119092 research outputs found
Sort by
Aletheia: What Makes RLVR For Code Verifiers Tick?
No full textMulti-domain thinking verifiers trained via Reinforcement Learning from Verifiable Rewards (RLVR) are a prominent fixture of the Large Language Model (LLM) post-training pipeline, owing to their ability to robustly rate and rerank model outputs. However, the adoption of such verifiers towards code generation has been comparatively sparse, with execution feedback constituting the dominant signal. Nonetheless, code verifiers remain valuable toward judging model outputs in scenarios where execution feedback is hard to obtain and are a potentially powerful addition to the code generation post-training toolbox. To this end, we create and open-source Aletheia, a controlled testbed that enables execution-grounded evaluation of code verifiers' robustness across disparate policy models and covariate shifts. We examine components of the RLVR-based verifier training recipe widely credited for its success: (1) intermediate thinking traces, (2) learning from negative samples, and (3) on-policy training. While experiments show the optimality of RLVR, we uncover important opportunities to simplify the recipe. Particularly, despite code verification exhibiting positive training- and inference-time scaling, on-policy learning stands out as the key component at small verifier sizes, and thinking-based training emerges as the most important component at larger scales
Kommunale Spitzenverbände in den Ausschüssen des Deutschen Bundestages: Eine Studie zur Messung ihres Lobbyingerfolgs
No full textDie kommunalen Spitzenverbände (KSV) sind das wichtigste Sprachrohr für die Interessenvertretung der Städte, Gemeinden und Landkreise in der Bundespolitik. Als Vereinigungen kommunaler Gebietskörperschaften nehmen sie einen Sonderstatus unter den Interessengruppen ein, weshalb ihnen spezielle Beteiligungsrechte im Gesetzgebungsverfahren des Bundes eingeräumt werden. Obwohl die KSV an der Ausgestaltung vieler Gesetze beteiligt sind, ist bislang ungeklärt, wie erfolgreich sie ihre Interessen tatsächlich durchsetzen können. Um diese Forschungslücke zu füllen, wird im vorliegenden Artikel ein Messkonzept für den Erfolg des Lobbyings der KSV in den Ausschüssen des 19. Deutschen Bundestages (2017–2021) entwickelt und angewendet. Im Rahmen einer qualitativen Inhaltsanalyse der Konsultationsdokumente aller 87 im Untersuchungszeitraum verabschiedeten Gesetze mit schriftlicher KSV-Beteiligung in öffentlichen Anhörungen werden insgesamt 555 Änderungswünsche identifiziert und mit dem Policy-Output vor und nach den parlamentarischen Beratungen verglichen. Die Befunde zeigen, dass es den KSV bei zwei Dritteln der abgegebenen Stellungnahmen mit konkreten Änderungswünschen gelungen ist, diese zumindest teilweise durchzusetzen. Generell gilt: Die KSV lobbyieren erfolgreicher, wenn sie sich auf eine gemeinsame Position verständigen und ihre Interessen in einer breiten Koalition aller Kommunen vertreten
Direct observation of nanoscale pinning centers in Ce(Co0.8Cu0.2)5.4 permanent magnets
No full textPermanent magnets containing rare earth elements are essential components for the electrification of society. Ce(Co1-xCux)5 permanent magnets are a model system known for their substantial coercivity, yet the underlying mechanism remains unclear. Here, we investigate Ce(Co0.8Cu0.2)5.4 magnets with a coercivity of ∼1 T. Using transmission electron microscopy (TEM) and atom probe tomography (APT), we identify a nanoscale cellular structure formed by spinodal decomposition. Cu-poor cylindrical cells (∼5–10 nm in diameter, ∼20 nm long) have a disordered CeCo5-type structure and a composition Ce(Co0.9Cu0.1)5.3. Cu-rich cell boundaries are ∼ 5 nm thick and exhibit a modified CeCo5 structure, with Cu ordered on the Co sites and a composition Ce(Co0.7Cu0.3)5.0. Micromagnetic simulations demonstrate that the intrinsic Cu concentration gradients up to 12 at.% Cu/nm lead to a spatial variation in magnetocrystalline anisotropy and domain wall energy, resulting in effective pinning and high coercivity. Compared to Sm2Co17-type magnets, Ce(Co0.8Cu0.2)5.4 displays a finer-scale variation of conventional pinning with lower structural and chemical contrast in its underlying nanostructure. The identification of nanoscale chemical segregation in nearly single-phase Ce(Co0.8Cu0.2)5.4 magnets provides a microstructural basis for the long-standing phenomenon of "giant intrinsic magnetic hardness" in systems such as SmCo5-xMx, highlighting avenues for designing rare-earth-lean permanent magnets via controlled nanoscale segregation
Anisotropic crack-domain wall interactions in ferroelectric single crystals
No full textThe crack-domain wall interaction is investigated with indentations and phase-field simulations. Cracks are introduced near the pre-existing 90◦ domain walls using the Vickers indenter to observe their interactions with
the domain wall during propagation. It was found that the domain wall can hinder, promote, or deflect the crack propagation. Phase-field simulations and configurational force theory are adopted to analyze the underlying
mechanisms of crack-domain wall interactions. The results reveal pronounced anisotropic characteristics dependent on the crack-domain wall interaction angles. The finding enriches the understanding of the influence of crack on the electromechanical response and failure of ferroelectric single crystals
GenSwarm: Scalable Multi-Robot Code-Policy Generation and Deployment via Language Models
No full textThe development of control policies for multi-robot systems traditionally follows a complex and labor-intensive process, often lacking the flexibility to adapt to dynamic tasks. This has motivated research on methods to automatically create control policies. However, these methods require iterative processes of manually crafting and refining objective functions, thereby prolonging the development cycle. This work introduces GenSwarm, an end-to-end system that leverages large language models to automatically generate and deploy control policies for real-world multi-robot systems based on user instructions in natural language. As a multi-language-agent system, GenSwarm achieves zero-shot learning, enabling rapid adaptation to altered or unseen tasks. The white-box nature of the code policies ensures strong reproducibility and interpretability. With its scalable software and hardware architectures, GenSwarm supports efficient and automated policy deployment on both simulated and real-world multi-robot systems, realizing an instruction-to-execution end-to-end functionality that may transform the development paradigm of multi-robot systems in the future
KGBauko Reader Konstruktives Gestalten, Ergebnisse des Wintersemesters 2025/2026
No full textLerninhalt
Der Schwerpunkt liegt auf der konstruktiven und gestalterischen Durcharbeitung zusammenhängender kleiner Projekte unter Zugrundelegung gezielter konstruktiver und wissenschaftlicher Aspekte (z.B. filigrane, leichte Tragwerke, sensible Strukturen, optimierter Materialeinsatz), Vorgänge beim Gestalten, Modell und Pläne, Leichtbau 1, Leichtbau 2, Bauen mit Textilien 1, Bauen mit Textilien 2, Bauen mit Luft, Bauen mit Glas 1, Bauen mit Glas 2, Bauen mit Stahl, Bauen mit Holz, Bauen mit Seilen, Bauen mit Papier, additive Fertigung
Framework selection for simulation-driven, web-based digital twins: methodology and practical implementation on a machine tool
No full textDigital twins, crucial in Industry 4.0 and IoT, provide data collection and simulation benefits. Web-based applications offer added accessibility and device compatibility. The primary challenges involve selecting an appropriate web framework and integrating simulation models while maintaining a user-friendly interface. This study focuses on a method for selecting effective web frameworks for digital twins and outlines a digital twin architecture suitable for web applications. It emphasizes the technical stack and design aspects essential for implementing a functional and efficient web-based digital twin. This approach is exemplified through a practical implementation of a web-based, intelligent digital twin for a CNC machine tool
SPEAK-SAFE: secure processing of electronic audio for knowledge in suicide assessment from therapeutic exchanges
No full textBackground: For therapists, the spoken word of their patients is among the most important foundations for clinical assessment. At the same time, it is hardly possible to monitor patients continuously and closely in sufficient numbers, for example, to ongoingly assess the risk of suicide in therapeutical conversations. Natural Language Processing (NLP) involves the use of Artificial Intelligence (AI) to analyze human language. Combining it with AI speech processing methods, we obtain multimodal methods which can automatically process large volumes of speech and language data to extract diagnostic information and therefore support individualized treatment plans. Thus, in NLP/multimodal methods, we see the opportunity to significantly improve patient care.Methods: The SPEAK-SAFE project, implemented by clinicians and clinical researchers from the University hospital in Frankfurt in collaboration with the AI experts from the TU Darmstadt, aims to create the first German psychiatric corpus for evaluating and developing multimodal and NLP models to optimize diagnostic processes in psychiatric, psychosomatic, and psychotherapeutic care. Therefore, we will collect therapist-patient dialogues during therapy sessions. This sensitive data necessitates robust privacy. To meet this requirement, all collected data is pseudonymized, to ensure that no personal data is part of the evaluation and training of the AI models.Discussion: During the implementation of our research project, we were faced with challenges regarding the security of patient privacy and the technical implementation of therapy recordings toreassure sufficient data quality for the data analysis. Therefore, in addition to improve the suicidality prediction with multimodal methods we will develop an end-to-end-workflow for further AI-research in the clinical context.
Clinical Trial Registration: https://drks.de/search/de/trial/DRKS00027878, identifier DRKS00027878
Restoration legacy, landscape context and elevation shape frugivory in a tropical landscape
No full textAssisted and passive natural restoration are widely applied strategies for forest restoration, yet their focus on tree recovery makes their effectiveness in restoring broader biodiversity unclear. Assuming that tree recruitment will rebuild whole ecosystems overlooks other taxa and their interactions, and thus underestimates key components of biodiversity like biotic interactions. To address this gap, we assessed two complementary questions: (i) how frugivory reestablishes and varies among assisted restoration and natural regeneration areas, and (ii) how local conditions (fruit availability, elevation, and time since restoration) and landscape context (forest cover and fragmentation) influence frugivory beyond restoration strategies. This was done by implementing a dummy fruit experiment in a tropical landscape in southeast Ecuador that considered two fruit sizes. Dummy fruit handling did not differ between assisted and naturally regenerated areas, indicating that neither restoration strategy was superior in promoting frugivory. Instead, landscape context played a central role. Fruit handling increased with old-growth forest cover and elevation, and declined with increasing forest fragmentation, highlighting the importance of habitat amount and connectivity for interaction recovery. Fruit size further shaped frugivory patterns, particularly in restored areas, where larger fruits were handled more frequently. Overall, our results show that restoration outcomes for frugivory depend less on restoration strategy alone and more on landscape structure and environmental context. By demonstrating the utility of artificial fruits as a rapid and practical tool to assess interaction recovery, this study highlights the need to move beyond vegetation-based metrics and explicitly incorporate biotic interactions into the evaluation of forest restoration success
Vom Fenster zur Ressource – Glas im Wandel zur zirkulären Baukomponente
No full textFrom windows to resources – How flat glass becomes a circular building material
Glass, due to its energy‐intensive production, has a high share of so‐called “embodied energy” and influences the operational energy demand of buildings through its building‐physical properties. In Germany, more than 150 million window units with uncoated double‐glazing are still installed, which will need to be replaced in the coming years. The processing, upgrading, and reuse of dismantled glass panes offer significant potential for reducing CO2 emissions. This paper examines, based on current case studies, how such an approach can be technically implemented. Key aspects include dealing with visual quality, determining the strength of dismantled glass panes, and producing new insulating glass units using these existing components. The results show that remanufacturing concepts represent an effective approach to closing material loops in the flat glass sector and conserving resources