[1/3] Happy to share that our paper "MapPFN: Learning Causal Perturbation Maps in Context" has been accepted at NeurIPS 2026! 🎉 TL;DR: We frame perturbation prediction as an in-context learning problem given a set of interventional experiments.

We thank her for sharing her work with our community. More BIFOLD Colloquia will follow soon. Photos: BIFOLD/Gereon Dusella

, and early work on foundation models and agents for predictive reasoning over tables. Together, these threads point toward smaller, specialized models that make insights from structured data more accessible and trustworthy.

assessing when a dataset is sensitive (a notion she argued goes beyond personally identifiable information and depends on context), grounding text-to-SQL models in real-world database schemas through the new SQaLe dataset

Many thanks to Madelon Hulsebos for her talk at the BIFOLD Colloquium. @madelonhulsebos.bsky.social @tuberlin.bsky.social In "On the Different Flavors of Tabular Reasoning," she showed that reasoning over tabular data spans several distinct challenges:

BIFOLD Co-Director Klaus-Robert Müller opened with a call for Safety by Design: safety must be built into AI systems from the outset, not added later. Keynotes by Judith Simon, Marc Toussaint, and Iyad Rahwan, plus a visit to Google AI Center Berlin.

Photo Recap: Berlin Summer School on AI and Society. From Sept 14-17, 50 students from 10 countries met in Berlin for four days on the responsible design of AI-driven autonomous systems, robotics, and physical AI. Organized by BIFOLD, the Weizenbaum Institute, and TU Berlin's LIS Lab.

Madelon Hulsebos, a tenured senior researcher at Centrum Wiskunde & Informatica (CWI) in Amsterdam, will explore how to judge the sensitivity of tabular data beyond personally identifiable information, and introduce SQaLe, a large-scale text-to-SQL dataset grounded in real-world database schemas.

Led by Dr. Mina Jamshidi Idaji (BIFOLD/TU Berlin), with Prof. Dr. Philipp Jurmeister (LMU Munich) and Prof. Dr. Michael Bockmayr (UKE Hamburg).

In reality, a patient's picture develops continuously as new findings arrive. The consortium is building AI models that integrate images, molecular data, and clinical information and adapt at each stage of the diagnostic journey.

When oncologists assess a cancer patient, they draw on histopathological images (tissue samples), molecular analyses of the tumor, and clinical data. Many AI systems examine these sources separately, or treat them as one fixed snapshot of the patient.

BIFOLD researcher Dr. Mina Jamshidi Idaji to lead a €1.5 million research consortium on AI in Oncology. The project brings together oncologists, pathologists, and ml researchers from BIFOLD/TU Berlin, LMU Munich, and UKE Hamburg. 🧵 @tuberlin.bsky.social @lmu.de @xai-berlin.bsky.social