Malte Schilling

Sebastian Kahl, Felix Löffler, Martin Maciol et al.

This study evaluates the performance of Large Language Models (LLMs) as an Artificial Intelligence-based tutor for a university course. In particular, different advanced techniques are utilized, such as prompt engineering, Retrieval-Augmented-Generation (RAG), and fine-tuning. We assessed the different models and applied techniques using common similarity metrics like BLEU-4, ROUGE, and BERTScore, complemented by a small human evaluation of helpfulness and trustworthiness. Our findings indicate that RAG combined with prompt engineering significantly enhances model responses and produces better factual answers. In the context of education, RAG appears as an ideal technique as it is based on enriching the input of the model with additional information and material which usually is already present for a university course. Fine-tuning, on the other hand, can produce quite small, still strong expert models, but poses the danger of overfitting. Our study further asks how we measure performance of LLMs and how well current measurements represent correctness or relevance? We find high correlation on similarity metrics and a bias of most of these metrics towards shorter responses. Overall, our research points to both the potential and challenges of integrating LLMs in educational settings, suggesting a need for balanced training approaches and advanced evaluation frameworks.

Fabian Ridder, Laurin Lessel, Malte Schilling

Retrieval-Augmented Generation (RAG) is widely used to augment the input to Large Language Models (LLMs) with external information, such as recent or domain-specific knowledge. Nonetheless, current models still produce closed-domain hallucinations and generate content that is unsupported by the retrieved context. Current detection approaches typically treat hallucination as a post-hoc problem, relying on black-box consistency checks or probes over frozen internal representations. In this work, we demonstrate that hallucination detection based on internal state representation can also serve as a direct training signal. We introduce RAGognize, a dataset of naturally occurring closed-domain hallucinations with token-level annotations, and RAGognizer, a hallucination-aware fine-tuning approach that integrates a lightweight detection head into an LLM, allowing for the joint optimization of language modeling and hallucination detection. This joint objective forces the model to improve the separability of its internal states regarding hallucinations while simultaneously learning to generate well-formed and meaningful responses. Across multiple benchmarks, RAGognizer achieves state-of-the-art token-level hallucination detection while substantially reducing hallucination rates during generation, without degrading language quality or relevance.

Fabian Ridder, Malte Schilling

Detecting hallucinations in large language models (LLMs) is critical for enhancing their reliability and trustworthiness. Most research focuses on hallucinations as deviations from information seen during training. However, the opaque nature of an LLM's parametric knowledge complicates the understanding of why generated texts appear ungrounded: The LLM might not have picked up the necessary knowledge from large and often inaccessible datasets, or the information might have been changed or contradicted during further training. Our focus is on hallucinations involving information not used in training, which we determine by using recency to ensure the information emerged after a cut-off date. This study investigates these hallucinations by detecting them at sentence level using different internal states of various LLMs. We present HalluRAG, a dataset designed to train classifiers on these hallucinations. Depending on the model and quantization, MLPs trained on HalluRAG detect hallucinations with test accuracies ranging up to 75 %, with Mistral-7B-Instruct-v0.1 achieving the highest test accuracies. Our results show that IAVs detect hallucinations as effectively as CEVs and reveal that answerable and unanswerable prompts are encoded differently as separate classifiers for these categories improved accuracy. However, HalluRAG showed some limited generalizability, advocating for more diversity in datasets on hallucinations.

Victoria Hankemeier, Malte Schilling

Developments in Deep Learning have significantly improved time series forecasting by enabling more accurate modeling of complex temporal dependencies inherent in sequential data. The effectiveness of such models is often demonstrated on limited sets of specific real-world data. Although this allows for comparative analysis, it still does not demonstrate how specific data characteristics align with the architectural strengths of individual models. Our research aims at uncovering clear connections between time series characteristics and particular models. We introduce a novel dataset generated using Gaussian Processes, specifically designed to display distinct, known characteristics for targeted evaluations of model adaptability to them. Furthermore, we present TimeFlex, a new model that incorporates a modular architecture tailored to handle diverse temporal dynamics, including trends and periodic patterns. This model is compared to current state-of-the-art models, offering a deeper understanding of how models perform under varied time series conditions.

Christian Eichenberger, Moritz Neun, Henry Martin et al.

The IARAI Traffic4cast competitions at NeurIPS 2019 and 2020 showed that neural networks can successfully predict future traffic conditions 1 hour into the future on simply aggregated GPS probe data in time and space bins. We thus reinterpreted the challenge of forecasting traffic conditions as a movie completion task. U-Nets proved to be the winning architecture, demonstrating an ability to extract relevant features in this complex real-world geo-spatial process. Building on the previous competitions, Traffic4cast 2021 now focuses on the question of model robustness and generalizability across time and space. Moving from one city to an entirely different city, or moving from pre-COVID times to times after COVID hit the world thus introduces a clear domain shift. We thus, for the first time, release data featuring such domain shifts. The competition now covers ten cities over 2 years, providing data compiled from over 10^12 GPS probe data. Winning solutions captured traffic dynamics sufficiently well to even cope with these complex domain shifts. Surprisingly, this seemed to require only the previous 1h traffic dynamic history and static road graph as input.

Luca Hermes, Barbara Hammer, Andrew Melnik et al.

Accurate traffic prediction is a key ingredient to enable traffic management like rerouting cars to reduce road congestion or regulating traffic via dynamic speed limits to maintain a steady flow. A way to represent traffic data is in the form of temporally changing heatmaps visualizing attributes of traffic, such as speed and volume. In recent works, U-Net models have shown SOTA performance on traffic forecasting from heatmaps. We propose to combine the U-Net architecture with graph layers which improves spatial generalization to unseen road networks compared to a Vanilla U-Net. In particular, we specialize existing graph operations to be sensitive to geographical topology and generalize pooling and upsampling operations to be applicable to graphs.

Luca Hermes, Barbara Hammer, Malte Schilling

Prediction of movements is essential for successful cooperation with intelligent systems. We propose a model that integrates organized spatial information as given through the moving body's skeletal structure. This inherent structure is exploited in our model through application of Graph Convolutions and we demonstrate how this allows leveraging the structured spatial information into competitive predictions that are based on a lightweight model that requires a comparatively small number of parameters.

André Artelt, Valerie Vaquet, Riza Velioglu et al.

Transparency is a fundamental requirement for decision making systems when these should be deployed in the real world. It is usually achieved by providing explanations of the system's behavior. A prominent and intuitive type of explanations are counterfactual explanations. Counterfactual explanations explain a behavior to the user by proposing actions -- as changes to the input -- that would cause a different (specified) behavior of the system. However, such explanation methods can be unstable with respect to small changes to the input -- i.e. even a small change in the input can lead to huge or arbitrary changes in the output and of the explanation. This could be problematic for counterfactual explanations, as two similar individuals might get very different explanations. Even worse, if the recommended actions differ considerably in their complexity, one would consider such unstable (counterfactual) explanations as individually unfair. In this work, we formally and empirically study the robustness of counterfactual explanations in general, as well as under different models and different kinds of perturbations. Furthermore, we propose that plausible counterfactual explanations can be used instead of closest counterfactual explanations to improve the robustness and consequently the individual fairness of counterfactual explanations.

Malte Schilling, Kai Konen, Frank W. Ohl et al.

Locomotion is a prime example for adaptive behavior in animals and biological control principles have inspired control architectures for legged robots. While machine learning has been successfully applied to many tasks in recent years, Deep Reinforcement Learning approaches still appear to struggle when applied to real world robots in continuous control tasks and in particular do not appear as robust solutions that can handle uncertainties well. Therefore, there is a new interest in incorporating biological principles into such learning architectures. While inducing a hierarchical organization as found in motor control has shown already some success, we here propose a decentralized organization as found in insect motor control for coordination of different legs. A decentralized and distributed architecture is introduced on a simulated hexapod robot and the details of the controller are learned through Deep Reinforcement Learning. We first show that such a concurrent local structure is able to learn better walking behavior. Secondly, that the simpler organization is learned faster compared to holistic approaches.

Timo Korthals, Malte Schilling, Jürgen Leitner

This contribution comprises the interplay between a multi-modal variational autoencoder and an environment to a perceived environment, on which an agent can act. Furthermore, we conclude our work with a comparison to curiosity-driven learning.

Malte Schilling, Helge Ritter, Frank W. Ohl

Recent developments in machine-learning algorithms have led to impressive performance increases in many traditional application scenarios of artificial intelligence research. In the area of deep reinforcement learning, deep learning functional architectures are combined with incremental learning schemes for sequential tasks that include interaction-based, but often delayed feedback. Despite their impressive successes, modern machine-learning approaches, including deep reinforcement learning, still perform weakly when compared to flexibly adaptive biological systems in certain naturally occurring scenarios. Such scenarios include transfers to environments different than the ones in which the training took place or environments that dynamically change, both of which are often mastered by biological systems through a capability that we here term "fluid adaptivity" to contrast it from the much slower adaptivity ("crystallized adaptivity") of the prior learning from which the behavior emerged. In this article, we derive and discuss research strategies, based on analyzes of fluid adaptivity in biological systems and its neuronal modeling, that might aid in equipping future artificially intelligent systems with capabilities of fluid adaptivity more similar to those seen in some biologically intelligent systems. A key component of this research strategy is the dynamization of the problem space itself and the implementation of this dynamization by suitably designed flexibly interacting modules.

Malte Schilling

An internal model of the own body can be assumed a fundamental and evolutionary-early representation as it is present throughout the animal kingdom. Such functional models are, on the one hand, required in motor control, for example solving the inverse kinematic or dynamic task in goal-directed movements or a forward task in ballistic movements. On the other hand, such models are recruited in cognitive tasks as are planning ahead or observation of actions of a conspecific. Here, we present a functional internal body model that is based on the Mean of Multiple Computations principle. For the first time such a model is completely realized in a recurrent neural network as necessary normalization steps are integrated into the neural model itself. Secondly, a dynamic extension is applied to the model. It is shown how the neural network solves a series of inverse tasks. Furthermore, emerging representation in transformational layers are analyzed that show a form of prototypical population-coding as found in place or direction cells.

Andrew Melnik, Sascha Fleer, Malte Schilling et al.

Complex environments and tasks pose a difficult problem for holistic end-to-end learning approaches. Decomposition of an environment into interacting controllable and non-controllable objects allows supervised learning for non-controllable objects and universal value function approximator learning for controllable objects. Such decomposition should lead to a shorter learning time and better generalisation capability. Here, we consider arcade-game environments as sets of interacting objects (controllable, non-controllable) and propose a set of functional modules that are specialized on mastering different types of interactions in a broad range of environments. The modules utilize regression, supervised learning, and reinforcement learning algorithms. Results of this case study in different Atari games suggest that human-level performance can be achieved by a learning agent within a human amount of game experience (10-15 minutes game time) when a proper decomposition of an environment or a task is provided. However, automatization of such decomposition remains a challenging problem. This case study shows how a model of a causal structure underlying an environment or a task can benefit learning time and generalization capability of the agent, and argues in favor of exploiting modular structure in contrast to using pure end-to-end learning approaches.

Łukasz Kidziński, Sharada Prasanna Mohanty, Carmichael Ong et al.

In the NIPS 2017 Learning to Run challenge, participants were tasked with building a controller for a musculoskeletal model to make it run as fast as possible through an obstacle course. Top participants were invited to describe their algorithms. In this work, we present eight solutions that used deep reinforcement learning approaches, based on algorithms such as Deep Deterministic Policy Gradient, Proximal Policy Optimization, and Trust Region Policy Optimization. Many solutions use similar relaxations and heuristics, such as reward shaping, frame skipping, discretization of the action space, symmetry, and policy blending. However, each of the eight teams implemented different modifications of the known algorithms.