Search Results (17,295)

The rapid advancement of large language models (LLMs) has created new opportunities for intelligent fault diagnosis, particularly in complex industrial systems, such as heating, ventilation, and air conditioning (HVAC) in urban rail transit. Although LLMs have shown strong general reasoning capabilities, adapting them to domain-specific fault diagnosis tasks remains challenging. This is particularly true for textual maintenance records, where sparse and brief entries can only provide limited context for effective knowledge adaptation. To address this challenge, we propose a multi-stage post-training framework based on LLMs. The framework consists of three components: (1) data augmentation via retrieval-augmented generation (RAG) to enrich brief maintenance records with domain knowledge and reasoning traces; (2) supervised fine-tuning (SFT) for domain-specific adaptation; and (3) reinforcement learning with group relative policy optimization (GRPO), using a task-specific reward that separately evaluates root-cause identification and maintenance action recommendation. The framework is applied to a real-world textual HVAC fault dataset derived from Ningbo Rail Transit, covering 33 equipment categories and over 100 fault types. With Qwen3-0.6B as the base model, the proposed method significantly improves diagnostic accuracy and reasoning quality, achieving a 97% increase in model-based diagnostic accuracy (from 0.323 to 0.635) and a 48% improvement in human expert evaluation scores (from 0.509 to 0.754). Moreover, conventional machine learning baselines, such as Support Vector Machine (SVM), Random Forest (RF), and Multi-Layer Perceptron (MLP), achieve accuracies below 0.4 in this task, further highlighting the superiority of the proposed LLM-based framework. These results indicate that the proposed multi-stage post-training framework effectively improves LLM performance on real-world text-based fault diagnosis. It provides a practical and extensible solution for intelligent maintenance decision support in complex electromechanical systems. Full article

Large language model (LLM)-powered retrieval-augmented generation (RAG) systems significantly reduce factual errors in question answering, but they present a novel and under-investigated attack surface: factual information in retrieved text can be exposed in ways that do not align with the disclosure norms of the information itself. In legal, medical, and enterprise environments, this leakage is not only a confidentiality violation but a contextual integrity (CI) violation, where privacy is understood as the appropriate flow of information between roles and for specific purposes. Current defenses are mostly input minimizers or post hoc output filters that are blind to the circumstances of the data source (sender, recipient, purpose), creating a structural gap between the document store and the model generation step. We introduce ContextGuard-RAG, a multi-agent privacy enforcement framework that embeds CI theory directly into the retrieval and generation pipeline. It consists of three tightly coupled components: a CI-policy encoder that attaches sender, recipient, subject, information-type, and transmission-principle norms to retrieved contexts from document metadata; a privacy-aware reranker that filters contexts violating inferred CI norms using a fine-tuned cross-encoder classifier; and a generative firewall agent that sanitizes output using reinforcement learning to avoid transitive leakage through document-derived hallucinations or indirect inferences. On PrivacyQA, MedQA, and LegalBench, ContextGuard-RAG reduces CI violations by about 34 percent relative to vanilla RAG, by 13.5 percent on average over AirGapAgent, and by 10 percent on average over the 1-2-3 Check multi-agent reasoning approach, while achieving parity with the best accuracy baseline in ROUGE-L (within 0.6 absolute points) and remaining stable under context-hijacking adversarial probes. Unlike prior approaches, ContextGuard-RAG performs norm-aware retrieval rather than post hoc filtering and benefits four backbone LLMs without retraining the upstream agents. Full article

The integration of Large Language Models is transforming recommender systems, offering unprecedented capabilities for complex reasoning and natural language generation. However, their propensity to generate hallucinations (incorrect or invented information) compromises reliability and user trust, limiting their adoption in critical domains. Grounding techniques, which link LLM outputs to external, verifiable knowledge sources, are emerging as a fundamental solution, yet the field’s literature remains fragmented. This study aims to conduct systematic literature mapping to characterize and structure knowledge of grounding techniques applied to LLM-based recommender systems. A systematic literature mapping was conducted following the PRISMA protocol. A search of seven academic databases yielded an initial corpus of 1669 documents. After applying inclusion and exclusion criteria, a final set of 57 primary studies was selected and analyzed. The analysis reveals that Retrieval-Augmented Generation is the most widely used grounding technique (29.5%), indicating a clear preference for architectures that decouple LLM reasoning from the knowledge store. A technological bifurcation is observed between proprietary models, such as the GPT family (37.2%), and open-source models, such as LLaMA (24.8%). The application of these systems is expanding from e-commerce to highly critical domains, such as healthcare and finance. Based on the results, a multidimensional taxonomy is proposed to classify grounding techniques by architectural paradigm, data source, underlying mechanism, and intervention phase. The frequent adoption of grounding mechanisms indicates they are transitioning from an optional improvement to a highly prioritized architectural component for LLM-based recommender systems. Full article

Background: Artificial intelligence (AI) is increasingly being used in radiology, prompting the development of reporting guidelines, checklists, position statements, and evaluation frameworks. This systematic review aimed to identify and characterize these documents, assess their methodological quality, and evaluate their applicability to AI tools supporting diagnostic imaging. Methods: PubMed, Scopus, Web of Science, Embase, and the Cochrane Library were searched in October 2025. English-language guidance documents first officially published or made available online between 2015 and 18 October 2025, including ahead-of-print articles, were eligible. Two reviewers independently screened records and assessed methodological quality using the Appraisal of Guidelines for Research and Evaluation II (AGREE II) with predefined project-specific interpretation guidance. The review was retrospectively registered in PROSPERO (CRD420261471153). Results: Sixteen guidance documents were included. Clarity of Presentation had the highest median AGREE II score (86.1%), whereas Rigour of Development had the lowest (35.4%). CLEAR, CLEAR-E3, the CLAIM 2024 Update and the European Society of Cardiovascular Radiology (ESCR) position statement were recommended for use; the remaining 12 documents were recommended with modifications. Conclusions: Radiology AI guidance is generally clear and applicable, but development methods are often insufficiently reported. No single imaging-specific document comprehensively addresses model development, validation, diagnostic workflow integration, and post-deployment monitoring; complementary guidance may therefore be required. Full article

Few-shot learning (FSL) and zero-shot learning (ZSL) are usually studied as separate problems, yet both require prediction when class-specific evidence is absent or scarce. This review analyzes their shared difficulty from an optimization perspective. Instead of grouping studies only by architecture, it tracks four common coordinates: the information available to the learner, the variables estimated from that information, the objectives and constraints, and the numerical solvers. These coordinates support a unified comparison of attribute-based ZSL, episodic meta-learning, metric and prototype estimators, graph and optimal-transport inference, generative any-shot models, and adaptation of vision–language models. The synthesis exposes recurring trade-offs rather than a universally preferable family: flexible updates increase estimator variance; tractable task-time solvers inherit representation bias; query batches can improve inference while changing the protocol; and strong pretrained priors reduce target-data requirements while making the origin of task evidence harder to audit. Canonical objectives are distinguished from simplified review formulations and prospective research targets. The framework also clarifies the progression from explicit semantic mappings to local adaptation around pretrained image–text representations. Three priorities emerge: model selection without extra validation labels, safe use of uncertain pretrained knowledge, and stable parameter-efficient adaptation. Under this view, FSL and ZSL are connected structured-estimation problems rather than an inventory of unrelated algorithms. Full article

Coastal spaces are important settings for recreation, public health, tourism, and human–environment interaction. However, their spatial attractiveness often changes substantially across seasons, especially in regions with strong climatic variability. This study examines the seasonal drivers of coastal spatial popularity in China’s temperate monsoon region using 440,857 geotagged Weibo posts across 13 coastal cities, aggregated into 2733 5 km grid cells. We integrate perception-related cultural ecosystem service (CES) indicators extracted through large language models with built-environment, natural-environment, and socio-economic variables at the 5 km grid level. Fixed-effects models are employed to identify overall and seasonal associations between environmental conditions and spatial popularity. The primary full model explains a substantial proportion of the variation in coastal spatial popularity (R² = 0.6330). The results show that coastal spatial popularity differs from conventional urban activity patterns. High building density and POI diversity are negatively associated with popularity, while accessibility, attractions, water-related environments, and economic conditions are positively associated with activity intensity. At the same time, substantial seasonal variation is observed. Visual perception remains important throughout the year, while the influence of other factors changes across seasons. For example, aquatic environmental conditions become particularly important in summer, whereas taste-related experiences become more relevant in autumn and recreational services in winter. These findings suggest that the drivers of coastal spatial popularity are not uniform over time but vary under different seasonal contexts. The study provides empirical evidence for seasonally adaptive coastal planning and demonstrates the potential of combining large language models with social media data to analyze perception-related spatial dynamics. Full article

Background: Cardiovascular magnetic resonance (CMR) is the reference non-invasive imaging modality for evaluating cardiomyopathies, providing comprehensive assessment of cardiac morphology, function, and tissue characterization. Radiomics has recently emerged as an advanced image analysis technique that extracts quantitative imaging biomarkers from routine CMR images, potentially enhancing disease characterization beyond conventional visual assessment. This systematic review summarizes current evidence on the role of CMR-based radiomics in the diagnosis, phenotypic characterization and risk stratification of cardiomyopathies. Methods: A structured search identified English-language, peer-reviewed studies published up to August 2026 investigating CMR-based radiomics in hypertrophic, dilated, arrhythmogenic, and infiltrative cardiomyopathies, including cardiac amyloidosis, Fabry disease, and cardiac sarcoidosis. Study selection followed PRISMA 2020 guidelines. Methodological quality was assessed using the Methodological Index for Non-Randomized Studies (MINORS) and the Radiomics Quality Score 2.0 (RQS 2.0) The authors have reviewed and edited the output and take full responsibility for the content of this publication. Results: Current evidence indicates that CMR radiomics provides incremental diagnostic information beyond conventional CMR by quantifying myocardial tissue heterogeneity. Radiomic features derived from cine imaging, late gadolinium enhancement, native T1/T2 mapping, and extracellular volume maps showed promising performance in differentiating cardiomyopathy subtypes, distinguishing pathological from physiological remodeling, and identifying infiltrative and inflammatory myocardial diseases. Multiparametric radiomic models generally outperformed individual imaging biomarkers, with preliminary evidence supporting applications in risk stratification and outcome prediction. However, studies were predominantly retrospective, involved small cohorts, used heterogeneous imaging and radiomics workflows, and rarely included external validation. RQS 2.0 assessment demonstrated low-to-moderate methodological quality. Conclusions: CMR-based radiomics shows considerable potential for improving diagnosis and phenotypic characterization of cardiomyopathies. However, methodological standardization, multicenter prospective validation, and demonstration of incremental clinical value are required before routine clinical implementation. Full article

Large language models (LLMs) are increasingly explored for biomedical research and clinical medicine, yet their use remains limited by hallucinations, missing provenance, domain drift, privacy constraints, and uncertain regulatory pathways. Retrieval-augmented generation (RAG) offers a pragmatic alternative to full model training or fine-tuning by grounding LLM outputs in curated, versioned, and auditable external sources. This narrative review examines RAG as a biomedical grounding infrastructure rather than a chatbot add-on. We clarify key terminology around databases, knowledge bases, vector stores, knowledge graphs, RAG systems, and grounding, and distinguish factual, contextual, analytical, provenance, normative, and operational grounding. Two anchor scenarios are used to structure the review: single-cell annotation and omics interpretation as an exploratory biomedical research setting, and EHR, PDF, and clinical free-text integration as a regulated clinical information setting. We compare how these scenarios differ in sources, retrieval units, risk profiles, evaluation targets, infrastructure needs, and governance requirements. We further discuss design choices across the biomedical RAG lifecycle, evaluation and benchmarking, semi-automated and agentic RAG construction, infrastructure integration, normative grounding, and deployer-side governance. We argue that biomedical LLM deployment will not be determined by model scale alone, but by the quality of grounding infrastructures that make outputs traceable, contextual, updateable, and accountable. Full article

Background/Objectives: This study aimed to examine gender-associated differences in extracorporeal shock-wave lithotripsy (ESWL) utilization as well as urolithiasis diagnostic evaluation, treatment strategy, and follow-up across a multinational sample of urologists and trainees. Methods: A multinational cross-sectional survey of members of the Société Internationale d’Urologie (SIU) was conducted from June to July 2022 in seven languages. The survey was completed by 3747 urologists and trainees (283 female and 3464 male respondents) from 108 countries. To reduce baseline differences, propensity-score matching was performed with exact matching on years of clinical practice, continent, and career stage, yielding a matched cohort of 1191 respondents (279 female and 912 male). Categorical comparisons used Cochran–Mantel–Haenszel analyses stratified by a matched-set identifier, and 0–10 scored items were compared using stratified Wilcoxon rank-sum analyses with the matched set as the stratum. Results: In the matched cohort, annual procedural-volume distributions differed by gender for ureterorenoscopy (URS) (p = 0.0009), open surgery (p = 0.002), laparoscopy/robotics (p = 0.010), and percutaneous nephrolithotomy (PCNL) (p = 0.018), whereas ESWL volume did not differ significantly (p = 0.077). Among ESWL-specific categorical practices, coupling-gel use differed by gender (p = 0.029), while alpha-blocker use, antibiotic prophylaxis, JJ-stent use, stone-fragment collection, ESWL-machine type, and operator did not. Among the 851 respondents that completed the 0–10 practice items (186 female, 665 male), urine-culture assessment (8.4 vs. 7.6; p = 0.030), complete blood count (8.0 vs. 7.5; p = 0.047), and repeating ESWL within 7 days (2.5 vs. 3.4; p = 0.048) showed nominal gender-associated differences. The majority of other scored items were similar, including computed tomography (CT) urography (p = 0.055). Because these analyses were exploratory and no multiplicity adjustment was applied, nominally significant findings should be interpreted as hypothesis-generating. Conclusions: In this multinational survey of SIU members, most self-reported ESWL-specific practices were similar between female and male respondents after matching on key demographic, geographic, and professional characteristics. Gender-associated differences were observed in procedural-volume distributions as well as a limited number of reported assessments and procedural items. These findings are exploratory associations rather than causal effects of gender and may reflect measured and unmeasured professional, institutional, geographic, and training-related factors. Full article

Background: Artificial intelligence (AI) is increasingly being used in interventions among individuals with neurodevelopmental and cognitive disorders, offering personalized and adaptive approaches that advance traditional therapeutic practices. Although previous reviews have focused on symptom detection or alleviation, less attention has been paid to the impact of AI in fostering the acquisition of higher-order skills essential for being functional and independent. This review uniquely addresses this gap by synthesizing evidence from randomized controlled trials on AI-driven skill acquisition across multiple domains. Objectives: The objective of this systematic review is to synthesize evidence from randomized controlled trials evaluating the effectiveness of AI-driven interventions in promoting skillfulness. More specifically, it investigates the acquisition of neurocognitive, self-regulation, and adaptive and related skills among individuals with neurodevelopmental and cognitive disorders, including attention deficit and hyperactivity disorder, autism spectrum disorder, dyslexia, dyscalculia, and cognitive impairment. Methods: A systematic search, according to the PRISMA 2020 guidelines, was conducted, selecting randomized controlled trials published between 2019 and 2026. Eligible technologies included conversational agents, intelligent tutoring systems, adaptive training platforms, and machine learning-based interventions. Risk of bias was assessed using the Cochrane Risk of Bias 2 tool. Results: Twenty-four randomized controlled trials met the inclusion criteria. The findings demonstrated improvements in a wide range of skills, such as attention, working memory, mental flexibility, metacognitive control, emotional regulation, inhibition control, and social and communication skills. Generative AI showed efficacy for language and communication skills, while machine learning-based systems demonstrated positive effects on attention regulation and self-regulation. Conclusions: This review concludes that artificial intelligence can effectively assist conventional interventions for individuals with neurodevelopmental and cognitive disorders. However, the heterogeneity in intervention designs, outcome measures, and participant populations limits generalizability and highlights the need for standardized assessment frameworks, larger-scale longitudinal trials, and mechanistic investigations to translate these preliminary gains into long-term functional improvements across diverse clinical and cultural contexts. Full article

Prompt engineering has emerged as a practical and resource-efficient alternative to fine-tuning large language models (LLMs), particularly as these methods have a lower computation cost than fine-tuning. In this paper, three widely adopted prompting techniques—Zero-Shot, Few-Shot, and Chain-of-Thought (CoT)—were assessed. While these prompting strategies are well established, practitioners still lack clear guidance on when each technique should be preferred, which types of tasks they fail to support reliably, and how performance trade-offs may affect the practical use of LLM-based systems. These techniques are tested across three benchmark tasks: sentiment classification (SST-2), multiple-choice questions (CommonsenseQA), and multi-step math problem solving (GSM8K) using Meta’s LLaMA-3 8B Instruct model. We provide a thorough performance comparison based on accuracy, F1 score, and solve rate. The solve rate is highlighted as a complementary metric for evaluating the usability of LLM outputs—a factor often overlooked in the existing literature. Experimental results showed that Few-Shot prompts are particularly effective in structured classification tasks, while CoT prompts excel in logic-heavy tasks that require multi-step reasoning. On the classification task, Few-Shot prompting improved the solve rate but achieved lower accuracy and F1 score than Zero-Shot prompting. On the multiple-choice questions, Zero-Shot, Few-Shot, and CoT prompting achieved a solve rate of 100%. On multi-step math problem solving, CoT improved the solve rate and interpretability compared to Zero-Shot but did not surpass Zero-Shot accuracy. Overall, the results demonstrate that the effectiveness of prompting strategies is task-dependent, with differences observed in both accuracy and output validity across the three benchmark tasks. Full article

Generative artificial intelligence (GenAI), such as ChatGPT 3.5, presents academic challenges to college student users. Past literature has emphasized using non-diverse student samples, which limits understanding of diverse students’ perspectives and experiences in using ChatGPT in the learning process. Diverse students face challenges in addressing their learning needs and seeking additional academic support. As such, this study aims to examine the experiences of diverse students using ChatGPT to identify learning challenges and opportunities in higher education, drawing on Intersectionality Theory and Connectivism Learning Theory. A qualitative methodological approach using a Generic Qualitative Inquiry (GQI) design was employed by interviewing 27 diverse student participants from a public higher education institution. In examining the first research question on learning challenges in using ChatGPT, the findings revealed that diverse students reported reduced critical thinking, research limitations, and difficulties in using it effectively. Findings also revealed that they were becoming dependent on AI answers. Diverse students also acknowledged issues such as inaccuracies and misinformation, as well as adherence to academic integrity. In examining the second research question about the benefits of using ChatGPT, the findings showed that ChatGPT was perceived as practical and accessible, valuable for time management, and useful for gaining a deeper understanding and for studying. ChatGPT was also helpful in language development and proficiency. Findings also revealed that students were using ChatGPT to brainstorm and understand assignments, and to guide them in completing assignments. Implications point to the recognition that ChatGPT provides academic value to the learning experience of diverse students, which they can obtain beyond the realms of the classroom and university settings. Given that ChatGPT also poses challenges for learning, implications for educators include considering strategies to help diverse students navigate their learning expectations of ChatGPT use in their classrooms. Broader theoretical implications are interpreted through Intersectionality Theory and Connectivism Learning Theory, raising considerations for learning equity, institutional policy, and the cultural responsiveness of AI tools. This study provides needed qualitative evidence of diverse students’ experiences of using ChatGPT for learning in courses, which can inform efforts toward learning inclusion in academic environments. Full article

Developing Japanese writing ability as a second language (L2) requires learners to simultaneously manage linguistic, cognitive, and metacognitive resources. Although the Think Aloud Protocol (TAP) has been widely employed to access writers’ cognitive processes, its pedagogical impact on L2 Japanese writing remains underexplored. This study investigates the effects of TAP on the writing performance of Indonesian learners of Japanese at the A2–B1 proficiency levels, as defined by the Japan Foundation Standard (JF Standard), by examining both group-level performance changes and individual developmental trajectories. Using a convergent mixed-methods case study design, data were collected from 30 university students across three writing sessions (90 compositions): Pre-TAP, TAP-1, and TAP-2. Quantitative analysis revealed a gradual increase of 12.7% in mean scores from Pre-TAP to TAP-2. However, the overall difference across the three sessions did not reach statistical significance (Friedman χ²(2) = 5.447, p = 0.066). The mean improvement from Pre-TAP to TAP-2 was 1.37 points (95% CI [0.33, 2.43]), although the difference was not significant after Holm correction (adjusted p = 0.100). Nonetheless, individual-level analysis demonstrated divergent performance trajectories, including improvement, stability, and decline. These findings suggest that TAP functions differently across learners, serving as cognitive scaffolding for some, while imposing cognitive pressure and providing diagnostic value by revealing vulnerabilities in L2 writing skills for others. Full article

The transition to adulthood is a critical developmental period for adolescents and young adults, yet little is known about how it is experienced by disabled immigrant youth. This exploratory qualitative study examines the lived experiences of disabled immigrant adolescents and young adults in Norway as they navigate the transition to adulthood. Semi-structured interviews were conducted with five participants aged 17–25, and the data were analysed using reflexive thematic analysis. Five interrelated themes were developed: language, identity, and belonging; friendship formation and social inclusion; education and employment; aspirations and pathways to independent adulthood; and negotiating autonomy within family relationships. The findings suggest that the transition to adulthood is not a linear progression, but a dynamic and co-constructed process shaped by the interplay of individual, familial, and broader social and structural factors. Participants expressed aspirations for belonging, autonomy, education, employment, and meaningful adult roles while also encountering social exclusion, limited peer networks, institutional barriers, and ongoing parental involvement. The study provides exploratory insights into how disability and immigrant background intersect to shape everyday pathways to adulthood, influencing opportunities for participation and inclusion. Full article

Background: At present, there are very limited studies evaluating autonomous artificial intelligence (AI) agents in oral oncology or healthcare education, and no studies have directly compared traditional frontier large language models (LLMs) with autonomous AI agents in the field of pediatric oral oncologic pathology. This study evaluated the performance of AI LLMs and an autonomous AI scientific agent in generating engaging and accessible patient information sheets for common pediatric oral pathologic conditions and rare head and neck tumors. The development of high-quality patient education materials is particularly important in pediatric pathology because parents must often navigate complex and emotionally sensitive diagnoses, including rare tumors and developmental lesions, for which accessible, patient-friendly educational resources are frequently unavailable. AI-generated patient information sheets therefore represent a potential strategy to improve communication, understanding, and shared decision-making for families facing these uncommon conditions. Methods: AI-generated patient information sheets from popular frontier chatbots on various pediatric pathological conditions and rare tumors were evaluated by five platforms (Eli5a v2.0, ChatGPT-5.4, Claude Sonnet 4.6, Perplexity and Doximity), in a blinded fashion, by three expert evaluators using the Global Quality Score (GQS), DISCERN score, understandability score and actionability score using PEMAT, and the Flesch-Kincaid Grade Level. Because the same 20 conditions were assessed on every platform, observations were paired; platforms were compared using Friedman tests with paired Wilcoxon signed-rank post hoc tests and Holm correction, and interrater reliability was quantified using an absolute-agreement intraclass correlation coefficient (ICC). Results: Platform differences were significant for all five outcomes (all p < 0.001). Claude Sonnet 4.6 obtained the best overall information quality (GQS 4.25 ± 0.39), scoring significantly higher than all other platforms, while Perplexity, Eli5a v2.0 and ChatGPT-5.4 performed similarly to one another and better than Doximity. For information reliability, Doximity (DISCERN 70.50 ± 3.20) and Claude Sonnet 4.6 (70.25 ± 3.02) performed best, while Eli5a v2.0 recorded the lowest DISCERN score (54.50 ± 4.26). Eli5a v2.0 obtained the best results for patient-centered communication, with the highest understandability (PEMAT-U 95.05 ± 1.23) and readability (FKGL 5.45 ± 0.51) and an actionability score among the highest of the five platforms (PEMAT-A 89.00 ± 2.62, not significantly different from Perplexity or Doximity). Interrater absolute agreement for GQS was poor to moderate (ICC(2,1) = 0.236; ICC(2,3) = 0.481). Conclusion: No single platform was superior across all evaluated domains. Claude Sonnet 4.6 led in overall quality whereas Eli5a v2.0 achieved the highest understandability and the most appropriatereading grade level. Eli5a’s actionability was high but did not differ significantly from Perplexity or Doximity. These findings indicate a trade-off between information quality and reliability on one hand and lay accessibility on the other. Platform selection should therefore be matched to the communication task, and all AI-generated patient materials require clinician review before use. Full article