CAIDAS Contributions at NeurIPS 2026
29.10.2026Three papers involving researchers from CAIDAS have been accepted at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026). NeurIPS is one of the world’s leading and most prestigious conferences in artificial intelligence and machine learning, bringing together researchers from academia and industry to present advances in areas such as machine learning, neural networks and data science.
The first contribution Weisfeiler and Leman Follow the Arrow of Time: Expressive Power of Message Passing in Temporal Event Graphs comes from CAIDAS researchers Franziska Heeg and Prof. Dr. Ingo Scholtes, together with Jonas Sauer and Prof. Dr. Petra Mutzel from the University of Bonn. Their work focuses on networks that change over time. Such dynamic networks appear in many areas, for example in communication, mobility or social interactions. The researchers develop new theoretical foundations for understanding what graph-based AI models can learn from temporal structures and use these insights to improve learning methods for time-dependent network data.
The second contribution Online Decision-Focused Learning under Semi-Bandit Feedback addresses machine learning for decision-making under uncertainty. Many real-world systems - from routing to energy markets to portfolio selection - make decisions continuously from uncertain forecasts and learn only from the option they chose. A natural approach is to fill in the unobserved costs with the model’s own predictions. The paper shows that this backfires: the errors compound. It introduces BayesianSPO, a method that avoids the problem. The work is led by doctoral researcher Aabhash Dhakal from the Chair of Information Systems and Business Analytics, together with Tim Lachner, and Prof. Dr. Christoph M. Flath from CAIDAS in cooperation with Jayanta Mandi from Loughborough University and Marco Foschini from KU Leuven.
The third contribution Quantifying Concentration Phenomena of Mean-Field Transformers in the Low-Temperature Regime from CAIDAS researcher Prof. Dr. Leon Bungert, together with Albert Alcalde from the FAU Erlangen, Konstantin Riedl from the University of Oxford and Tim Roith from TU Munich, investigates the mathematical foundations of transformer models, which underpin many modern AI systems. The work studies how information is processed within self-attention transformer architectures and provides new insights into how the internal representations of tokens evolve and concentrate as information passes through the layers.
The three contributions highlight the broad range of AI research at CAIDAS and the University of Würzburg, spanning graph learning, decision-making under uncertainty and the mathematical foundations of modern machine-learning models.
The works will be presented at NeurIPS 2026 in December. This year, the conference will take place across three locations in Sydney, Atlanta and Paris.
