This paper introduces a novel benchmark dataset designed to evaluate the capabilities of Vision Language Models (VLMs) on tasks that combine visual reasoning with subject-specific background knowledge in the German language. In contrast to widely used English-language benchmarks that often rely on artificially difficult or decontextualized problems, this dataset draws from real middle school curricula across nine domains including mathematics, history, biology, and religion. The benchmark includes over 2,000 open-ended questions grounded in 486 images, ensuring that models must integrate visual interpretation with factual reasoning rather than rely on superficial textual cues. We evaluate thirteen state-of-the-art open-weight VLMs across multiple dimensions, including domain-specific accuracy and performance on adversarial crafted questions. Our findings reveal that even the strongest models achieve less than 45% overall accuracy, with particularly poor performance in music, mathematics, and adversarial settings. Furthermore, the results indicate significant discrepancies between success on popular benchmarks and real-world multimodal understanding. We conclude that middle school-level tasks offer a meaningful and underutilized avenue for stress-testing VLMs, especially in non-English contexts. The dataset and evaluation protocol serve as a rigorous testbed to better understand and improve the visual and linguistic reasoning capabilities of future AI systems.

In this paper, we introduce a novel approach to reliably verifying DRAM addressing functions and function components from software. We perform the first systematic analysis of 5 DRAM function reverse-engineering tools on 2 different DDR3, 4 DDR4, and 4 DDR5 system configurations, revealing a significant variance in the success rate of these tools, from 0% to 92.9%. We discover the previously unknown rank selection side channel and reverse engineer its function on two DDR4 and two DDR5 systems. These results enable novel DDR5 row-conflict side-channel attacks, which we demonstrate in two scenarios: First, we evaluate the DDR5 row-conflict side channel in a covert channel with 1.39 Mbit/s. Second, we evaluate the channel in a website fingerprinting attack with an F1 score of 84% on DDR4 and 74% on DDR5.

This paper systematically analyzes 32 offensive Rowhammer papers, including 48 experiments. However, we avoid finger-pointing but identify 6 threats to the validity and relevance of Rowhammer research results and give multiple examples. The threats include small sample sizes, over-estimated attacker capabilities, unrealistic attack scenarios, non-comparability of the results, age and wear of hardware, and sub-optimal attack performance metrics. Additionally, we provide recommendations with detailed justification to the scientific community to mitigate those threats: (1) pre-experimental testing of DIMM integrity, (2) increasing and broadening the DIMM sample size, (3) expanding reproduction studies of published work, (4) defining attacks in real-world conditions and distinguishing them from theoretical ones, (5) publishing DIMM manufacturing data, (6) documenting DIMM wear and, (7) leveraging multiple metrics for bit flip evaluations.

Industrie 4.0 hat mit der Verschmelzung von physischen Systemen und digitalen Lösungen wie Künstliche Intelligenz (KI), Internet of Things (IoT) und Cloud-Computing einen tiefgreifenden Wandel in der Fertigung ausgelöst [1]. Im Zentrum dieser Entwicklung steht der Digitale Zwilling (DZ) [2] – eine virtuelle Abbildung physischer Objekte und Prozesse, welche Echtzeitüberwachung, Simulation und Optimierung ermöglichen. Trotz ihrer Vorteile stehen DZ in verschiedenen Sektoren vor erheblichen Herausforderungen bei der Einführung. Hohe Implementierungskosten sowie technische und organisatorische Probleme hindern die Industrie häufig daran, den vollen Nutzen aus diesen Lösungen zu schöpfen [3].

Ziel dieser Studie ist es, die Lücke zwischen neuen Technologien und deren praktischer Anwendung zu schließen. Sie verfolgt den Ansatz, ein kostengünstiges und leicht umsetzbares DZ-Modell mit TRL von 5 zu entwickeln, allein mithilfe frei verfügbarer Open-Source-Plattformen und gängiger kommerzieller Software.

Digitale Zwillinge verbessern die Echtzeit-Überwachung, die vorausschauende Wartung und die Automatisierung. Trotz des weitverbreiteten Einsatzes in der Fertigung, der Automobilindustrie und im Gesundheitswesen wird die Technologie in der Landwirtschaft nur begrenzt genutzt [4]. Generelle Hindernisse für die Einführung in Deutschland sind strenge Datenschutzgesetze (77 %), ein Mangel an Fachkräften (64 %) und finanzielle Zwänge [5]. In der Landwirtschaft kommen Unsicherheiten bezüglich der Rentabilität, hoher Anschaffungskosten und begrenztem technischen Support hinzu [6].

Trotz dieser Herausforderungen können DZ die Digitalisierung der Landwirtschaft verändern, indem sie virtuelle Darstellungen von Anlagen und Prozessen in Echtzeit liefern und so die Landwirtschaft nachhaltig unterstützen. Angesichts des hohen Anteils der Landwirtschaft an Treibhausgasemissionen und Energieverbrauch [7] können DZ dabei helfen, CO₂-Emissionen, Biodiversität und Bodenqualität systematisch zu erfassen und so die nachhaltige Entwicklung fördern [8].  

Virtual reality has proven to be a valuable addition in the tool belt of teachers. Immersive learning environments are applied in various settings, including, but not limited to the medical and nursing domain. In this study we present “We care in VR”, a simulation for practicing nursing tasks for care at home, a part of nursing that is currently underrepresented in available VR applications. We investigate how realistic interactions are perceived by end users compared to consistent usage of buttons on the controllers and how they affect the ease of use of the simulation. We conduct an empirical study with 50 participants from three vocational schools of nursing and a university of applied sciences. Results suggest that our simulation already works quite well and is accepted by the target group, but still needs improvement regarding ease of use, especially for users without any previous experience with VR applications.

Digital media have become an integral part of everyday life, education and work. However, intensive and problematic media use, and in particular problematic smartphone use has been shown to reliably predict reduced well-being and increased stress. Therefore, it is essential to investigate the factors that lead to problematic smartphone use and respective negative well-being outcomes and to develop interventions that effectively address these factors. Considering self-control and fear of missing out (FoMO) as key psychological factors promoting problematic smartphone use, we present a conceptual model explaining the emergence of digital stress due to problematic smartphone use, upon which we built a blended training intervention against digital stress. A controlled trial evaluation of the training intervention provided data at multiple time points for multilevel regression analyses on a sample of 175 university students. The results indicated that the intervention was effective in reducing FoMO (fear of missing out) and increasing self-control. Indirect effects suggested that both the reduction in FoMO and the gain in self-control effectively reduced emotional irritation and perceived stress via a reduction in problematic smartphone use. Conclusively, the findings identify key psychological factors that promote problematic smartphone use and demonstrate that these factors can be effectively addressed through appropriate psychological interventions.

This study examines how Large Language Models (LLMs) can reduce biases in text-to-image generation systems by modifying user prompts. We define bias as a model's unfair deviation from population statistics given neutral prompts. Our experiments with Stable Diffusion XL, 3.5 and Flux demonstrate that LLM-modified prompts significantly increase image diversity and reduce bias without the need to change the image generators themselves. While occasionally producing results that diverge from original user intent for elaborate prompts, this approach generally provides more varied interpretations of underspecified requests rather than superficial variations. The method works particularly well for less advanced image generators, though limitations persist for certain contexts like disability representation. All prompts and generated images are available at https://iisys-hof.github.io/llm-prompt-img-gen/

This paper considers three fundamental approaches to software development, namely manual coding, modeldriven software engineering, and code generation by large language models. All of these approaches have their individual pros and cons, motivating the desire for an integrated approach. We present MoProCo, a technical solution to integrate the three approaches into a single tool chain, allowing the developer to split a software engineering task into modeling, prompting or coding sub-tasks. From a single input file consisting of static model structure, natural language prompts and/or source code fragments, Java source code is generated using a two-stage approach. A case study demonstrates that the MoProCo approach combines the desirable properties of the three development approaches by offering the appropriate level of abstraction, determinism, and dynamism for each specific software engineering sub-task.

Similar to LLMs, the development of vision language models is mainly driven by English datasets and models trained in English and Chinese language, whereas support for other languages, even those considered high-resource languages such as German, remains significantly weaker. In this work we present an analysis of open-weight VLMs on factual knowledge in the German and English language. We disentangle the image-related aspects from the textual ones by analyzing accuracy with jury-as-a-judge in both prompt languages and images from German and international contexts. We found that for celebrities and sights, VLMs struggle because they are lacking visual cognition of German image contents. For animals and plants, the tested models can often correctly identify the image contents according to the scientific name or English common name but fail in German language. Cars and supermarket products were identified equally well in English and German images across both prompt languages.