ABSTRACT. Embedding methods are among the most efficient approaches for learning to reason about relational knowledge. In this talk, I will focus on the framework of region-based embeddings, where relations are encoded as geometric regions. The spatial arrangement of these regions allows such models to capture symbolic rules, which enables them to faithfully simulate some forms of relational reasoning. This region-based encoding allows for a tight coupling between learning and reasoning and naturally supports various forms of commonsense reasoning. An important practical consideration, however, is that the parameterisation of the regions needs to remain as simple as possible, to keep the framework efficient and help prevent overfitting.

Most approaches use convex regions which are defined in terms of coordinate-wise comparisons. This makes them highly efficient, but the resulting models have inherent theoretical limitations, meaning that they are incapable of capturing certain types of rule bases. One possibility to escape these limitations is to allow for cross-coordinate comparisons, while keeping the model otherwise as simple as possible, which is the approach that is taken by the RESHUFFLE model. Another possibility is to keep the restriction to coordinate-wise comparisons but allow non-convex regions.

Despite recent progress, many important challenges remain. Effectively learning region-based models from small amounts of data is highly non-trivial, as these models lack the strong inductive biases of simpler embedding-based models. Furthermore, it remains unclear to what extent region-based models can go beyond classical reasoning (e.g. non-monotonic or probabilistic reasoning). Finally, neural reasoning models, for instance those based on Graph Neural Networks, also implicitly rely on geometric encodings of relational knowledge. There is some evidence that such models can be improved by incorporating more explicit region-based representations, but this complicates gradient-based learning. As a result, this strategy has thus far only been applied with extremely simple types of regions.

ABSTRACT. Vision classifiers can exploit spurious correlations, achieving high in-distribution accuracy yet failing under distribution shift. Existing approaches to bias mitigation and analysis often depend on curated datasets, spurious-attribute or group labels, or retraining, which may be infeasible once a model is deployed or the relevant bias is unknown. We present a bias-label-free, post-hoc method for identifying spurious concepts in frozen vision models, relying only on standard class labels from a held-out audit dataset. For each target class, we collect patches from inputs predicted as that class and apply non-negative matrix factorization to intermediate activations to obtain a bank of interpretable concept vectors. Candidate concepts are then ranked with a bias estimator derived from their interaction with backpropagated gradients on misclassified examples: bias concepts tend to get activated when correcting false negatives and suppressed when correcting false positives. On Colored MNIST and Waterbirds the method recovers concepts aligned with the known spurious cue, and on CelebA it surfaces decision-relevant directions that only partially coincide with the annotated gender attribute; suppressing the top-ranked concepts at inference time improves worst-group accuracy by up to 17.9 percentage points on Waterbirds and 10.4 on CelebA without any retraining or parameter updates. Our method identifies decision-relevant spurious directions that need not coincide with annotated ones, providing both an interpretable auditing tool and an actionable debiasing handle for frozen vision models.

ABSTRACT. Conditional belief bases with ranking-function semantics provide a powerful framework for non-monotonic reasoning and belief revision. c-Representations and c-revisions based on conditional preservation satisfy strong rationality requirements, but generally admit multiple admissible models or revision operators. In this paper, we introduce core c-revisions, a subclass of c-revisions obtained by simplifying the underlying constraint systems. We show that core c-revisions inherit the desirable properties of general c-revisions, including compliance with the DP-postulates, while being stratified and computable via revision-aware tolerance partitions. This yields the existence of a minimal core c-revision, which allows us to define the first fully specified strategic c-revision operator selecting minimal revision outcomes.

ABSTRACT. This paper provides convergence and divergence results for weighted argumentation graphs that are bipolar, i.e. contain both attacks and supports for arguments. We study a notion of a \emph{modular acceptability semantics}, which cleanly separates \emph{aggregation} of attacking and supporting arguments from the computation of their \emph{influence} on an argument's initial weight. Based on this modular framework, we study characteristics of acceptability semantics and prove general convergence and divergence theorems.

ABSTRACT. This work establishes that trigger-word data poisoning of VLAs is practical, while at the same time the open-source robotics ecosystem holds trust assumptions about community contributions. A few poisoned samples can silently embed a backdoor that disables a robot on command. We evaluate this threat against smolVLA on a real-world pick-and-place task, training on three poison ratios and evaluating across different prompts on the LeRobot platform. Three poisoned episodes in 320 clean episodes suffice for a complete denial of service. Success rate drops to 0.0±0.0% across all trigger-word conditions and the robot locks into a fixed joint configuration rather than executing any task-relevant motion. Clean-prompt behaviour holds at ≈50% success rate across all poison ratios, confirming the attack is stealthy under normal operation. A single poisoned episode already reduces success rate to 6.7 ± 6.7%. The robot still moves, but no longer completes the task. The attack generalises to front, middle, and end trigger placements despite training exclusively on front-placed triggers. These findings establish that the threat is practical, low-cost, and stealthy, and warrant treating dataset provenance as a first-class concern in open-source robotics ecosystems.

ABSTRACT. Industrial acoustic monitoring in real manufacturing environments is challenging due to strong background noise, heterogeneous operating conditions, and limited labeled fault data. In such settings, the normal operating conditions often vary acoustically. Background sounds may resemble malfunction events. All of these factors may lead to unstable decision boundaries and increased false alarm rates.To address this challenge, we formulate industrial audio anomaly detection as a multi-state problem rather than a binary normal vs. anomalous task. We propose to adapt the PatchCore retrieval framework for industrial audio, using log-Mel spectrogram representations and class-specific mem- ory banks to model different operating states. The approach is evaluated under multiple temporal segmentation strategies using a pretrained CNN14 encoder on AudioSet, and compared with a supervised CNN baseline. Our experiments show that fault state detection performance depends strongly on the alignment between temporal segmentation and the encoder receptive field, rather than on the temporal context length alone. When this alignment is preserved, the retrieval-based approach produces well-separated embeddings of operating states and achieves strong ranking performance while maintaining competitive classification accuracy.

ABSTRACT. Magnetic resonance imaging-proton density fat fraction (MRI-PDFF) measures the concentration of fat within tissue and is an established quantitative biomarker for metabolic and musculoskeletal disorders. The acquisition requires, however, dedicated quantitative MRI sequences that are not routinely available. We investigate whether PDFF can be reliably estimated from widely available T2-weighted MRI using image-to-image translation. We further assess the necessity of complex generative models for this task and compare a lightweight 4-level U-Net to a Denoising Diffusion Probabilistic Model (DDPM). Experiments are conducted on a large-scale dataset from the German National Cohort (NAKO) comprising 207,138 paired images.

We evaluate clinical relevance beyond pixel-wise metrics via downstream PDFF quantification in four muscle compartments based on segmentations produced by VIBESegmentator. Both approaches significantly outperform the identity baseline, which achieves a Pearson correlation of r = 0.764 and a mean absolute error of MAE = 0.070 ± 0.055, confirming that the models learn a non-trivial cross-modal mapping. Interestingly, the lightweight U-Net outperforms the DDPM in both correlation (r = 0.974 vs. 0.961) and error (MAE = 0.014 ± 0.015 vs. 0.017 ± 0.018), while reducing inference time by a factor of 208 (25.2 ms vs. 5,227.2 ms per image).

Our results indicate that for deterministic, paired medical image translation tasks, lightweight architectures can match or even surpass the performance of state-of-the-art diffusion models, while drastically reducing computational cost, enabling real-time clinical deployment.

ABSTRACT. Within the last years the usage of generative artificial intelligence tools rose significantly. These tools are now able to fabricate convincing human-like language and texts. Hence they are widely used which becomes problematic in areas like academia, where integrity matters most. We can already see a flood of machine-generated papers as well as peer reviews, which endangers academic processes of becoming meaningless. Also we face a significant amount of machine-generated student submissions in our courses. Hence ways must be found to reliably determine authorship.

We developed a system that addresses this issue. It involves four individual classifiers as well as gradient boosting. Its classification pipeline processes texts sentence-wise and applies two tests afterwards. A consecutive-sentences text looks for at least three flagged sentences in a row, allowing the detection of generated paragraphs. Here, a custom cutoff can be set to make the test search for generated texts more greedy or permissive. Furthermore we apply a gauss test that compares the amount of flagged sentences in a document with an expectation derived from 15,000 arXiv papers published between January 2010 and December 2014. If the proportion of flagged sentences in a document is significantly higher than the expectation, the assumption that the text is human-written is rejected and it is assumed the text is machine-generated. As the expected distribution that is being tested against already includes the effects of potential false positives, they are virtually eliminated from the final result.

Overall this system shows to be highly capable of identifying machine-generated texts with a false-positive rate of less than 1%. We further validated our system with an analysis of 30,000 arXiv papers published between 2015 and 2025. Here we can clearly see a significant shift towards machine-generated texts with the release of ChatGPT in November 2022. Furthermore we can see significant rises in machine-generated texts with each new generation of generative artificial intelligence. Our contribution documents the system we developed as well as the conducted validation study in detail.

ABSTRACT. This paper introduces a robust pipeline enabling Large Language Models (LLMs) to perform precise Table Question Answering (Table-QA) on complex statistical tables. Processing complex, visually oriented table formats remains a major challenge for large language models, as they exhibit inherent multi-dimensional and hierarchical structures. Despite their visual complexity, these formats adhere to strict structural conventions. Our approach leverages these regularities—such as specific cell border styles and hierarchical indentations—to transform these files into a unified JSON structure using a deterministic extraction method. Instead of generating direct answers, the LLM is fine-tuned to produce Python scripts that programmatically extract information from the generated JSON files. To achieve this, we generated a synthetic dataset comprising approximately 13,000 high-quality question-code pairs. Our fine- tuned Microsoft Phi-4 model achieves a functional success score of 88.5% (peaking at 95.2% on XLSX files), on par with the most advanced proprietary models, such as GPT-4.1. Furthermore, we demonstrate that a smaller 8B parameter model (Llama-3.1) is a highly efficient alternative, achieving a functional success score of 78.6% in resource-constrained environments.