Nico Scherf

Sasha Brenner, Thomas R. Knösche, Nico Scherf

Next-token predictors often appear to develop internal representations of the latent world and its rules. The probabilistic nature of these models suggests a deep connection between the structure of the world and the geometry of probability distributions. In order to understand this link more precisely, we use a minimal stochastic process as a controlled setting: constrained random walks on a two-dimensional lattice that must reach a fixed endpoint after a predetermined number of steps. Optimal prediction of this process solely depends on a sufficient vector determined by the walker's position relative to the target and the remaining time horizon; in other words, the probability distributions are parametrized by the world's geometry. We train decoder-only transformers on prefixes sampled from the exact distribution of these walks and compare their hidden activations to the analytically derived sufficient vectors. Across models and layers, the learned representations align strongly with the ground-truth predictive vectors and are often low-dimensional. This provides a concrete example in which world-model-like representations can be directly traced back to the predictive geometry of the data itself. Although demonstrated in a simplified toy system, the analysis suggests that geometric representations supporting optimal prediction may provide a useful lens for studying how neural networks internalize grammatical and other structural constraints.

Nikola Milosevic, Johannes Müller, Nico Scherf

Reinforcement Learning (RL) agents can solve diverse tasks but often exhibit unsafe behavior. Constrained Markov Decision Processes (CMDPs) address this by enforcing safety constraints, yet existing methods either sacrifice reward maximization or allow unsafe training. We introduce Constrained Trust Region Policy Optimization (C-TRPO), which reshapes the policy space geometry to ensure trust regions contain only safe policies, guaranteeing constraint satisfaction throughout training. We analyze its theoretical properties and connections to TRPO, Natural Policy Gradient (NPG), and Constrained Policy Optimization (CPO). Experiments show that C-TRPO reduces constraint violations while maintaining competitive returns.

Nikola Milosevic, Johannes Müller, Nico Scherf

In constrained Markov decision processes, enforcing constraints during training is often thought of as decreasing the final return. Recently, it was shown that constraints can be incorporated directly into the policy geometry, yielding an optimization trajectory close to the central path of a barrier method, which does not compromise final return. Building on this idea, we introduce Central Path Proximal Policy Optimization (C3PO), a simple modification of the PPO loss that produces policy iterates, that stay close to the central path of the constrained optimization problem. Compared to existing on-policy methods, C3PO delivers improved performance with tighter constraint enforcement, suggesting that central path-guided updates offer a promising direction for constrained policy optimization.

Semih Eren, Deniz Kucukahmetler, Nico Scherf

Accurately predicting distributed cortical responses to naturalistic stimuli requires models that integrate visual, auditory and semantic information over time. We present a hierarchical multimodal recurrent ensemble that maps pretrained video, audio, and language embeddings to fMRI time series recorded while four subjects watched almost 80 hours of movies provided by the Algonauts 2025 challenge. Modality-specific bidirectional RNNs encode temporal dynamics; their hidden states are fused and passed to a second recurrent layer, and lightweight subject-specific heads output responses for 1000 cortical parcels. Training relies on a composite MSE-correlation loss and a curriculum that gradually shifts emphasis from early sensory to late association regions. Averaging 100 model variants further boosts robustness. The resulting system ranked third on the competition leaderboard, achieving an overall Pearson r = 0.2094 and the highest single-parcel peak score (mean r = 0.63) among all participants, with particularly strong gains for the most challenging subject (Subject 5). The approach establishes a simple, extensible baseline for future multimodal brain-encoding benchmarks.

Nahid Torbati, Michael Gaebler, Simon M. Hofmann et al.

Representational similarity analysis (RSA) is widely used to analyze the alignment between humans and neural networks; however, conclusions based on this approach can be misleading without considering the underlying representational geometry. Our work introduces a framework using Ollivier Ricci Curvature and Ricci Flow to analyze the fine-grained local structure of representations. This approach is agnostic to the source of the representational space, enabling a direct geometric comparison between human behavioral judgments and a model's vector embeddings. We apply it to compare human similarity judgments for 2D and 3D face stimuli with a baseline 2D native network (VGG-Face) and a variant of it aligned to human behavior. Our results suggest that geometry-aware analysis provides a more sensitive characterization of discrepancies and geometric dissimilarities in the underlying representations that remain only partially captured by RSA. Notably, we reveal geometric inconsistencies in the alignment when moving from 2D to 3D viewing conditions.This highlights how incorporating geometric information can expose alignment differences missed by traditional metrics, offering deeper insight into representational organization.

Charlotte Beylier, Hannah Selder, Arthur Fleig et al.

While deep reinforcement learning agents demonstrate high performance across domains, their internal decision processes remain difficult to interpret when evaluated only through performance metrics. In particular, it is poorly understood which input features agents rely on, how these dependencies evolve during training, and how they relate to behavior. We introduce a scientific methodology for analyzing the learning process through quantitative analysis of saliency. This approach aggregates saliency information at the object and modality level into hierarchical attention profiles, quantifying how agents allocate attention over time, thereby forming attention trajectories throughout training. Applied to Atari benchmarks, custom Pong environments, and muscle-actuated biomechanical user simulations in visuomotor interactive tasks, this methodology uncovers algorithm-specific attention biases, reveals unintended reward-driven strategies, and diagnoses overfitting to redundant sensory channels. These patterns correspond to measurable behavioral differences, demonstrating empirical links between attention profiles, learning dynamics, and agent behavior. To assess robustness of the attention profiles, we validate our findings across multiple saliency methods and environments. The results establish attention trajectories as a promising diagnostic axis for tracing how feature reliance develops during training and for identifying biases and vulnerabilities invisible to performance metrics alone.

Nikola Milosevic, Nico Scherf

Reward maximization, safe exploration, and intrinsic motivation are often studied as separate objectives in reinforcement learning (RL). We present a unified geometric framework, that views these goals as instances of a single optimization problem on the space of achievable long-term behavior in an environment. Within this framework, classical methods such as policy mirror descent, natural policy gradient, and trust-region algorithms naturally generalize to nonlinear utilities and convex constraints. We illustrate how this perspective captures robustness, safety, exploration, and diversity objectives, and outline open challenges at the interface of geometry and deep RL.

Sebastian Niehaus, Ingo Roeder, Nico Scherf

Decentralised learning enables the training of deep learning algorithms without centralising data sets, resulting in benefits such as improved data privacy, operational efficiency and the fostering of data ownership policies. However, significant data imbalances pose a challenge in this framework. Participants with smaller datasets in distributed learning environments often achieve poorer results than participants with larger datasets. Data imbalances are particularly pronounced in medical fields and are caused by different patient populations, technological inequalities and divergent data collection practices. In this paper, we consider distributed learning as an Stackelberg evolutionary game. We present two algorithms for setting the weights of each node's contribution to the global model in each training round: the Deterministic Stackelberg Weighting Model (DSWM) and the Adaptive Stackelberg Weighting Model (ASWM). We use three medical datasets to highlight the impact of dynamic weighting on underrepresented nodes in distributed learning. Our results show that the ASWM significantly favours underrepresented nodes by improving their performance by 2.713% in AUC. Meanwhile, nodes with larger datasets experience only a modest average performance decrease of 0.441%.

Charlotte Beylier, Simon M. Hofmann, Nico Scherf

The learning process of a reinforcement learning (RL) agent remains poorly understood beyond the mathematical formulation of its learning algorithm. To address this gap, we introduce attention-oriented metrics (ATOMs) to investigate the development of an RL agent's attention during training. In a controlled experiment, we tested ATOMs on three variations of a Pong game, each designed to teach the agent distinct behaviours, complemented by a behavioural assessment. ATOMs successfully delineate the attention patterns of an agent trained on each game variation, and that these differences in attention patterns translate into differences in the agent's behaviour. Through continuous monitoring of ATOMs during training, we observed that the agent's attention developed in phases, and that these phases were consistent across game variations. Overall, we believe that ATOM could help improve our understanding of the learning processes of RL agents and better understand the relationship between attention and learning.

Nikola Milosevic, Gesine Müller, Jan Huisken et al.

In this work, we introduce the concept of Active Representation Learning, a novel class of problems that intertwines exploration and representation learning within partially observable environments. We extend ideas from Active Simultaneous Localization and Mapping (active SLAM), and translate them to scientific discovery problems, exemplified by adaptive microscopy. We explore the need for a framework that derives exploration skills from representations that are in some sense actionable, aiming to enhance the efficiency and effectiveness of data collection and model building in the natural sciences.

Nastassya Horlava, Alisa Mironenko, Sebastian Niehaus et al.

In recent years, Convolutional Neural Networks (CNNs) have become the state-of-the-art method for biomedical image analysis. However, these networks are usually trained in a supervised manner, requiring large amounts of labelled training data. These labelled data sets are often difficult to acquire in the biomedical domain. In this work, we validate alternative ways to train CNNs with fewer labels for biomedical image segmentation using. We adapt two semi- and self-supervised image classification methods and analyse their performance for semantic segmentation of biomedical microscopy images.

Marie Kloenne, Sebastian Niehaus, Leonie Lampe et al.

Machine Learning has considerably improved medical image analysis in the past years. Although data-driven approaches are intrinsically adaptive and thus, generic, they often do not perform the same way on data from different imaging modalities. In particular Computed tomography (CT) data poses many challenges to medical image segmentation based on convolutional neural networks (CNNs), mostly due to the broad dynamic range of intensities and the varying number of recorded slices of CT volumes. In this paper, we address these issues with a framework that combines domain-specific data preprocessing and augmentation with state-of-the-art CNN architectures. The focus is not limited to optimise the score, but also to stabilise the prediction performance since this is a mandatory requirement for use in automated and semi-automated workflows in the clinical environment. The framework is validated with an architecture comparison to show CNN architecture-independent effects of our framework functionality. We compare a modified U-Net and a modified Mixed-Scale Dense Network (MS-D Net) to compare dilated convolutions for parallel multi-scale processing to the U-Net approach based on traditional scaling operations. Finally, we propose an ensemble model combining the strengths of different individual methods. The framework performs well on a range of tasks such as liver and kidney segmentation, without significant differences in prediction performance on strongly differing volume sizes and varying slice thickness. Thus our framework is an essential step towards performing robust segmentation of unknown real-world samples.