To evaluate the accuracy and reliability of four artificial intelligence (AI) models-ChatGPT, Copilot, DeepSeek, and Gemini-in generating PubMed citations for literature related to lens disease, cataracts, iris disorders, and anterior chamber pathology. Comparative accuracy assessment study. Forty standardized clinical paragraphs from The Review of Ophthalmology (4 th edition) were used as test inputs. Each AI model was prompted to generate AMA-11-style PubMed references. Citation accuracy was assessed using predefined criteria, including PubMed verifiability, DOI concordance, and bibliographic accuracy. Two expert reviewers independently classified the citations as fully cited, partially cited, or not cited, and assessed inter-rater reliability. The citation accuracy varied significantly among the models. DeepSeek demonstrated the highest accuracy (52.5%), followed by ChatGPT (32.5%) and Copilot (20.0%), whereas Gemini demonstrated the lowest accuracy (2.5%) ( P < 0.001). DOI mismatches were the most common errors across all models. Expert validation confirmed these findings, with DeepSeek producing the highest number of fully cited references. Inter-rater agreement was substantial (Cohen's κ = 0.65). Domain-specific AI models, particularly DeepSeek, outperform general-purpose models in generating PubMed citations from ophthalmic literature. However, all the evaluated models exhibited citation errors, underscoring the necessity of human verification. AI tools may enhance academic workflows as assistive systems but should not be used autonomously for reference generation in medical research.

Large language models (LLMs) have demonstrated remarkable reasoning capabilities across various natural language tasks. However, comparable breakthroughs in scientific discovery remain limited, as understanding complex chemical and physical phenomena demands multidimensional representations that extend far beyond language alone. Transition metal complexes (TMCs) are a well-recognized paradigm, essential for the development of catalysts and functional materials. The exploration of their vast and intricate design space, characterized by diverse coordination geometries and topological structures, poses significant challenges within language-based representations interpretable by LLMs. To address this limitation, we introduce 3DTMC-LLM, the first multimodal LLM designed specifically for TMCs. 3DTMC-LLM achieves efficient alignment of structural and textual spaces through a pretrained 3D encoder trained on 12 million TMCs, combined with a lightweight single-token projection layer. In downstream tasks, including knowledge/description generation, property prediction, and the more challenging reactivity modeling, 3DTMC-LLM was benchmarked against state-of-the-art closed-source LLMs (e.g., GPT-5.2) as well as domain-specific machine learning models. It achieved competitive or improved performance on several tasks, particularly those with strong three-dimensional dependencies. This framework highlights the potential of multimodal approaches to accelerate research in TMCs and suggests broader opportunities for advancing the development of the general-purpose chemistry model.

Graphene and carbon nanotube (CNT) membranes are promising for filtration, desalination, and water treatment, yet their performance requires the joint interpretation of their architecture, nanoconfined transport, selectivity, fouling, swelling, defects, stability, and operating conditions. Here, GCMembrane-LLM was developed as an evidence-grounded domain-specific large language model. A curated 582-paper corpus generated 12,208 cleaned membrane-specific question-answer pairs for Low-Rank Adaptation (LoRA)-based supervised fine-tuning of Llama-3.1-8B-Instruct, and retrieval-augmented generation provided article-title and page-level traceability. GCMembraneBench included 100 application-oriented questions on graphene oxide (GO) membranes, CNT membranes, GO/CNT hybrids, and cross-material reasoning. Under direct answering without retrieval context, the anonymized and shuffled automatic evaluation showed that GCMembrane-LLM achieved a mean weighted score of 4.237/5.0, exceeding Llama-3.1-8B-Instruct and Doubao-1.5-lite. A stratified 30-question blinded manual assessment showed the same ranking. The application cases further yielded membrane science conclusions: CNT-assisted GO/CNT transport should be evaluated with dispersion, interfacial compatibility, defects, and stability; GO desalination depends on swelling control, interlayer spacing, and defect suppression; and CNT high flux requires joint examination of pore diameter, entrance chemistry, hydration barriers, ion rejection, and operating conditions. GCMembrane-LLM supports source-traceable evidence organization and preliminary hypothesis formulation before experimental validation.

The integration of multi-stain histopathology images through deep learning poses a significant challenge. Current approaches struggle with data heterogeneity and missing data, as concatenating multi-stain features may not effectively model stain-specific and cross-stain interactions. We introduce UNICORN (UNiversal stain Integration network for CORonary classificatioN), a two-stage, end-to-end trainable model comprising transformer self-attention blocks to process multi-stain histopathology for atherosclerosis severity prediction. The initial stage employs domain-specific expert models to extract features from each staining. An aggregation expert model then integrates features by learning their interactions. On a multi-class, multi-stain whole slide images (WSIs) dataset of atherosclerotic lesions from Munich Cardiovascular Studies Biobank (MISSION), UNICORN achieved a classification accuracy of 0.68, significantly outperforming state-of-the-art models. UNICORN identifies relevant tissue phenotypes across stainings and implicitly models disease progression. Its explainability and effectiveness in predicting atherosclerosis progression highlight the potential for broader applications in medical research and decision support.

ABSTRACT Physical education is an important setting for developing students’ social and emotional competence. Yet research in physical education has largely taken a skill focused approach, giving limited attention to motivation and to the domain-specific characteristics of physical education that shape how competence is learned and sustained. Drawing on social and emotional competence scholarship and self-determination theory, we propose a domain-specific model of social and emotional competence in physical education organized around three core mechanisms. The model offers a coherent account of how social and emotional competence is generated, internalized, and sustained through an iterative process linking domain-specific characteristics of PE, need support and satisfaction, and motivational regulation and its proximal and distal outcomes. It also offers conceptual guidance for designing, implementing, and evaluating sustainable social and emotional competence focused practice in physical education.

Abstract This work introduces a data-centric framework for answering analytical questions using network models, transcending domain-specific modeling conventions. The unified network modeling framework (UNMF) provides an interdisciplinary strategy for principled modeling and analysis of complex networks, including multilayer and temporal networks that arise in sustainability-driven applications. UNMF connects dynamic analysis, pattern discovery, and network-grounded integration of heterogeneous sources (structured and unstructured). We present guided instantiations of UNMF in urban development, mobility, ecosystems, and social-network settings to show how explicit modeling choices can be documented, compared, and assessed within a shared evaluative framework. In doing so, we formalize and systematize Networked Data Science as a field at the intersection of network science and data science, and we define its scope and applications. This work contributes to network science by providing an auditable design-and-evaluation procedure for studying complex, evolving systems and for making representation choices more explicit, inspectable, and reusable across domains.

Can one model fit all? Evaluating foundation models for time series forecasting across clinical medicine.

Caesarean section (CS) is a life-saving intervention when medically indicated; however, its use has increased substantially worldwide, including in high-access healthcare settings. Emerging evidence suggests that women's perceptions, prior childbirth experiences, and counselling practices may contribute to persistently elevated CS rates. This study assessed the prevalence of reported history of CS and examined sociodemographic, health-system, and perception-related factors associated with prior CS among women in the United Arab Emirates (UAE). A cross-sectional questionnaire-based survey was conducted among women aged ≥ 18 years residing in the UAE between February and October 2025. Data were collected using a structured self-administered questionnaire assessing sociodemographic characteristics, obstetric history, delivery-related perceptions, and factors influencing decision-making regarding mode of delivery. Analyses examining CS history were restricted to women with at least one prior delivery, and the outcome was defined as reporting at least one previous CS. Descriptive statistics, chi-square tests, and multivariable binary logistic regression analyses were performed. Separate domain-specific regression models and an additional combined multivariable model were constructed to assess potential confounding between sociodemographic and perception-related variables. Among 351 women with previous deliveries, 219 (62.4%) reported a history of at least one CS. The most commonly reported reasons for CS included foetal medical indications (28.3%), healthcare provider recommendation (18.7%), perceived foetal safety (18.7%), and fear of labour pain (16.0%). In the combined multivariable model, age ≥ 31 years (AOR 1.73; 95% CI 1.06-2.81) and undergraduate/postgraduate educational attainment (AOR 2.04; 95% CI 1.13-3.69) were independently associated with prior CS. Perception-related variables associated with prior CS included disagreement that vaginal delivery leads to better maternal recovery (AOR 18.50; 95% CI 2.37-144.51), disagreement that vaginal birth promotes better mother-baby bonding (AOR 2.37; 95% CI 1.28-4.38), and the belief that CS impacts infant immune outcomes (AOR 2.57; 95% CI 1.35-4.90). Women identified improved counselling from healthcare providers as the most important factor perceived to support better delivery-related decision-making. A reported history of CS was high within this UAE study sample. While most sociodemographic and health-system characteristics demonstrated limited independent association with prior CS, selected perception-related and experiential factors remained associated after adjustment. These findings suggest that in high-access healthcare settings, communication- and perception-related influences may contribute meaningfully to delivery-related preferences and experiences. Strengthening balanced, evidence-based, and woman-centred antenatal counselling may support informed decision-making regarding mode of delivery.

BackgroundExploratory applications of large language models within the specialized field of metabolic and bariatric surgery have begun to emerge. Nevertheless, existing research remains fragmented, lacking comprehensive integration.ObjectiveTo conduct a scoping review of studies on the application of large language models in the field of metabolic and bariatric surgery, aiming to provide a reference for clinical practice and future research.MethodsThis scoping review adhered to the Joanna Briggs Institute methodological framework and followed the preferred reporting items for systematic reviews and meta-Analyses extension for scoping reviews (PRISMA-ScR) guidelines.PubMed, Web of Science, The Cochrane Library, Embase, CINAHL, CNKI, Wanfang, and VIP databases were searched for relevant studies, with the search timeframe from database inception to November 2025. The included literature was summarized and analyzed.ResultsA total of 21 English-language studies were included. LLMs were primarily applied in scenarios such as patient education and information consultation, clinical decision support, and professional knowledge assessment. While LLMs performed well in information-provision tasks, they showed low consistency with expert opinions in complex clinical tasks such as individualized surgical recommendations. Performance varied across different models, with GPT-4 generally demonstrating superior performance, and domain-specific models showing professional potential. Current research still faces challenges regarding information accuracy, readability, and clinical applicability.ConclusionLarge language models hold auxiliary potential in the field of metabolic and bariatric surgery, particularly for knowledge dissemination and patient education. However, their reliability in complex clinical decision-making remains limited. Future efforts should focus on conducting high-quality studies, advancing model specialization and standardized evaluation, and exploring safe and effective human-AI collaboration models.

The management of safety in wastewater treatment plants (WWTPs) is faced with fundamental challenges, including sparse domain knowledge, dynamic evolution of safety protocols, and the necessity for highly reliable decision-making. While traditional risk assessment methods and expert systems provide essential support, they struggle to integrate multi-source heterogeneous knowledge to mitigate high-consequence, low-frequency(HCLF) risks. Existing general-purpose large language models (LLMs) demonstrate significant deficiencies in domain-specific knowledge, meanwhile, traditional fine-tuning methods are susceptible to catastrophic forgetting and knowledge conflicts during continual learning, rendering them unsuitable for direct application in this context. To address these challenges, this study proposes a progressive fine-tuning strategy to develop a domain-specific LLM tailored specifically for WWTP safety management. First, a domain-specific dataset is constructed through specialized dataset engineering. Subsequently, the proposed progressive fine-tuning strategy partitions domain knowledge into sequential stages, enabling the model to gradually learn and consolidate core knowledge at each stage before proceeding to the next. This orderly accumulation process ensures the deep integration of knowledge. The model is deployed and continuously optimized using vLLM, and direct preference optimization (DPO). The experimental results demonstrate that the progressive fine-tuning strategy effectively mitigates knowledge conflicts arising from multi-task fine-tuning. This approach not only ensures precise adherence to bottom-line safety protocols and enhances the model's depth of domain understanding in WWTP safety management, but also facilitates more coordinated capability allocation and knowledge integration across different professional tasks. By progressively refining the model, the proposed approach achieves superior task-specific performance even compared to models with substantially larger parameter scales, offering an effective and reproducible pathway for addressing analogous domain adaptation challenges.

Education for Sustainable Development (ESD) in teacher education is widely recognized as a key context for advancing sustainability, yet empirical research has often examined sustainability-related attitudes and dispositions in isolation. This study adopts an integrative approach to examine how empathy, prosocial behavior, sustainable consumer behavior, and responsibility are associated with attitudes toward sustainable development (ASD). Using validated self-report instruments, data were collected from 330 student teachers and analyzed through reliability testing, correlational analyses, and domain-specific multiple regression models. The findings confirm that ASD is a multidimensional construct characterized by distinct yet overlapping patterns of psychological and behavioral correlates across environmental, economic, social, and educational domains. Empathy (both affective and cognitive), prosocial behavioral tendencies, environmentally conscious consumption, and feelings of moral responsibility were consistently associated with stronger ASD, although their relative contributions varied by domain. Prosocial behavior emerged as a salient correlate of the social and educational dimensions of ASD, while responsibility and sustainable consumption were more strongly linked to environmental and economic orientations. Gender showed a small but significant association with the educational domain, underscoring heterogeneity within student teacher populations. By modeling multiple correlates simultaneously, this study extends existing ESD research beyond fragmented analyses. The results highlight the value of operationalizing prosociality as a multidimensional construct and offer original empirical insights to inform the design of responsive and inclusive ESD curricula in teacher education.

The worldwide effort to reach carbon peak and neutrality objectives alongside energy market expansion has sped up renewable energy integration, like wind and solar power. The shift towards renewable energy integration introduces substantial uncertainties in power system scheduling and control processes, which test the limits of existing theoretical methods. The advanced reasoning and data-processing capabilities of Large Language Models (LLMs), with particular reference to their ability to analyze multimodal data, provide transformative potential for managing and controlling smart grids. This review examines how LLMs can tackle modern power system challenges while confirming their fit with the power sector’s expanding dependency on Artificial Intelligence (AI) technologies. We assess the requirements of modern power systems for such AI-based solutions, while evaluating how LLMs shape grid management and exploring their enabling technologies, such as model architecture and training methods, along with necessary data. Our review investigates how multimodal LLM technology serves different smart grids’ functions, including generation, transmission, distribution, consumption, and equipment management, to exhibit its adaptable nature in strengthening grid resilience and efficiency. • This review explores the role of multimodal Large Language Models (LLMs) in smart grid management, showing how their ability to integrate and process different types of data, including sensor readings, text logs, weather forecasts, and equipment images, can significantly improve decision-making, fault diagnosis, and operational planning in power systems. • The study analyzes the architectural and training aspects of multimodal LLMs, including the use of pretrained modular encoders, efficient fine-tuning methods such as Low-Rank adaptation (LoRA), and specialized loss functions, highlighting how these techniques enable adaptation to the specific needs of smart grid applications without lengthy retraining. • Practical considerations for industrial implementation are examined, covering multimodal data collection and preprocessing, domain-specific knowledge integration, intelligent task decomposition, and system-level integration, illustrating how LLMs can be seamlessly integrated into power system operating environments. • The review highlights the potential of multimodal LLMs to improve the resilience of the power grid, optimize the integration of renewable energy, and support human-machine collaboration, while outlining future research directions, such as domain-specific base models, physics-based architectures, and human-in-the-loop feedback, in order to further improve reliability and interpretability in critical infrastructure applications.

Transforming oncology clinical trial matching through neuro-symbolic, multi-agent AI and an oncology-specific knowledge graph: a prospective evaluation in 3804 patients.

Cross-domain sequential recommendation via interest-guided knowledge migration.

SNOMED CT is one of the largest and most widely used medical ontologies. It is a collection of medical concepts with a large number of relationships between concept pairs. One of the most laborious tasks for SNOMED CT curators is placing concepts in the correct hierarchical positions with accurate relationships. We present a method that utilizes domain-specific BERT models integrated with a Relational Graph Convolutional Network (RGCN) for classifying the relationships between SNOMED CT concept pairs. We generate node embeddings using five domain-specific BERT models, BioBERT, ClinicalBERT, SapBERT, SciBERT, and BioMedBERT, and feed them individually into an RGCN model for relationship classification. We apply a three-layer RGCN to capture and utilize the graph-structured dependencies among concepts, enabling the model to propagate relational information across connected nodes before performing classification. We demonstrate that each BERT embedding integrated with RGCN outperforms both the Message-Passing Neural Network (MPNN) and Neural Network (NN) baselines. Our work shows that SapBERT-RGCN achieves the best performance across all tested encoders, with an accuracy of 0.9394 and an F1-weighted score of 0.9412. Compared to the best-performing domain-specific BERT-NN baseline, this is a 4% accuracy gain with a notable improvement of F1-macro.

ABSTRACT In this article, we introduce EduAssist-LLM — a large-scale domain-specific intelligent learning assistant designed specifically for classroom environments. The underlying technology is supported by the Llama 3.1 8B model and integrates a hybrid modeling strategy of Low-Rank Adaptation (LoRA) along with a Retrieval-Augmented Generation (RAG) approach that utilizes structured knowledge graph representations of the NCERT curriculum. EduAssist-LLM provides students with adaptive tutoring capabilities via a multi-turn conversational function, enabling learners to receive personalized Socratic dialogue feedback, generate curriculum-aligned content, and identify error patterns in their responses. The development pipeline incorporates curated textbook data, teacher-student dialogs, and synthetic question-and-answer pairs used for fine-tuning and continual learning while preventing catastrophic forgetting. Evaluation using the MMLU-Edu benchmark and simulated A/B classroom tests demonstrates 25–30 percent improvement in student engagement, approximately 20 percent higher knowledge retention, and substantially reduced hallucination rates compared to baseline models including GPT-4o and TeachLM. The model is further optimized for low-resource hardware deployment and incorporates bilingual support for Hindi and English, making it particularly suitable for under-resourced educational settings in India.

Heterogeneous domain adaptation survival analysis with partially observed outcomes via dictionary learning and distribution alignment.

General-domain sentiment models have been found ineffective in distinguishing positivity and negativity in Intensive Care Unit (ICU) clinical notes and domain-specific sentiment models are recommended. Although there are multiple common approaches to sentiment analysis, there has been little work comparing and evaluating the effectiveness of specialized models, largely due to the difficulty of recruiting clinical annotators and of accessing clinical sentiment data. This study has three contributions: (1) MIMIC-III-Ext-Notes-Sentiment: the first public ICU-specific ground- truth dataset labeled by clinicians for investigating clinical sentiment polarity in ICU clinical notes. (2) SentimentICUModel: an effective model for classifying clinical sentiment in ICU narratives on the ground truth. (3) A guiding comparison of the effectiveness of a range of approaches to clinical sentiment classification on the dataset. We recruited five clinicians to annotate notes for the ground truth. Annotators indicated which pieces of note text influenced their labeling. Six clinical-specific models were compared on the ground truth. The task of annotation was challenging due to clinicians' workload and spanned 15 months. The ground truth data was formed based on inter-annotator agreement analysis. Clinicians' extracts similarity aligned with their agreement level. Clinical language models provide comparable accuracy (up to 82%), with top score achieved by ClinicalT5 which is being released as SentimentICUModel. They outperform keyword-based lexicon (p[Formula: see text], [95% CI, -0.47, -0.28]). Clinical language models have demonstrated effectiveness in identifying clinical sentiment within clinical notes, enabling early detection of sudden changes and exploring different patterns in patients' ICU stays.

Background/Objectives: General-purpose and domain-specific multimodal foundation models show considerable promise in medical image analysis. In this study, we evaluated the classification accuracy of diabetic retinopathy vs. normal fundus images using general-purpose conversational models (Gemini 3 Flash, GPT-5.2, and Pixtral-Large), a medical conversational model (MedGemma-1.5), and its image-encoder (MedSigLIP), as well as ophthalmology-specific models (RETFound and EyeCLIP). Methods: We applied zero-/few-shot to general-purpose conversational models, linear probing, and fine-tuning approaches to domain-specific models for evaluation purposes. Results: We found that the zero-shot accuracies for Pixtral-Large (70.7%) and fine-tuned RETFound (77.1%) were comparable but lower than those of GPT-5.2 (77.9%), MedGemma-1.5 (88.2%), and Gemini 3 (88.5%) as well as the fine-tuned EyeCLIP (85.8%) and MedSigLIP (94.8%). The accuracy gains from few-shot prompting were substantial for Pixtral-Large (+7.4%) but were limited for GPT-5.2 (+3.6%), Gemini 3 (−3.4%), and MedGemma-1.5 (−1.1%). Embedding-based linear probing further improved accuracy over fine-tuning for RETFound (+9.7%) and yielded only marginal gains for EyeCLIP (+2.3%) but did not benefit MedSigLIP (−0.8%). Overall, with minimal prompting enhancement, general-purpose conversational models such as Gemini 3 and GPT-5.2 achieved performance comparable to ophthalmology-specific models that were either fine-tuned or enhanced via embedding-based linear probing, but remained inferior to MedSigLIP and its conversational counterpart, MedGemma-1.5. Conclusions: The findings highlight a trade-off between specialization and flexibility, where domain-specific models provide higher accuracy and stability, while general-purpose multimodal models offer greater accessibility, adaptability, and interactive reasoning, serving as complementary tools for retinal disease screening and clinical decision support.

The exponential growth of scholarly literature necessitates automated, scalable systems for organizing knowledge domains. However, text classification of academic abstracts presents distinct challenges due to specialized terminology and diverse discourse structures across disciplines. This study proposes a resource efficient deep learning methodology to categorize academic abstracts, scaling from coarse grained domains (arXiv) to fine grained disciplinary hierarchies (Web of Science). Systematic comparative analysis of Recurrent Neural Networks (Attention-GRU) and Transformer based architectures (BERT, SciBERT) are conducted, specifically focusing on the trade-off between predictive accuracy and computational efficiency. Extensive experiments on massive benchmarks that include the WOS-46985 dataset with 134 sub-disciplines, reveal a notable finding: Our proposed Attention-based GRU model utilizing static GloVe embeddings achieved a Macro-F1 score of 0.920, achieving higher performance than leading domain specific models such as SciBERT (F1: 0.867). Furthermore, this accuracy was achieved with over 3× faster training times and significantly lower estimated energy proxy compared to Transformer variants. This research contributes to the field by providing a systematic evaluation of "Green AI" architectures, demonstrating that computationally efficient models can robustly handle the linguistic diversity of high cardinality, fine grained scientific taxonomies without the prohibitive estimated energy costs of Large Language Models.