (York University) revisited interpretability through the lens of universal concepts, Robert Geirhos (Google DeepMind) offered insights into how video models can help spatial understanding, Frank Hutter (ELLIS Tuebingen) highlighted the growing role of foundation models in tabular data, and Ekaterina Shutova (ELLIS Amsterdam) spoke about the challenges of developing multilingual LLMs. On the final day, Nuria Oliver (ELLIS Alicante) began by reflecting on how human biases manifest in foundation models—illustrated by the example of lookism—before presenting work on how effectively LLMs can jailbreak one another. Elisa Ricci (ELLIS Trento) closed the winter school by emphasizing the need to move beyond accuracy and toward trust.

During the panel discussion, ‘From Research to Real-World Impact’, academics seemed to shift their focus when speaking as startup founders, emphasising the constraints of European regulations rather than the advantages of designing systems that reflect our values. Hopefully, future founders will take a different approach. On a more positive note, a poster session featuring PhD students’ research offered a glimpse into exciting ongoing work and left me feeling optimistic about what’s ahead.

Finally, regarding safety: while many speakers touched on the topic, the focus often leaned more toward finding ways to break systems than toward building architectures that ensure safety. Safety assurance for foundation models is clearly difficult, but that’s exactly why it deserves more attention, not less.