Albert Catalan-Tatjer

Hi, I'm Albert. I am a PhD candidate at the safety- & efficiency- aligned learning (🦭) group advised by Jonas Geiping at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems. I am fortunate to be funded by IMPRS-IS and Amazon Science Hub. Before this, I spent time at the Universities of Tübingen and Barcelona.

I am very interested in understanding how to build intelligent systems that can adapt after their release. Ideally, we could modify existing AIs to the demands of new technologies as they are developed—hardware requirements, legislation, use cases, etc. For instance, the analyses in Training dynamics and Learning-forgetting serve to study how to ameliorate robustness to weight perturbations with interventions during the pretraining stage, or the posttraining stage respectively. Furthermore, at deployment time, in-context learning (ICL) adapts the system to its context history; in C3 we propose how to persist the effect of the context history to the system beyond the KV cache.

I am following the standing invitation: if you want to talk to me, I want to talk to you. Send me an email or message me on twitter.

News

• Jun 2026 🇰🇷 Presented Learning-Forgetting Optimality in Supervised Finetuning: A Cliff Perspective at HiLD ICML
• Apr 2026 🇧🇷 Presented Training Dynamics Impact Post-training Quantization Robustness at ICLR 2026
• Mar 2026 🎙️ Gave a talk at Prof. Simon Lacoste-Julien research group at MILA.
• Mar 2026 🎙️ Gave a talk at Friday Talks Tübingen.
• Dec 2025 🇺🇸 At NeurIPS San Diego
• Jul 2025 🎉 Started my PhD

Selected Publications

* equal contribution