Search Results (1,058)

Large Language Models (LLMs) have demonstrated impressive capabilities in structured reasoning and symbolic tasks, with coding emerging as a particularly successful application. This progress has naturally motivated efforts to extend these models to mathematics, both in its traditional form, expressed through natural-style mathematical language, and in its formalized counterpart, expressed in a symbolic syntax suitable for automatic verification. Yet, despite apparent parallels between programming and proof construction, advances in formalized mathematics have proven significantly more challenging. This gap raises fundamental questions about the nature of reasoning in current LLM architectures, the role of supervision and feedback, and the extent to which such models maintain an internal notion of computational or deductive state. In this article, we review the current state-of-the-art in mathematical reasoning with LLMs, focusing on recent models and benchmarks. We explore three central issues at the intersection of machine learning and mathematical cognition: (i) the trade-offs between traditional and formalized mathematics as training and evaluation domains; (ii) the structural and methodological reasons why proof synthesis remains more brittle than code generation; and (iii) whether LLMs genuinely represent or merely emulate a notion of evolving logical state. Our goal is not to draw rigid distinctions but to clarify the present boundaries of these systems and outline promising directions for their extension. Full article

Introduction: Massage therapy delivers structured mechanosensory input that can influence brain function, yet the central mechanisms and potential for neuroplastic change have not been synthesized across neuroimaging modalities. This mechanistic review integrates evidence from electroencephalography (EEG), functional MRI (fMRI), and functional near-infrared spectroscopy (fNIRS) to map how massage alters human brain activity acutely and over time and to identify signals of longitudinal adaptation. Materials and Methods: We conducted a scoping, mechanistic review informed by PRISMA/PRISMA-ScR principles. PubMed/MEDLINE, Cochrane Library, Google Scholar, and ResearchGate were queried for English-language human trials (January 1990–July 2025) that (1) delivered a practitioner-applied manual massage (e.g., Swedish, Thai, shiatsu, tuina, reflexology, myofascial techniques) and (2) measured brain activity with EEG, fMRI, or fNIRS pre/post or between groups. Non-manual stimulation, structural-only imaging, protocols, and non-English reports were excluded. Two reviewers independently screened and extracted study, intervention, and neuroimaging details; heterogeneity precluded meta-analysis, so results were narratively synthesized by modality and linked to putative mechanisms and longitudinal effects. Results: Forty-seven studies met the criteria: 30 EEG, 12 fMRI, and 5 fNIRS. Results: Regarding EEG, massage commonly increased alpha across single sessions with reductions in beta/gamma, alongside pressure-dependent autonomic shifts; moderate pressure favored a parasympathetic/relaxation profile. Connectivity effects were state- and modality-specific (e.g., reduced inter-occipital alpha coherence after facial massage, preserved or reorganized coupling with hands-on vs. mechanical delivery). Frontal alpha asymmetry frequently shifted leftward (approach/positive affect). Pain cohorts showed decreased cortical entropy and a shift toward slower rhythms, which tracked analgesia. Somatotopy emerged during unilateral treatments (contralateral central beta suppression). Adjuncts (e.g., binaural beats) enhanced anti-fatigue indices. Longitudinally, repeated programs showed attenuation of acute EEG/cortisol responses yet improvements in stress and performance; in one program, BDNF increased across weeks. In preterm infants, twice-daily massage accelerated EEG maturation (higher alpha/beta, lower delta) in a dose-responsive fashion; the EEG background was more continuous. In fMRI studies, in-scanner touch and reflexology engaged the insula, anterior cingulate, striatum, and periaqueductal gray; somatotopic specificity was observed for mapped foot areas. Resting-state studies in chronic pain reported normalization of regional homogeneity and/or connectivity within default-mode and salience/interoceptive networks after multi-session tuina or osteopathic interventions, paralleling symptom improvement; some task-based effects persisted at delayed follow-up. fNIRS studies generally showed increased prefrontal oxygenation during/after massage; in motor-impaired cohorts, acupressure/massage enhanced lateralized sensorimotor activation, consistent with use-dependent plasticity. Some reports paired hemodynamic changes with oxytocin and autonomic markers. Conclusions: Across modalities, massage reliably modulates central activity acutely and shows convergent signals of neuroplastic adaptation with repeated dosing and in developmental windows. Evidence supports (i) rapid induction of relaxed/analgesic states (alpha increases, network rebalancing) and (ii) longer-horizon changes—network normalization in chronic pain, EEG maturation in preterm infants, and neurotrophic up-shifts—consistent with trait-level recalibration of stress, interoception, and pain circuits. These findings justify integrating massage into rehabilitation, pain management, mental health, and neonatal care and motivate larger, standardized, multimodal longitudinal trials to define dose–response relationships, durability, and mechanistic mediators (e.g., connectivity targets, neuropeptides). Full article

We present a combined ML/NLP (Machine Learning, Natural Language Processing) pipeline for protecting cloud-based APIs (Application Programming Interfaces), which works both at the level of individual HTTP (Hypertext Transfer Protocol) requests and at the access log file reading mode, linking explicitly technical anomalies with business risks. The system processes each event/access log through parallel numerical and textual branches: a set of anomaly detectors trained on traffic engineering characteristics and a hybrid NLP stack that combines rules, TF-IDF (Term Frequency-Inverse Document Frequency), and character-level models trained on enriched security datasets. Their results are integrated using a risk-aware policy that takes into account endpoint type, data sensitivity, exposure, and authentication status, and creates a discrete risk level with human-readable explanations and recommended SOC (Security Operations Center) actions. We implement this design as a containerized microservice pipeline (input, preprocessing, ML, NLP, merging, alerting, and retraining services), orchestrated using Docker Compose and instrumented using OpenSearch Dashboards. Experiments with OWASP-like (Open Worldwide Application Security Project) attack scenarios show a high detection rate for injections, SSRF (Server-Side Request Forgery), Data Exposure, and Business Logic Abuse, while the processing time for each request remains within real-time limits even in sequential testing mode. Thus, the pipeline bridges the gap between ML/NLP research for security and practical API protection channels that can evolve over time through feedback and retraining. Full article

Background and Objectives: Degenerative lumbar spinal stenosis (DLSS) frequently manifests as lower leg radiating pain (LLRP), requiring selective nerve root block (SNRB). Comprehensive and Integrative Medical Services (CIMS)—a multimodal program consisting of acupuncture, cupping, and manual therapy—have been increasingly incorporated into clinical practice in Korea. However, randomized evidence remains limited. This study evaluated the efficacy and safety of adjunctive CIMS in patients with DLSS presenting neuropathic LLRP requiring SNRB. Materials and Methods: In a single-center, parallel-group, assessor-blinded randomized controlled trial (CRIS KCT0006036), adults with DLSS (LANSS > 7; VAS > 5) were randomized 1:1 to experimental or control groups (n = 77; experimental 38, control 39). All participants received SNRB plus pharmacotherapy (limaprost, pregabalin). The experimental group additionally received CIMS, delivered eight times over 4 weeks. The primary outcome was pain intensity (VAS) at baseline and weeks 4, 8, and 12. Secondary outcomes included SF-36, ODI, and RMDQ at baseline and weeks 4, 8, and 12. Repeated-measures two-factor ANOVA assessed the main effects and time × group interaction. Results: Mean VAS (experimental vs. control) was 4.73 ± 1.67 vs. 4.70 ± 1.95 at baseline; 3.74 ± 1.68 vs. 4.66 ± 1.60 at week 4; 3.93 ± 2.03 vs. 4.79 ± 1.55 at week 8; and 3.98 ± 1.98 vs. 4.98 ± 1.68 at week 12. The significant time × group interaction was identified (p = 0.040), indicating a greater pain reduction with CIMS. No significant time × group interactions were observed across SF-36 domains. Adherence to CIMS modalities was high, and no unexpected adverse events occurred. Conclusions: In DLSS patients receiving SNRB and pharmacotherapy, adjunctive CIMS resulted in greater pain reduction over 12 weeks compared with standard care alone, without introducing new safety concerns. These findings support the clinical utility of CIMS as an effective adjunctive treatment option for DLSS. Full article

Ovarian cancer remains a top mortality contributor within gynecological cancers because patients receive diagnoses late in the disease course and conventional treatment resistance along with high recurrence rates cause poor outcomes. Aberrant regulation of autophagy and apoptosis has a critical role in the development, progression, chemoresistance, and immune escape from ovarian cancer. Recent evidence has demonstrated a complicated and dynamic crosstalk between autophagy and apoptosis, during which autophagy can act as a cytoprotective or cell death-promoting process depending on tumor stage and therapeutic context. In parallel, apoptosis functions as a tightly regulated form of programmed cell death that is essential for eliminating damaged or malignant cells and serves as a major tumor-suppressive mechanism in ovarian cancer. The PI3K/AKT/mTOR signaling pathway is the most active and clinically relevant pathway in the management of ovarian cancer as a master regulator of both autophagy and apoptosis, suppressing apoptotic cell death while promoting cytoprotective autophagy under chemotherapeutic stress. Bioactive natural products derived from plants, marine sources, and dietary intake have emerged as potential modulators of the autophagy-apoptosis crosstalk. Curcumin, resveratrol, quercetin, berberine, and epigallocatechin gallate are known to have the ability to restore apoptotic signaling, block pro-survival autophagy, and sensitize ovarian cancer cells to chemotherapy through the regulation of key pathways including PI3K/AKT/mTOR, AMPK, MAPK, p53, and Bcl-2 family proteins. In this review, we provide an updated understanding of the molecular mechanisms through which bioactive natural products modulate autophagy–apoptosis crosstalk in ovarian cancer. We also highlight the translational challenges, therapeutic potential, and future directions for the integration of natural product-based strategies in precision medicine for ovarian cancer. Full article

Metastasis is still the leading cause of cancer-related death. It happens when disseminated tumor cells (DTCs) successfully navigate a series of steps and adapt to the unique conditions of distant organs. In this review, key molecular and immune mechanisms that shape metastatic spread, long-term survival, and eventual outgrowth are examined, with a focus on how tumor-intrinsic programs interact with extracellular matrix (ECM) remodeling, angiogenesis, and immune regulation. Gene networks that sustain tumor-cell plasticity and invasion are described, including EMT-linked transcription factors such as SNAIL and TWIST, as well as broader transcriptional regulators like SP1. Also, how epigenetic mechanisms, such as EZH2 activity, DNA methylation, chromatin remodeling, and noncoding RNAs, lock in pro-metastatic states and support adaptation under therapeutic pressure. Finally, proteases and matrix-modifying enzymes that physically and biochemically reshape tissues, including MMPs, uPA, cathepsins, LOX/LOXL2, and heparinase, are discussed for their roles in releasing stored growth signals and building permissive niches that enable seeding and colonization. In parallel, immune-evasion strategies that protect circulating and newly seeded tumor cells are discussed, including platelet-mediated shielding, suppressive myeloid populations, checkpoint signaling, and stromal barriers that exclude effector lymphocytes. A major focus is metastatic dormancy, cellular, angiogenic, and immune-mediated, framed as a reversible survival state regulated by stress signaling, adhesion cues, metabolic rewiring, and niche constraints, and as a key determinant of late relapse. Tumor-specific metastatic programs across mesenchymal malignancies (osteosarcoma, chondrosarcoma, and liposarcoma) and selected high-burden cancers (melanoma, hepatocellular carcinoma, glioblastoma, and breast cancer) are highlighted, emphasizing shared principles and divergent organotropisms. Emerging therapeutic strategies that target both the “seed” and the “soil” are also discussed, including immunotherapy combinations, stromal/ECM normalization, chemokine-axis inhibition, epigenetic reprogramming, and liquid-biopsy-enabled minimal residual disease monitoring, to prevent reactivation and improve durable control of metastatic disease. Full article

Cervical cancer remains a major global public health concern, with persistent infection by high-risk human papillomavirus (hrHPV) types recognized as the primary etiological factor. This review explores the multifactorial nature of the disease, focusing on the complex interplay between host genetic susceptibility and epigenetic alterations that drive cervical carcinogenesis. Evidence from genome-wide association studies (GWAS) is discussed, highlighting the contribution of specific genetic loci, predominantly within the HLA region, to susceptibility to HPV infection and disease progression. In parallel, the review examines the molecular mechanisms by which the viral oncoproteins E6 and E7 promote genetic instability and epigenetic reprogramming, including histone modifications and dysregulation of non-coding RNAs. Particular emphasis is placed on DNA methylation, affecting both the viral genome and host genes such as FAM19A4, CADM1, PAX1, and MAL, as a promising biomarker for triage and detection of high-grade intraepithelial lesions (CIN2+). Finally, the review evaluates currently available methylation-based assays and self-sampling devices, highlighting their potential to enhance diagnostic accuracy and increase participation in cervical cancer screening programs. Full article

Background: Strength and the strength–power continuum may increase athletic performance, although data are scarce regarding the effects of long-term periodized training on the athletic performance of adolescent track and field athletes. The purpose of this study was to investigate performance modifications following 8 weeks of strength and strength–power resistance training, focusing on the athletic performance of adolescent track and field athletes. Methods: Following an equivalent single-arm pre–post intervention design, 16 adolescent athletes (age: 16.3 ± 0.5 years; mass: 56.5 ± 10.4 kg; height: 1.67 ± 0.07 m) participated in the study. Athletes followed an 8-week periodized resistance training program aiming to increase strength and strength–power. Measurements were performed before (T1), at the middle (T2) and at the end of the training period (T3) and included the standing long jump, single-leg standing long jump, five-step long jump, seated medicine ball throw, 0–80 m sprint and 1RM in the bench press and parallel squat. Results: The standing long jump (F_(2,14) = 109.564; η² = 0.940; p = 0.001), single-leg long jump (F_(2,14) > 41.801; η² = 0.857; p = 0.001) and five-step long jump (F_(2,14) = 148.564; η² = 0.955; p = 0.001) improved significantly from T1 to T2 (p < 0.001) and from T2 to T3 (p < 0.001). The seated medicine ball throw (F_(2,14) = 124.305; η² = 0.947; p = 0.001) and sprinting performance (F_(2,14) = 51.581; η² = 0.828; p = 0.001) were significantly enhanced from T1 to T2 (p < 0.001) and from T2 to T3 (p < 0.001). The 1RM in the bench press (F_(2,14) = 36.280; η² = 0.838, p = 0.001) and in the parallel squat (F_(2,14) = 48.165; η² = 0.873, p = 0.001) increased significantly from T1 to T2 (p < 0.001) and from T2 to T3 (p < 0.01). Conclusions: Strength and the strength–power continuum appear to have a positive effect on the physical fitness of adolescent track and field athletes, which highlights the importance of strength-based resistance training programs in adolescent athletes. Full article

This study investigates how master’s student teachers (MSTs) conceptualize and integrate programming and computational thinking within mathematics education. Grounded in the Anthropological Theory of the Didactic, and specifically its notion of praxeology, the study analyses 39 written reflections produced by MSTs who completed a compulsory programming-based mathematics task. The analysis identifies both mathematical and didactic praxeologies, revealing how MSTs’ engagement with programming reflects their development both as learners and as future teachers. The findings demonstrate that MSTs’ personal learning strategies, such as exploration, iteration, and productive struggle, closely parallel their envisioned classroom practices. The findings also show that many participants framed programming itself as the central learning object, highlighting a need to develop confidence and competence before applying programming as a tool for mathematical inquiry. The study argues that programming tasks provide a productive arena for bridging theory and practice in teacher education by fostering an interplay between praxis (know-how) and logos (know-why). Finally, the results indicate that MSTs require institutional support specifically aimed at developing basic programming fluency (e.g., handling syntax, debugging, and programming environments), so that computational thinking can be mobilized for mathematical exploration rather than being overshadowed by technical challenges. Full article

Background/Objectives: One of the pillars of optimal footballer performance is the gradual preparation of the body for physical exertion in terms of intensity. The aim of this study was to evaluate the impact of a structured warm-up and cool-down program on flexibility, perceived fatigue, and injury prevention in young football players. Methods: Participants were 60 junior football players (U17), with a mean age of 16.5 ± 0.5 years, mean height of 172.5 ± 6.7 cm, and mean body mass of 70.2 ± 6.4 kg. The participants were assigned to experimental (EXP; n = 30) and control (CON; n = 30) groups during 8 mesocycles. A 4-week training stimulus was applied in parallel, consisting of an author-designed exercise routine with a profiled intensity (warm-up and cool-down parts) for the EXP group and standard exercises for the CON group. Selected variables (motor, endurance, injuries) were assessed before, during, and after the intervention. Additionally, the profile of selected correlations was analysed. Statistical analysis was performed using t-tests with a significance level set at p < 0.05. Results: In the EXP group (post-test), a significant improvement in flexibility was observed in the forward trunk flexion test (d = 1.13 cm; p < 0.001; d_c2 = 1.05). Simultaneously, participants reported lower levels of subjective fatigue (RPE = 6.86 ± 0.82 points) compared to the CON group (p = 0.016; d_c = 0.46) and demonstrated fewer injuries during the annual cycle (0.97 ± 0.83 vs. 1.33 ± 0.66; p = 0.026; d_c = 0.48). Both groups showed a strong negative correlation between flexibility and the number of injuries in the annual cycle, training experience and the number of injuries, as well as training experience and RPE (all r_p > −0.50). A strong positive correlation was found between RPE and the number of injuries (r_p > 0.60). Conclusions: The results demonstrate that the structured warm-up and cool-down program significantly improved flexibility, reduced perceived fatigue, and decreased injury occurrence in the participants. Full article

This study examines the effectiveness of an eight-week AI training program aimed at enhancing teacher candidates’ pedagogical competence and AI literacy in rapidly changing and evolving educational environments. As the modern world continues to change and develop, the transformation of education, which is one of the most important elements of our lives, cannot be ignored. Accordingly, the integration of teacher candidates, who constitute key education stakeholders, into technological developments is very important in terms of both efficiency and sustainability. The “parallel–simultaneous design”, one of the mixed research methods in which quantitative and qualitative research methods are used together, was employed. In line with the stated purpose, the study started with a needs analysis conducted with 33 teacher candidates studying in different branches at the faculty of education. As a result of the needs analysis, knowledge gaps, digital skill levels and readiness for integration of artificial intelligence tools in future classrooms were determined. Its application to teacher candidates, instead of teachers in the profession, was determined by the needs analysis. The results indicate that it would be more beneficial to apply the education of the future to the teachers of the future and that they will find it easier to adapt to such training. Accordingly, a pre-test–post-test design was applied to observe how the participants changed, and an artificial intelligence literacy scale was also used. QDA Miner Lite was used for the analysis of the qualitative data, and SPSS 29.0 was used for the analysis of the quantitative data. During the eight-week training, Gamma programs were used for the presentation, Suno for audio, Midjourney for visuals and ChatGPT-4 for a descriptive search in order to provide better quality education to the participants. While practicing with these applications, the aim is to provide more up-to-date education that reveals problem-solving skills that include critical thinking exercises. According to the results, the teacher candidates who expressed that they were undecided or had insufficient knowledge reached a sufficient level in the post-test. In the light of these results, it can be stated that artificial-intelligence-oriented education is effective in developing sustainable pedagogical skills, digital literacy, readiness and professional self-confidence. The study also offers evidence-based recommendations for the design of future teacher training programs. Full article

Influenza D virus (IDV), an emerging orthomyxovirus with zoonotic potential, infects diverse hosts, causes respiratory disease, and remains poorly characterized in China despite its global expansion. From October 2023 to January 2025, we collected 563 nasal swabs from cattle across 28 farms in Jilin Province, Northeast China, and identified seven IDV-positive samples (1.2%), recovering two viable isolates (JL/YB2024 and JL/CC2024). Full-genome sequencing revealed complete, stable seven-segment genomes with high nucleotide identity (up to 99.9%) to contemporary Chinese D/Yamagata/2019 strains and no evidence of reassortment. Maximum-likelihood and time-resolved Bayesian phylogenies of 231 global hemagglutinin-esterase-fusion (HEF) sequences placed the Jilin isolates within the East Asian D/Yamagata/2019 clade and traced their most recent common ancestor to approximately 2017 (95% highest posterior density: 2016–2018), suggesting a cross-border introduction likely associated with regional cattle movement. No IDV was detected in parallel surveillance of swine, underscoring cattle as the principal reservoir and amplifying host. Bayesian skyline analysis demonstrated a marked decline in global IDV genetic diversity during 2020–2022, coinciding with livestock-movement restrictions imposed during the COVID-19 pandemic. Collectively, these findings indicate that IDV circulation in China is sporadic and geographically localized, dominated by the D/Yamagata/2019 lineage, and shaped by multiple independent incursions rather than a single emergence. Both the incorporation of IDV diagnostics into routine bovine respiratory disease surveillance and cattle-import quarantine programs, and the adoption of a One Health framework to monitor potential human spillover and future viral evolution, were recommend. Full article

The liver stage of Plasmodium infection represents a critical bottleneck in malaria pathogenesis and a unique interface between parasite development and hepatocyte-intrinsic immunity. Recent evidence suggests that hepatocytes do not eliminate liver-stage parasites through canonical xenophagy, as previously assumed, but instead employ a noncanonical autophagy response known as the conjugation of ATG8 to single membranes (CASM). CASM drives rapid lipidation of LC3 onto the parasitophorous vacuole membrane (PVM) via a V-ATPase-ATG16L1-dependent mechanism, thereby activating the Plasmodium-associated autophagy-related (PAAR) response. This process represents a major hepatocyte-intrinsic mechanism that limits early liver-stage parasite development. Plasmodium liver-stage parasites have evolved specialized strategies to counteract this host defense. The PVM proteins UIS3 and UIS4 enable parasite evasion by sequestering LC3 and remodeling perivacuolar actin, thereby preventing endolysosomal fusion and inhibiting PAAR execution. In parallel, parasites selectively exploit host autophagy components—particularly GABARAP paralogs—to activate TFEB, promoting lysosomal biogenesis and improving access to host-derived nutrients. These interactions highlight autophagy as both a protective and parasite-supportive pathway, depending on the molecular context. Understanding how CASM, PAAR, and parasite evasion mechanisms intersect is crucial for designing pathway-selective interventions that amplify hepatocyte-intrinsic clearance while avoiding the inadvertent enhancement of parasite-supportive autophagy programs. Selective modulation of noncanonical autophagy offers a promising avenue for host-directed therapies that restrict liver-stage development while limiting the emergence of antimalarial resistance. This review synthesizes recent advances in the mechanistic interplay between Plasmodium liver stages and hepatocyte autophagy, identifies major knowledge gaps, and outlines future directions for translating these discoveries into therapeutic innovation. Full article

Diabetes self-management education (DSME) programs are evidence-based interventions that improve glycemic control and self-care behaviors, yet their effectiveness may be limited by unaddressed health-related social needs (HRSN) (e.g., food insecurity, housing or utility instability, transportation barriers). This qualitative multiple case study examined the feasibility of integrating HRSN assessments into DSME delivery within three community-based organizations (CBOs) across urban and rural U.S. settings. Guided by the Consolidated Framework for Implementation Research, semi-structured interviews were conducted with 15 DSME facilitators and program leadership to identify contextual factors influencing implementation. Findings revealed that while DSME’s structured, manualized design promotes fidelity and client autonomy, it constrains responsiveness to the client’s HRSN. Facilitators expressed openness to integrating HRSN screening, particularly during intake, yet cited limited infrastructure, role clarity, and training as key barriers. CBOs were recognized as trusted, accessible spaces for holistic care, but growing expectations to address HRSN without adequate resources for referral created sustainability concerns. Participants recommended a parallel support model involving navigators or community health workers to manage HRSN screening and referrals alongside DSME sessions. Integrating HRSN assessment processes into DSME may enhance engagement, reduce attrition, and extend the reach of diabetes education to populations most affected by HRSN. However, successful implementation requires dedicated funding, workforce development, and cross-sector coordination. Findings underscore the importance of supporting CBOs as critical partners in bridging diabetes education and social care to advance whole-person, chronic disease management. Full article

In an increasingly technologized and automated world, workplace safety and health remain a major global challenge. After decades of regulatory frameworks and substantial technical and organizational advances, the expanding interaction between humans and machines and the growing complexity of work systems are gaining importance. In parallel, the digitalization of Industry 4.0/5.0 is generating unprecedented volumes of safety-relevant data and new opportunities to move from reactive analysis to proactive, data-driven prevention. This review maps how artificial intelligence (AI), with a specific focus on natural language processing (NLP) and large language models (LLMs), is being applied to occupational risk prevention across sectors. A structured search of the Web of Science Core Collection (2013–October 2025), combined OSH-related terms with AI, NLP and LLM terms. After screening and full-text assessment, 123 studies were discussed. Early work relied on text mining and traditional machine learning to classify accident types and causes, extract risk factors and support incident analysis from free-text narratives. More recent contributions use deep learning to predict injury severity, potential serious injuries and fatalities (PSIF) and field risk control program (FRCP) levels and to fuse textual data with process, environmental and sensor information in multi-source risk models. The latest wave of studies deploys LLMs, retrieval-augmented generation and vision–language architectures to generate task-specific safety guidance, support accident investigation, map occupations and job tasks and monitor personal protective equipment (PPE) compliance. Together, these developments show that AI-, NLP- and LLM-based systems can exploit unstructured OSH information to provide more granular, timely and predictive safety insights. However, the field is still constrained by data quality and bias, limited external validation, opacity, hallucinations and emerging regulatory and ethical requirements. In conclusion, this review positions AI and LLMs as tools to support human decision-making in OSH and outlines a research agenda centered on high-quality datasets and rigorous evaluation of fairness, robustness, explainability and governance. Full article