Daniel Durstewitz
Authors: Florian Hess, Florian Götz, Daniel Durstewitz
Presentation: Spotlight
Develops an approach to parallelize nonlinear recurrent neural network training, achieving over 100-fold speedups for reconstructing dynamical systems, including chaotic dynamics.
https://arxiv.org/abs/2605.12683
Authors: Christoph Jürgen Hemmer, Florian Plaswig, Daniel Durstewitz
Investigates how to simplify dynamical systems foundation models while retaining their ability to reproduce long-term statistical and geometric properties of previously unseen systems.
https://arxiv.org/abs/2607.14937
Authors: Georg Trede, Charlotte Ricarda Doll, Elias Weber, Daniel Durstewitz
Addresses generalization across fundamentally different dynamical regimes, such as transitions from cyclic to chaotic behavior, rather than merely predicting new trajectories within familiar regimes.
https://arxiv.org/abs/2606.22969
Georgia Koppe
The Dynamic-Probabilistic Consistency Gap in Chaotic Surrogate Modeling
Authors: Andre Herz, Matthijs Pals, Daniel Durstewitz, Georgia Koppe
Examines discrepancies between dynamical fidelity and probabilistic consistency in surrogate models of chaotic systems.
https://arxiv.org/abs/2605.31547
Oliver Stegle
Authors: Abdul Moeed, Stefan Schrod, Martin Rohbeck, Marc Jan Bonder, Pavlo Lutsik, Oliver Stegle, Daniel Dimitrov
Introduces Cellina, a framework for predicting how a cell's gene expression changes when its surrounding cellular environment is altered. The method separates intrinsic cell states from spatial context and was evaluated on more than 2.5 million cells, including colorectal cancer and mouse brain datasets.
Authors: Hendrik Alexander Mehrtens, Florian Buettner, Oliver Stegle
Presents a systematic benchmark of classical and neural temporal point process models, including large-scale medical datasets from UK Biobank and MIMIC-IV. The study challenges previous claims of neural models' superiority, showing that classical Hawkes processes remain competitive when equipped with flexible, learnable kernels. It also introduces an improved transformer-based approach and an open-source framework for benchmarking temporal point processes.
Klaus Maier-Hein
Dataset Collections: Challenges of Large-Scale Data Aggregation in 3D Medical Image Datasets
Authors: Yannick Kirchhoff, Saikat Roy, Elisa Stegmeier, Hamideh Haghiri, Constantin Ulrich, Maximilian R. Rokuss, Tassilo Wald, Karol Gotkowski, Benjamin Hamm, Raphael Stock, Nico Disch, Michael Baumgartner, Dimitrios Bounias, Anand Deshpande, Stefan Dinkelacker, Stefan Dvoretskii, Katharina Eckstein, Selen Erkan, Jessica Kächele, Kim-Celine Kahl, Balint Kovacs, Lucas Kulla, Moritz Langenberg, Philipp Schader, Stephen Schaumann, Darya Trofimova, Shuhan Xiao, Sebastian Ziegler, David Zimmerer, Marco Nolden, Fabian Isensee, Klaus Maier-Hein, Ralf Floca
Presentation: Poster - NeurIPS 2026 Evaluations & Datasets Track.
Examines the challenges of combining large collections of 3D medical images into datasets suitable for developing and evaluating AI models.
Ullrich Köthe
Show Me What You Don’t Know: Efficient Sampling from Invariant Sets for Model Validation
Authors: Armand Rousselot, Joran Wendebourg, Ullrich Köthe
Introduces invariance auditing, a method for examining which differences between inputs a machine learning model ignores. By generating diverse inputs that produce identical feature representations, the approach helps researchers identify whether a model has learned appropriate invariances or is overlooking information essential to the task.
https://arxiv.org/abs/2603.21782
Carsten Rother:
Do Image Editing Models Understand Lighting?
Authors: Tim Küchler, Johann-Friedrich Feiden, Matthias Niessner, Carsten Rother
Introduces the 3D-anchored Light Probe (3DLP) benchmark to evaluate whether generative image editing models accurately reproduce real-world lighting effects. Using 1,000 pairs of real-world images with controlled lighting changes, the study assesses how faithfully AI models reproduce shadows, reflections and other light interactions. Results show that leading models achieve impressive physical consistency but still struggle with subtle lighting effects, while vision-language models prove unsuitable for detailed pixel-level evaluation.
https://arxiv.org/abs/2606.26738
