ABSTRACT Evaluations of AI systems for alloy design are built on commercial data sources in which metallurgically identical grades circulate under different vendor trade names, a leakage that name‐based deduplication cannot detect. We document this hazard and propose compositional‐distance stratification as the corrective, demonstrated on AlloyGraph, an open‐source nickel‐based superalloy design‐support system coupling an RDF knowledge graph, whose OWL 2 DL ontology serves as schema and ingestion‐time data validation, with physics‐informed machine learning and a multiagent LLM layer, with processing route and wrought processability constraints enforced deterministically inside the design pipeline, before any design is evaluated or reported. A compositional audit found 13 of 88 nominally unseen evaluation alloys to be near‐duplicates of training alloys under vendor branding, so all results are stratified by compositional novelty. The stratification attributes accuracy to distinct mechanisms: knowledge‐graph anchoring acts as exact retrieval, reducing yield strength error by 30% on near‐duplicate alloys through a few sparse anchoring events on five alloys, and essentially nothing beyond them; the agent layer, arbitrating between machine‐learning and physics estimates, improves yield strength by 18% on compositionally new alloys, stable across five seeded campaigns, a gain concentrated in solid‐solution‐strengthened alloys and absent for ‐strengthened alloys beyond the near‐duplicates; and deterministic physics corrections carry elastic modulus. Negative results are reported with equal prominence: the agents degrade elastic modulus, and in a surrogate‐scored in silico evaluation the LLM Designer's contribution to inverse design is not demonstrated against an informed sampling baseline. The deployed platform uses open‐access data and open‐weights models throughout, with per‐prediction conformal uncertainty intervals and automated provenance verification; commercial models appear only as external baselines and as an evaluation judge.
ABSTRACT Biomedical image segmentation is a critical task in medical diagnosis and treatment planning, enabling precise delineation of anatomical structures and pathological regions. Despite significant advancements, challenges persist due to the inherent variability, noise and complex morphology present in diverse medical imaging modalities. This paper introduces ConvNeXt‐FD, a novel deep learning architecture for robust biomedical image segmentation, built upon a U‐Net‐like encoder‐decoder framework leveraging the powerful ConvNeXt backbone. Our approach integrates a hybrid loss function combining the Dice coefficient with a boundary‐aware regularization term inspired by a differentiable formulation of fractal dimension, designed to enhance the model's sensitivity to object boundaries and shape fidelity. We rigorously evaluate ConvNeXt‐FD across six distinct biomedical datasets: BUSI (breast ultrasound images), DDTI (thyroid ultrasound images), FluoCells (fluorescent cell images), IDRiD (diabetic retinopathy images for optic disc segmentation), ISIC2018 (skin lesion images) and MoNuSeg (nuclei segmentation). Experimental results demonstrate that ConvNeXt‐FD, particularly when initialized with ImageNet pre‐trained weights, achieves competitive and often superior performance compared to existing state‐of‐the‐art methods across various metrics, including Dice, Jaccard, Accuracy, Sensitivity, Specificity and False Positive Rate. The integration of ConvNeXt as a strong encoder, coupled with the boundary‐aware regularization, proves effective in capturing both high‐level semantic features and fine‐grained boundary details, leading to more accurate and reliable segmentations in challenging biomedical contexts.
ABSTRACT Automating the extraction of domain models from natural language requirements remains a central challenge in model‐driven engineering (MDE), particularly at the level of the computation‐independent model (CIM), the most abstract layer of the model‐driven architecture (MDA). This paper introduces an approach that uses large language models (LLMs) to generate domain models from software requirements. Unlike prior work that relies on a single monolithic prompt or on generic task decomposition, our approach organises the modelling task as a modular chain of prompts with dedicated stages for enumerations, entities, and relationships, each producing an explicit intermediate output that feeds the next stage before a final consolidation step. The approach is implemented as a Visual Studio Code extension called SmartModel, which integrates generation, visualisation, and requirement‐level traceability into the developer's environment. To evaluate the approach, a case study on smart tourism software requirements was conducted, comparing the LLM‐generated domain model with an expert‐constructed reference model using both qualitative expert assessments and quantitative metrics. Results show that LLMs achieve high precision in identifying entities and enumerations, whereas relationship and attribute extraction remain considerably less reliable and require systematic human review. The findings underscore the value of LLMs in accelerating early‐stage MDE tasks, while highlighting the necessity of human feedback to ensure semantic integrity. This work contributes a practical and extensible tool and empirical insights that inform future research into hybrid AI‐human modelling workflows and advanced LLM techniques for software modelling.
ABSTRACT Consumer AR smart glasses are evolving from passive displays into proactive, intelligent assistants. Currently, these ‘Agentic AR’ systems rely heavily on cloud‐based AI. They continuously upload First‐Person View (FPV) video streams and multimodal sensor data. However, this massive data offloading creates severe problems. It introduces critical privacy vulnerabilities, unacceptable network latency and major security risks. To address these challenges, this paper reviews state‐of‐the‐art secure and lightweight Edge‐AI technologies for wearable AIoT devices. We propose a ‘Data‐Never‐Leaves‐Device’ architecture. This fully localized approach entirely eliminates cloud dependency. Our system synergizes three key components: optimized native inference frameworks (e.g., NCNN), ultra‐lightweight perception networks (e.g., YOLO26) and compressed on‐device Large Language Models (LLMs). Together, they execute complex cognitive reasoning directly on the edge. Furthermore, we systematically evaluate the necessary architectural trade‐offs. We analyse the balance between model compression, autonomous usability and strict resource constraints like computation and energy. Ultimately, this paper concludes that migrating AI inference to a secure edge is not just a performance upgrade. It is a fundamental requirement for ensuring user privacy, robust security and the sustainable future of spatial computing.
ABSTRACT The integration of legacy industrial systems with modern intelligent technologies poses significant challenges for preserving and reusing engineering knowledge embedded in unstructured technical documentation. This limitation affects maintenance, diagnostics and decision‐making processes in industrial environments. This study investigates the use of Retrieval‐Augmented Generation (RAG) architectures for knowledge continuity in legacy industrial automation, using PROFIBUS documentation as a representative case of legacy industrial communication systems. Three configurations are comparatively evaluated: (i) a standalone large language model (LLM), (ii) a Single‐Agent RAG architecture and (iii) a Multi‐Agent RAG architecture with explicit supervisor‐worker orchestration. The evaluation combines quantitative lexical and semantic metrics (ROUGE, METEOR, MATTR and BERTScore) with qualitative assessment based on the LLM‐as‐a‐Judge paradigm, complemented by human expert judgement. Results indicate that the proposed Multi‐Agent RAG architecture with supervisor‐worker, which combines domain‐specialized retrieval with explicit multi‐agent orchestration, achieves higher contextual precision, factual accuracy and faithfulness than the evaluated baselines. These gains, however, come at the cost of increased computational complexity, highlighting a trade‐off between response quality and efficiency. Overall, the findings demonstrate that the proposed architecture provides a robust approach for reliable knowledge synthesis from legacy technical documentation, supporting maintenance activities and decision‐making in industrial environments.
